Company News

A scale model of Runjian's Wuxiang Cloud Valley intelligent computing centre in Nanning, China.

Across Southeast Asia, companies have put AI to work on ordinary tasks: recommending products, answering queries, screening transactions for fraud. A February survey of 330 companies in six regional markets, by McKinsey with Singapore's Economic Development Board and Tech in Asia, found nearly half had moved past trials, putting 11 to 40 percent of technology budgets into AI.

The computing to run it is thin on the ground. Indonesia's digital-infrastructure directorate counted about 185 data centres totalling 274 megawatts late last year, against a government target above 2,000 megawatts by 2029, and its data centre association reckons only about 30 percent of national capacity is AI-ready. In Singapore, the region's largest hub, tight supply has pushed operators across the border in search of room to grow.

Build it or buy it

AI is not sold by the machine or by the month. It is metered by output, in units called tokens — the fragments of words a model reads and produces. A short customer-service reply runs to a few hundred; a contract, summarised, a few thousand. The bill arrives as an electricity bill does, though as Liu Jian of the Energy Research Institute under China's state planner told a Shanghai forum this month, a kilowatt-hour has a settled value while a token's worth depends on which model produced it.

A power station need not sit beside the house it lights. Training a model takes a large facility and months of running time; using one is far lighter, and need not happen in the country where the question is asked. That leaves two routes. The first is to build. Vietnam's FPT, Indonesia's Indosat with its $200 million "GPU Merdeka" project, Malaysia's YTL and Singtel's RE:AI are sovereign AI facilities whose promise is that data never leaves the country. Each needs hundreds of millions of dollars and a year or more of construction, and each depends on chips that Washington controls. US export restrictions have stretched procurement times for high-end accelerators in Vietnam from eight weeks to 26 and pushed up spot prices, splitting the market between operators with offshore access to chips and those without, Asia Times reported this month.

A 512-chip AI computing cluster on display at the 2026 World Artificial Intelligence Conference in Shanghai [China Economic Net/ Xiong Weisheng]

The second is to buy the answers from machines elsewhere. That route runs around all three constraints at once: nothing is built, no chips are imported, and the bill starts small. Price is the other half of the appeal. Chinese open-weights models have nearly closed the capability gap — Stanford's 2026 AI Index put the leading US model 2.7 percent ahead in March — while costing a fraction as much per task on Artificial Analysis's independent index.

The offer from Guangxi

At a forum alongside the World Artificial Intelligence Conference (WAIC) in Shanghai on July 18, Guangxi-based Runjian Co., Ltd. launched what it calls a token factory and signed agreements with Indonesian and Malaysian partners, among them a research centre head at Indonesia's National Research and Innovation Agency and Malaysia's National Aerospace and Defence Industries.

The phrase describes a change in what is sold, not what is installed: a conventional data centre rents out servers by the hour; a token factory keeps the machines, runs the models and sells what comes out. The company told China Economic Net (CEN) its Nanning site runs on green electricity and connects to Southeast Asia over a dedicated cross-border line the city obtained in June.

Location does some of the work. Guangxi borders Vietnam, and its capital, Nanning, sits a few hundred kilometres from Hanoi. Power does the rest: electricity accounts for 40 to 60 percent of operating costs at some data centres, by Goldman Sachs's estimate, and Guangxi has hydropower to spare.

Buyers ask first where their data goes. Chu Ge, Runjian's chief AI security scientist, told China Economic Net (CEN) that users could have systems installed on their own premises, and that the company was working on encryption that would let a provider process a request it cannot read. Under China's cross-border pilot scheme, queries arrive over a dedicated line and only results are returned.

Liew Soung Yue, dean of information and communication technology at Universiti Tunku Abdul Rahman in Malaysia, said most countries in the region have neither high-end computing equipment nor advanced applications of their own, and that buying by the token lets smaller firms reach systems they could not otherwise afford.

The harder part

Metered answers add no megawatts inside Southeast Asia, and buying them does not produce people who can build with the technology. That is being worked on separately: the ASEAN Foundation's AI Ready ASEAN programme, funded by Google.org, aims to reach 5.5 million learners across the bloc, while China and ASEAN have agreed on a digital academy and a joint AI innovation centre.

Indonesia is pressing that point. Andreas Bondan Satriadi, a deputy director at its development planning ministry, told CEN on the sidelines of WAIC that partnerships with China should extend to cultivating AI talent, noting that high digital literacy remains concentrated in Jakarta. The rules are unsettled too: data-residency requirements are spreading, and ASEAN only finished negotiating its digital economy pact in May.