Study brief: ThaiLLM, local AI hardware and a Chiang Mai data center
Prepared October 27, 2026, by Alexander Grabowski, Chief of Operations, B4L (buildfor.life). The research is in progress; nothing has been decided or ordered.
Roles. The investment decision and the money are the investor's. B4L (buildfor.life) works for the investor as consultant: it researches, plans and prepares the decision. B4L does not invest in, own or operate the hardware or the data center.
Why this study
An investor wants to put money into AI hardware and open-weight models hosted in Chiang Mai rather than depend only on foreign clouds and token APIs. This study surveys the field and sets out the options for that investment. Thailand's national effort to build Thai language models (ThaiLLM) is the obvious reference. What it builds on, what it runs on and where it is hosted show what a Thai deployment needs and who the partners could be.
The questions
- ThaiLLM. Who runs it, which models does it release or build on, what hardware was it trained on, and where and how is it served?
- Thai-capable open models. What sizes and licenses do Typhoon, OpenThaiGPT, Pathumma, THaLLE, WangchanX, SEA-LION and the strong general models come in, may they be hosted commercially, and how well do they handle Thai?
- Where Thai AI runs today. What do the national supercomputer LANTA, the Thai GPU clouds, the hyperscaler regions and the government cloud offer? Which BOI incentives apply, what does GPU rental cost, and what do the US export rules for AI chips say?
- Option A, the investor's servers at CMU ITSC. The investor would buy GPU servers and host them in Chiang Mai University's data center (colocation).
- Option B, a new data center in Chiang Mai. It would be built to the highest standards of reliability and security, with the best energy efficiency achievable in a tropical climate. The investor would prefer this direction if the numbers carry it.
- The baseline, renting. The investor would buy GPU hours from a cloud or tokens from an API. Every option is compared against it.
What decides between the options
Seven criteria decide between the options:
| Criterion | What we measure |
|---|---|
| Cost | cost per month and per million tokens against renting, over 3 to 5 years, with power at the PEA tariff |
| Control and data | data stays in Thailand and on hardware the investor owns (PDPA, customers' trust) |
| Reliability | certified tier or availability class, redundancy, who is on call |
| Energy | PUE and water use, share of solar and other renewable power, heat reuse |
| Time | months until the first model serves users |
| Capability | which model sizes can be served, for how many users at once |
| Partners and revenue | who else would use or pay for capacity: CMU, Thai companies, the ThaiLLM partners |
Related B4L work
- B4L's agents already use an open-weight model (Qwen) on the p0
inference endpoint. The plan is in
docs/src/assistants.mdin cnx-network-clan, on a work branch as of October 2026. - The CNX Vinyl plant assistant plans a self-hosted Typhoon model with an optional single-GPU server; see cnx_vinyl, Plant assistant.
- B4L's office is at CMU's Science and Technology Park (CMU STeP), next to the university whose data center is option A, so B4L can be the investor's local contact for CMU.
Method
Six research streams collect the evidence, from primary sources first: ThaiLLM; Thai models; Thai compute; CMU and Chiang Mai; data-center standards; and hardware and prices. Every outside figure in these documents carries a numbered citation to the sources list, and our own estimates are marked "est." Money is in THB first, with USD in parentheses.