Chips, cognitive labour, and control in Southeast Asia

Chapter 1 · 2025 to 2026

Two forces arrive at once: a tightening of who may buy the best chips, and a flood of capable models that anyone may download. The region discovers it is not a bystander.

Kim Seng had spent nineteen years learning that a factory's most important document was its inventory sheet. On the third Monday of March he printed his and took it to the window, where the light was better and nobody could read over his shoulder.

The numbers were good in a way that made him uneasy. Substrate on hand: eleven weeks. Bonders: fully committed through the following March. Two customers had quietly asked whether the plant could take a third shift, and a third had asked whether it could take a whole building. In the trade press these were called capacity constraints. On the floor they felt like something closer to gravity: everyone in the world wanted the same narrow set of operations performed on the same narrow set of parts, and there were perhaps a dozen places on earth where those operations could be performed at volume.

For decades, advanced packaging was the unglamorous end of the semiconductor business. It was the back end, the work that came after the clever people in the fab had finished. Now they needed it more than it needed them. Stacking memory next to logic, at yield and at scale, had become the step that limited AI accelerator production. Malaysia had roughly a tenth of the world's assembly, test, and packaging capacity and a workforce that had done this work since the 1970s.

Kim Seng understood exactly what that meant, and it did not make him feel powerful. A chokepoint you cannot use still leaves you exposed. The tools on his line, the software that ran them, and the customers who bought its output all sat under a licensing regime written in Washington and a supply chain that ran through Tokyo, Eindhoven, and Hsinchu. He had a chokepoint but no hand on the valve.

Ratna's version of the same problem arrived as a PDF at 4:40 in the morning, which is how most of her problems arrived.

The document proposed sorting countries into tiers for the purchase of advanced computing hardware. A handful of allies would face no meaningful limits. Most of Southeast Asia would sit in a middle band with caps, licences, and conditions. A short list would get nothing at all. Ministries across the region read it like a medical result they had not asked for.

The rule was rewritten within months, then rewritten again, then partly replaced by bilateral arrangements. The churn mattered more than any particular version. Access to frontier compute became a diplomatic variable that could shift at short notice in response to events far away.

Ratna started keeping two lists. The first was of things her governments could not control: model capability, chip supply, the mood of a foreign legislature. The second was of things they could: where datacentres were sited, what the grid connection agreements required, which workloads got priority, how procurement was written, what a public agency was permitted to send to a foreign inference endpoint. The second list was shorter and much more interesting.

While governments argued about who could buy accelerators, the other half of the story was being given away.

Open-weight models were downloadable, modifiable, and runnable on hardware a mid-sized firm could afford. For ordinary commercial tasks, they closed much of the gap with frontier models. They were good enough to summarise a contract in Bahasa Indonesia, triage a support ticket in Taglish, extract fields from a Vietnamese invoice, or draft a compliance memo that would have taken a junior associate an afternoon.

This produced an odd asymmetry that shaped everything afterwards. The capability to disrupt ASEAN's largest white-collar employment sectors arrived faster and cheaper than the capability to train competitive models domestically. Disruption was democratised. Frontier development was not.

Mara's floor had not changed yet. Sixteen hundred seats across four floors, three shifts, the same fluorescent hum. But the sales conversations had changed, and she heard about those before the floor did.

Clients had stopped asking about cost per seat. They asked about cost per resolution, then about resolution without a seat. Pilots in other buildings deflected simple tickets, drafted replies, and routed only hard cases to people. On a slide, the savings looked like free money.

The sector employed well over a million Filipinos directly and supported several million more indirectly. It was, by a wide margin, the country's most successful bet on cognitive work. Studies circulating that year estimated that roughly eighty million workers across ASEAN held jobs with meaningful exposure to generative AI, concentrated in exactly the clerical and digital-service categories the region had spent thirty years building.

Mara read one of those studies on her phone during a break, standing by the window on the eighteenth floor. Then she went back inside, because the queue was long and the night was young, and exposure is a word for something that has not happened yet.

Chapter 2 · 2027 to 2028

The automation of routine cognitive work stops being a pilot and becomes a quarterly target. At the same time, packaging capacity becomes a bargaining chip in someone else's negotiation.

The fluorescent lights never slept.

From the 18th floor, Mara could see Quezon City's towers blinking through the haze, a lattice of reds and whites cutting into the monsoon clouds. Inside, the contact-centre floor hummed the way it always had: overlapping voices, the soft thud of headsets against desks, the occasional whoop from the QA team when someone got a perfect score.

Tonight, the hum sounded different.

On the right side of every agent's screen, a new pane flickered green. It listened to the customer, transcribed in real time, pulled account details from the CRM, and drafted replies before the human could even start typing. The company called it the copilot panel. The vendor called it a "next-generation agentic assistant." Mara called it trouble.

"Ma'am, this is so much easier," Jen said during their first break. "I just click 'accept' and it fills in half the form."

"Easier is good," Mara said. "As long as you're still checking it."

She didn't mention the slide deck she'd seen from regional HQ. Phase 1: augmentation. Phase 2: "productivity optimisation." A bullet buried near the bottom: headcount reduction through AI-enabled efficiency.

For a few months, the story was simple. Average handle time down by more than 20 percent. Customer satisfaction up. Error rates cut in half. Supervisors loved the new dashboards. One of the US clients sent a video of a cheerful VP saying how proud they were to be "on the frontier of AI-powered customer care."

Mara's spreadsheet told a different story.

She tracked the tasks the copilot handled almost alone: password resets, address changes, simple refunds, and appointment scheduling. These had trained new hires in the old world. They were starter calls that taught rhythm and empathy. Now, people barely touched them.

In late 2027, a regional director flew in from Singapore. Navy suit, perfect hair, a watch that probably cost more than an agent's annual salary. In the town hall, he framed everything as opportunity.

"We're moving into a new era," he said. "Our clients want strategic partnerships. Our AI systems will handle routine work. You will focus on complex cases and client relationships."

Someone in the back asked what Mara had been thinking.

"And what happens to the people whose work was...routine?"

The director smiled the well-practised smile of someone with a script.

"We'll manage this transition responsibly," he said. "There will be reskilling opportunities. We're talking to government partners about future-skills vouchers. Nobody will be left behind."

Two weeks later, a memo announced that one of their largest US accounts was "optimising its footprint." It would move 60 percent of Tier-1 support to a fully automated system from the vendor behind the copilot panel. Night-shift seats would go first.

When Mara opened the spreadsheet with the affected staff, her stomach clenched. She knew the stories behind the employee IDs: Ana, whose husband's tricycle had broken down; Luis, who sent remittances to a sick mother in Iloilo; Jam, who was saving to finish nursing school. The memo's "30 percent reduction over 18 months" was abstract. The names were not.

Government policy arrived like a press release: a "Future of Work" program with stock photos and a minister surrounded by tech executives. The training modules taught generic "AI literacy" and basic Python. They had nothing on evaluating customer-service models, spotting subtle failure modes, or working with regulators. The gap between political language and the floor's reality felt like a chasm.

One Tuesday, halfway through a graveyard shift, an unfamiliar address appeared in her inbox: asean-transition@...

We're a regional group working on AI and labour transitions in ASEAN. We've heard about changes in your sector. We're looking for partners to pilot a program that retrains BPO workers into AI deployment, evaluation, and governance support roles. Would you be willing to talk?

She almost deleted it.

For months, her life had been all fire-fighting and paperwork: negotiating voluntary-separation packages, smoothing over client calls, explaining to agents why some got retraining offers while others didn't. Another "initiative" sounded like more meetings and more broken promises.

But something about the phrasing made her pause. They weren't selling motivational slogans or generic "digital skills." They were asking about concrete capacities: red-teaming customer-service models; designing realistic tests for AI systems; monitoring early warning signs of fraud and abuse in AI-mediated workflows.

She clicked Reply.

A week later, she was on a video call with three people: a policy analyst in Jakarta, a labour economist in Bangkok, and a soft-spoken engineer dialling in from Penang who said he worked on "advanced packaging" for AI chips. They asked sharp, practical questions.

"What fraction of your tickets do you think could be fully automated today?"

"How many agents have experience in QA or process improvement?"

"Would some of them be interested in monitoring how the AI behaves as well as using it?"

For the first time since the copilot panel appeared, someone treated her people as experts in how messy, real-world systems fail.

The pilot they proposed was modest: fifty workers from her company and two others; a twelve-week program split between technical training, scenario exercises, and placements with regulators, NGOs, or internal AI-governance teams. The group had secured small grants. They needed an industry partner willing to share data and co-design the curriculum.

"What's in it for you?" she asked.

"If ASEAN does not build its own evaluation and governance capacity," the analyst said, "we will import other people's judgments about safety and fairness. Your people touch those judgments every day."

Through the glass wall, Mara watched the floor: headsets on, agents working through queues under the constant glow of the green copilot panel. She imagined some of them, a year from now, sitting across from regulators explaining where the model failed; or pushing back when a client wanted to roll out a risky automation.

"I can't promise the board will approve it," she said. "But I can promise a meeting."

"Sometimes that's the hardest part," the analyst replied.

Two months, a dozen memos, and one supportive CFO later, the pilot was approved. It was small and precarious, tied to funding cycles and politics she could not control. Looking at the first fifty names, anxious, hopeful, curious, she felt something that had been missing since the copilot panel lit up.

It was not optimism. It was agency.

The AI systems were not going away. Neither were the export-control fights over chip shipments to Penang or the grid negotiations around new datacentres in Johor. For the first time, though, the region was building people who could see how AI behaved in the real world and push back when it failed.

On the evening before the first training session in Jakarta, she wheeled her suitcase through NAIA's departures hall, past a billboard advertising a new "AI-enabled super-app." She thought of the thousands of workers who wouldn't be on her flight, and the thousands more who might follow if this worked.

At the gate, she opened her laptop and drafted a message to her night-shift supervisors.

I'll be gone for two weeks. While I'm away, watch where the copilot helps and where it fails. Write the failures down. That is how we make our case.

She hit Send, then boarded.

Somewhere above the South China Sea, in that thin layer where radio waves carry futures back and forth, the first cohort of ASEAN's AI-governance workforce was on its way.

Kim Seng's plant became newsworthy on a Tuesday, which is when trade actions are usually announced.

The measure itself was narrow: additional licensing on a class of high-bandwidth memory and on the equipment used to integrate it. What it did to his order book was not narrow at all. Three customers suspended releases the same week, not because they were prohibited from shipping but because their lawyers could not say within thirty days whether they were. Compliance ambiguity does the work of a ban at a fraction of the political cost.

Then the second call came, from a different direction. A large customer wanted assurances that the facility would not process parts for a named competitor. In exchange: a five-year volume commitment and co-investment in two new lines.

Kim Seng put the phone down and sat with the thing he had avoided since March 2025. His plant was no longer a supplier. It was something other people wanted to take control of.

He drafted a note to the industry association proposing something that sounded naive when he wrote it and less naive every year afterwards: that packaging and test capacity in the region should be offered on published, non-discriminatory terms, coupled with hard commitments on provenance, end-use transparency, and safety-relevant disclosure. Neutrality as a product, sold at a premium, backed by a shared regional position rather than by one company's nerve.

Eleven of the forty-two recipients replied. Four said it was interesting. Two said it was dangerous. One forwarded it to Ratna.

Chapter 3 · 2029 to 2031

Governments start buying compute the way they once bought power stations. They discover that owning the building is not the same as controlling the system.

The site visit was scheduled for eleven and Ratna arrived early, which meant standing in the heat looking at a substation while an engineer explained load curves.

The campus behind him was the fourth of its kind within forty kilometres. Together, the sites represented several gigawatts of contracted load. A decade earlier, national energy plans would have called that a decade's growth. Those plans had already been rewritten twice in three years and were about to be rewritten again.

"What runs here?" Ratna asked.

The engineer was honest, which she appreciated and did not expect. "We don't know. We provide the shell, the power, the cooling, the security. The tenant provides the racks and the workloads. We see utilisation. We don't see purpose."

That sentence stayed with her longer than any briefing she read that year. The region had absorbed enormous physical infrastructure, jobs, tax revenue, and strain on water and grid. It had almost no visibility into what the sites computed, for whom, or under whose law.

The politics of compute stopped being about chips and became about electricity bills, which is a much more dangerous kind of politics.

In two markets, tariff adjustments landed in the same quarter as new datacentre approvals. The connection was contested and mostly wrong. Fuel costs and legacy subsidy reform drove the increases. The timing was still politically fatal. Opposition parties discovered that "foreign machines, our blackouts" fit on a placard. One state assembly passed a moratorium, then quietly relaxed it eleven months later after the investment moved next door.

The governments that came through this best had done something unglamorous years earlier: they had made new large loads pay for their own firm renewable generation and their own network reinforcement, published water draw, and reserved a defined share of capacity for domestic public-interest use. Where those terms existed, the political argument had an answer. Where they did not, it had only a press release.

Mara's pilot of fifty had become a network of about nine hundred across five countries. It was still small. It was no longer ignorable.

What made it useful was not that its members could train models. It was that they could break them in ways that mattered locally. They knew what a fraudulent loan application looked like in three languages and two dialects. They knew which polite refusals a system produced that would be read as an insult in Manila and as ordinary bureaucracy in Singapore. They knew the specific failure that occurs when a model trained mostly on English confidently mishandles a code-switched sentence in a dispute about money.

The regional incident-reporting scheme drafted in 2030 was voluntary, thin, and much weaker than its authors wanted. Former contact-centre supervisors wrote a disproportionate share of its technical annexes. That was not a heartwarming detail. It was why the annexes described real failure modes instead of imported abstractions.

Ana, whose husband's tricycle had broken down in 2028, spent 2031 running acceptance testing for a national agency's benefits chatbot. She failed it twice. The second time, she wrote a four-page memo about what happens when a system trained to reduce handling time meets a claimant who is frightened. The agency deployed it anyway, with her escalation path attached. Both facts are true and neither cancels the other.

The proposal that arrived at the end of 2031 was generous, which was the problem.

A frontier provider offered subsidised inference capacity across the region for public-sector health triage, education, agricultural advice, and disaster response. It would run below cost for five years. The conditions looked modest: telemetry, standard terms of service, and no local-model requirement. Their effect was not. Accepting would let ministries deploy at a price no domestic alternative could match. It would also make public services dependent on an inference endpoint governed by a foreign contract and revocable at the provider's discretion.

Ratna's counterproposal took four months and looked like a procurement schedule. That was the form control took. Accept the capacity. Require workload portability and exportable records. Require pre-deployment evaluation by accredited regional evaluators. There were now more than nine hundred. Require incident disclosure within seventy-two hours. Reserve the right to run critical services on regional capacity within eighteen months of notice.

Three governments signed the amended version. Two signed the original. One signed nothing and spent the next three years explaining why. That split, not the technology, shaped the following decade.

Chapter 4 · 2032 to 2035

No grand bargain, no single architecture. What emerges instead is a set of unglamorous, interlocking commitments that turn out to be worth more than any of them looked.

The ASEAN Compute Grid was never built, in the sense that no one ever cut a ribbon on it.

What was built was a reciprocal capacity arrangement between four national programmes, an agreed workload-portability specification, a shared queue for public-interest research, and a small standing secretariat with a budget roughly equal to one mid-sized hospital's annual running costs. It was routinely described in the press as a disappointment relative to the announcements of 2029.

It was also load-bearing. When a member state's largest foreign provider suffered a nine-day regional outage in 2033, health triage and disaster coordination in two countries ran on borrowed capacity from the shared queue. Degraded, slower, workable. The arrangement's value had never been in the capacity it added. It was in the option it preserved.

Kim Seng's naive note from 2027 took seven years to become ordinary.

By 2034 a majority of the region's packaging and test capacity operated under a published-terms framework: non-discriminatory access by declared end-use category, verified provenance, audited compliance, and disclosure obligations for safety-relevant integration work. It had not been adopted out of idealism. It had been adopted because exclusive-supply pressure from three directions at once was commercially intolerable, and a shared rule was the only defence smaller firms could afford.

It did not make the region neutral in any grand geopolitical sense. Licences still came from elsewhere. Tools still came from elsewhere. What it did was make the terms of access legible and collective rather than negotiated plant by plant, in private, under pressure. Kim Seng, who was sixty-one and thinking about a garden, regarded this as the most useful thing he had done with his career, and continued to find it insufficient.

Set against the decade, the accounts do not balance cleanly.

• Cognitive-labour displacement in voice and Tier-1 digital services was real, large, and concentrated in 2027 to 2030. Transition programmes reached a minority of affected workers. The ones that worked were small, specific, and employer-linked.
• Compute capacity in the region grew enormously. Control over it grew far less. The gap between the two is the decade's central unresolved problem.
• Grid and water stress became a first-order political constraint, and the jurisdictions that had priced it into approvals early kept both their investment and their public consent.
• Regional evaluation and governance capacity grew from near zero into a professional community of a few tens of thousands. It mattered more than any other investment and was underfunded throughout.
• The entry-level ladder into professional work never fully reformed. A generation entered the labour market with fewer of the ordinary jobs that used to teach judgement, and the consequences of that are still ahead of this story.

Ratna, now working for a smaller organisation with a shorter title, was asked in 2035 whether the region had won or lost. She said that was the wrong frame, then gave a better answer than she meant to.

"In 2025 almost every consequential decision about AI in this region was made somewhere else and arrived here as a fact. In 2035 a meaningful share of them are made here, badly, slowly, with insufficient money, by people who can be voted out. I would not call that a win. I would call it the difference between weather and politics."

Mara did not attend that panel. She was in a training room in Cebu with thirty-one people who had been made redundant by a logistics automation rollout, teaching the first session of a course that did not exist eight years earlier, which begins with the same instruction she sent her supervisors from an airport gate in 2028.

Write down where it fails. That is how we make our case.