China’s new AI export: Tokens, not chips

At the sidelines of the World Artificial Intelligence Conference in Shanghai in July, a Guangxi-based company launched a “token factory” — an industrial-scale plant that bypasses local infrastructure by running models centrally and selling only the final algorithmic output. The company also announced partnerships with Indonesia and Malaysia.

Two days earlier, five ASEAN members, Cambodia, Indonesia, Laos, Malaysia and Myanmar, had become founding members of the Chinese-initiated World Artificial Intelligence Cooperation Organization (WAICO). For a region that has spent two years debating how to build its own artificial intelligence (AI), the quieter development is that it is starting to decide it may not need to.

Why build when you can rent?

Southeast Asia’s costly domestic data centre builds, aimed at anchoring national infrastructure, have always been framed as construction projects. Partnerships with Nvidia by Vietnam’s FPT, Indonesia’s Indosat, Malaysia’s YTL and Singtel’s RE:AI each promise a national facility where a country’s data never leaves its borders, and each needs hundreds of millions of dollars a year for building and chips that Washington controls. 

China is now offering a way around all of it. Do not build the plant, buy the output.

Rather than trading in physical data centres or heavy hardware, this model shifts the transaction to a service economy: Southeast Asian enterprises buy finished algorithmic outputs directly from cross-border Chinese data hubs via application programming interfaces (APIs). Metered strictly in “tokens”— the computational fragments of text read and generated — the transaction bypasses infrastructure entirely. The buyer purchases neither raw electricity nor servers, but ready-made cognitive judgment delivered instantly across a fibre-optic line from across the border. 

Start with the economics, because that is what settles the matter in Jakarta and Hanoi. Unlike a data centre that rents servers by the hour, a token factory “keeps the machines, runs the models and sells only what comes out”. This output is the inference result, metered in tokens. Under this arrangement, a Guangxi-based hub running on cheap regional hydropower connects directly to Southeast Asian enterprises over a dedicated cross-border fibre link. No physical hardware or restricted chips change hands, no import bottlenecks are triggered, and the commercial billing scales strictly with consumption. 

The cost of building AI

By contrast, the domestic build route runs headfirst into hardware friction; in Vietnam, tighter US export controls have stretched the procurement timeline for high-end accelerators from eight weeks to 26. 

Even the sourcing of hardware is tightening. In July, it was reported that Nvidia had dropped more than half of its approved buyers in Singapore, Malaysia and Japan under a new compliance white list meant to keep its chips from China. Electricity runs to between 40% and 60% of a data centre’s operating costs by some estimates. Indonesia’s own digital-infrastructure directorate counted about 185 data centres totalling 274 megawatts late last year, against a government target above 2,000 by 2029, and reckons only about a third of that capacity is ready for AI work.

Moreover, the capability argument that once favoured building has narrowed to almost nothing. Stanford’s 2026 AI Index put the leading American model just 2.7% ahead of the best Chinese one, down from a gap of more than 17 points in 2023, even as the US outspent China on private AI investment by a factor of 23. 

DeepSeek’s V4, released in April, returns benchmark scores near the American frontier at roughly one-fiftieth of the price: a task that costs US$15 on a top US model runs about US$0.50. As of early 2026, Chinese models had become the most-used on OpenRouter, the developer marketplace, accounting for a clear majority of the tokens consumed there. 

Thus, for a firm in Surabaya screening transactions for fraud or summarising legal contracts, the calculus is absolute: it has little reason to wait a year and spend a fortune building domestic infrastructure when it can rent the exact same capability this afternoon for cents, in near-parity quality, over a dedicated fibre line from Guangxi. 

The dependency problem

A chip is a countable object an inspector can find at the border, and governments tackling this issue are beginning to realise that regulatory compliance cannot stop at data residency alone. For cash-strapped enterprises across Jakarta or Manila concerned with monthly operating budgets, the payment model is irresistibly cheap: instead of capital-intensive outlays for server architecture, firms pay fractions of a cent per thousand tokens via low-cost API subscriptions, making adoption economically trivial to bear. Yet this convenience masks a deeper vulnerability. A model called via an API is invisible, updated on someone else’s schedule, priced by someone else’s meter, and answered according to algorithmic weights a buyer will never audit. 

None of this is settled. Data residency rules are spreading in Indonesia and Vietnam and may keep the most sensitive workloads off servers inside China, which is precisely the appeal the token factory is selling. ASEAN only finished negotiating its digital economy pact in May, and while some buyers will baulk at letting a Chinese provider anywhere near their data, the broader prospect of renting intelligence introduces a profound regulatory dilemma. 

Rather than halting the trend, these frameworks create a dual-track market where strict compliance rules govern sensitive state data while commercial enterprises quietly opt for cheap, cross-border token flows. Cost and near-parity would pull the region toward buying even without Washington’s controls; the controls only sharpen a movement that the economics had already begun. 

Who builds the future?

While the token factory is a direction, not a done deal, the regional trajectory is increasingly visible in national procurement strategies and institutional alignments. Whichever route the region takes, it imports the thing that runs and skips the thing that builds.

Metered answers add no engineers inside Southeast Asia. Indonesia’s own officials have said as much, noting that high digital literacy remains concentrated in Jakarta and that partnerships with China ought to extend to cultivating talent rather than just delivering service. The programmes meant to close that gap are real but thin against the need.

The ASEAN Foundation’s AI Ready ASEAN, funded by Google.org, aims to reach 5.5 million learners across a bloc of some 680 million people. China has offered a joint innovation centre and a digital academy, and Xi Jinping used the Shanghai stage to pledge 5,000 training places for developing countries over five years. Useful, and also a way to bind customers to a supplier.

Beyond individual corporate tie-ups, the practical uptake of renting intelligence is finding a receptive audience among ASEAN’s commercial sectors. Governments from Bangkok to Jakarta are discovering that trying to mandate local infrastructure for every small-to-medium enterprise creates an unworkable bottleneck, making API-led foreign cloud consumption a path of least resistance.

While state ministries debate data sovereignty and draft digital blueprints, the sheer commercial pressure to remain competitive against Western and Chinese tech baselines is tilting corporate behaviour toward outsourced inference. Yet this path risks embedding a permanent structural division: a region fast adopting advanced cognitive tools while remaining entirely dependent on external powers to maintain and master the code. 

The larger pattern is a familiar one. 

China has spent a decade turning surplus industrial capacity into exports the world found hard to refuse: solar panels, then electric vehicles, now tokens. Its daily token processing has climbed from about 100 billion in early 2024 to 140 trillion by this spring. Washington’s export controls, written to deny Beijing advanced compute, have helped point a new market toward it and handed China a governance club, WAICO, in which to set the terms its customers will follow.

The region can rent its way past the compute squeeze this year, and most of it will, because the arithmetic is not close. But every token bought abroad is a model a country did not train and an engineer it did not grow. When Indonesia measures itself against its 2029 target, the number that will matter is not the megawatts it installed. It is how many of its own people can build the thing everyone else is renting.

Related: China tech firms drive AI data centre boom in Southeast Asia | Why hyper-efficient Singapore is Asia’s most AI-anxious society