AMRO Says ASEAN+3 Will Rent the Generative AI Dividend

ASEAN+3 will rent most of its frontier generative AI from foreign providers, the region’s official research office said on September 8, 2026. The ASEAN+3 Macroeconomic Research Office released a policy perspectives paper on AI-as-a-service that treats that rental model as a live issue for productivity, jobs, external accounts, and resilience.

Economists Toàn Long Quách and Xianguo (Jerry) Huang wrote the study. They say the job is to make the reliance productive, diversified, measurable, and reversible.

AMRO Treats Frontier AI as an Import

AMRO is the surveillance arm of the ASEAN+3 finance process, set up in 2011. Its members are the ten ASEAN states plus China, Hong Kong, Japan, and Korea. The new paper sits in the Policy Perspectives series, which is written for finance ministries rather than for model labs.

Quách and Huang map a concentrated global AI value chain and the region’s uneven place in it. Frontier capability, in their account, is something ASEAN+3 mostly buys as a service. The paper’s test is not whether every capital should train a national model. It is whether the rental can be used, switched, and accounted for.

The objective should be neither technological self-sufficiency nor passive dependence, but to make that reliance productive, diversified, measurable, and reversible.

Toàn Long Quách and Xianguo (Jerry) Huang, AMRO Policy Perspectives, September 8, 2026

Huang is deputy group head of AMRO’s Macro-Financial Research Group and has spent 2026 writing about generative AI from several sides: labor demand, the dollar, and now this service model. Quách has been the office contact on that labor work. The September paper pulls those threads into one brief for the region’s finance officials.

The Region Already Sells the Chips and the Servers

The rental story sits on top of a boom the same office has been marking all year. In its July 2026 quarterly update, AMRO said around half of global AI-related trade already runs through ASEAN+3. AI-related exports contributed around two-thirds of the region’s export growth in the first quarter of 2026.

That is why the growth forecast moved. AMRO raised 2026 growth to 4.1 percent, with inflation at 1.6 percent, and it holds 2027 at 4.0 percent growth and 1.6 percent inflation. Chief Economist Dong He tied the upgrade to domestic demand and to the region’s place in AI supply chains.

The same update carries the other side of that bet. If global AI investment growth falls back to its 2024 pace, regional growth could slow to 2.5 percent in 2027, the weakest rate outside the pandemic years since the Asian Financial Crisis. The region is long AI hardware. It is also long the cycle that pays for that hardware.

WHO ALREADY EXPORTS THE PHYSICAL STACK

World Bank figures, cited by AMRO economist Abdurohman, draw that split. Korea dominates memory chips. Malaysia leads in packaging and testing and is scaling as a data hub. Vietnam is a large electronics assembly base. Singapore holds advanced fabrication, research, and regional headquarters. Those are export roles. They are not the same as owning the models that run on the racks.

Model Access Shows Up as a Recurring Import

AI-as-a-service is the product that fills that gap. A bank in Jakarta or a hospital in Manila does not need to train a frontier model. It buys access through a cloud API, pays for tokens or seats, and plugs the output into existing software. The weights stay with the vendor. The bill shows up as a service import.

That is already how the industry is packaging the work. In the dollar-loop note Huang wrote with Chenxu Fu, they point out that OpenAI launched a Deployment Company to put engineers inside client firms and rebuild workflows around the models. Once those systems are in place, they wrote, compute becomes a recurring operating expense rather than a one-off experiment.

The meter is getting larger. Boston Consulting Group analysis projects the AI compute market will climb from $360 billion in 2025 to about $2.3 trillion in 2030. Most of that spend still clears through a short list of chip designers, cloud operators, and model vendors. For an ASEAN+3 finance ministry, the relevant fact is that the invoice is concentrated, dollar-priced, and easy to grow and hard to unwind.

Local data centers do not automatically break that pattern. A campus in Johor or Batam can store data under local law and still run someone else’s model over someone else’s API. The building is local. The capability is rented. That is the distinction the AMRO paper is asking officials to measure.

How AI-as-a-Service Hits External Accounts

The second-order hit is on the balance of payments and on the currency in which the service is settled. Productivity can rise at the firm even as the economy pays a larger, stickier import bill for the tool that produced the gain, and those payments are being wired into a dollar system that ASEAN+3 has spent years trying to dilute.

Compute Becomes a Recurring Bill

Abdurohman put the current-account version of this in plain language in his September 3, 2026 note on Indonesia. Countries that import costly chips and cloud services without generating tech exports of their own can widen the trade deficit and put pressure on the local currency. For Indonesia that currency is the rupiah. For the wider region the same arithmetic applies wherever the AI stack is consumed but not sold.

Hardware exporters have a natural hedge: they buy advanced inputs and they earn foreign exchange by shipping the finished electronics. Service-only adopters do not. Their AI dividend arrives as higher output and leaves again as a cloud invoice. Huang and Fu add that the invoice is only the visible layer.

A Payment Rail That Favors the Dollar

On July 30, 2026, the two AMRO economists described an energy-compute dollar loop they describe as forming through commercial contracts, not through a new official accord. Electricity feeds data centers. Data centers sell compute. Compute automates payments. Those payments, they argue, are being pulled onto dollar-priced rails and, increasingly, onto dollar-pegged stablecoins whose reserves buy US Treasuries.

Electricity powers data centers; data centers produce compute; compute enables the automation of business activity, including payments that favor programmable settlement; and stablecoin reserves flow into US Treasuries.

Chenxu Fu and Xianguo Huang, AMRO economists, July 30, 2026

They point to a 20-year data-center lease Anthropic signed with TeraWulf as the supply-side wager, and to Open USD, a dollar-pegged stablecoin backed by more than 140 payment, finance, and crypto firms, as the settlement bet. ASEAN+3 has been linking local payment systems and trying to settle more trade in local currency. An AI stack that invoices in dollars and settles on dollar tokens would cut across that project without anyone putting it on a summit agenda.

Fu and Huang’s policy line matches the September paper. The region cannot block the loop on its own. It can limit how much of its digital production has to travel through it, by pairing local compute, cleaner power, and local-currency tokenized settlement so that agentic payments do not default to the dollar rail.

Most ASEAN Jobs Sit Outside the High-Risk Slice

Labor is the other ledger the paper flags, and the best regional count still comes from the International Labour Organization’s July 2026 brief. Occupations with more than a minimal degree of generative AI exposure hold 22.9 percent of ASEAN employment. That is a large number of people and a much smaller share of the highest-risk work.

GENAI EXPOSURE ACROSS ASEAN JOBS

Only 3.3 percent of the workforce, 11.7 million people, sit in the highest-exposure group. Around 67 percent of employment remains in occupations with no identified exposure. The ILO also found that employment in highly exposed jobs has kept growing. Widespread disruption is not visible yet. Adoption is still concentrated in tech-heavy roles, not in the office jobs that look exposed on paper.

Huang’s own labor work with World Bank economist Shu Yu points the same way, and it is a warning about what happens as adoption deepens. Using 555 million Lightcast job postings across 84 countries from 2021Q1 to 2025Q2, they found that in high-income countries, postings for above-median-vulnerability occupations fell 5.8 percent relative to less substitutable work. In middle- and low-income countries the effect was much smaller and not statistically significant. Displacement was larger where GenAI adoption, schooling, English, and digital-services trade were already high.

AMRO’s April 2024 Regional Economic Outlook had already taken a similar stance, calling it qualified optimism: more ASEAN+3 jobs looked likely to be augmented than automated. The 2026 paper does not throw that out. It adds a condition. If the tools arrive as a foreign service, the productivity gain still has to be absorbed by firms and workers who can actually use them, or the rent leaves and the wage gain never shows up.

Indonesia Risks Paying for AI Without Making It

Indonesia is the clearest case of scale without a maker role. Abdurohman wrote on September 3, 2026, that CoreWeave plans three data centers there, and that a market of more than 280 million people will not be short of AI demand. High-tech goods were only 8.7 percent of its manufactured exports in 2024, lower than a decade earlier, against about 60 percent in Singapore and Malaysia.

The country built its growth on commodities, resource downstreaming, and consumer digital platforms. Malaysia spent decades on testing and packaging. Vietnam folded assembly into a wider industrial network. Indonesia entered the AI cycle with a thin mid-tech base, low research spending, a tight pool of advanced engineers, and a grid that still has to deliver large volumes of clean, reliable power.

Abdurohman’s three shifts are specific. Move downstream policy from raw minerals into advanced materials and electronics. Compete on predictable rules, licensing, and talent, not only on tax breaks. Force data-center and chip investment to train engineers, build local suppliers, and work with universities, because those linkages will not appear on their own. The government’s 2026-2029 AI road map is, in his account, a start. It is not yet a single industrial, energy, education, and trade policy aimed at climbing the stack.

The rest of ASEAN+3 is not Indonesia, and that is the point of AMRO’s “uneven position.” Korea, Japan, China, Singapore, Malaysia, and Vietnam already occupy rungs of the physical chain. Several ASEAN members do not. A regional AI-as-a-service boom that is only an adoption boom will widen that gap: the same invoice, very different offsets.

What Reversible AI Reliance Looks Like for ASEAN+3

Reversible reliance, in the paper’s terms, is a procurement and industrial test rather than a slogan about sovereignty. Officials would ask whether a given model, cloud, or payment rail can be used productively, swapped for a rival, measured on the external accounts, and turned off without freezing the domestic firms that have come to depend on it.

THE FOUR TESTS IN THE PAPER

• Productive: Domestic firms and workers have to absorb the tools, which means skills, data, and process change, not only licenses.
• Diversified: No single vendor, cloud, or currency should be the only path to frontier capability.
• Measurable: Finance ministries need the service import on the books, not buried inside a bundled software line.
• Reversible: Contracts, open weights, and regional capacity have to make an exit possible if a provider, a rule, or a payment rail changes.

The tools they name for getting there are stronger domestic absorptive capacity, credible alternatives, and regional cooperation. That is a lower bar than a full national model program, and a higher bar than signing the first hyperscaler memorandum that comes in. It also lines up with the earlier AMRO file on the same subject.

THE AMRO AI FILE BEHIND THE PAPER

1. April 2024: The Regional Economic Outlook treats generative AI with qualified optimism and says more regional jobs look likely to be augmented than displaced.
2. January 28, 2026: An AMRO note asks whether ASEAN+3 can still escape the middle-income trap when AI rewards early movers with infrastructure, data, and talent.
3. July 29, 2026: The quarterly outlook raises growth on AI-related demand and warns that a slower investment cycle could cut 2027 growth to 2.5 percent.
4. July 30, 2026: Fu and Huang set out the energy-compute dollar loop and tell the region to treat energy, AI, and payments as one agenda.
5. July 31, 2026: Huang presents labor-demand evidence that high-income, high-adoption markets are already seeing the hit in vulnerable job postings.
6. September 3, 2026: Abdurohman warns Indonesia against remaining a consumer market in an AI boom its neighbors are manufacturing.
7. September 8, 2026: Quách and Huang publish the AI-as-a-service paper that ties those strands to external accounts and resilience.

Finance ministries already have the hardware cycle on their export ledgers. The paper is asking them to put the model cycle on the import ledger, with an exit clause. If the next API a state bank or a hospital buys cannot be measured, swapped, or shut off, the generative AI dividend is someone else’s recurring revenue.

Frequently Asked Questions

What is ASEAN+3 and what does AMRO do?

ASEAN+3 is the ten ASEAN member states plus China, Hong Kong, Japan, and Korea, with Hong Kong listed as its own member economy in AMRO’s mandate. Finance ministers created the ASEAN+3 Macroeconomic Research Office in 2011 to watch regional growth and financial stability, and in May 2026 they marked its tenth year as an international organization while asking it to keep deepening surveillance and the Chiang Mai Initiative Multilateralisation.

What is AI-as-a-service in AMRO’s paper?

It is frontier model capability sold on a meter, usually through a cloud API, so a firm pays for tokens or seats instead of training and hosting its own weights. Huang and Fu wrote that once companies rebuild daily operations around those systems, officials may end up talking about dollars per unit of compute the way they already talk about dollars per barrel, which is why the service shows up as a recurring operating cost and as an import.

Which ASEAN jobs sit in the highest GenAI exposure group?

The ILO’s highest-exposure, or Gradient 4, slice is only 3.3 percent of ASEAN employment, about 11.7 million workers, and the occupation list in the brief’s annex includes financial analysts, web and multimedia developers, securities and finance dealers and brokers, credit and loans officers, and general office clerks. Those are task-heavy information jobs, which is why Singapore’s 42.2 percent exposure share sits so far above the 22.9 percent ASEAN average.

Does AMRO want ASEAN+3 to build its own frontier models?

No. The September paper’s tests are about making foreign supply usable and replaceable, and in the July dollar note Huang and Fu wrote that full technological self-sufficiency is probably not possible any time soon. Their aim is to join digital production without taking a new layer of dollar dependence as the price of entry, and they note that, unlike old oil-dollar bargains, the emerging AI system offers other countries no seat at a summit table.