Companies across Southeast Asia are approving AI pilots before answering a more basic question: can the underlying workflow actually support automation at scale?

For operators, that question matters more than the model being tested. If invoices, contracts, procurement records, and customer data are still scattered across paper, PDFs, spreadsheets, emails, and legacy systems, AI can quickly become another layer on top of an already fragmented process.

Fukuda pointed to one Sansan client, a construction company in Thailand that was processing more than 2,000 invoices every month, with over 90% still exchanged on paper. Each invoice took about 20 minutes to process manually, delaying the monthly closing and making it difficult for headquarters to monitor activity across construction sites.

After digitizing the workflow, the company was able to consolidate invoices from different offices and sites into a single system. The shift cut processing time to eight minutes per invoice and saved around 4,800 work hours a year.

The case is a success story, but its starting point is the more important lesson. Before the company could automate invoice processing, it first had to make a messy paper-heavy workflow visible, structured, and usable.

The issue is not necessarily that companies lack AI tools. Many are trying to layer AI onto workflows that were never properly standardized, governed, or prepared for automation in the first place.

The Real Bottleneck: Fragmented Data

For many enterprises, the biggest obstacle to AI adoption is not model capability, but the condition of the underlying data itself.

Fukuda frames data readiness as a precondition for scaling AI, not a problem to fix after deployment.

Invoices, contracts, business cards, procurement records, and customer information are often stored across disconnected systems, handled manually by different departments, or exchanged in inconsistent formats. Even companies that have digitized parts of their operations may still rely heavily on scanned PDFs, spreadsheets, emails, and legacy approval processes.

According to Fukuda, organizations often assume that having large amounts of data automatically makes them ready for AI deployment. In practice, fragmented and poorly structured data can make AI outputs unreliable from the beginning.

The problem becomes more visible in Southeast Asia’s emerging markets, where digitization maturity varies widely across industries and the supply chain. External vendors, suppliers, and contractors may still submit documents manually or use incompatible systems, making it difficult to create standardized workflows that AI systems can process consistently.

Even Sansan ran into this problem while developing AI-driven document management tools. Fukuda said the company found that general-purpose AI models struggled with the variety and complexity of real business documents, leading to delays and inaccurate outputs.

Sansan’s response was to move away from relying on general-purpose AI alone and toward models trained for business-document structures. The broader lesson for operators is not product-specific: generic AI will struggle when the workflow depends on messy documents, inconsistent formats, and business-specific exceptions.

The experience highlights a broader challenge for enterprises adopting AI. Vendor demos and pilot environments are often cleaner than production reality. Once AI systems encounter fragmented workflows, inconsistent formats, incomplete data, and edge cases at scale, performance can deteriorate quickly.

For operators, that is the lesson to take into vendor selection. A successful demo does not prove that a system can handle real document variety, messy supplier inputs, or the exceptions that appear in day-to-day operations.

The Human and Workflow Problem

Fukuda adds that AI failures are not only caused by technical limitations, but also by how new systems fit into existing workflows.

If AI tools disrupt how teams already work, employees may see them as additional friction rather than productivity tools. In some cases, teams revert to manual processes when AI outputs become inconsistent or difficult to trust.

In most cases, these problems stem from the same root issue: AI initiatives were launched before the underlying data and workflows were properly prepared. Without proper training and integration into day-to-day workflows, AI initiatives can struggle to move beyond experimentation.

That makes adoption a pre-scaling test, not a post-launch training issue. Before expanding AI across departments, operators need to know who owns the workflow, who monitors the output, who investigates errors, and how teams will use the system when results are imperfect.

The same applies to governance. In regulated sectors, the question is not only whether policies exist, but whether companies can prove that controls are working inside AI-supported workflows day-to-day. That includes access controls, activity logs, monitoring, auditability, and audit trails that show how information is processed and who has accessed it. They also expect incident response procedures and regular security assessments to be in place before AI-assisted workflow changes are approved. What is often missing is operational evidence that these controls are consistently applied in daily workflow, not just written into policy.

Before the Next AI Budget Gets Approved

For Fukuda, the bigger risk for Southeast Asian enterprises is not moving too slowly on AI, but moving too quickly without fixing the operational foundations underneath.

The safer path is to start with narrow operational problems where the business pain is clear, the data can be prepared, and the outcome can be measured. Only then should companies expand AI across more complex workflows.

That approach is especially relevant in Southeast Asia’s asset-heavy industries, where many operational systems remain fragmented across sites, suppliers, and legacy processes. In these environments, the companies that benefit most from AI may not be the ones deploying the most tools, but the ones that spend more time preparing their operational foundations before scaling them.

Before the next AI line item lands in the budget, operators should ask a narrower set of questions. What workflow is this supposed to fix? Is the data usable? Who owns the output? How will employees use the tool? What happens when the system gets it wrong and how will success be measured?

The companies that get this right will not be the ones that moved fastest on AI. They will be the ones that were honest enough to fix their operations first.

More From Asia Tech Lens

• Why AI Agents Still Struggle Inside Southeast Asia’s Enterprises

  Why enterprise AI adoption in Southeast Asia runs into hierarchy, trust, workflow ownership, and accountability.

• Agentic AI Can Act. Singapore’s New Rulebook Says: Prove You Can Stop It

  Once AI moves from assistance to action, operators need bounded autonomy, audit trails, oversight, and rollback plans before deployment can be trusted.

• Budget 2026 Puts AI Into Execution Mode. Operators Need To Sequence It Carefully

  Singapore’s AI push shows why enterprises need to sequence deployment around sector readiness, operational capacity, and measurable use cases rather than treating AI adoption as a broad transformation mandate.

• India’s AI Push Is Real. Production Access Is the Constraint

  Why AI ambition does not equal deployment readiness: regulated operators need reservable capacity, auditable controls, and portability before pilots become usable production infrastructure.

• Why Quantum Pilots Fail Before They Start And What To Do About It

  Different technology, same operator lesson: pilots fail when teams start with what the technology can do instead of defining the costly decision, measurable baseline, owner, and procurement path upfront.

When enterprises across Asia start a quantum pilot, they often begin by asking what the technology can do for them. The entry point can produce interesting pilots but they will likely go nowhere—no procurement decision, no budget owner, no path to deployment, no business impact.

Alexandra Beckstein, CEO of QAI Ventures, has spent enough time inside these programs to know exactly where that happens. “Teams ask where they can use quantum, rather than asking which decision is costly, complex, and worth improving. That usually leads to a pilot that sounds exciting but is too vague to succeed,” she says.

Beckstein runs QAI Ventures, a Switzerland-founded firm with its Asia-Pacific headquarters in Singapore. In February, the firm announced an industry cluster program backed by SoftBank Corp. and HorizonX. Beckstein has built quantum startup ecosystems across Europe, North America, and Asia Pacific. She has seen where the pipeline breaks for many enterprise quantum programs. She shares her insights in a written interview with Asia Tech Lens.

The Wrong Question Is Being Asked First

What she keeps finding is that the decision that determines a pilot’s fate is made at the very beginning. Teams enter quantum programs by asking where the technology can be applied rather than which specific, costly business decision needs to be improved. That produces pilots that are technically interesting but commercially vague.

The most reliable early signal that a pilot is incorrectly structured is the absence of a P&L owner. “A pilot needs a defined P&L owner who is accountable for outcomes,” Beckstein says, “not just an innovation or IT team exploring new tools.” If the initiative lives inside an innovation function without a business unit accountable for the result, it is not a business pilot. It is an experiment with a business-shaped label. In regulated industries—banking, insurance, or telecommunications—a pilot without a business owner rarely makes it past procurement, regardless of how technically promising the results are.

The failure pattern has a structural cause. On the supply side, researchers with strong IP often lack business scaffolding. They don’t have the capacity to translate technical capability into a deployable system. On the demand side, enterprises are aware of quantum as a category but have no capacity to structure a pilot that produces evidence procurement can act on. QAI Ventures is positioned around both sides of that gap. Most of what is deployable today is quantum-inspired or hybrid, not full quantum computing, which remains dependent on hardware breakthroughs that have not yet arrived. While quantum research is maturing, enterprise deployability is a separate endeavor.

Beckstein recalls a case where a company attempted to tackle a very large planning problem too early: the data was inconsistent, the business team had not agreed on a success metric, and the project became too broad to produce a clean result. “The lesson was simple,” she says. “Start with a smaller problem, cleaner data, and a tighter commercial goal.”

What A Correctly Framed Pilot Looks Like

Three cases Beckstein points to illustrate what problem-first framing produces in practice; the figures that follow are drawn from her account. The following cases are not arguments for quantum broadly. They are examples of when a specific problem is matched to the right method.

Multiverse Computing applied its Singularity framework to compress AI models for a customer service network. The outcome was an 80% reduction in model size and up to 75% lower energy consumption, with no degradation in response quality. The business case—cost reduction and sustainability—was defined before the quantum-inspired approach was selected, not after.

Fujitsu’s Digital Annealer, deployed at the Port of Hamburg, optimized vehicle traffic flows and increased average travel speed by 20% while cutting CO₂ emissions by 10%. More operationally significant: a calculation that previously took days now runs in seconds. The framing was a logistics bottleneck the port already owned. The quantum-inspired approach was chosen because classical methods had hit a scaling ceiling.

QTFT, a quantum software startup founded in Thailand, built a routing solution for supermarket goods deliveries that generates several strong, viable alternatives rather than a single theoretically optimal route. For a logistics operator, that distinction matters at the moment a route fails and a decision needs to be made in minutes, not hours.

As Beckstein puts it, these cases share something fundamental: “A focus on solving genuine operational constraints and producing results that procurement and operations leaders can directly compare against their existing benchmarks. That is ultimately what separates promising pilots from deployments that stick.”

Her point is that success criteria need to be quantified before a pilot begins. For example, a 3% or greater improvement over a well-understood classical baseline. Without that number agreed in advance, the pilot has no natural endpoint, and no moment at which anyone is obligated to act on the result.

How To Run It, And When To Stop

The second failure mode Beckstein flags is less about framing and more about discipline. Pilots run too long because nobody agreed on exit criteria before they started. By the time results are inconclusive, the budget has been spent and the business owner has moved on.

Her stop rules are unambiguous: “A pilot should stop if the data is not good enough, if the business owner is not engaged, or if the result is not clearly better than the current approach. It should also stop if costs keep rising without stronger evidence.”

In regulated industries, the governance layer deserves particular attention—it is a condition of deployment. In banks, insurers, telcos, and critical infrastructure, the demo is not the hard part. Auditability, reproducibility, vendor risk, security, compliance, and long-term support are what procurement will flag. Pilot teams that do not document these early should not be surprised when a technically promising project stalls at the procurement stage.

Not every category is ready for enterprise piloting. The categories Beckstein sees as decision-relevant in the next 12 to 24 months are specific: quantum-safe security planning in regulated sectors, financial services workflows including pricing, risk, and fraud detection, and operational planning in logistics, energy, and supply chains. What she is explicitly not backing are categories that still depend on hardware breakthroughs before they can deliver enterprise value. For operators choosing between vendors, that boundary is a useful screen: if a vendor’s pitch depends on technology that does not yet work at enterprise scale, no amount of careful pilot design will produce a deployable result.

To move from exploration to a budget line-item, Beckstein is clear about what needs to be in place: “A clear owner, clear business value, usable data, and a realistic path to implementation.” If any of those four are absent, the pilot is not ready.

Before funding a quantum pilot, operators should be able to answer the following: which expensive decision is worth improving, why current methods are no longer sufficient, and whether a quantum-inspired or hybrid approach can beat a defined baseline under real conditions. If that cannot be answered upfront, the pilot should wait.

More From Asia Tech Lens

• Can India Build Quantum Computers That Matter Globally?
  India’s quantum push shows why national ambition is only the first step; the harder test is turning research capacity into commercially relevant systems.

• Why Japan’s Quantum Strategy Starts With Algorithms, Not Qubits
  Japan’s approach underlines the same near-term lesson: quantum value may arrive first through applied software and workflow improvements, not hardware breakthroughs.

• Singapore’s Quantum Bet: Where AI Meets the Next Compute Revolution
  Singapore’s quantum ecosystem helps explain why the region is trying to close the gap between research investment and enterprise deployment.

• Agentic AI Can Act. Singapore’s New Rulebook Says: Prove You Can Stop It.
  Like quantum pilots, agentic AI deployments show that advanced technology only earns enterprise trust when governance, control, and accountability are built in early.

• India’s AI Push Is Real. Production Access Is the Constraint
  This piece echoes the same deployment problem: emerging technology only matters when it can move from promise to production infrastructure.

Editor’s Note: Asia Tech Lens turns one this week. The goal when we started was straightforward: track Asia’s technology rise with perspective, without the hype, and free from rhetoric. A year in, readers from nearly 60 countries have found their way here—many in the U.S., the U.K., and across Europe, alongside strong communities in Singapore and India. Different geographies, one shared question: how technology actually gets built and scaled in this part of the world.

That readership has clarified what this publication is.

Asia Tech Lens is defined less by where the stories come from than by how we approach them. We look at Asia the way a builder would—execution, trade-offs, and the systems beneath the headlines and explain what we find in a way that travels across markets.

The direction is set. The next year is about going deeper.

Thank you for being part of our journey. Here’s to the next chapter.

Much of the discussion around AI in cybersecurity has focused on faster, more scalable attacks. But that framing is incomplete. As AI shortens the gap between a weakness being found and being exploited, the bigger question for operators is what happens next. Once core systems are hit, can they regain control and keep the business running?

There is a gap in recovery. Veeam’s 2025 ransomware research found that nearly 7 in 10 organizations experienced at least one cyberattack in the previous year, but only 10% recovered more than 90% of their data, while 57% recovered less than half.

Many organizations believe they are ready because they rely on plans, backups, and recovery targets. In practice, those do not guarantee that a critical system can be restored under real conditions.

As the window to respond shrinks, the problem shifts. Restoring systems alone is not enough when identity, access, backups, and dependencies are uncertain. Operators also have to rebuild trust in what comes back online.

AI Is Shrinking The Time Between Vulnerability and Compromise

AI is making it easier to surface vulnerabilities across systems and applications. In Singapore, cybersecurity awareness is relatively mature. Still, AI is already changing how enterprises think about exposure, according to Joey Lim, Country Manager at Exclusive Networks Singapore, a global cybersecurity specialist.

“On the ground, we see a shift from periodic security assessments toward a more continuous posture,” Lim told Asia Tech Lens. “Organizations are asking harder questions about their attack surface, not just what they know about, but what they don’t. And that’s the right instinct.”

That compression changes which capability matters most. When the window between vulnerability and exploitation was measured in weeks, detection and patching kept most incidents from reaching the recovery phase. As that window shrinks, more incidents will get through. Recovery stops being the fallback—it becomes the front line.

That leaves operators with a harder question. When something gets through, can they respond quickly, regain control, and recover before the damage spreads?

The Real Bottleneck Is Recovery

Backups can make organizations feel safer than they are. A completed backup job shows that data exists somewhere, but it does not prove the business can recover.

For Gareth Russell, Field CTO, APAC, at Commvault, the starting point is no longer how fast an organization can restore systems. “In a cyber incident, speed without trust is a huge risk,” he told Asia Tech Lens.

The more important question is how quickly a company can identify a known clean state, reestablish trusted control, and bring back a service it can rely on. That is the difference between trusted recovery and simply powering systems back on.

Traditional metrics such as recovery time objective (RTO) and recovery point objective (RPO) still matter, but Russell said they often reflect intent rather than reality. What matters most is whether organizations can recover a clean, usable service end-to-end without reintroducing the threat.

“When I talk to CIOs and CISOs about recovery readiness today, we look at things like time to clean recovery, coverage of immutable and verified data, the ability to regain control of identity systems, and whether recovery has been tested under realistic conditions,” Russell said.

Across Asia, Commvault has found that 85% of organizations have incident response plans, but only 30% test all mission-critical workloads. When a real incident hits, those plans often do not hold, and recovery takes longer than expected.

The gap shows up in execution. Restoring a database or application is one thing. Getting the business running again during a real incident is another.

Recovery, in this sense, means testing the full process before a real incident leaves no room to guess.

Identity Is Where Recovery Often Breaks

If identity is compromised, recovery can’t start with simply restoring workloads. The organization first has to decide who and what can still be trusted. Otherwise, bringing systems back may also bring back the attacker’s access. Russell said identity failure changes the nature of recovery.

“In most incidents, it is not just that access is lost, it is that you cannot trust who or what is accessing anything,” he said. Federation fails, tokens may still be valid, service accounts may keep running, and organizations can lose control of the control plane.

“Teams often try to recover workloads before identity is stable. That is where things fall apart. If identity is not clean, nothing you bring back can be trusted,” Russell added.

Lim described how this failure unfolds in real time. A threat is detected, but the scope is unclear, so escalation is delayed. By the time leadership is engaged, critical hours have passed. The response team then realizes the incident response plan no longer matches the current environment. Systems have changed, contacts are outdated, and dependencies are unclear.

At the same time, attackers move laterally and target credentials. The team now faces a more dangerous problem. They no longer know which accounts, systems, or access paths can be trusted. Recovery slows as every action needs to be verified.

This is where the incident turns. Shut down too broadly and disrupt operations, or move too carefully, and the attackers may still be active. “This is often where a serious incident becomes a catastrophic one,” said Lim.

When identity is uncertain, recovery becomes a series of high-risk decisions made without clear visibility of what is safe.

Operators’ Takeaways

Do Now

• Prove recovery. Test one critical service end-to-end, with clean data, trusted access, and dependencies restored in the right order. Russell said the minimum proof is a full recovery under real conditions, where users can log in, and the service works.

• Map your exposure. Understand not just known assets, but shadow IT, cloud workloads, and third-party integrations. Lim said the assets that organizations do not track are often the ones that get exploited first.

• Harden identity. Reduce over-privileged accounts, clean up service accounts, and enforce consistent multi-factor authentication. Lim warned that many teams are still “working blind” when identity is compromised.

• Run real drills. Test recovery under realistic conditions, not just tabletop exercises. Recovery needs to be proven in execution, not assumed because plans exist.

Wait

• Hold off on more AI-security tools until recovery basics are proven. Lim said the bigger issue is not adding AI features, but whether the response is built for machine-speed attacks.

• Be realistic about in-house response. A full 24/7 detection-and-response capability is “not realistic for most” organizations.

• Delay framework rewrites if plans have not been tested. According to Russell, recovery readiness comes from repeatable execution, not documentation.

Avoid

• Do not assume backups guarantee recovery. Russell said untested and compromised backups are common failure points.

• Do not restore systems before identity is trusted. If identity is still compromised, restored systems cannot be trusted. “Break glass only works if it is genuinely separate, operationally ready, and trusted when everything else is not,” said Russell.

• Do not treat compliance as resilience. Compliance sets a baseline but “doesn’t give you resilience,” said Lim.

• Do not treat cybersecurity as only a technology problem. Communication and decision-making often fail first.

Related Reads On Asia Tech Lens

• AI Is Accelerating Cybercrime—and Southeast Asia Feels It First

  AI is making cybercrime faster, cheaper, and easier to scale across Southeast Asia. This piece explains why organizations have less time to detect, contain, and recover from attacks.

• Agentic AI Can Act. Singapore’s New Rulebook Says It Needs Guardrails

  As AI systems gain more autonomy, the risks shift from what they can generate to what they can do. We look at why permissions, oversight, and recovery planning matter before AI agents are allowed into real workflows.

• Two AI Phones. Two Access Models. One Critical Difference.

  AI access is becoming an operational risk, not just a product feature. This piece compares two approaches to AI control and shows why permissions, identity, and trust layers matter as systems become more automated.

• Why ByteDance’s AI Phone Hit a Wall

  When AI agents start acting across apps and services, security and accountability become the real constraints. The piece looks at why uncontrolled access can quickly turn AI capability into operational risk.

• India’s AI Push Is Real. Production Access Is the Constraint

  AI ambition only matters if systems can work under real operating conditions. This piece examines why production access, auditability, and incident ownership are becoming the true tests of AI readiness.

Nearly eight weeks after US-Israeli strikes on Iran effectively closed the Strait of Hormuz, most coverage has focused on oil prices, stranded tankers, and ceasefire negotiations. That is the visible story. There is a quieter one running underneath it.

The chemicals used to build battery cells trace back to petrochemicals that flow from Gulf refineries through the Strait of Hormuz. That supply is now effectively cut off. Prices have already moved sharply. But the more consequential impact arrives later, with a lag of several weeks between a feedstock disruption and a factory-floor shortage.

That lag is where the danger sits. Most battery and EV manufacturers across Asia are still running on inventory built before the crisis. The disruption feels distant, but it isn’t. And for procurement and operations leads at battery cell manufacturers and EV companies across the region, the window to act before that lag resolves is open now—and closing. If availability tightens, no amount of money resolves the problem quickly. The bottleneck becomes time.

The Chemicals Inside Every Battery Cell

The key components of a battery cell—the separator, the electrolyte solvent, and the binder—are all derived from petrochemicals. Most of these petrochemicals trace back to a raw material called naphtha. Asia’s petrochemical producers rely on the Middle East for 60 to 70% of their naphtha imports, most of which transits the Strait of Hormuz. When the strait effectively closed in early March, Gulf naphtha exports to Asia came to a halt.

The price signal arrived immediately. Naphtha was trading at around US$776 per metric ton before the disruption. It surged past US$1,000 at its peak and remains above US$870 today. But the more consequential number is the lag. When a feedstock like naphtha becomes difficult to source, the effect moves through the supply chain in stages—from refinery to petrochemical plant, from petrochemical plant to component manufacturer, from component manufacturer to battery cell maker. Asian petrochemical producers typically carry only a few weeks of naphtha inventory. The disruption began in late February. Factory-floor shortages were already visible by mid-March, and continue to deepen.

The problem is compounded by substitutability, or the lack of it. Switching to a different supplier or material isn’t just a procurement decision. It’s an engineering call as well. As Maithri Venkat, Battery Cell Technical Specialist at Lucid Motors, puts it, every single change to a cell must be scrutinized, because even a minor process change from a supplier can have unintended and severe impacts downstream. That scrutiny takes time that procurement teams don’t control. For operators who haven’t started that process, the window is already narrowing.

How Exposed You Are Depends on Where You Sit

South Korea is the most directly exposed manufacturing base in Asia to the current disruption. The country imports roughly 70% of its crude from the Middle East, with most of that transiting Hormuz. The impact is already visible on the ground: Yeochun NCC declared force majeure on naphtha supply in early April, with Lotte Chemical, LG Chem, and Hanwha Solutions all subsequently notifying customers of potential supply disruptions. For South Korean battery cell manufacturers sourcing petrochemical inputs from these suppliers, this is no longer a future risk. It is a current one.

China’s exposure is different in character but no less serious. Analysts estimate around 45% of China’s oil imports transit Hormuz—lower than South Korea’s share, but compounded by a geopolitical dimension that makes it harder to manage. In late March, Iran granted selective passage to vessels from a handful of nations including China, framing it as a diplomatic gesture toward non-hostile states. That arrangement has since collapsed. That brief window of selective passage was never a durable solution, and no operator should plan around it now. For Chinese manufacturers, the risk is not just cost inflation. It is that a deeply optimized supply chain is less flexible than it looks when upstream access turns unstable.

Japan moved earliest. Mitsubishi Chemical began cutting output within days, and Mitsui confirmed it was actively sourcing naphtha from non-Middle Eastern suppliers. That process costs more. Alternative naphtha from US Gulf Coast or Southeast Asian refiners carries a meaningful premium. But Japan’s early action bought optionality. If the disruption extends, Japanese manufacturers are already positioned. If it resolves sooner, they wind down the alternative contracts.

Across all three markets, the dividing line is simple: the manufacturers who know where their supply chain touches the Gulf are managing a problem. Those who don’t are walking into one.

The Decision Window

For procurement and operations leads across the region, one decision remains open, but not for long.

The choice is this: lock in alternative petrochemical supply now, at a visible premium, or hold and wait for Hormuz to normalize. The case for waiting has largely fallen away. Daily transits have collapsed back to near zero, against a pre-war average of 135. The ceasefire announced on April 8 collapsed within days, with Iran re-closing the strait on April 18 after the US refused to lift its blockade of Iranian ports. QatarEnergy has extended force majeure through at least mid-June 2026. Commercial normalization is not expected before July at the earliest, and only if US-Iran peace talks produce a concrete agreement.

The asymmetry is worth stating plainly. If you lock in alternative supply and the strait normalizes next month, you absorb a premium you can recover from. If you wait and the disruption extends through the second quarter, you face a shortage that money cannot solve quickly—because the constraint is not price, it is time. Production stoppages caused by input unavailability are not fixed with a larger purchase order. They require months of requalification work that should have started earlier.

The signal to act is simple: talks have stalled, the strait is shut, and your suppliers are starting to tell you they can’t guarantee delivery—not just that prices are up.

The Hormuz strait will reopen. When it does, the vulnerability it exposed will still be there. Asia’s battery and EV supply chains were built on Gulf petrochemicals long before this crisis. The manufacturers who use this window to map that exposure and build alternative sourcing relationships are fixing an assumption that was always fragile.

More from Asia Tech Lens

When Battery Economics Shift: What Gets Stress-Tested in China’s EV Strategy?
A look at what breaks when battery assumptions stop holding.

Not Just EVs: China Leads the World in Battery Production
A battery-first read on how chemistry, materials, and manufacturing scale shape competitiveness.

VinFast Pivots to Southeast Asia. But Can It Outrun BYD in Indonesia?
A piece on EV competition shaped by localization, industrial constraints, and execution risk.

The Chip War’s New Reality: A View from the Crossroads
A read on how geopolitical pressure reshapes commercial supply chains.

Asia’s High-Stakes Chip Game: What To Do When You’re Caught Between Washington and Beijing
A read on how geopolitical pressure turns cross-border chip dependency into an operating problem for Asian firms caught between the US and China.

Vietnam’s 5G network is expanding quickly. For operators building industrial operations in the country, the harder question is not whether the network will reach them—it almost certainly will—but whether the infrastructure underneath it can be trusted over time.

As Vietnamese telecom operators roll out 5G using a mix of domestic and Chinese vendors—including Huawei and ZTE—businesses are taking on a structural risk that goes beyond connectivity. Early infrastructure choices are shaping data exposure, compliance posture, and switching costs for years to come. Once systems are built on top of a network stack, changing course later gets expensive and complicated, especially if regulation tightens or vendor compatibility becomes a problem.

For manufacturers, logistics operators, and industrial infrastructure players in Vietnam, this is the question that should be driving infrastructure decisions right now—not rollout speed, not coverage maps, but whether the vendor stack underneath the network is one they can build on without regret.

The Rollout Is Real. Enterprise Uptake Is Not

Vietnam’s 5G rollout is moving fast by any regional standard. According to the Vietnam Telecommunications Authority, the country had installed nearly 40,000 base transceiver stations as of February 2026, covering about 90% of the population and serving close to 23 million subscribers. Independent benchmarks from OpenSignal and GSMA broadly support the pace of expansion, though public estimates of enterprise adoption remain limited.

But enterprise uptake has been slower to follow. Private 5G networks—often positioned as the business payoff of the technology—have yet to take off in a meaningful way. Deployment in specialized settings like seaports and industrial facilities has been particularly slow, with businesses remaining cautious over technical integration challenges and regulatory issues such as spectrum allocation. Citing comments from former deputy minister of post and telecommunications Mai Liem Truc in a January 2026 Vietnamnet Global report, private networks were expected to become a major revenue driver for telco operators, but consumer users still account for most demand.

The Real Risk Is the Vendor Stack

With Vietnam expanding its 5G rollout, vendor choice is becoming the more consequential question. Telco operators Viettel and VNPT have signed 5G equipment deals with Huawei and ZTE, while Mobifone is exploring similar partnerships. Cost is clearly part of the rationale—Reuters reported that Vietnamese officials described Chinese telco equipment as reliable and cost-effective.

But for operators building on top of this infrastructure, the trade-offs go well beyond price. European officials have warned that the involvement of Chinese vendors in Vietnam’s advanced networks could raise questions for foreign investors about data security and the reliability of the underlying infrastructure.

For operators running data-heavy industrial operations, this translates into three specific risks.

Compliance exposure.

Vietnam’s own regulatory environment is tightening. The Personal Data Protection Decree (Decree 13/2023), effective since July 2023, requires cross-border data transfer impact assessments filed with the Ministry of Public Security, and the Law on Personal Data Protection (Law 91/2025) that took effect in January 2026 introduces revenue-based penalties—up to 5% of annual revenue for cross-border transfer violations. Operators whose systems sit on networks built with vendors flagged as high-risk by their trading partners face a compounding problem: Vietnam’s own data rules are getting stricter at the same time that the EU and US are tightening restrictions on Chinese-origin ICT infrastructure. The EU’s 5G Cybersecurity Toolbox has driven several member states to restrict or exclude high-risk vendors from core and sensitive network functions, and the European Commission has signaled it may make these measures binding. In the US, the ICTS supply chain final rule (effective February 2025) gives the Commerce Department broad authority to prohibit transactions involving ICT from foreign adversaries. An operator in Vietnam working with a European automotive OEM or a US logistics partner may find that vendor choices made at the network layer create compliance friction they did not anticipate.

Interoperability constraints.

Integrating with global systems, cloud providers, and cross-border platforms becomes more complex when the underlying network infrastructure is built on vendor-specific architectures. Open RAN—the set of standards designed to enable multi-vendor interoperability—is gaining momentum globally, with operators like AT&T targeting 70% of RAN traffic on open platforms, and initial deployments appearing in Southeast Asia through the Orex SAI–Surge rollout in Indonesia. But Vietnam’s 5G buildout has largely followed the traditional single-vendor model. For operators who need their Vietnamese sites to integrate cleanly with global cloud infrastructure, enterprise platforms, or cross-border data systems, a proprietary network stack can mean additional integration layers, workarounds, and dependencies that compound over time.

Switching costs.

Once operations are built on top of a network stack, moving away from it requires reconfiguring systems, vendors, and workflows across the entire dependent layer. The scale of this problem is not hypothetical. In the US, the FCC’s “rip and replace” program to remove Huawei and ZTE equipment from domestic networks generated reimbursement claims of over $5 billion against an initial appropriation of $1.9 billion—and that was for relatively small carriers, not large-scale industrial operations. For operators in Vietnam who build factory automation, logistics tracking, or port management systems on top of a 5G network stack, the cost of switching later would involve not just the network equipment itself but the entire operational technology layer built above it.

Operator Takeaways

• Do Now

  Focus on targeted 5G deployments in controlled settings where the use case is already clear—smart factories, seaports, airports, logistics hubs. These are areas where Vietnamese operators are already testing private networks and where value is easier to measure. Where possible, negotiate flexibility into infrastructure decisions: vendor diversification clauses, interoperability requirements, or exit options that reduce long-term lock-in risk. Wait too long, and operators may lose the chance to shape early network partnerships in industrial zones where infrastructure terms are already being set.

• Wait

  Hold back on scaling private 5G too broadly across sites until integration maturity, enterprise demand, and regulatory clarity improve. Vietnam’s enterprise rollout has been slower than expected, especially outside pilot environments. Push too quickly beyond pilots, and operators risk managing parallel legacy and 5G systems simultaneously—higher costs, engineering teams stretched thin, and uncertain returns on infrastructure that may need to be reworked later.

• Avoid

  Do not treat fast rollout or early 6G ambition as a signal of full industrial readiness. Operators who over-commit now risk locking into a vendor stack that becomes a liability when compliance requirements tighten, when global partners demand higher standards for data handling and interoperability, or when investment committees ask why the infrastructure cannot be adapted without a seven-figure rearchitecting program. The US rip-and-replace experience showed what happens when vendor choices made for cost reasons become a compliance problem years later. In Vietnam, the same dynamic is forming—but operators still have time to build with flexibility if they act now.

Vietnam’s 5G network is moving ahead, and the opportunity for industrial operators is real. But the opportunity comes with a structural question that rollout statistics cannot answer: whether the vendor stack underneath the network is one that operators can build on for the long term—or one they will eventually need to build around. The smarter move is to deploy selectively, negotiate flexibility, and treat the country’s longer-term ambitions, including 6G, as a signal of direction rather than a decision point.

More from Asia Tech Lens

OpenClaw in China: What the AI Agent Frenzy Actually Means for Enterprise Deployment in Asia
Why fast enterprise adoption can hide deeper integration, dependency, and lock-in risks.

Two AI Phones. Two Access Models. One Critical Difference
How underlying system design shapes trust, control, and the long-term risks of technological dependence.

Why Smartphone Prices Could Rise in 2026 as RAM Costs Surge The AI phone race is also a hardware story. As memory costs reshape device pricing and features, the pressure is exposing how upstream component dependence can ripple through the wider tech stack.

US vs China AI Showdown: How Southeast Asia Is Quietly Choosing Open-Source Over Closed Model
How regional operators are making stack decisions based on cost and usability, even when those choices carry longer-term strategic consequences.

Can South Korea Replicate Its K-Pop Success in AI Chips?
Why geography, supply-chain resilience, and strategic dependence matter as much as technical performance when countries try to build alternative tech stacks.

If you are a manufacturing or operations director in Southeast Asia, there is a reasonable chance someone has recently put China’s 15th Five-Year Plan in front of you as evidence that now is the time to move on robotics. The pitch is not wrong, exactly. But it is incomplete in a way that could cost you a capex cycle.

China’s plan frames industrial automation and humanoid robots under the same umbrella—”new quality productive forces”—governed by the same political language, eligible for the same funding mechanisms, and pointed at the same national transformation goals.

The framing is deliberate and effective at the macro level. At the procurement level, it creates a specific problem: the plan draws no meaningful line between what is deployable today and what is still being figured out.

That line is the operator’s job to draw.

The Stack That’s Ready

In 2024, China installed 295,000 industrial robots—54% of the global total—with a world record of 2 million units working in factories. Domestic brands now supply 58.5% of that market, up from 31.4% in 2020.

The ROI case for industrial robotics is well-established, with cobots—collaborative robots designed to work alongside humans on the same floor without safety caging—typically offering a payback period between 12 to 18 months.

For a manufacturing director, this is a procurement decision with quantifiable ROI available now. The 15th FYP is confirmation that China will keep compounding this advantage.

But “ready” does not mean frictionless.

Integration remains the constraint most operators underestimate. Retrofitting robotics into legacy production lines requires compatible control systems, reliable integrators, and carefully planned downtime. A cobot that performs well in a controlled demo still needs calibration against the variability of real factory inputs.

Trade policy is another variable emerging in the region. Some Southeast Asian markets are tightening scrutiny around Chinese industrial imports in sensitive sectors, and tariffs or certification delays can stretch deployment timelines. Even where trade barriers remain low, labor relations can complicate adoption when automation arrives in labor-intensive industries.

Even with a stack that works, successful deployments still depend on execution. The constraints are real but manageable. The question for most operators is not whether to engage this stack, but how quickly.

The Stack That Isn’t

The plan’s unified language covers humanoid robots with the same confidence. The deployment reality is categorically different.

Chinese firms shipped roughly 90% of the world’s humanoid robot units in 2025, according to research firm Omdia, led by AgiBot at 5,168 units and Unitree at over 4,200—though Unitree’s own reported figures put its total at 5,500. The numbers sound significant until context is applied: global humanoid robot shipments reached just 13,317 units in 2025 in total, and it remains unclear how many of those represent genuine commercial sales versus demonstration models or pilot deployments. The headline showcase deployments—UBTECH at Zeekr, for example—involve materials handling and quality inspection: tasks that purpose-built industrial arms already handle, often better and for less.

The technical constraints are specific and unresolved. Battery life is the hardest wall. Agility’s Digit, currently one of the first commercially deployed humanoids in the world, operates in warehouse environments with battery life reaching approximately four hours depending on task intensity. A standard factory shift runs eight to twelve hours. The gap requires either a charging rotation system or continuous human supervision, both of which erode the labor-saving economics being sold.

This is where the timeline question matters.

When operators say humanoids might become relevant in “18 to 24 months,” the claim is not about hype cycles. It refers to three concrete thresholds: first, whether machines can sustain something close to a full factory shift; second, whether unit economics support pilot-to-production conversion; and third, whether the vendor ecosystem stabilizes enough to trust multi-year procurement decisions.

Until at least some of those thresholds are crossed, humanoids remain a monitoring exercise rather than a deployment plan.

What Happens When Operators Don’t Separate Them

The risk is not theoretical. In November 2025, K-Scale Labs—a humanoid startup that had received over $2 million in orders, and launched two products—collapsed on the verge of mass production. The CEO cited failed financing and an out-of-control burn rate.

K-Scale is one data point, but the NDRC’s warning that over 150 humanoid companies are now competing in China on largely identical products suggests the consolidation risk is sector-wide.

For a director who has signed an MOU or committed integration resources to a humanoid vendor, this is vendor survival risk, not just product maturity risk.

The second failure mode is subtler. Key suppliers in China’s humanoid robot supply chain are making preemptive investments in production capacity ranging from 100,000 to 1 million robot-equivalent units annually—despite no company having confirmed large-scale orders or a clear production timeline.

A manufacturing director who runs a premature pilot off the back of that supply-side confidence doesn’t just lose the capex, they make it harder to get the next automation proposal approved internally.

That’s because when companies run pilots around immature technology, the cost rarely ends with the pilot budget itself. Engineering teams divert time from deployable automation projects to support experimental trials. Management attention shifts toward solving integration problems that cannot yet be solved. And when the pilot inevitably stalls, the experience often hardens internal skepticism toward robotics more broadly. A failed experiment becomes the board’s reference point the next time someone proposes automation spending, even if the next proposal involves technology that is already commercially proven.

Two Decision Tracks, Not One

The 15th FYP groups these two stacks together because that serves China’s industrial strategy. It does not resolve the decision a manufacturing director in Southeast Asia is actually facing.

Proven industrial automation in structured environments is a procurement decision for now. The ROI is quantifiable, the supply chain is mature, and the sourcing window is open. Operators who wait for the humanoid narrative to settle before moving on this risk losing ground to competitors who already have.

Humanoid robotics is a market to monitor, not a capex line item, for at least the next 18 to 24 months. Assign someone to track battery milestones, safety certification progress, and vendor order books.

Watch for a humanoid platform sustaining something close to an eight-hour operational cycle in a real production environment, not a staged demo. Watch for second-tier manufacturers—not just headline startups—entering serial production with confirmed customer orders. And watch for safety certification frameworks that allow legged robots to operate routinely in shared factory workspaces. Note when the constraints actually resolve. Don’t budget ahead of that.

Beijing’s plan is a credible signal of long-term strategic commitment to both. It is not a procurement signal for both. The plan doesn’t draw that line.

That’s your job.

Apple’s once-unified iOS ecosystem is splintering into a geography-dependent maze of permissions and protocols, driven by a patchwork of regional mandates—from the European Union’s Digital Markets Act (DMA) to Japan’s Mobile Software Competition Act (MSCA). Even in markets where formal legislation hasn’t yet passed, the pressure is forcing a retreat; in March 2026, Apple preemptively slashed its App Store commission in China to 25% following supposed discussions with the State Administration for Market Regulation (SAMR).

By breaking this model, regulators have handed developers a marginal discount in exchange for a significant administrative burden, leaving them to decide if the thin profit increases are worth the weight of managing their own payment infrastructure and security vetting. In the EU, this means navigating Core Technology Fees; in Japan, it involves third-party billing protocols; and in China, it requires balancing the new 12% “Mini App” rate against the technical requirements of the Declared Age Range API.

Ultimately, the true measure of success won’t be found in commission fee reductions, but in whether the market remains functional once the friction of fraud and compliance is fully priced into the user experience.

The Illusion of Savings

This connection between market functionality and institutional behavior is perhaps best viewed through the lens of economic incentives. According to Lazar Radic Boskovic, a PhD in law and digital competition expert, Apple remains entitled to charge for access to its ecosystem regardless of the regulatory framework.

“When regulators constrain one monetization channel—IAP commission, for example—Apple has strong incentives to reprice elsewhere—different commissions, per-install charges, developer services fees, and entitlements,”

he explained in an email interview with Asia Tech Lens.

This suggests that some laws change the form of the rake more than the existence of the rake itself, leading to what Lazar Radic Boskovic calls “compliance by reclassification.” He argues this is the natural response of a company under constraint: when a primary price is regulated directly, the business inevitably moves toward add-on fees and more complex pricing structures.

Consequently, the fee reshuffling in the EU and Japan functions as a technical pivot rather than a financial gain. This unbundling of the App Store experience into individual, billable components establishes a regulatory blueprint that enterprises should expect to see repeated across other vertically integrated platforms.

However, the redistribution of labor—including the hidden overhead of self-managed hosting, security, and payment processing—quickly erodes the savings of lower commissions. In this landscape, margins remain essentially flat while platform authority shifts from an inherent right into a series of negotiated, jurisdictional arrangements.

That means, in practice, the true expense lies not in the commission delta, but in the operational weight enterprises inherit as universal platform guarantees fragment into regional mandates.

The Operational Breakdown: Managing The Unbundled Platform

This transition is not merely a legal abstraction; it can be an immediate operational turning point for enterprise IT. When security is “unbundled” from the hardware, the burden of proof shifts from Apple to the enterprise.

For enterprises subject to the DMA and MSCA, compliance demands a rewrite of four core pillars: MDM architecture, BYOD boundaries, audit scope, and procurement strategy.

MDM & Configuration

Previously, Mobile Device Management (MDM) on iOS was largely about enabling features. But in a post-DMA and post-MSCA environment, MDM is a defensive shield used to disable regional openings.

Recent iOS updates have introduced specific MDM keys that allow administrators to prohibit the installation of alternative app marketplaces. By using these keys to manage Manual Configuration Profiles, enterprises can navigate region-specific functional changes, effectively splitting the ecosystem into “Managed iOS” and “Consumer iOS.” Managing this landscape demands an MDM architecture sophisticated enough to toggle keys based on an employee’s precise legal location, which in turn drives a granular approach to audit logs and financial compliance.

This means managing a fleet of devices is no longer about giving every unit the exact same rules. Instead, MDM profiles must now be conditionally applied based on user identity and legal jurisdiction, with every override logged against the specific regulatory framework it enforces.

Moreover, since software companies now charge for individual features instead of one set fee, the audit log has become a billing verification tool. This now requires IT managers to track specific actions and payments to make sure their bills are correct.

BYOD & The “Managed Open In”

The “Managed Open In” protocol has long been a standard for Bring Your Own Device (BYOD) security, ensuring work data stays in work apps. However, the introduction of third-party marketplaces breaks the closed ecosystem assumption that underpinned this protocol.

When employees download tools from third-party marketplaces on personal devices, they bypass Apple’s rigorous malware scanning and “Managed Open In” protections, creating significant gaps in data isolation and patch management. Moreover, these tools from third-party marketplaces may use alternative frameworks or private APIs that haven’t been audited for how they interact with the system clipboard or file providers. This increases the risk of leakage, where corporate data is accidentally moved into an unvetted environment.

As a result, enterprises must now grapple with the decision to either ban all third-party marketplaces on BYOD devices or accept that corporate data may be put at risk in apps whose provenance hasn’t been verified by Apple’s traditional App Review team.

Recommended Action: IT admins must now manually deploy specific MDM keys, such as allowMarketplaceAppInstallation, to lock down corporate devices. For a regional office in Singapore, this key might be “Allow,” while for an office in Japan, it might be “Disallow with Exceptions,” creating a fragmented security posture across the same company.

Audit Scope

For enterprises maintaining ISO 27001 compliance certifications, iOS was previously treated as an “inherited control.” Auditors accepted that because Apple managed the App Store, the platform was secure by default.

But by allowing alternative payment processors and marketplaces, the “scope” of a corporate audit expands. For example, if a financial services company in Japan uses an app that utilizes a third-party payment link—permitted under MSCA—that payment gateway now enters the firm’s audit scope.

This expansion can be problematic because it replaces the single “Chain of Trust” formerly guaranteed by Apple with a fragmented web of unvetted third-party providers. Auditors must now verify the encryption, access controls, and data handling practices of these external entities, saddling the enterprise with a significant operational and financial burden of auditing every link in their new, jurisdictional supply chain.

Apart from that, CISOs must now develop internal “Approved Marketplace Lists,” effectively building their own mini-App Stores. These administrative tasks require vetting overhead, which is the hidden cost of Apple’s commission discount.

Recommended Action: IT and Compliance teams must now vet every third-party marketplace and payment processor used by employees. This involves deploying a Mobile Threat Defense (MTD), often integrated with Unified Endpoint Management (UEM), that can automate security checks on every app and payment service accessed on managed devices. If a service doesn’t meet the specific legal and security standards for that user’s current country, the software can automatically block the connection.

Procurement

Procurement was once a volume-discount exercise. It is now a compliance architecture decision.

The traditional concept of “Global Procurement” is changing as the regulatory gap between regions widens. A device purchased in a non-regulated market may lack the system-level APIs required to run certain localized enterprise apps that rely on alternative frameworks. In this example, if a Tokyo-based employee needs a specialized Japanese enterprise app that is only distributed via a local third-party marketplace, a “Global SKU” iPhone purchased outside of Japan may refuse to install it.

This then forces enterprises to pivot toward “Sovereign Fleet Management,” where procurement is tied strictly to the legal jurisdiction of the employee, not the lowest global hardware price.

Using a non-regulated device in a regulated market exposes enterprises to compliance failures. If a Tokyo-based enterprise provides an employee with an iPhone sourced from a US procurement contract, for example, that device will lack the jurisdictional logic to trigger the MSCA-mandated selection screens.

Recommended Action: IT departments must mandate region-specific SKUS in their purchase orders to ensure the hardware includes the built-in digital permission required to trigger local features, such as the mandatory browser-selection screens in regulated markets. This means shifting from “buy 1,000 iPhones” to “buy 1,000 Japan (J/A) models,” ensuring the device identity matches the laws of the country where the employee works.

The Restrictive Default

Apple faces an engineering dilemma: how can it maintain a global codebase without falling into a state of compliance failure. The result is a shift toward regulatory arbitrage, where Apple as the platform holder has incentive to move the default toward the most restrictive security settings in every country unless a local law specifically requires it to do otherwise. This allows Apple to minimize legal risks and maintain control over the platform.

The strategy, however, creates a state of jurisdictional isolation. Geofencing “open” features—such as alternative browser engines, third-party app stores, and external payment links—to specifically regulated zones like the EU and Japan leaves the likes of India, Singapore, and South Korea, among others, on the “closed” global baseline.

For enterprises operating across both regulated and non-regulated Asian markets, this legislative delay is not a reprieve but a source of operational complexity. They face an asymmetric fleet—some devices are more open than others, not because of company policy, but because of where the phone was activated.

The burden of “platform integrity” now shifts from Apple to the enterprise. And IT teams must implement continuous attestation to ensure that a device from a regulated jurisdiction doesn’t compromise the network of an office operating on the global security baseline.

This results in “compliance by reclassification,” where the enterprise’s primary expense is no longer the device itself, but the massive operational overhead required to audit and manage a fleet that is no longer uniform.

Currently, Apple has no commercial incentive to provide the “unbundled” features mandated by Japan’s MSCA—such as alternative browser engines or third-party payment links—to other Asian markets. Because these features are built as “jurisdictional entitlements” Apple can technically geofence them. Without the threat of legislation, Apple will likely maintain its traditional closed ecosystem to protect its original commission structure and security branding.

After all, curation, security, and a clear allocation of responsibility are fundamental features that define Apple’s products and services, as Lazar Radic Boskovic notes.

“Many consumers choose Apple precisely because they value a single, reliable intermediary and know where accountability lies when something goes wrong,” he said.

If other Asian countries follow Japan’s lead, the region will not likely have a single set of rules. Instead, every country will have slightly different requirements for what Apple must allow.

Conclusion

The end of a global security baseline for iOS reveals that the most significant cost for the modern enterprise is not Apple’s commission, but the massive internal overhead of compliance and configuration. The use of third-party marketplaces and alternative payment processors force IT teams to assume direct responsibility for app provenance and data isolation. Enterprises can no longer rely on platform holders like Apple for security; they must instead conduct checks and audits of third-party providers and manage the manual MDM keys required to defend their corporate data.

In short, the real price of platform “openness” is the transfer of risk and responsibility from Apple’s engineers to the enterprise’s own balance sheet.

Enterprises must watch out for Asia’s rapidly shifting legislative map. Based on current legislative momentum, India could be one of the first to introduce similar “unbundling” mandates to Japan and the EU.

The country’s Digital Competition Bill is the most notable looming threat to closed ecosystems because it specifically targets “core digital services”—including operating systems and app stores—and would likely force Apple to allow third-party marketplaces and alternative payment systems. Following a series of “gatekeeper” investigations by the CCI, the bill is now reportedly a top priority in the 2026 legislative queue.

While technically “Oceania,” Australia has signaled it will introduce legislation in 2026 that mirrors the EU’s DMA. The Australian Competition and Consumer Commission (ACCC) has been conducting a multi-year inquiry that is expected to conclude later this year with formal legislative recommendations to the Treasury. The Commission’s fifth and seventh reports specifically recommend new “service-specific” rules to address the power of mobile OS providers and app stores.

This move would likely turn the “Australia-New Zealand” corridor into an “open” zone, further fragmenting the fleet for multinational enterprises in the South Pacific. Australia exerts significant regulatory gravity within the Asia-Pacific; the enactment of such legislation serves as a bellwether for shifting digital mandates across the region.

As iOS becomes more fragmented, enterprises should be prepared to allot a 6 to 12-month lead time for adopting a new fleet management model. This process is a structural overhaul, not a configuration update, which begins with a shift in perspective: IT directors must stop managing devices as uniform hardware and start treating them as “Regulatory Units.” They must collaborate with Legal to map every device in the inventory to its specific jurisdictional mandate and Finance to reconcile platform invoices against actual regional usage. Ultimately, success depends on treating every device as a specific legal commitment to the country where it is used, requiring IT, Legal, and Finance to work as a single unit.

Go Deeper on Asia Tech Lens

Why ByteDance’s AI Phone Hit a Wall: Security, Fair Play, and the Economics of Attention

ByteDance launched an AI assistant on ZTE's Nubia M153 that could operate across apps by reading the screen and tapping like a human. WeChat, Taobao, and Alipay pushed back within days. The piece examines why platforms draw hard lines when an outside agent starts executing inside their ecosystems—and what guardrails need to exist before phone-level AI goes mainstream.

AI Phones Explained: The Two Models Shaping the Next Smartphone Battle

Samsung's Galaxy S26 with Gemini and ByteDance's Doubao phone both promise AI that acts on your behalf—but they gain device access in fundamentally different ways. One works through approved APIs and permissions; the other drives the screen like a user. The gap between those models determines what scales, what breaks, and what enterprise IT will eventually need to govern.

The Dependency Economy of AI: Sovereignty, Chips, and Global Chokepoints

A 25-country analysis of national AI strategies reveals that only the US and China run anything close to a full-stack AI ecosystem. Everyone else is managing dependencies they don't control—on GPUs, cloud infrastructure, and model APIs. The piece argues enterprises should treat AI like a geopolitically exposed supply chain, mapping dependencies and stress-testing for export controls and vendor disruptions.

The Prediction Market Boom Is Real. In Asia, So Is the Ban Hammer

Prediction markets are surging globally but hitting legal walls across Asia. Singapore, Taiwan, Thailand, and China have blocked access to platforms like Polymarket, classifying them as illegal gambling. Local platforms are emerging through offshore entities with opaque headquarters. The piece traces a regulatory cat-and-mouse dynamic where demand is real, but the legal landscape offers no clear path to legitimacy.

Editor’s Note: The frenzy around OpenClaw dominated your news feed this week. But behind the noise, several other stories were quietly unfolding in Asia.

A Chinese appliance giant committed another 60 billion yuan to AI and robotics. A Singapore startup raised $20 million to build enterprise AI models in-house. And major chip packaging expansions in Taiwan and the Chinese Mainland showed that capacity is still racing to catch up with AI hardware demand.

Taken together, these moves point to a more structural question: who will build Asia’s AI infrastructure, and who will control it?

Midea Pledges $8.7 Billion on Robots and AI

Midea Group, the Chinese appliance conglomerate that owns German robotics maker Kuka, announced it will invest 60 billion yuan (roughly $8.7 billion) in AI and robotics R&D over the next three years. That roughly matches what the company spent on R&D over the previous five years combined.

This isn’t a new direction. Midea has been building toward this since acquiring Kuka in 2017. What’s new is the scale of the commitment, and the fact that its six-armed humanoid robot, Miro U, is already being used on a production line at its Wuxi factory, where the company says it improved changeover efficiency by 30%.

Signals To Watch

• Midea has said it will roll out its AI factory model globally. If that happens at pace, it sets a new efficiency baseline that competitors and suppliers will be measured against.

• Other Chinese appliance and manufacturing giants are making similar moves. Watch for a cluster of similar announcements in the coming months.

• Miro U is already on the factory floor. The gap between announcement and deployment is shrinking, which means that entities treating humanoid robotics as a 2028 problem may need to revisit that timeline.

New AI Packaging Capacity on Both Sides of the Strait

Two significant chip packaging developments this week, both aimed at the same hardware demand wave, but serving very different strategic purposes.

In Taiwan, ASE Holdings, the world's largest semiconductor packaging and testing company, broke ground on a new $540 million facility in Kaohsiung dedicated to high-end AI and HPC packaging and testing. Construction starts this year, but completion is targeted for Q2 2028.

In Chinese Mainland, JCET, China’s largest chip packaging and testing company, opened its new automotive and robotics chip packaging plant in Shanghai. It is one of China’s first facilities dedicated specifically to automotive-grade and robotics chip packaging and testing, with AI-assisted defect detection and full production traceability built in.

Together, the two announcements show that advanced packaging capacity is being built in parallel across geographies because AI demand is outrunning what any one region can supply.

Signals To Watch

• ASE’s 2028 completion date is the number to hold onto. Procurement plans built around new AI packaging capacity coming online before then need to be pressure-tested.

• JCET’s plant is purpose-built for automotive and robotics grades, the same grades that China’s expanding robot production, including facilities like Midea’s, will require. A domestic supply chain is being quietly assembled.

• New capacity on paper does not mean available allocation. Watch how quickly both facilities reach full utilization before adjusting sourcing assumptions.

Singapore Bets on Its Own AI Models

Sapiens AI, a Singapore-based startup, raised $20 million this week to build large language models and enterprise AI tools in-house, targeting sectors including finance, healthcare, and government services. The funding will go toward model development, engineering hires, and computing infrastructure.

The company launched Agnes AI in April 2025. Reported user growth has been fast, with more than 5 million total users and 150,000 daily active users in under a year.

The bigger bet is strategic: if regulated enterprises in Asia grow more cautious about routing sensitive workloads through foreign model providers, local alternatives start to look less like nationalism and more like procurement logic.

Signals To Watch

• Agnes AI already has 5 million users and 150,000 daily active users after less than a year. The adoption numbers are the signal worth watching, not just the funding.

• Malaysia, Indonesia, and Singapore all have government-backed homegrown model programs already underway. Sapiens is the first private commercial raise of this scale in the region. Watch whether others follow with VC-backed bets.

• The data residency and API cost argument is gaining ground in regulated industries across Southeast Asia. This raise suggests investors believe the market is real, not just the narrative.

Also This Week: The OpenClaw Frenzy

We covered OpenClaw in depth earlier this week. But one thing shifted after we published.

Chinese authorities moved to restrict the tool’s use on office devices across government agencies and state-linked entities, including major banks, over security concerns. Bloomberg also reported that some employees were warned against installing it on personal phones while connected to company networks, and were asked to report prior installations for checks or removal.

Chinese AI and tech stocks slid on the news. Recent debutantes MiniMax and Zhipu fell more than 6%.

Read the full piece here: China’s OpenClaw Wave: Signal or Noise?

Takeaway

The common thread across all three stories this week is timing. Midea has robots on the factory floor, but its $8.7 billion commitment is a three-year pledge, not a delivery. ASE's new AI packaging facility in Kaohsiung will not be ready until 2028. Sapiens has users, but whether regulated enterprises will trust a homegrown model over an established foreign one is still open.

In other words: announcements are accelerating faster than operational readiness. The key question is not who is moving first, but who will be ready when the bottlenecks actually matter.

Sources

• SCMP: China’s Midea pledges US$8.7 billion for AI and robotics in pivot to automation

• Taiwan News: ASE breaks ground on third Nanzih site in Kaohsiung

• Futu Bull: JCET’s automotive-grade chip packaging and testing factory inaugurated in Shanghai

• Bloomberg: China Moves to Curb OpenClaw AI Use at Banks, State Agencies

China may already have much of the AI infrastructure it needs. Data centers across the country are still underused, with many running at roughly 30% capacity according to Reuters, even as Beijing pushes a new AI-heavy plan.

The real task now is absorption. Whether that surplus becomes useful infrastructure for industrial operators across Asia, or remains largely locked inside China’s own policy-driven ecosystem, will decide whether the plan reshapes the regional AI stack or simply deepens a domestic one.

The plan is built around that absorption problem. Beijing’s latest Five-Year Plan makes clear how policymakers intend to tackle that mismatch.

The plan calls for a unified national computing network, larger intelligent computing clusters, and tighter coordination of where computing power is built and used. In practice, that means focusing on initiatives like East Data and West Computing, designed to better match data center capacity with energy resources and demand.

The Compute Boom Mostly Stays Domestic

China’s surplus compute is unlikely to become a regional utility story. Data governance rules, AI security systems, and local vendor ecosystems mean most of this infrastructure will primarily support China’s own industrial AI deployment.

Beijing has made that goal explicit. The government’s new “AI+” push is meant to embed AI across sectors such as manufacturing, logistics, and healthcare to lift productivity across the economy.

However, there are a few narrow channels where the effects may spill over.

Industrial companies already running on Chinese technology stacks may see lower compute costs. This includes manufacturing systems using Chinese automation platforms, logistics infrastructure linked to Chinese ports and supply chains, and industrial software tied to Chinese vendors.

For most operators outside China, the impact will be more indirect than transformative. The Five-Year Plan is designed first to absorb China’s own infrastructure buildout, not to export compute capacity to the region.

Where The Stack Actually Shifts

China’s compute push will be felt first inside sectors already tied to Chinese technology ecosystems.

That matters because many multinational and Asian companies have operations in China or rely on Chinese suppliers, vendors, and industrial platforms. For operators already running inside Chinese vendor ecosystems, the first change may simply be cost. If China can put more of its surplus capacity to work, AI workloads should get cheaper inside Chinese vendor stacks. Port logistics shows the pattern. Many terminals already use Chinese automation systems for crane operations and yard management.

If computing gets cheaper, AI tools for routing, scheduling, and predictive maintenance could become easier to roll out on top of those platforms.

Data access may also start to shift for companies operating in China’s industrial systems. China’s push for a national data market may make it easier to organize and use industrial data inside the country. That matters for factory automation systems that already rely on Chinese industrial software.

These systems run on large volumes of operational data for quality inspection or error detection. Local governments are already experimenting with this model. Reuters reported that Jiangsu province alone plans 50 pilot AI applications in logistics and infrastructure, along with 186 smart production lines. That suggests these are some of the first places where Beijing wants AI to move from policy language into day-to-day operations.

Regulation is the third place where operators may start to feel change. Beijing is tightening AI oversight through new security and monitoring rules. For industrial systems linked to Chinese infrastructure, that could mean stricter compliance and reporting requirements. Energy grid management is one example, as tighter governance could shape how Chinese monitoring and optimization systems are deployed and run.

What Companies Should Do Now?

The key question is simple: where does the data stay and whose system is it running on? If the data is generated and used within one market, and the vendor stack is already Chinese, the workload is more likely to benefit if compute gets cheaper inside that ecosystem. That is why use cases like predictive maintenance, warehouse automation, and on-site quality inspection look safer for now.

Riskier bets are projects that assume China’s surplus computing will quickly become an open regional utility. Business models that rely on running cross-border AI workloads through Chinese infrastructure may run into barriers from data governance rules, security reviews, and platform restrictions.

The clearest thing to watch is whether China’s national integrated computing network starts to behave like a real market. One sign would be major cloud players like Alibaba Cloud or Huawei Cloud making compute pricing easier to compare across regions. Another would be clearer tools for moving workloads across provinces or easier access for companies outside China. That would suggest China’s compute system is becoming more unified, not just a patchwork of local clusters.

When The Story Changes

The most common mistake is assuming China’s AI infrastructure will evolve like the US cloud market, where platforms expanded globally and created a shared compute layer across regions.

China’s system is more likely to remain tied to domestic regulation and industrial policy. Treating it as a regional compute utility too early could lead companies to build systems that depend on infrastructure they cannot easily access.

For most operators outside China, the compute surplus is not yet a direct opportunity; it matters mainly for companies already tied to the Chinese supply chain or tech platforms.

The story changes if China turns its scattered infrastructure into a unified compute market, making cheaper AI capacity relevant across Asia.

That said, if you’re not already on Chinese tech stacks, this plan doesn’t change your priorities yet; if you are, the cost and compliance environment around your existing systems is about to move, and you should be mapping that exposure now.

Related Reading On Asia Tech Lens

• China’s Robot Spectacle Is an Industrial Strategy

  How Beijing used the Spring Festival Gala’s humanoid showcase to manufacture procurement cover and unlock budgets.

• OpenClaw in China: What the AI Agent Frenzy Actually Means for Enterprise Deployment in Asia

  China's current AI adoption wave is substantially a domestic story and why operators outside China risk expensive, premature commitments if they treat it as a universal signal.

• Inside China’s Bold Push to Build Humanoid Robots

  A ground-level look at how policy coordination, dense supply chains, and state-backed testbeds are compressing China’s path from demo to deployable.

• The Chinese New Year AI Gateway War

  How China’s Big Four used red packets and subsidies to manufacture habit formation around AI, and what happens when the incentives end.

• US vs China AI Showdown: How Southeast Asia Is Quietly Choosing Open-Source Over Closed Model

  How countries like Malaysia, Indonesia and Singapore are already running Chinese open-source models in production, and what that tells you about where Chinese tech stacks are actually taking root outside China.

On the morning of March 6, Tencent’s cloud engineers set up free installation booths outside their Shenzhen headquarters for OpenClaw, an open-source AI agent that can autonomously run tasks on a computer. Students stood alongside retired engineers. Someone had flown in from Hangzhou. Meanwhile, across the city, Taobao sellers were charging up to 500 yuan per remote install, with the busiest stores logging over a thousand orders. ByteDance, Alibaba, and other major cloud providers had already launched their own one-click deployment services. China’s own cybersecurity regulator had issued a formal warning about the same software just weeks before.

What OpenClaw Actually Does

OpenClaw started as a hobbyist GitHub project in November 2025 and within weeks had become one of the fastest-growing AI agent projects in the world. It links a large language model to tools such as messaging, browser control, file systems, scheduling, and command execution. In practice, that means it can handle tasks like filing reports, managing files, running shell commands, booking calendar slots, and writing code with limited human intervention.

What makes it notable is not that it replaces APIs altogether, but that in some workflows it can operate through the interface layer rather than relying on clean, purpose-built integrations. That matters in legacy-heavy environments, where AI projects often stall because every new deployment requires API work, permission redesign, vendor coordination, and system-by-system integration. OpenClaw changes that conversation by offering a way to work across messy stacks without waiting for all of that to be rebuilt first.

In China, that integration problem is especially acute because many large enterprises still run on fragmented IT estates built up over years of legacy systems, custom fixes, and weak documentation. Traditional agent deployment requires remapping those systems from the ground up. OpenClaw bypasses that but it is considerably slower. Each step takes 15 to 30 seconds compared to 1 to 3 seconds for a properly integrated agent. For enterprises blocked by integration costs, however, that is a price worth paying.

Reading the Subsidies Correctly

The frenzy did not go unnoticed by local governments. Shenzhen’s Longgang district, home to China’s first AI and robotics bureau, released a draft policy on March 7 proposing financial support of up to 2 million yuan (USD 276,000) for approved OpenClaw application projects, with larger commitments of up to 10 million yuan (USD 1.4 million) for more substantial ones, alongside free compute credits and discounted office space. Wuxi’s high-tech district followed two days later, offering between 1 and 5 million yuan (USD 138,000 to USD 690,000) for industrial applications including quality inspection and equipment maintenance.

In the same news cycle, China’s Ministry of Industry and Information Technology issued a formal warning that default or poorly configured OpenClaw deployments create serious exposure risks, and that public access and permissions must be tightly controlled. The Chinese state was simultaneously encouraging and cautioning against the same technology. That is not a contradiction to explain away. It is the most honest signal available about where OpenClaw actually sits: useful enough to promote, risky enough to warn against, and not yet resolved into either. Everything else in this piece follows from that.

These are district-level draft proposals, not central government policy. Longgang’s measures are still open for public comment until April 6. The goal is to attract developers and startups into a nascent ecosystem, not to signal that operators have tested and validated the technology. No one has signed off on a return on investment yet.

The cloud vendors are making a similar calculation. Alibaba Cloud, Baidu Intelligent Cloud, ByteDance’s Volcano Engine, JD Cloud, and Tencent Cloud all launched one-click OpenClaw deployment within weeks of the project going viral. Installation is free, but cloud compute, bandwidth, and API calls are chargeable, meaning every task OpenClaw runs on a cloud server generates a bill.

On March 9, several China cloud and software-linked stocks jumped around 20% amid policy support and OpenClaw enthusiasm. That looked more like a bet on infrastructure demand than on proven enterprise deployment revenue.

The Constraint Map for Regulated Industries

Access and Audit Gaps

OpenClaw includes some security controls and tool restrictions, but it does not natively provide the kind of enterprise-grade RBAC, centralized policy enforcement, and regulator-friendly audit architecture that banks, hospitals, and critical-infrastructure operators typically require. For firms in regulated industries, such as a bank that must demonstrate data access controls to its regulator, a hospital managing patient records, or an energy company operating under national security protocols…who accessed what data and when is a compliance requirement. And these are not gaps that better configuration will fix - they are baked into how the tool is built.

Security is Already a Flagged Problem

SecurityScorecard identified over 135,000 OpenClaw instances exposed to the public internet as of February 2026. A separate independent study found that 42,665 of those exhibited authentication bypass conditions, meaning they could be accessed without any credentials at all. Three critical security vulnerabilities have been formally catalogued in the software, each with publicly available attack code, meaning the tools to exploit them are already in circulation. One of them, even when OpenClaw is configured to run only on a local machine rather than over the internet, allows an attacker to remotely execute their own commands on that machine and intercept the access credentials the agent holds. A further six vulnerabilities cover issues including missing authentication and unauthorized file access. More than 800 confirmed malicious add-ons have been found within OpenClaw’s own plugin registry.

China’s Ministry of Industry and Information Technology had already moved. In February, it issued a formal warning that default or poorly configured deployments create serious exposure risks and that public access and permissions must be tightly controlled. The advisory was not issued in isolation. It landed in the same week as the district subsidy announcements.

The Continuity Risk

In February, OpenClaw’s founder announced he was joining OpenAI and that the project would move to an independent open-source foundation, with OpenAI committed to supporting it. There are no reports of a foundation being formally established. OpenAI has publicly backed the arrangement, but the company has its own complicated history with open source commitments, and is currently facing litigation over its transition from a nonprofit to a for-profit entity.

For operators considering dependency on this tool, that context matters. The governance structure that will oversee security patches, licensing terms, and the development roadmap does not yet exist in a defined form. Until it does, there is no clear entity to hold accountable if something changes.

The Data Exposure Problem

Data volume is a further concern. Each task OpenClaw completes generates a trail of processed text, measured in units called tokens. A single active session can exceed 200,000 tokens. A power user can consume up to 50 million in a day. That data moves through the underlying model, through the messaging platform OpenClaw connects to, and in many configurations through a third-party cloud server. For firms handling sensitive information, the exposure surface widens with every task the agent runs, with no guaranteed boundary around what leaves the organization.

The Rest of Asia Has a Different Calculation

For operators in regulated Asian markets outside China, the calculation is different. The subsidy schemes are district-level Chinese policy, not a regional endorsement. The one-click deployment push is also running through Chinese cloud providers, which means the speed, cost, and convenience of adoption are being shaped by a domestic ecosystem rather than by a model that travels easily across borders.

Part of what has made OpenClaw compelling in China is the cost structure around it. OpenRouter data in late February showed Chinese-developed models accounting for 61% of token consumption among the platform’s top ten models, pointing to a pricing and usage dynamic built on Chinese infrastructure, intense domestic competition, and Chinese regulatory conditions. Those economics are part of the story, not a backdrop to it.

None of that carries cleanly into Singapore, Jakarta, Manila, or Mumbai. Operators in those markets would be adopting the same architecture without the subsidy window that lowers early risk, without the cloud ecosystem that makes deployment friction lighter, and without the cost structure that has helped make the trade-off look attractive in China. They would still be taking on the same security, governance, and continuity risks, but with fewer of the offsetting advantages.

The frenzy is therefore substantially a China-domestic story. Treating China’s current adoption wave as a universal signal is where operators outside mainland China are most likely to make a premature and expensive commitment.

What to Do With This, Right Now

The architecture underneath OpenClaw is real and the direction of travel is not in doubt. But the signals surrounding the current frenzy are a different matter. Stock moves indicate compute demand, not deployment proof. Subsidies signal ecosystem building, not validated operator returns. Cloud vendor moves signal infrastructure opportunity, not technological readiness for regulated environments.

First, understand the architecture properly before the next board conversation arrives.

OpenClaw’s screen-level approach to legacy integration is a genuinely new capability, and knowing what it can and cannot do is table stakes for any technology leader in Asia over the next 12 months.

Second, identify one contained pilot candidate: a workflow that sits behind a legacy interface, involves no sensitive data, and has a clear enough output to measure whether the agent actually performed. Run it. Learn from it. Do not scale it.

Third, watch the governance transition closely. The establishment of the OpenClaw foundation and the terms under which OpenAI supports it will signal whether this project is heading toward enterprise readiness or toward becoming an OpenAI product feature. Those are very different outcomes for an operator evaluating dependency.

Fourth, do not let the China frenzy set the timeline. The subsidy window is short and the obsolescence risk is real. The vendors currently subsidizing adoption are building toward native agent layers that may make OpenClaw’s middleware position redundant. The evaluation window exists independently of their promotional cycle.

The operators who come out ahead will be the ones who use this window to build understanding rather than dependency. The current moment rewards speed in headlines and punishes it in production.

Related Reading On Asia Tech Lens

• At WAIC Hong Kong, the AI Conversation Has Moved Past the Model Race The two-stack decision frame for operators in third market.

• China's Robot Spectacle Is an Industrial Strategy

  China's deployment playbook and subsidy cycle logic.

• The Chinese New Year AI Gateway War

  Cloud vendor pile-on and what happens when incentives end.

• What Tencent’s Yuanbao PAI Reveals About Its AI Strategy

  A read on how Tencent uses mass adoption moments to force trial and legitimize new technology behaviors at scale.

• India's AI Push Is Real. Production Access Is the Constraint

  Why India's AI ambitions are running ahead of its deployment infrastructure, and what that means for operators building now.

Two models are emerging in the race to build “AI phones.”

One works through approved integrations with apps and operating systems, relying on APIs and permissions that developers explicitly grant.

The other works by reading the screen and acting through the interface like a human user, tapping through apps even without formal integrations.

Both approaches can complete the same task. But they represent fundamentally different ways for AI to gain authority on a device.

And that difference is how the AI gets access and that may ultimately determine which model actually scales.

One Task, Two Paths

Say you tell your phone: “Get me a ride to the hotel.”

Two AI phones can do it.

In one approach, the assistant connects to ride-hailing apps through official integrations. It uses permissions granted by the user and the app developer. The system stays inside Android’s normal app boundaries, and the user can intervene or confirm sensitive steps.

In the other approach, the AI assistant reads what is on the screen and taps through apps like a person would, often using accessibility features. Because it works through the interface itself, it can move across apps even when those apps do not provide official integrations.

That second model entered the spotlight in December 2025, when ByteDance introduced its Doubao assistant on ZTE’s Nubia M153 handset. The device demonstrated an AI agent capable of navigating apps through UI automation.

When Samsung launched the Galaxy S26 with Gemini in February 2026, comparisons quickly followed. Some Chinese tech media even dubbed it the “global version” of the Doubao phone.

It’s a catchy headline. But it collapses a crucial difference.

Prof Bo An, President’s Chair Professor and Head of Division of Artificial Intelligence at the College of Computing and Data Science, NTU Singapore, draws a clear line: “The fundamental difference is where the AI assistant operates.”

One approach works through official system functions and app integrations, while the other operates at the GUI level, interacting with apps the way a person would.

Architecture Determines Control

For many researchers, the real question in the AI phone debate is not features, but system design.

Robert Dahlke, Managing Partner at German firm TNG Technology Consulting, says the core issue is the “role the AI agent plays within the system.”

Dahlke, who also works on building leading open source large language models at TNG, says that in some implementations, he explains, the AI agent effectively acts “as a substitute for the human user.”

“The agent gains privileges close to those of a human user and can tap, navigate the interface, and execute cross-application tasks on the phone in the same way a person would.”

That makes it one of the most powerful ways to enable AI on a device, and also one of the most permissive. The challenge, Dahlke says, is control.

When people use a phone, actions happen step by step. In a fully agentic setup, the AI can replace the user in executing those actions, reducing the control layers that normally sit between intention and execution.

Samsung + Gemini and similar systems take a more constrained approach.

In those designs, the functions an AI agent can perform are defined through “explicit permission mechanisms,” such as APIs, MCP or other structured integrations. This establishes clear capability boundaries and ensures the AI operates only within authorized functions.

Professor Wei Lu from the College of Computing & Data Science at the Nanyang Technological University in Singapore says the issue is not just what the AI can do, but what happens when the user no longer directs each step. In a conventional smartphone setup, the user interacts with individual apps while the operating system enforces the rules around what each app can access. A permission-based assistant still works within that structure.

But when an AI system begins interpreting a user’s broader goal and carrying out the intermediate steps on its behalf, responsibility becomes harder to trace.

As Wei puts it, “Whether that concierge faithfully represents the user’s intent—and who is accountable when it does not—is precisely what this architectural difference makes harder to determine.”

Dahlke notes that similar high-authority agents are already being explored in desktop environments. Tools such as the popular experimental agent systems OpenClaw are sometimes only run on employees’ separate machines, allowing organizations to contain potential risks.

“AI agents may operate with high levels of authority, but they should do so within controlled environments.”

Smartphones make that balance harder to maintain. They are deeply personal devices containing banking information, private communications, identity credentials, and payment systems.

Allowing an autonomous agent to operate freely on such a device, Dahlke warns, could introduce risks such as unintended actions or accidental transactions.

The real challenge, he says, is “finding the right balance between expanding AI capabilities and maintaining system control.”

Reliability is the Real Benchmark

Even before governance questions arise, there is a practical challenge: reliability.

Complex tasks require what researchers call “long-horizon planning.” The agent has to break one goal into many smaller steps, then keep track of what it has already done as it moves across screens and apps.

Prof Bo An explains that this is where problems begin to accumulate.

“Probabilistic errors compound with each action.”

As tasks stretch across multiple screens, systems must implement robust error-recovery mechanisms and maintain strong state management to track progress and user intent.

The difficulty is that mobile interfaces are not designed for automation. Buttons move, pop-ups appear, and security checks interrupt flows.

Those variations can easily confuse an AI agent relying on screen interpretation, causing it to tap the wrong button or get stuck halfway through a task.

Chai Yeow Yeoh, Senior Specialist in Cybersecurity at Singapore Polytechnic’s School of Computing, frames the divide as one between a protocol-driven model and a vision-driven one.

“The main technological difference between these two methods lies in the interaction layer.”

One model, he says, uses a protocol-driven approach that lets AI communicate with apps through structured code and APIs in the background. The other, by contrast, relies on a vision-driven method in which the AI interprets the front end by reading pixels and simulating human actions on the screen.

To simplify, Yeoh says, “consider an AI agent that works mainly through app and system permissions as having backstage access,” while the UI-driven agent is “a highly advanced robot standing in front of the phone.”

That distinction has practical consequences. APIs behave more like contracts. They usually remain stable even when an app’s visual design changes. A UI-driven model is more fragile. Move a button, change a color, or insert a new prompt, and the automation can fail.

Yeoh notes that “APIs act like contracts that don’t change every time an app’s appearance alters.” By contrast, if a developer moves a Submit button or changes its color, “the AI’s visual model might fail to recognize it, leading to failed automations.”

Pradeep Reddy Varakantham, Professor of Computer Science in the School of Computing and Information Systems at Singapore Management University, says the UI-driving approach can be “brittle and not as reliable.”

He adds that if apps worry their data is being used without explicit permission, “they may start behaving adversarially and start adding UI features to fool the OS agent.”

In other words, interface automation risks becoming a cat-and-mouse game between AI agents and app developers.

Accountability Is the Harder Problem

Reliability is only part of the challenge. The deeper issue is accountability.

Ding Xuhua, Professor of Computer Science and Co-Director, Centre on Security, Mobile Applications and Cryptography, Singapore Management University, calls it a trust boundary problem.

“When it is at the OS level, you have to fully trust it as it becomes the boss of the phone,” he says.

If the AI remains an application, the operating system can still regulate its behavior.

But when the agent sits deeper in the system, the bar for oversight rises significantly.

“If the AI system is at the OS level, it becomes much harder to reliably attain accountability and auditability,” Ding explains.

That matters because when something goes wrong, the question is not just whether the AI made a mistake. It is whether the system can clearly show what it did, why it did it, and whether the user properly authorized the action in the first place.

As Prof Wei Lu from NTU notes, “that interpretation is not itself an auditable record in the same way that a permission grant is.”

In other words, when a user approves a goal instead of a specific step, more of the execution path is determined by the AI itself. That can make it harder to reconstruct decisions afterward and harder to assign responsibility when outcomes go wrong.

Such safeguards become especially important in services involving banking, government systems, and identity verification, where unclear access models can quickly become unacceptable.

Ding does not argue that interface automation should never exist. But he says it must operate under strict guardrails and clear accountability frameworks.

Chai Yeow Yeoh, Senior Specialist in Cybersecurity at Singapore Polytechnic’s School of Computing, says systems with this level of autonomy should not run without stronger user oversight. “An AI agent should operate under a Human-in-the-Loop framework that emphasizes observability and final decision-making.”

In practice, that means users should be able to see what the agent is doing, stop it immediately, and explicitly approve high-risk actions before they are completed.

Without those, trust becomes difficult to establish.

Privacy and Exposure

The access model also shapes what the AI is able to observe.

“When an AI agent can operate across multiple apps, it raises concerns about privacy and data exposure,” Yeoh says.

A UI-driven model may require broad visibility into whatever appears on the screen. A permission-based model is narrower by design, accessing only the fields or functions explicitly made available through approved interfaces.

As Yeoh puts it, the Samsung/Gemini approach follows “the principle of Least Privilege by only accessing specific data fields through secure APIs.”

That distinction matters because privacy risk begins not only when an agent acts, but when it sees more than it needs to.

The Ecosystem Decides What Scales

Beyond architecture and reliability lies another factor: the ecosystem.

Kyle Chan, an American researcher and fellow at the Brookings Institution’s China Center, a prominent tech voice on X following the AI race between China and the US, argues that platform partnerships may ultimately matter more than raw AI capability.

“Google approach to AI agents for Android offers a more durable approach than the one by ByteDance. Google is making a longer-term investment in app partnerships. ByteDance is trying to rush ahead without permission and has already run into walls.”

“Google’s approach is more comparable to Huawei’s agent-to-agent framework approach. Because both Google and Huawei build mobile operating systems, they can integrate AI agents far more deeply into their devices.”

He further compares the competition to an earlier platform race.

“Building out a network of app partners will likely be even more important than sheer agentic AI performance. This is like the race to become the next Apple app store, but even bigger.”

That comparison matters because distribution layers shape which services users actually adopt. If everyday tasks increasingly begin with an AI assistant, that layer could influence which apps are called by default and who ultimately owns the user relationship.

Wei Lu says the assistant layer could become more powerful as users shift from opening apps directly to simply stating goals.

When that happens, the competitive question is no longer just which app performs a task best, but which systems the assistant chooses to call first.

One model asks platforms to participate. The other tries to work around them.

That distinction helps explain why app ecosystems may accept one approach more readily than the other—and why scaling an AI phone may depend as much on developer cooperation as on the model itself.

For now, the AI phone race is coalescing around two access models: approved integrations with defined permissions, and interface-level automation with broader system access. Both promise powerful automation.

But in areas where trust, traceability, and responsibility matter, such as banking, government services, and identity systems, that architectural difference becomes decisive.

An AI phone becomes truly viable only when existing services are willing to work with it, tasks complete reliably, and actions can be traced when something goes wrong. That is the standard these systems will ultimately have to meet.

Related Reading On Asia Tech Lens

Why ByteDance’s AI Phone Hit a Wall: Security, Fair Play, and the Economics of Attention
The rollout showed how quickly platforms push back when a phone-level agent starts acting across apps without approved access.

The Chinese New Year AI Gateway War: The Big Four’s Fight for Daily Habit
How China’s tech giants are competing to become the default AI gateway, and why habit, distribution and control of the user layer matter as much as model power.

What Tencent’s “Yuanbao PAI” Reveals About Its AI Strategy
A consumer AI read on how Tencent is using product design, distribution and social mechanics to make AI part of everyday behavior.

Why Smartphone Prices Could Rise in 2026 as RAM Costs Surge
Why the AI phone race is also a hardware story, with rising memory costs starting to reshape smartphone pricing and device trade-offs.

Agentic AI Can Act. Singapore’s New Rulebook Says: Prove You Can Stop It
A governance companion on why autonomy needs limits, visibility and the ability to step in when something goes wrong.

Editor’s Note: The world is watching China this week as the Two Sessions begin. This is where the year’s economic and industrial direction gets set, with direct consequences for trade, supply chains, and capital allocation across Asia. Beijing is using the government work report and the 15th Five-Year Plan outline to signal an unambiguous priority: technology as national capacity. The emphasis is not AI as a standalone sector, but AI and robotics as the operating layer for manufacturing and supply chains, supported by the rails that make deployment governable, including data markets, security systems, and standards. This week’s wrap follows that logic from policy intent to the bottlenecks that will decide who can actually ship: packaging, test, and integration capacity.

China’s Five-Year Plan Puts AI And Robotics At The Center

China has set out a five-year roadmap to turbocharge scientific breakthroughs and embed AI across its industrial economic machine, positioning technology and scientific research as top national priorities. According to reports, the plan name-checks AI more than 50 times and describes an AI-driven industrial future where robots plug labor shortages and factories operate with little human oversight. It also pushes commercialization, including AI-powered humanoid robots and wider deployment across the full supply chain.

The plan also signals how broad this tech push is. It name-checks frontier areas including biomedicine, quantum technology, atomic-scale manufacturing, hyperscale computing clusters, nuclear fusion, and brain-computer interfaces. Watch for where this turns into procurement logic, especially in industrial automation, robotics, and domestic AI stack adoption.

Signals To Watch:

• Whether specific industries get near-term procurement targets for “AI+” and factory automation.

• How fast humanoid robots move from pilots into subsidized, scaled deployments in industrial settings.

• Whether SOEs are directed to buy domestic stacks for industrial AI, robots, and automation systems.

China’s Five-Year Plan Builds A National Data Market And An AI Security System

The draft outline of the 15th Five-Year Plan (2026–2030) also focuses on the rails that make AI adoption workable at scale. It aims to raise the value-added of “core digital economy industries” to 12.5% of GDP. It calls for new policies for an integrated national data market, AI adoption across the full supply chain, and an AI security system.

For operators, this is the governance and control layer. A national data market changes how data is accessed, priced, shared, and enforced. An AI security system changes what gets approved, audited, and monitored in production.

For firms operating in China-linked industrial ecosystems, data governance and model controls may increasingly become commercial requirements, not just compliance issues. Logging, monitoring, change visibility, and approval traceability could become critical for deployment.

Signals To Watch:

• Whether the national data market produces real standards and enforcement, not just a policy slogan.

• What the AI security system becomes in practice: standards, certification, audits, or incident reporting.

• Whether “security” starts to mean required model controls, logging, and monitoring for enterprise deployments.

Micron Opens India’s First Semiconductor Assembly And Test Facility

Micron has opened its semiconductor assembly and test facility in Sanand, Gujarat. The facility is designed to take DRAM and NAND wafers from Micron’s global network and turn them into finished memory and storage products. The project represents a combined investment of approximately US$2.75 billion by Micron and its government partners. This is the most concrete step so far in India’s push to build real semiconductor capacity, starting with assembly, test, and packaging rather than leading-edge fabrication.

This matters because packaging and test is where supply chains can diversify first. It also signals where ecosystem bets will cluster next: chemicals, substrates, equipment maintenance, logistics, and the technician pipeline. The facility does not change next quarter’s supply constraints, but it does change 12- to 36-month planning for anyone qualifying alternate routes for memory and components.

Companies qualifying alternate semiconductor supply routes should begin paying attention not only to fabs, but to where packaging, testing, maintenance, and logistics ecosystems are becoming reliable enough to support mainstream programs.

Signals To Watch:

• Whether large OEMs start qualifying India-linked packaging and test for mainstream programs, not just pilots.

• Whether the surrounding supplier ecosystem forms fast enough to keep timelines predictable.

• Whether India becomes a serious node for advanced packaging over time, not just basic assembly and test.

ASE Bets Bigger On Advanced Packaging To Meet AI Demand

ASE, the world’s largest chip packaging and testing group, says demand is rising fast for advanced packaging tied to AI chips. It expects advanced packaging sales to double in 2026 to about US$3.2 billion, and says it plans to lift investment from last year’s US$5.5 billion level.

TSMC gets most of the attention in the AI supply chain, but this is a reminder that packaging and integration capacity is not only a foundry story. A lot of the work still sits with specialist packaging players, and when those queues tighten, delivery timelines slip and costs show up in places buyers do not always track.

Signals To Watch:

• Whether packaging lead times lengthen even if chip supply improves.

• Whether cost pressure shifts from GPUs to packaging, testing, and integration.

• Whether more capacity comes online outside a few dominant packaging players.

Takeaway

This week made one thing clear: the AI push is moving from narrative to industrial buildout. China is signaling direction and building the governance rails around deployment. India is adding real semiconductor capacity in packaging and test. ASE is warning that advanced packaging is already becoming a pressure point.

The practical lesson is to plan around the full deployment chain, not just access to chips. Packaging, testing, and integration capacity can now shape rollout timelines as much as compute itself. Operators should ask vendors where they sit in the packaging queue, what capacity they have locked in, and what contingencies exist if lead times stretch. The issue is no longer just whether AI systems can be built. It is whether the surrounding infrastructure is ready to support deployment at scale.

Sources

• Reuters: China ramps up ‘high stakes’ tech race with US as economic imbalances deepen

• EE Times Asia: ASE Ramps Up Investment as AI Packaging Demand Accelerates

• CRN: Micron opens India’s first semiconductor assembly and test facility in Gujarat

When humanoid robots danced across the stage of the Spring Festival Gala this Lunar New Year, the performance appeared at first glance to be technological theater. Broadcast live by state broadcaster China Central Television to an average of 325 million viewers per minute, the machines executed synchronized martial arts choreography alongside human performers. It was a guaranteed viral moment.

But to read this as a scheme to gain virality would risk missing the bigger picture. Virality, in this case, is a means to an end. The robot performance was not meant to be mere entertainment. It was a showcase of how China uses spectacle as an industrial instrument to accelerate the conditions under which emerging technologies can be deployed, tested and improved. With the humanoids market projected to surpass a value of US$5 trillion by 2050, it seems strategic for a country to realign priorities around these machines.

More importantly, for industrial operators, the issue is less whether China’s robots impress on stage than whether China is accelerating the path to affordable, repeatable deployments faster than its competitors—and what response that demands: early pilots, strategic partnerships, cautious observation, or avoidance.

Spectacle as Industrial Legitimacy

The Spring Festival Gala occupies a unique position in China’s political economy. Beyond entertainment, it functions as a national signaling platform. Technologies granted prominence on the broadcast often acquire informal policy legitimacy, shaping how investors, local governments and state-owned enterprises interpret industrial priorities.

That legitimacy translates into specific institutional mechanisms. First, national exposure provides procurement cover: officials responsible for purchasing decisions face career risk when adopting immature technologies, but a robot showcased on a state-endorsed broadcast reframes experimentation as policy alignment rather than personal judgment. Second, visibility helps unlock budgets. Pilot projects that might otherwise struggle for approval can be categorized as supporting nationally prioritized industries, allowing funding to move through existing innovation or industrial modernization programs. Third, signaling cascades downward. Following high-profile endorsements, local governments frequently respond by launching demonstration zones, issuing targeted tenders, or incorporating emerging technologies into industrial park development plans.

For companies and public-sector buyers, legitimacy lowers institutional risk. While the Gala does not cause factories to automate overnight, it makes pilot programs easier to approve within bureaucratic procurement systems. That’s because in China, legitimacy is a procurement primitive. In other words, technologies are often purchased once they are seen as legitimate or aligned with national priorities.

Government entities are already among the largest early buyers. In recent years, they have increasingly shown interest in adopting humanoid robots. Reuters reported that state procurement of humanoid robots and related tech increased to 214 million yuan (US$31 million) in 2024 from 4.7 million yuan (around US$680,000) in 2023. The Economist recently argued that the Chinese government will likely remain the biggest source of demand for some time to come.

On top of procurement, Chinese authorities have been handing out generous subsidies for humanoid firms, with more than US$20 billion allocated to the humanoid sector in recent years.

China’s hands-on approach to humanoid robotics mirrors its earlier push into electric vehicles, where aggressive state support and large-scale deployment helped drive costs down and produce globally competitive products. Already, several Chinese humanoid models are significantly more affordable than Western models—the former priced between roughly US$5,000 and US$15,000, while the latter often exceeded US$150,000.

But more importantly, these state-backed environments function less as profit centers than as testbeds. They are practice fields where the machine can learn and improve before serving commercial clients.

Building the Data Flywheel

Most humanoid robots today are not commercially efficient. That is precisely why China is deploying them early.

Chinese robotics firms have placed humanoid systems in pilot programs across different facilities, where robots perform limited repetitive tasks while generating operational data. This creates what industry analysts describe as a data flywheel: deployment produces data, data improves algorithms, and improved systems justify further deployment. Rather than subsidizing finished products, China is effectively subsidizing learning itself.

For operators, this phase resembles large-scale field testing rather than commercialization. The critical metrics are not unit sales or performance demonstrations but operational evidence: how long robots can function without interruption, how frequently they require maintenance, and how easily they integrate into existing workflows. Each deployment reveals hidden costs that rarely appear in demonstrations but ultimately determine economic viability. They can be costs related to software updates, supervision requirements, or even safety protocols and retraining.

The Gala helps build a case for this learning loop by making experimentation politically safe.

But that doesn’t mean all stakes are eliminated. If humanoids fail to transition from demonstrations to reliable productivity tools, the industry could drift into what critics describe as “demo-ware”: impressive performances without durable business models.

Commercial Reality and the Limits of Spectacle

Despite rapid progress, commercial viability remains uncertain. Many humanoid robots operate for only a few hours under heavy workloads due to battery constraints, with some models’ runtimes limited between one to two hours. These runtimes fall far below the eight- to twelve-hour work shifts typical of human workers. On top of that, dexterous robotic hands remain expensive, and most systems still require supervision in unpredictable environments.

For that reason, the distinction between demonstration and deployment is critical. A robot capable of dancing on stage may still struggle to operate continuously on factory floors. For potential buyers, the relevant question is no longer whether robots can move convincingly, but whether deployments produce measurable returns. Can machines operate through full production shifts? Is maintenance predictable? Are responsibilities clear if failures cause disruption or damage?

China’s approach distributes these risks across state-supported pilots. It’s quite a unique approach compared with other countries. The United States typically emphasizes profitability before scale, while Japan and South Korea have historically focused on socially acceptable robotics applications such as caregiving and assistance.

While each approach carries its own rationale, China’s deployment-led strategy seems to secure it an early global lead. Out of the 13,000 humanoid robots delivered worldwide in 2025, most of them came from Chinese companies, according to research firm Omdia.

Whether China’s strategy succeeds in the long run depends on whether early pilots evolve into repeatable return-on-investment use cases. If deployments generate reliable productivity gains, the country could industrialize embodied AI much as it did electric vehicles and solar manufacturing. If not, humanoid robotics risks remaining trapped in a cycle of impressive demonstrations without sustainable demand.

The dancing robots of the Spring Festival Gala were therefore less a declaration of technological arrival than an early move in an industrial experiment. China is betting that putting imperfect machines into the real world sooner will allow it to learn faster than rivals.

China is optimizing for iteration speed: more deployments, more data, faster cost-down. The open question is whether that converts into repeatable ROI use cases outside state-supported pilots.

Related Reading On Asia Tech Lens

Inside China’s Bold Push to Build Humanoid Robots
How policy coordination, component dominance, and real-world testbeds are compressing the path from “demo” to deployable machines.

The Chinese New Year AI Gateway War
How the Big Four use red packets and subsidies to manufacture habit formation, and what happens when incentives end.

What Tencent’s “Yuanbao PAI” Reveals About Its AI Strategy
A Spring Festival playbook read on how China uses mass moments to force trial and legitimize new consumer tech behaviors at scale.

Singapore’s Robots Don’t Go Viral—They Go Global
A contrast case: why Singapore’s winners optimize for reliability and exportability, not spectacle, and what “quiet scale” looks like.

Plush, Playful, and Powered by AI: How China’s Toys are Evolving
A “fast-deploy, fast-learn” consumer parallel to humanoids: how open models + cheap hardware + mass-market distribution create an iteration flywheel, and why privacy and failure modes show up only after deployment.

A chatbot flow gets updated. A vendor proposes a “slightly better” model. A team automates a step in approvals.

Nothing dramatic. Just the steady drip of AI becoming normal.

Then comes the question businesses rarely prepare for: can you prove what this system is doing, who it affects, and what happens when it changes?

Vietnam’s AI Law makes those questions unavoidable. It took effect on 1 March 2026, but businesses running existing AI systems have a transition window: 18 months (1 Sept 2027) for healthcare, education, and finance, and 12 months (1 March 2027) for other sectors. But that window is not a pause button. It is the time you have to build documentation, logging, vendor change control, and disclosure workflows before procurement, audits, or regulators can block deployment.

The law’s core mechanic is straightforward: classify AI by risk, then apply obligations accordingly. Systems fall into high-risk, medium-risk, or low-risk, and providers must self-classify, keep documentation supporting that classification, and for medium- and high-risk systems, notify the Ministry of Science and Technology through the national AI portal before deployment.

The International Association of Privacy Professionals (IAPP) describes this as a “risk-based management” approach, similar in structure to the EU’s AI Act, which in practice means organizations need to explain, consistently, how systems were classified and what controls sit around them.

This will not fail because a model is unsafe. It will fail because the business cannot prove what model ran, on what data, with what oversight, and what changed. In practice, the pressure shows up in three choke points. Procurement blocks deployments when vendors cannot provide audit trails or change notes. Audit blocks when the organization cannot reconstruct decisions, including what version ran and what oversight existed. Disclosure blocks when labeling exists only as policy and the workflow does not enforce it consistently.

Vietnam is borrowing two enforcement instincts: the EU forces pre-deployment discipline; China forces labeling/platform responsibility. Vietnam borrows from both, so the first friction will appear in procurement, auditability, and disclosure workflows.

A defining feature of the AI Law is that it regulates by role, not by industry. It distinguishes between the developer, provider, deployer, user, and affected person. That role mapping matters because it determines who is responsible for documentation, oversight, vendor obligations, and disclosure.

Build the AI Inventory

AI already appears in more places than internal reporting suggests. It is embedded in customer chat and automated replies, fraud alerts and transaction monitoring, queue triage and routing, recommendation systems, and content workflows. Some uses are explicit. Others are bundled into vendor products and automation features that may not be labeled as AI.

The point of the inventory is to make it clear what AI is in use, where it matters, and who owns it, so businesses can move on to risk triage, minimum controls, vendor requirements, and disclosure decisions.

For each system in use, capture:

• System name and business owner.

• Where it is used (customer interaction, approvals or eligibility, prioritization/triage, safety monitoring, fraud/AML flags, content generation).

• What it influences (recommendation only, or automated action).

• Sourcing (internal build, vendor system, or hybrid).

• Data exposure (customer, health, financial, operational).

• Change history (customization, fine-tuning, new data sources, expanded use cases).

That final point is often underestimated. In a role-based regulatory structure, incremental modifications can shift responsibilities and trigger re-assessment expectations. A strong inventory makes such changes visible early, before a vendor update, an incident, or an audit forces the issue.

Triage What Will Become High-Risk

A workable triage starts with two questions.

First: does the system influence consequential outcomes?
If an AI system can affect safety, legal rights, access to services, or material financial outcomes, treat it as high-risk until assessed otherwise. This commonly includes approvals and eligibility decisions, credit or insurance outcomes, medical or safety-related triage, and automated enforcement actions.

High-risk status matters because conformity assessment is a condition for being put into use. Assume modifications trigger reclassification duties, especially new data sources, new user groups, or expanded decision authority.

Also, most companies are deployers even when they think they are “just using software,” and deployer obligations do not disappear because the model is “the vendor’s problem.”

The cliff edge is the Prime Minister’s high-risk list. It will specify which high-risk categories require mandatory pre-use conformity certification before being put into service, and it may arrive late in the transition window. Do not treat the absence of the list as clearance. Treat any high-risk likely system as if it could land on the certified subset and build the evidence trail now: classification memo, logs and versioning, human oversight and rollback, and vendor change control that survives procurement and audit.

Second: could it reasonably mislead, influence, or manipulate people who do not realize they are dealing with AI?
Many customer-facing systems land here. Medium-risk systems include situations where customer support chat that sounds human and provides account or service guidance without explicit disclosure that it is AI.

The inventory should already reveal which systems will be hardest to defend later. The goal of triage is to surface those systems early, before a vendor update, expanded use case, or external scrutiny turns a workable deployment into a stop-ship.

Minimum Viable Compliance

• AI inventory exists (named owner + system purpose + data class + deployment surface)

• Risk classification memo exists (why it’s low/medium/high + what would change that)

• Evidence capture works (logs/versioning/inputs/outputs stored and retrievable)

• Human override path exists (who/when/how recorded)

• Rollback path exists (pause/revert/manual route)

• Vendor change control exists (notification + audit trail access + incident support)

• Disclosure is enforced in workflow (UI/script/publishing gate + audit evidence)

Install Controls That Survive Audits

Logging and Traceability

Businesses should be able to reconstruct what happened when a decision is questioned. At minimum, that means recording the key inputs used, the output produced, the version of the model or system, and the business action taken as a result.

Human Oversight

It should be clear who is authorized to intervene or override, when intervention is required, how overrides are recorded, and what happens when the system behaves outside expected boundaries.

Rollback Capability

Medium- and high-impact systems should not be treated as one-way deployments. There needs to be a defined method to pause the system, revert to a prior configuration, or route decisions back to manual handling when performance degrades, unexpected behavior is detected, or risk thresholds are breached.

Incident handling is treated as a shared obligation across the AI value chain. That means setting incident thresholds, escalation procedures, and operational readiness to suspend or withdraw systems when required, including content generation incidents where outputs impersonate a real person or create deepfake-like media.

These controls only work if the business can access the underlying evidence consistently. For many deployments, that evidence sits with third-party suppliers, which is why contract terms become the next practical control surface.

Vendor Discipline

Many AI systems used in business workflows are vendor products or vendor models embedded into platforms, which means the information needed for traceability and incident handling may not be available by default.

Before renewing, expanding, or embedding a vendor system into a regulated workflow, operators need to ask four questions

• Can the third party provide model and version change logs?

• Can the third party provide audit evidence without requiring source code disclosure?

• What is the third party’s incident notification SLA, and what investigation support is provided?

• Who is the third party’s lawful local contact point or authorized representative in Vietnam, if the system is high-risk?

For high-risk systems supplied by foreign providers, vendor due diligence also needs a Vietnam-specific check. Tilleke & Gibbins notes that foreign providers of high-risk AI systems must establish a lawful local contact point, and systems subject to mandatory pre-use conformity certification may require a commercial presence or an authorized representative in Vietnam.

Operational contract clauses to prioritize:

• Model change notifications: Versioning and change notes that let you assess whether classification, controls or disclosures must be revisited.

• Audit trail access: Logs, documentation, and functional explanations that support auditability without requiring source code disclosure.

• Incident disclosure: What counts as an incident, notification timelines, and obligations to support investigation and remediation.

Transparency Duties and Disclosure

Providers must ensure that users are aware when they are interacting with an AI system, and must ensure that AI-generated audio, images, and videos are marked in a machine-readable format as prescribed by the Government. Deployers must clearly notify the public when AI-generated or AI-edited content could cause confusion about authenticity, and must apply easy-to-recognize labels for content that simulates real persons or recreates events.

The failure mode is predictable. Labeling is written into a policy, but nothing in the product or publishing workflow enforces it. The result is uneven compliance across teams, vendors, and channels.

A workable transparency workflow answers three questions:

• Who decides what gets labeled and what exceptions exist.

• Where it is enforced (product UI, customer support scripts, publishing gates, creative pipelines, metadata steps).

• What happens when it fails (escalation, correction or removal, evidence retention).

What This Looks Like in Three Common Settings

Banks

• Where does AI influence eligibility, pricing, credit limits, fraud escalation, or collections prioritization, and what is automated versus advisory?

• For any disputed outcome, can the business reconstruct inputs, model/version, output, and any override decision, including who overrode it and why?

• Do third parties commit contractually to change logs, audit evidence, and incident notification SLAs, including investigation support?

Hospitals

• Which clinical workflows use AI for triage, imaging support, prioritization, or safety-related decisions, and which of those should be treated as high-risk likely?

• If an outcome is questioned, can the hospital reproduce what ran and when, including model/version, key inputs, human oversight, and escalation actions?

• Is there an operational rollback path to manual care pathways, and an incident escalation process that works 24/7 rather than relying on a single person

Logistics players

• Where does AI influence routing, warehouse vision, safety monitoring, or prioritization decisions, and what are the operational consequences when it is wrong?

• Can the business trace an operational decision end to end, including inputs, model/version, output, and the action taken, and can it pause or revert the system quickly?

• Where do disclosure and labeling obligations show up in customer-facing or worker-facing workflows, and is enforcement built into the process rather than left to policy?

Vietnam’s AI Law does not require businesses to stop using AI. It requires businesses to be able to stand behind it.

That means risk classification that holds up under scrutiny, baseline controls that make systems traceable and interruptible, vendor contracts that prevent silent model changes and missing audit trails, and transparency that is enforced where real work happens.

If you run medium/high-risk AI and can’t produce evidence on demand, treat the next 90 days as a compliance build sprint—because the first enforcement you’ll meet is procurement and audit, not a courtroom.

Related Reading On Asia Tech Lens

Agentic AI Can Act. Singapore’s New Rulebook Says: Prove You Can Stop It
Why “prove oversight and rollback” is becoming a procurement and audit baseline, even before hard law.

Indonesia Is Racing To Regulate AI. The Messy Part Is Implementation
A practical read on why fast regulation creates ambiguity, and how operators should pace deployments when enforcement detail lags policy intent.

AI Is Accelerating Cybercrime and Southeast Asia Feels It First
A risk-and-controls lens on fraud, abuse, and why incident thresholds and escalation paths need to mature alongside AI adoption.

The AI Battleground: How Southeast Asia Is Forging a New Path Between Superpowers
A regional framing on how Southeast Asian governments balance adoption and governance, and what that means for enterprise deployment constraints.

Why ByteDance’s AI Phone Hit a Wall: Security, Fair Play, and the Economics of Attention
A deployment-constraints case study on what blocks “working” AI in the real world when security, platform rules, and operational evidence are not ready.

Editor’s Note: Another AI heavy week. The biggest moves came from incumbents setting the terms on price, capacity, and where workloads can run. Across Asia, the questions are growing sharper: how cheap can AI get, how quickly can compute expand, and what constraints are already showing up in power, networks, and governance.

Alibaba Bundles China’s Top Models into A Low-Cost Coding Tool

Alibaba is pushing into AI coding tools by selling a subscription that lets users switch between several Chinese models, including Qwen 3.5 plus models from Zhipu, Moonshot, and MiniMax. The pricing is deliberately mass-market: the lite tier is 7.9 yuan (about US$1.15) for the first month, then 40 yuan (about US$5.82) thereafter. The pro tier starts at 39.9 yuan (about US$5.81) and goes up to 200 yuan (about US$29.11).

By bundling multiple models behind one workflow and one price point, Alibaba is trying to make AI-assisted coding a default habit, not a special purchase. For operators, this is the practical version of ‘priced by incumbents: AI becomes a line item you can roll out widely, but it also increases the need for controls, because cheap access makes it easier for usage to sprawl across teams without clear governance.

Signals To Watch:

• Whether other clouds copy the bundle model (multi-model switching, one subscription).

• Whether enterprises start treating coding assistants like a standard seat licence.

China Models Take The Lead As MiniMax And Moonshot Top Token Usage

Chinese open-source models from MiniMax and Moonshot topped OpenRouter’s latest token-usage ranking, after a wave of new releases pushed Chinese models ahead of U.S. labs on the platform. MiniMax’s M2.5 (launched about two weeks ago) ranked first with 4.55 trillion tokens, while Moonshot’s Kimi K2.5 (released last month) ranked second with 4.02 trillion tokens. MiniMax, Moonshot, and DeepSeek together made up nearly two thirds of total token usage among the top five models.

Token usage on a neutral hosting platform tells you where developers are actually spending time and money, which tends to shape tooling, integrations, and what shows up on enterprise shortlists next. It also tightens the set of “default” choices, which is helpful for speed but risky for resilience. If a small group of models becomes the default across teams, you inherit concentration risk: pricing power shifts to the winners, and you need a clear plan for governance, support, and switching costs before dependency hardens.

Signal To Watch:

• Whether enterprise software vendors start supporting these models by default.

• Whether pricing tightens as usage concentrates and capacity gets scarce.

• Whether buyers start requiring stronger operational guarantees: audit logs, incident response, uptime commitments, and clear data boundaries.

Malaysia Freezes New Non-AI Data Centers

Malaysia has stopped approving new data centers that are not related to AI because of rising electricity and water demands. Prime Minister Anwar Ibrahim said Malaysia’s current energy supply is sufficient for the next one to two years, but the country will need more sources longer term, including the ASEAN Power Grid. The policy has effectively been in place for about 1.5 to 2 years.

For operators, this is the hard constraint behind AI scale in Southeast Asia. Compute is starting to be treated like a limited resource that gets allocated. That changes rollout planning, contract terms, and where you place workloads. It also shifts leverage: when capacity is scarce or gated, vendors and landlords set the rules.

Signals To Watch:

• Clearer definitions of what qualifies as “AI-related” for approvals.

• Power and water requirements showing up as stricter conditions in contracts and permits.

• More workloads designed to run across multiple countries from day one.

SK Hynix Plans A US$15 Billion Expansion Of Chip Production In Yongin

SK Hynix said it will invest 21.6 trillion won (about US$15.07 billion) to build new semiconductor production lines in Yongin by 2030. For operators, this is a reminder that AI budgets are being shaped by the supply chain. The biggest costs in AI systems are still hardware and availability, especially for memory-heavy configurations. Announcements like this do not help your next purchase cycle, but they do matter for 12–36 month planning because they are one of the clearest signals on whether component shortages and pricing pressure might ease, or whether AI infrastructure stays expensive and capacity-constrained.

Signals To Watch:

• Whether lead times improve for AI-grade components and systems.

• Whether pricing pressure eases or just moves to a different bottleneck.

• More large capex commitments across the Korea–Taiwan–Japan semiconductor supply web.

Japan’s DOCOMO Uses Its Mobile Network Servers To Run AI Apps

NTT DOCOMO, Japan’s largest mobile operator, ran an AI application on the same general-purpose servers it already uses to run its virtualized mobile network (vRAN), inside its commercial network.

For operators, this is a shift in where AI can live. If telcos turn spare network compute into a managed AI layer, AI execution starts to sit alongside connectivity, not just inside hyperscaler clouds. That can help for latency-sensitive use cases, distributed operations, and situations where you want tighter control over where data is processed. The trade-off is dependency: you are buying into a telco platform, with telco tooling and telco terms. If this becomes a real product, buyers will need clear performance guarantees and a practical exit path.

Signals To Watch:

• Whether telcos productize this with pricing.

• What workloads it can support in production.

• Portability: whether the setup can move across telcos/vendors without major rework.

Takeaway

This week’s signals point in the same direction: AI is getting packaged into low-cost seats, usage is concentrating around a few models, hardware investment is ramping, and networks are being positioned as a place to run workloads. The operator move is to make governance practical: set budget guardrails, standardize logging and approvals, and design for at least one credible fallback option from the start.

Sources

• The Edge Singapore: Alibaba pushes deeper into AI coding tools with low-cost access

• SCMP: China’s MiniMax, Moonshot top AI token use ranking, ending year of US dominance

• Malay Mail: Malaysia freezes new non-AI data centres over power and water concerns, says Anwar

• Evertiq: SK hynix to Invest $15 billion in Yongin Semiconductor Cluster

• The Fast Mode: DOCOMO Demos AI Applications Running Directly on CPUs in Commercial vRAN Network

China’s EV mix is shifting: EV growth is slowing, battery prices remain higher than many expected, and hybrids are gaining market share.

At first glance, this appears to be a straightforward sales story. Within the broader EV category, the demand mix—meaning the breakdown of what consumers are choosing to buy—is tilting away from fully battery electric vehicles (BEVs) and toward hybrids, which combine a battery with a traditional combustion engine.

Hybrid models differ based on how much they rely on electricity: a HEV (Hybrid Electric Vehicle) runs mainly on gasoline and typically cannot be plugged in; a PHEV (Plug-in Hybrid Electric Vehicle) has a larger battery that can be charged from an external power source, allowing for sustained all-electric driving before the gasoline engine kicks in; an EREV (Extended-Range Electric Vehicle) operates primarily as an electric car, with a gasoline engine serving mainly as a backup generator.

Analysts estimate that BEVs could be up to 20 percent more expensive than comparable PHEVs. Lithium price pressures could add up to 3,800 yuan to the production cost of a midsize BEV, compared with roughly 2,000 yuan for a PHEV, according to recent industry research. In January this year, BEVs accounted for 58 percent of total EV sales, compared with 62 percent for 2025 as a whole. Hybrids’ market share rose accordingly.

But to focus on these sales trends is to miss the bigger picture.

For the past decade or so, China’s EV push has been built on several assumptions: battery costs would steadily fall, fully electric cars would soon cost roughly the same as gasoline vehicles, and scale would deliver sustainable profits. Companies invested billions into electric-only platforms, battery plants, and supply chains on that premise.

Now those investments are put under scrutiny. Against the backdrop of less than predictable battery prices, decision-makers are forced to rethink their capital allocation—where money has been committed, how those investments were justified, and what happens when the assumptions behind them begin to shift.

Automakers, battery manufacturers, and policymakers would, soon enough, have to grapple with the following question: if hybrids remain more popular and more economically practical than fully electric cars for the next three to five years, what part of their EV strategy should they reconsider?

Reconsider: Battery Costs and the Cost Parity Assumption

At the center of China’s EV expansion was the idea of cost parity. In simple terms, cost parity means a fully electric vehicle can be produced and sold at roughly the same cost as a comparable gasoline car, without heavy reliance on government subsidies. Once that point is reached, electric cars can compete naturally on price while preserving healthy profit margins.

The challenge is that batteries make up a large portion of an EV’s total cost. If parity slips, OEMs face a bad choice: protect margin and lose BEV share, or protect share and bleed cash through pricing. That is why hybrids matter. They are not just a consumer preference; they are a way to sell electrification with less exposure to cell cost.

Long-term procurement contracts further complicate the picture. Automakers typically sign multi-year agreements with battery suppliers to lock in volume and pricing. These contracts are meant to provide stability. But if BEV demand grows more slowly than forecast, or if battery input costs fluctuate in unexpected ways, such agreements can pressure margins instead of protecting them.

In this context, hybrids gaining share sends an important signal. Hybrids use smaller batteries and therefore rely less heavily on falling battery prices to make economic sense. If consumers—and manufacturers—find hybrids more practical in the near term, it suggests the economics of full electrification are maturing more slowly than anticipated.

Who Wins and Who’s Exposed if Hybrids Outperform BEVs

Winners:

• Flexible-platform OEMs that can swing mix between BEV, PHEV, and EREV without stranding plants or tooling.

• OEMs with pricing power and cash buffers that can slow BEV expansion, keep utilization healthy, and avoid margin-destructive price cuts.

• Battery makers with demand diversity, meaning hybrid packs, energy storage, and export volumes that keep factories loaded even if domestic BEV growth softens.
  Players with contract optionality, including volume-flex terms and indexation bands that reduce take-or-pay exposure when forecasts slip.

Exposed:

• BEV-only platform bets that require sustained BEV volume to justify fixed costs and can’t easily pivot the line.

• Margin-fragile OEMs that need BEV growth to cover high fixed costs and are forced into price cuts to protect share.

• Cell makers concentrated on large-pack BEV demand, especially where expansion phases assume high utilization to hit unit economics.

• Regions that planned infrastructure and incentives around BEV-only acceleration, where a slower BEV curve creates stranded public and private capex.

The dividing line is optionality. Flexible platforms and flexible contracts turn a mix shift into a planning problem. BEV-only capacity turns it into a balance-sheet problem.

Reconsider: Platform Strategy and Utilization Rates

The next layer of impact appears in platform strategy. A vehicle platform is the underlying structural design that determines how cars are engineered and manufactured.

Some automakers have invested in dedicated BEV platforms built exclusively for electric vehicles. These designs are optimized for large battery packs and electric drivetrains. They can be highly efficient—but only if produced at sufficient scale. They require substantial upfront capital and depend on strong, sustained demand to justify the investment.

Others rely on flexible architectures that can produce hybrids, plug-in hybrids, and fully electric vehicles on the same production line. While not as optimized for pure EV performance, these platforms offer adaptability. If demand leans toward hybrids for several years, companies with flexible systems can adjust production more easily.

This distinction matters because profits rise and fall with changes in sales volume. EV manufacturing involves high fixed costs: factories, tooling, and battery facilities. When volumes are strong, profits can increase rapidly. But when sales fall short of expectations, those same fixed costs weigh heavily on earnings.

The dynamic extends to gigafactory capacity planning. A gigafactory is a massive battery production facility designed to supply millions of EVs annually. These plants require enormous upfront investment and only generate strong returns when running close to full capacity. If hybrids require smaller batteries and BEV demand underperforms, utilization rates may decline. Lower utilization means higher costs per unit and thinner margins.

Platform choices and gigafactory buildouts are long-cycle decisions. Once capital is committed, it is difficult to reverse. When battery economics shift, these bets come under pressure.

Reconsider: Industrial Pacing

Zooming out, the issue becomes one of industrial pacing—the speed at which the entire EV ecosystem expands. This includes mining, battery production, vehicle assembly, and charging infrastructure. China accelerated this ecosystem aggressively, seeking technological leadership and scale advantages.

If BEV demand grows more slowly than expected, the question becomes one of timing. Expanding too quickly can create excess capacity and compress margins. Expanding too slowly risks ceding competitive advantage. Policymakers may need to recalibrate subsidy schedules or electrification targets to better align ambition with economic reality.

In short, hybrids gaining share is not, in itself, the central story.

The deeper story is what happens when the economic foundations of China’s EV expansion—falling battery costs, rapid cost parity, and sustained BEV scale—are tested by reality. When battery economics shift, it is not consumer sentiment alone that matters. It is capital allocation, platform design, factory utilization, and industrial pacing that determine who absorbs the shock—and who adapts.

Decision Guide Sheet

If hybrids remain more popular and more economically practical than fully electric vehicles (BEVs) for the next three to five years, what concrete decisions should automakers, battery manufacturers, and policymakers reconsider—and under what conditions?

For Automakers

• If BEV sales growth underperforms expectations for 2–3 consecutive quarters: Reassess planned expansion of electric-only factories. Delay or phase gigafactory-linked capacity additions. Rebalance production targets toward hybrid or extended-range models.

• If BEV factory utilization drops below ~70–75%: Shift production mix toward hybrid models where feasible. Accelerate transition to flexible platforms capable of producing multiple powertrains. Freeze new capex tied exclusively to BEV scale until demand visibility improves.

• If hybrid gross margins exceed BEV margins for multiple quarters: Reevaluate product mix strategy. Prioritize profitability over pure BEV volume growth. Adjust investor guidance away from “scale at all costs” narratives.

• If battery input prices remain volatile or elevated beyond forecast bands: Renegotiate long-term procurement contracts for flexibility. Reduce exposure to rigid volume commitments. Diversify battery chemistry sourcing where feasible.

For Battery Manufacturers

• If BEV order growth slows while hybrid penetration rises: Reassess long-term capacity expansion timelines. Delay additional gigafactory phases not yet under construction. Increase development of smaller-pack solutions for hybrid platforms.

• If utilization falls below ~80% across major plants: Consolidate production lines. Pursue export markets more aggressively. Shift production toward energy storage or non-auto segments.

• If pricing pressure intensifies and margins compress: Tighten capex discipline. Prioritize higher-efficiency chemistries or cost-optimized LFP lines. Reevaluate expansion assumptions embedded in prior investment cycles.

For Policymakers

• If BEV adoption plateaus while hybrids structurally gain share: Reassess electrification timelines. Consider recalibrating policy targets to reflect economic pacing. Avoid forcing demand beyond cost readiness.

• If industry-wide capacity utilization weakens: Moderate further supply-side incentives. Shift policy support toward R&D and efficiency improvements rather than pure volume expansion.

• If OEM margins deteriorate despite rising EV penetration: Evaluate whether subsidy withdrawal or regulatory acceleration is misaligned with industry economics. Reexamine incentive structures for hybrid vs BEV vehicles.

Related Reading On Asia Tech Lens

• Not Just EVs: China Leads the World in Battery Production and Technology for all Vehicles
  A battery-first look at how chemistry, materials, and scale shape cost curves, margins, and the pace of electrification.

• VinFast Pivots to Southeast Asia - But Can it Outrun BYD in Indonesia?
  A demand-and-economics read on how Chinese incumbents pressure challengers through pricing, localization, and infrastructure constraints.

• Singapore’s Driverless Proving Ground: How Chinese AVs Are Turning Pilots Into Proof
  A production lens on mobility deployment: regulation, operating constraints, and what it takes to move beyond pilots.

• The Chip War’s New Reality: A View from the Crossroads
  A broader industrial framing on how China’s manufacturing scale and supply chains compound advantages across strategic sectors, including batteries and EVs.

• China’s Bet on Domestic Tech
  A capital-allocation lens on China’s “scale-first” approach to strategic stacks (chips + compute), including the efficiency tradeoffs China is willing to accept to sustain momentum under constraints.

Agentic AI, advanced autonomous systems that can act without human supervision, is becoming more common and sophisticated.

An agent that can draft an email is one thing. An agent that can send it, trigger a payout, or change settings in a live system is a very different kind of risk. The operator question is no longer whether AI can help with everyday tasks, but can we safely let it execute actions in regulated systems this year, or keep it in recommend-only mode until controls are provable?

This sounds practical and convenient, but it also comes with serious risks. Failures can threaten safety and financial stability, making governance and control mechanisms central to deployment decisions.

IMDA launched Singapore’s Agentic AI Guidelines in January to promote responsible deployment. The voluntary framework is intended to move faster than legislation and shape what companies can defend in audits, risk reviews, and procurement discussions.

Prof Wei Lu from NTU’s College of Computing & Data Science bluntly describes the shift. “At a fundamental level, the shift from generative AI to agentic AI marks a move from AI as an ‘advisory co-pilot’ to an ‘operational actor’,” he tells Asia Tech Lens.

The biggest risk in early AI rollouts is giving an agent too much power over old, clunky systems, where small mismatches can cascade once automation is introduced. As Lu pointed out, the fix is to set strict boundaries and only give the AI the bare minimum access it needs to do its job.

Before an agent gets permission to act, operators need to answer three things: do you know exactly what the agent is allowed to do, can you audit every action, and can you roll it back quickly when needed? If you cannot answer yes to all three, you are not making a technical choice; you’re taking a governance risk.

Charmian Aw, partner (Data, Privacy, and Cybersecurity) at Hogan Lovells, believes that operators need to be extra careful when using agentic AI in sectors such as finance and healthcare.

“These are repeatedly identified in the Guidelines as high‑impact areas where erroneous or unauthorized autonomous actions can lead to significant harm for individuals,” she tells Asia Tech Lens.

For anyone running these systems, the message is clear: pick exactly which tasks the AI can touch, and make sure humans are always in the loop to spot and stop mistakes before they spread.

Why Voluntary Rules Still Matter

While Singapore’s guidelines are not law yet, they give companies a head start on future regulations.

Aw points out that in a market without strict AI laws, these guidelines serve as a much-needed benchmark. Companies can use them to assess their own risks and conduct audits, demonstrating they’re keeping up with emerging regulatory expectations.

“In practice, commercial entities frequently adopt non‑binding frameworks as procurement and vendor‑management baselines, which may create commercial pressure for service providers to align with the guidelines even without legal compulsion,” Aw said.

For operators running the system, use the guidelines as a checklist. Ask yourself if your AI’s actions are restricted, easy to track, and most importantly, reversible.

“We see the guidelines as likely to influence deployment and rollout, with organizations in highly-regulated sectors adopting tighter controls such as autonomy limits and layered approvals, consistent with the guidelines’ emphasis on restricting bounded autonomy and meaningful human oversight,” Ciara O’Leary, associate (data, privacy, and cybersecurity) at Hogan Lovells, explains.

Even though it is not binding, the framework raises the bar. Teams will need to show they can monitor what an agent does, explain why it did it, and roll back mistakes before giving it more freedom.

What Breaks First

Early failures in agentic AI deployments often occur at the interfaces among agents, tools, and human workflows, especially in legacy environments that rely on implicit judgment rather than explicit, machine-readable rules. In such settings, small mismatches can quickly escalate once automation is introduced, according to Prof Lu.

Most incidents will not look like sci-fi autonomy. They will look like silent misrouting: the agent writes to the wrong field, triggers the right workflow with the wrong parameters, or sends the right message to the wrong customer.

This makes observability an important bottleneck.

“Agentic systems operate at machine speed across multiple tools and environments, generating large volumes of unstructured reasoning traces and execution logs that are difficult to monitor or interpret in real time using existing enterprise tooling,” Lu said, adding that this makes it challenging to detect abnormal behavior early or to reconstruct what went wrong after an incident.

Accountability is messy when too many people are involved. To stay safe, design rollback paths wherever you can. The operational rule is simple: don’t give the systems more power than your safety nets can catch. If you can’t detect and fix a mistake quickly, keep it recommend-only.

Controls Before Scale

Early deployments of agentic AI will likely incorporate human-based accountability mechanisms, which the guidelines explicitly emphasize and that are reflected across global regulatory approaches, according to Aw.

“Given the relative immaturity of agentic systems, sophisticated organizations will likely prioritize controls that strengthen accountability, observability, and reversibility,” she says.

In practical teams, prioritize human approvals. Don’t let the AI make big calls like making a payment or changing safety settings without a human signing off first. Keep a clear paper trail of everything it does, test the system regularly, and make sure you’re still the one in control.

Echoing this view, Prof Lu said that governance must evolve from content moderation toward more demanding priorities, such as enforcing behavioral guardrails with clearly defined action-space boundaries, and extending the principle of least privilege from human users to agent identities.

“Critically, these boundaries must be defined at the design stage, not retrofitted after deployment,” he said.

For operators, think of agentic AI as a step-by-step process. Start with low-risk, easy-to-undo tasks. Only give the AI more freedom once you’ve seen it work and make sure you know how to catch any mistakes before they spread.

Regional Implications

Singapore is ahead of its neighboring Southeast Asian countries in guiding agentic AI, but that doesn’t mean the guidelines will become a single regional standard.

“Southeast Asia reflects a highly heterogeneous and decentralized regulatory landscape. Jurisdictions are taking divergent, and in many cases, deliberately sovereign approaches to AI governance, shaped by differing political priorities, institutional capacities, and levels of digital‑economy maturity,” said Aw.

That said, while the guidelines may inform regional thinking, they seem unlikely to replace domestic regulatory preferences.

For Chinese tech companies operating in Singapore or Southeast Asia, Aw believes that the implications are limited. She points out that China already operates under a set of binding and prescriptive AI regulations, including the Algorithm Recommendation Rules, the Deep Synthesis Rules, and the Generative AI Measures.

ByteDance’s Doubao “AI phone” episode is a reminder that agents are arriving before the scaffolding is ready. If the guardrails do not align with platform rules, autonomy becomes untraceable and reversibility becomes theoretical. For operators, strategy comes first, define controls and boundaries before deployment.

Singapore has set a reference point, not a universal template. For operators deploying agentic AI across Asia, treat Singapore’s framework as the reference procurement baseline, then tune autonomy and evidence requirements to each regulator’s enforcement posture. Do not assume portability across markets unless your control plane is portable.

Related Reading On Asia Tech Lens

• Budget 2026 Puts AI Into Execution Mode. Operators Need To Fund Foundations Before Features
  Singapore’s AI push is shifting from experimentation to production.

• The Chinese New Year AI Gateway War: The Big Four’s Fight for Daily Habit
  A distribution-first AI race: red packets, subsidies, and bundling tactics designed to force trial at scale.

• What Tencent’s ‘Yuanbao PAI’ Reveals About Its AI Strategy
  Tencent’s bet is “Social + AI” - one that rides the existing WeChat social mechanics rather than trying to build a new habit from scratch.

• Why ByteDance’s AI Phone Hit a Wall: Security, Fair Play, and the Economics of Attention
  A cautionary tale of cross-app agency. When an assistant behaves like an operator across apps, platforms start treating it as automation abuse, triggering security and policy constraints that can kill distribution even if the product works.

• Asia’s Agentic Moment: The Manus Interview
  The interview frames the real operator risks (permission creep, brittle tool integrations, and silent workflow errors) and the controls that separate a pilot from production: scoped identities, least privilege, full action trails, and fast rollback.

• At WAIC Hong Kong, the AI Conversation Has Moved Past the Model Race
  A conversation with Steven Hoffman on where the market is actually heading: away from “who has the best model” and towards who can monetize, distribute, and operate AI at scale.

• AI Is Accelerating Cybercrime, and Southeast Asia Is Where The Damage Shows Up
  Maps AI capability to attacker advantage: faster fraud, better impersonation, and identity compromise at scale. Reinforces why agentic deployment must be paired with monitoring, controls, and incident response, not just model upgrades.

Editor’s Note: This week, Asia’s AI story was told through two showcases, one in India and one in China, and they shone a light on where the region is heading.

In India, the spotlight came through the AI summit week, with the country selling itself as a global AI hub. But the message was not only ambition. As the summit played out, the regulatory posture tightened in parallel, a reminder that scale comes with oversight.

In China, the Lunar New Year window and the Spring Festival Gala acted like a national distribution moment. AI and humanoid robots were placed on the main stage, not as a novelty but as a signal of where the country wants momentum to sit. Around it, the holiday period doubled as a live trial of how quickly AI can be pushed into everyday use.

Put together, these two moments made the subtext clear: the next phase of AI in Asia will be decided as much by distribution and governance as by model capability.

AI Impact Summit Week Signals India’s AI Ambition

India’s AI Impact Summit has become one of the week’s defining showcases, drawing global tech leaders to New Delhi and framing India as the next major theatre for AI adoption and governance.

The summit’s biggest announcements were about infrastructure and capacity, not a single product drop. Google rolled out a new India–US subsea cable initiative, launched a $30 million AI-for-Science Impact Challenge and a second $30 million challenge focused on AI for government innovation, and expanded its skilling push, including training aimed at public servants. Microsoft used the week to reinforce its access-and-adoption pitch, saying it is on pace to invest $50 billion by 2030 to expand AI access across the “Global South.” On domestic capacity, Yotta announced a $2 billion investment in Nvidia’s latest chips to build a major AI computing hub, including a New Delhi “supercluster” targeted to go live by August 2026. The Indian government also underlined its own compute intent, announcing plans to add 20,000 GPUs under its national AI push.

Running in parallel with the summit, though, was the sharper signal: a tightening compliance posture. India’s IT minister told global platforms they must operate within India’s constitutional framework and cultural context. Earlier this month, the government updated its rules to formally define AI-generated content and make platforms responsible for ensuring it is clearly labelled, while also cutting the deadline to remove unlawful content to three hours after notification, down from 36 hours. In finance, the Reserve Bank of India (RBI) has asked lenders to adopt board-approved policies for AI use, keep model information available for internal and external audits when required, ensure human oversight for models used in financial decision-making, and disclose where AI is used in products and services as adoption accelerates.

Signals To Watch:

• From pledges to projects: which summit announcements translate into deployments you can point to, not just MOUs and skilling headlines.

• Three-hour takedown becomes real: whether enforcement is strict and consistent, and how quickly platforms build 24/7 escalation, legal review, and audit trails to match.

• Labeling starts to bite: whether “AI-generated” labeling becomes standardized in products and platforms, and how aggressively non-compliance is penalized.

• Finance rules harden: whether RBI guidance turns into concrete supervisory expectations, including board-level governance, audit readiness, human oversight, and mandatory disclosure.

• Capacity buildout changes the map: whether public-sector GPU additions and private builds (like Yotta) materially shift where large AI workloads run in India over the next 6–12 months.

• Who wins the ecosystem layer: whether cloud, telco, or domestic infrastructure players become the default on-ramp for enterprises and government agencies adopting AI.

Lunar New Year Turns Into an AI Stress Test in China

In China, the Spring Festival Gala is not just entertainment. It is the country’s biggest attention node and a reliable signal of what the system wants to elevate. When humanoid robots move from a side segment to the main stage, it is a message about industrial direction, status, and momentum. The Gala is where national priorities get packaged into something mass audiences can absorb, and where companies get to borrow institutional credibility at population scale.

What mattered this week was what sat underneath that spectacle. The Spring Festival window has become a structured launch season for consumer AI. Major platforms timed upgrades, releases, and distribution pushes around the holiday peak, using it to drive trial and to test whether assistants can turn into daily routines. The push toward “agent” positioning, and the effort to embed AI inside core apps, were part of the same pattern: the fight is moving from who has the best model to who becomes the default layer people reach for first.

For operators, the Gala week compresses the future into a few days. It shows what “mass adoption conditions” actually look like: sudden traffic concentration, high expectations for uptime and safety, and a user base that will churn fast once incentives or novelty fades. If you sell AI products or rely on AI workflows, treat these attention spikes as real-world stress tests. Build for resilience, guardrails, and escalation, and watch the companies that convert a national showcase into repeat use after the holiday peak.

Signals To Watch

• Post-holiday retention: who keeps users once the peak traffic and promotions fade.

• Agent reality vs agent branding: which products move from “agent” marketing to reliable multi-step task completion inside real workflows.

• Distribution shifts: whether AI features get pulled deeper into default surfaces (search, messaging, payment, short video, super-app entry points) rather than living as standalone apps.

• Operational safety as a differentiator: how players handle misfires at scale, including content risk, model drift, and escalation speed under public scrutiny.

• Robots beyond spectacle: whether humanoids show up in repeatable deployments in factories, logistics, retail, or public services instead of remaining choreographed demonstrations.

  ---

Takeaway

If you are buying or building, carry forward the practical questions: what parts of your rollout depend on temporary attention spikes versus durable integration, whether your operating model can meet faster enforcement rhythms, and whether your governance, labeling, oversight, and audit trails are strong enough for AI to live inside everyday workflows without eroding trust.

Sources

• Reuters: From OpenAI to Google, India hosts global AI summit

• Reuters: India File: AI fanfare meets policy pushback

• Al Jazeera: India’s hosts AI Impact Summit, drawing world leaders, tech giants

• The Economic Times: Google announces $30 million science fund; new subsea connectivity initiative

• Microsoft Blog: We need to act with urgency to address the growing AI divide

• SCMP: Kung fu, somersaults and scale: Unitree eyes 20,000-robot output in 2026 after gala

• 36Kr: Doubao Officially Announces Interactive Games for CCTV Spring Festival Gala, ByteDance’s AI Continues Aggressive Advance

• Global Times: AI giants ramp up user competition with pre-Spring Festival digital red envelope campaigns

Singapore’s Budget 2026 elevates AI from emerging tech to a national priority. A new National AI Council, chaired by PM Lawrence Wong, will coordinate strategy, regulation, and resources.

National AI Missions will target four sectors: advanced manufacturing, connectivity including logistics, finance, and healthcare. Enablers include a new one-north AI park for testing and scaling, expanded TeSA program to build AI literacy in non-tech roles such as accountancy and legal, and a Champions of AI program to support enterprise transformation.

For operators in banking, logistics, manufacturing, and similar sectors, this is not a signal to rush AI purchases. Long procurement cycles, legacy cores, and regulatory scrutiny mean early spending choices determine whether deployments scale safely or create operational risk.

The temptation is to fund visible, demo-friendly tools, leveraging government incentives, missions, and talent programs. These feel aligned and low-risk but often fail under production load, audit pressure, or incident response when deeper foundations are absent.

So the question is simple: are you funding AI tools, or the foundations that let them survive production?

The Operator Gate: Three Tests Before Scaling

Before scaling any AI system, operators should test these three basics:

• Reliability: Can networks, integrations, and core systems handle real-world usage without degradation or failure? If not, AI will falter at peak traffic or during incidents.

• Evidence: Can you rapidly trace data inputs, model changes, and decision rationale? Without this, audits, compliance reviews, and post-incident probes become major risks.

• Portability: Can you switch vendors or roll back without rewriting core workflows? Vendor lock-in turns every update into a costly, high-risk event.

Foundations, Not Models, Drive ROI

Most AI projects fail not because models underperform, but because surrounding systems are unprepared. In regulated sectors, success hinges on legacy integration, clean data flows, robust identity/access controls, monitoring/incident response, and disciplined change management.

As Adeline Liew, Country Business Leader, Singapore, Alcatel-Lucent Enterprise, said in her notes to media, the real gap is between ambitious AI applications and the aging infrastructure required to run them. For many enterprises, the bottleneck is not the model itself but the reliability and speed of internal networks. If underlying systems are slow or disconnected, AI investments struggle to deliver returns.

Budget 2026 incentives are sequencing tools, not transformation funding. The enhanced Enterprise Innovation Scheme (EIS) offers 400% tax deductions on qualifying AI spend, capped at S$50,000 per year of assessment for YA2027 and YA2028. It helps offset early experiments and foundation work, but it will not remove enterprise constraints on delivery.

“To realize the full economic potential of these national investments, infrastructure must be viewed as a strategic business asset rather than a back-office expense,” said Liew.

In this view, AI readiness becomes a core business capability: investing in modernized foundations enables organizations to move beyond testing and achieve sustained gains in productivity and service delivery.

Use these incentives for low-regret priorities: foundational cleanup such as data lineage and monitoring upgrades, and tightly controlled pilots, not broad deployments.

Skills Boost Adoption; Operating Models Ensure Survival

Budget 2026 expands AI training across the workforce, including non-tech roles, raising baseline comfort and usage. Yet training alone does not redefine decision rights, escalation paths, or risk ownership. In regulated environments, redesigning workflows, governance, and accountability determines whether systems survive audits, outages, and other future problems.

PM Wong framed the Budget as a shift from isolated pilots to scaled deployment at national speed. For operators, speed and scale only work if operating models evolve too. As KPMG partner Edmund Heng notes, “Clear accountability, a risk-based approach, and early governance will be critical for sustainable AI implementation at scale. Good governance empowers AI adoption with confidence, while unclear governance hinders it.”

What to Fund Now, Pilot, or Delay: By Sector

National AI Missions provide sequencing signals. Each sector carries different evidence burdens and incident tolerances.

Finance and healthcare

• Fund now: Data lineage, audit trails, model validation, and rollback controls meeting regulatory standards.

• Pilot with guardrails: Narrow decision-support use cases under human oversight with clear incident playbooks.

• Delay: High-impact automation in core banking or clinical workflows until auditability, explainability, and safe rollback are proven. In practice, this means decisions that change customer or patient outcomes without an auditable human sign-off path. In healthcare, delay automation that touches triage, diagnostic support, or patient routing unless model behavior can be demonstrated and safely rolled back.

Advanced manufacturing and connectivity/logistics

• Fund now: Legacy system integrations, network upgrades, and monitoring for stable data flows under load.

• Pilot with guardrails: Limited automation on specific lines, routes, or segments with manual fallback.

• Delay: End-to-end autonomy across systems of record unless integrations have proven SLOs and tested fallback paths.

In all cases, prioritize low-regret foundations, run scoped pilots with rollback, and scale only after reliability, governance, and accountability prove themselves in live conditions.

Budget 2026 makes AI a national imperative but leaves execution risks inside enterprises. Competitive advantage will come not from adopting fastest, but from sequencing investments more deliberately than peers: fund foundations first, treat missions as leading indicators, and move with disciplined speed.

Related Reading On Asia Tech Lens

• The Chinese New Year AI Gateway War: The Big Four’s Fight for Daily Habit
  How red packets, subsidies, and new AI apps are being used to force trial at scale, and what happens when incentives end.

• AI Is Accelerating Cybercrime, and Southeast Asia Feels It
  Why AI raises the baseline threat level, and why resilience, monitoring, and incident response matter as much as model capability.

• Why ByteDance’s AI Phone Hit a Wall: Security, Fair Play, and the Economics of Attention
  A useful case study on what breaks at scale when security, platform rules, and operational constraints collide with fast product rollout.

• AI for Global Equity Begins With Local Realities
  Why the hardest AI problems are trust, language, and deployment, and why “responsible” starts with real user constraints.

• The Chip War’s New Reality: A View from the Crossroads
  Why AI and semiconductor decisions are becoming geopolitical choices from Singapore’s vantage point, and how ecosystem fragmentation raises the stakes for vendor lock-in and resilience.

For China’s tech giants, the CCTV Spring Festival Gala is a once-a-year distribution super node. It is the closest thing China has to the Super Bowl, and the sponsorship optics carry institutional weight in a system where industrial priorities are set from the top. The Gala made that subtext visible. Humanoid robots were not a side act. It signaled what platforms want to be associated with: industrial capability, not just consumer entertainment.

While robots took center stage on screen, a parallel battle ran underneath it in the infrastructure layer. ByteDance’s Volcano Engine secured an ‘exclusive AI cloud partnership’ for the 2026 Gala, following Alibaba Cloud’s exclusive cloud-and-AI role for the 2025 show. That matters in a market where demand for AI workloads is rising and cloud competition is tightening. In AI, the company that wins the infrastructure relationship can shape the rest of the lifecycle, from deployment and performance to cost, distribution, and commercial leverage.

The Gala is also the sharpest peak in a wider Lunar New Year campaign window that stretches across late January through February, where red packets, perks, and sharing mechanics are used to force trial at scale and test who stays once incentives fade. Asha Hemrajani, Senior Fellow, S. Rajaratnam School of International Studies at Nanyang Technological University, sees it as a sign that consumer AI is moving into what she calls the age of “agentic commerce,” shifting from assistants that respond to queries to systems that “act on the consumer/customer’s behalf,” including automating decisions about what to buy.

The shift in tone is hard to miss. After years of talking up cost-cutting and efficiency, the biggest platforms are back to paying for behavior change. Liang Chen, an associate professor of strategy and entrepreneurship at Singapore Management University, calls the giveaways “a subsidy on user behavior,” arguing the companies are “buying the right to be users’ default interface for the next decade.”

That bet is now visible in the incentive pools and the distribution mechanics. Alibaba’s Lunar New Year campaign, launched on February 6, will distribute RMB 3 billion (US$431 million) in incentives through Qwen. It has packaged the push around high-frequency consumption, including AI-driven ordering for everyday services. Tencent put RMB 1 billion (about US$140 million) behind Yuanbao’s cash red-packet campaign. Alongside the incentives, Tencent promoted Yuanbao PAI, a group-based feature built for coordination and shared use. Think of it as a shared “room” rather than a chat thread: users join a PAI space and the assistant participates in the group, designed to reduce friction in multi-person planning through quick summaries, reminders and timed updates, while also enabling lightweight creative generation that can be shared back into the group. Baidu is running an RMB 500 million (about US$70 million) incentive programme to pull users into its assistant experience over a longer window. ByteDance is anchoring its push on the Gala’s credibility stage through Volcano Engine, alongside a broader Doubao holiday promotion drive.

That “brute force” matters because the real goal is to change what people do every day. If an AI assistant becomes the first place people go to ask a question, plan something, or trigger an action, it starts to function like a new interface layer on top of apps. Steven Hoffman, the Chairman & CEO of Founders Space, who tracks platform strategy in China, says the harder part is keeping users engaged once the giveaways stop and the novelty wears off.

The winner among them will be the one that can turn a Chinese New Year download spike into a daily routine, by making its AI feel like the easiest, most natural way to get things done.

From Territory Wars to Gateway Wars

China’s platforms have fought brutal battles before. Ride hailing, food delivery, e-commerce, short video, payments. Those wars were about owning a vertical. Liang Chen, Associate Professor of Strategy & Entrepreneurship at Singapore Management University, says this round is different because it targets the layer above the verticals.

Liang describes this round as “a battle for the gateway.” If one player becomes the default AI entry point, it can start routing users into everything else, effectively brokering traffic and transactions across other apps and services. He compares it to earlier distribution resets created by control of the entry layer, from WeChat’s super-app gravity to Google’s Android advantage.

He says “the Spring Festival provides a window for resetting user behavior” when hundreds of millions of people pause routines at the same time. That pause creates a brief opportunity to install new behaviors quickly. In that sense, the subsidies are a temporary discount on habit change, buying the first wave of trial at national scale.

Liang argues the real goal is a forced migration from ‘search’ to ‘agent,’ paying for the first-mover chance to become the next operating system rather than just another app.

Ashley Dudarenok argues Chinese New Year is exposing a shift in what companies are building. “This Spring Festival marks the death of ‘AI as a feature,’” she says. The big spend is pushing standalone assistants toward habit formation, with integration into everyday workflows as the differentiator, from commerce use cases like Qwen’s ordering hooks to social coordination features like Yuanbao PAI.

Four Plays, Four “DNAs”

The giveaway war may look uniform on the surface, but the strategies underneath are not.

“Each company is playing to its unique strengths, revealing their core DNA,” Ashley Dudarenok says.

• ByteDance (Doubao): Ashley Dudarenok says ByteDance is “all about platform independence,” using Douyin and its CCTV tie-up to bypass rival ecosystems. The Gala badge matters here. Volcano Engine, ByteDance’s cloud arm, was named the show’s exclusive AI cloud partner, a signal that ByteDance can run national-scale AI workloads on its own stack rather than renting credibility from someone else’s ecosystem. Alongside that infrastructure flex, ByteDance used the holiday window to push product momentum. It promoted Seedance 2.0, its AI video-generation model, as a marquee capability inside Doubao, reinforcing the message that Doubao is not just another chatbot, but an interface for creating and doing. Beyond these headlines, ByteDance also ran a broader Lunar New Year giveaway push to drive trial and sharing at scale.

• Alibaba (Qwen): As per Ashley, Alibaba is focused on “consumption integration,” tying AI to real-world transactions so it feels immediately useful. That’s why it has put RMB 3 billion (about US$431 million) behind Qwen. The simplest way to see it is this: Qwen is being pushed less as a chatbot and more as a tool that can trigger a real purchase. The proof point is the milk-tea giveaway. Within hours of launch, the surge pushed Qwen past Tencent’s Yuanbao to the top of China’s Apple App Store free chart. Alibaba’s Qwen team said more than 10 million free orders were placed within nine hours, using vouchers capped at 25 yuan. The spike overwhelmed many stores, with users reporting overload and disruption as traffic surged.

• Tencent (Yuanbao): Ashley calls Tencent’s approach “the most nuanced play,” because its real edge is not a bigger model, but AI + social. Tencent put RMB 1 billion behind Yuanbao’s cash red-packet campaign. In figures released on Feb 18, Tencent said Yuanbao’s daily active users (DAU) exceeded 50 million and monthly active users (MAU) reached 114 million, with users participating in more than 3.6 billion lucky draws and making over 1 billion AI creations. But Tencent is trying to make sure Yuanbao is remembered for more than cash. That is where Yuanbao PAI comes in: a group feature built around shared activity rather than solo prompting. PAI is closer to a shared room than a chat thread, designed for coordination and group utility, from lightweight recaps and timed info drops to simple creative outputs that can be generated and shared in-context. The architecture is the tell. Tencent is not dropping an AI bot directly into WeChat group chats as a visible participant. It is ringfencing the experiment inside Yuanbao, using PAI as a bridge to Tencent’s social graph without risking backlash inside its core messaging product. Steven Hoffman argues that is precisely the advantage: “the big advantage of Yuanbao PAI is that it is literally injecting AI into the space where people live, communicate, socialize, and do everything they need to get done.” But he also flags the constraint: in group spaces, AI can feel intrusive fast. If Tencent wins this round, it will be by finding a workable social contract for AI in groups, useful enough to reduce friction, quiet enough to keep the conversation human.

• Baidu: Ashley describes Baidu as being on a “quiet scale quest.” With RMB 500 million behind its holiday push, Baidu is trying to convert existing habits into AI habits through a spread of seasonal features and incentives. This push is built for endurance. Baidu’s rewards run from January 26 to March 12, with prizes up to 10,000 yuan, and the company is bundling the push into the Baidu App via nearly a hundred Spring Festival-themed activities tied to the Beijing Radio and Television Spring Festival Gala. What makes Baidu different is the installed base it is trying to redirect. The Ernie AI Assistant has reportedly surpassed 200 million monthly active users, embedded inside Baidu’s core surfaces and connected into services such as JD.com, Meituan and Trip.com. Baidu is betting that distribution plus integration can outlast the holiday spike.

The Retention Cliff

The red packets are doing what they have always done in China’s internet wars. They pull people through the door fast. The harder part starts when the door prize disappears.

Ashley Dudarenok calls the escalation a warning sign in itself. “Firstly, incentive inflation,” she says. “The budget escalated from RMB 500 million to RMB 3 billion in a matter of days. This is an unsustainable subsidy war.” The spending can force trial quickly, but it also sets a trap: if users arrive for the prize rather than the product, they leave the same way.

Steven Hoffman is even more direct about what happens after the holiday glow fades. “People will try it out,” he says, “but they won’t see the value there, and then they will revert to doing what they did before.” In his view, the deepest challenge is not distribution. It is the fact that “what people really want and what you think they want are often two entirely different things.” The Spring Festival window can manufacture curiosity, but it cannot manufacture conviction.

That gap is especially sharp in social settings, where platforms can’t afford to break the vibe. “In these group chats, people don’t really want AI interfering,” Hoffman says. “They don’t want AI taking a lead role.” That, in his view, is why Tencent is proceeding cautiously rather than pushing a fully integrated experience too fast.

The Cost of Failure

The second half of the story is not only about retention. It is about risk.

Asha Hemrajani worries most about transaction-heavy workflows because one small error can spread. She points to a pattern described by OWASP, a widely used set of application-security guidelines: “Cascading Failures,” where a single initial error, “from a malicious prompt injection, a bug in the agent, or a misunderstanding of the end-user’s command,” can propagate across multiple systems before a human has the chance to intervene. She also flags “deliberate or accidental leakage of sensitive personal information which can then be used to propagate other crimes such as scams.”

Trust constraints are not limited to transaction safety. They also extend to provenance and rights in generative media. ByteDance’s Seedance 2.0 drew attention for its capabilities, but also sparked controversy over training data and content similarity claims, a reminder that create features can turn into governance and reputational liabilities.

In other words, the moment the assistant can place an order, move money, or trigger services, trust becomes a hard gate. If users believe the system can misfire at speed, they stop delegating. At that point, no incentive pool can buy them back.

Hemrajani adds that if these consumer AI playbooks are exported into Southeast Asia, the dependency question becomes political as well as technical. Chinese apps remain subject to China’s cyber and data laws, including the National Intelligence Law, which has shaped regional scrutiny over data access and sovereignty.

Liang Chen adds a different kind of constraint, one that does not show up on App Store charts. He argues the strategic risk is building a massive user base that is economically negative to serve because inference costs are variable. Unlike traditional apps, where serving an extra user is close to free, AI usage scales with cost. If a platform buys a user base that is there for rewards rather than real demand for the intelligence, it can end up carrying high servicing costs without durable value in return.

Regulation Is Tightening the Arena

Hemrajani’s view is that this consumer AI escalation is also happening in a narrower arena. She points to two concrete signals: “In September 2025, the Chinese State Administration for Market Regulation (SAMR) intervened in the food delivery marketplace and warned e-commerce platforms about too high subsidies. In January 2026, Chinese regulator banned platforms from forcing their online merchants to launch promos and give discounts. This new rule is effective this month (February). The aim of these new regulations is to cool down intense competition between the platforms. There is good reason to do so because this intense competition has been eating into the profit margins of these platforms and thus has a deflationary impact on the Chinese economy.”

This is why she reads the AI app push as more than product ambition. “These platforms launching AI apps is yet another escalation in their battle for market share,” she says, “given that other routes have been closed to them due to new government regulations.”

What to Watch After the Festivities End

The subsidies cannot last forever. There is a definite spike right now because the incentives and attention are concentrated, but that limited-time surge alone doesn’t prove the plan worked.

What matters is what usage looks like when the festival ends and routine returns. Ashley Dudarenok says she would watch for a retention cliff, warning that Spring Festival incentive apps have historically seen very low carryover, sometimes dropping to single-digit retention after the first week.

Her scoreboard is simple:

• DAU/MAU ratio: the stickiness score. Does daily usage hold up, or does it collapse once the holiday rush passes?

• Session frequency: are users opening the app multiple times a day for different tasks, or just once to check for leftover rewards?

• Core feature engagement: what share of daily users are actually using the assistant’s capabilities rather than spending most of their time in the rewards layer? If the traffic stays concentrated in red packet mechanics, Dudarenok warns, “you haven’t built an AI product; you’ve built a temporary casino.”

Liang Chen frames the same test in behavioral terms. Look for longer, more complex interactions, not just quick prompts. Look for evidence that queries lead to completed actions inside the ecosystem, not dead-end chat. And watch whether usage holds after people return to work, when patience drops and utility has to be obvious.

In Liang’s view, retention will not come from social virality but from integration depth. The agent that holds payment credentials, knows your location, and has a history of successful transactions becomes hard to replace.

If this holiday sprint shows anything, it is that China’s platforms can still manufacture mass trial on command. What they cannot manufacture is habit. That has to be earned after the festival ends, when users return to normal routines and start judging these assistants on whether they save time, reduce friction, and complete tasks reliably.

The direction of travel is clear. Consumer AI is shifting from being something people talk to, toward something people delegate to. That shift concentrates power in the layer that can turn intent into action safely, across payments, services, and coordination, without breaking trust.

The next few months will decide whether this was a seasonal spike or the start of a new default. The winner will not be the platform that bought the most attention. It will be the one that users keep opening when there is nothing left to collect.

Related Reading On Asia Tech Lens

• What Tencent’s “Yuanbao PAI” Reveals About Its AI Strategy

  Why Tencent is testing “social AI” during Spring Festival; explains how group coordination features could turn AI into a daily habit layer.

• AI for Global Equity Begins With Local Realities
  Why the hardest AI problems are trust, language, and deployment, and why “responsible” starts with real user constraints.

• Thailand’s Digital Transformation
  A ground-level look at how platforms, payments, and infrastructure shape adoption, and how China-linked tech shows up in everyday systems.

• Inside China’s Bold Push to Build Humanoid Robots
  Why humanoids are becoming an industrial signal; looks at policy support, public demos, and the race to move from spectacle to scale.

• Plush, Playful, and Powered by AI: How China’s Toys Are Evolving
  Why China’s AI-toy boom matters for consumer adoption; compares China’s fast product cycles with a more cautious Western approach.

• Why Southeast Asia’s Super Apps Won’t Look Like China’s
  Why WeChat-style dominance is hard to copy, and why distribution in SEA looks more fragmented, regulated, and competitive.