Findings and recommendation

Research note, October 27, 2026, by Alexander Grabowski (B4L). Sources were retrieved on October 27, 2026. Figures are cited, and our own estimates are marked. Money is in THB with USD in parentheses at 33 THB per USD, the B4L planning rate.

Roles. The investment decision, the money, the hardware and any data center are the investor's. B4L (buildfor.life) works for the investor as consultant: it researches, plans and prepares the decision.

The short answer

ThaiLLM has released two models, of 8 and 30 billion parameters, both built on Alibaba's Qwen3. Its two foundation models, ThaiLLM-8B and ThaiLLM-30B, are Alibaba Qwen3 models trained further on about 63 billion tokens and released under Apache 2.0.[1][2] They were trained on LANTA, the national supercomputer, which has 704 NVIDIA A100 GPUs from 2022.[3][4] The free Playground and API run on compute from National Telecom (NT) and SIAM.AI CLOUD, at sites that no source names.[5] The ThaiLLM note has the details.

Thai AI does not need big hardware, either to train or to serve. Every Thai model of 2025 and 2026 continues a foreign open model, and the teams trained on one 8-GPU server or less.[6][7] The best Thai models, such as Typhoon 2.5 (30B), each run on one 80 GB GPU.[8] The note on Thai models compares them.

No Thai model maker sells a paid, production API hosted in Thailand, which leaves a gap in the market. Typhoon's free beta is capped at 5 requests per second, and its AWS listing still reads "Coming soon."[9] AWS and Google run Thai cloud regions but list no GPU instances there,[10][11] and the Thai GPU clouds sit in Bangkok and the Eastern Economic Corridor (EEC). Chiang Mai has no GPU cloud. NT's site there has 25 racks, and the ERAWAN cluster of Chiang Mai University (CMU) is the only sizeable GPU system we found.[12][13] The note on Thai compute maps the rest.

Option A, the investor's servers at CMU ITSC, is cheap and fast but small. ITSC is the university's Information Technology Service Center. A full rack there costs 20,000 THB (USD 606) a month, outside organizations need an academic cooperation agreement with CMU, and the whole room runs on backup power systems of 120 kVA and 125 kVA and one 320 kVA generator.[14][15] In our reading, one or two GPU servers fit, and a cluster does not. The note on option A gives the details.

Option B, a new data center, can be built to a high standard, but its scale is the hard part. We would define "highest standard" as a certified Uptime Tier III facility with selected Tier IV features, ISO 27001, 50001 and 14001, and an annual power usage effectiveness (PUE, total facility power over IT power) of 1.15 to 1.2 with warm-water liquid cooling; that definition is our judgment. NVIDIA's reference design for its largest systems asks for Tier 3.[16] We estimate 320 to 510 million THB (USD 9.7 to 15.4 million) per MW of IT before any GPU is bought, and the data-center incentive of the Board of Investment (BOI) starts at 2 MW of IT load.[17][18] The note on option B has the full estimate.

The Korat and Chaiyaphum wind belt is a better place than Chiang Mai for option B. Chiang Mai sits in seismic zone 3, and the whole Upper North has almost no wind. The candidates in the Northeast lie in no seismic zone,[314][307] farmland there is listed at about 130,000 to 300,000 THB per rai,[345][340] and outside the EEC a qualifying data center gets 8 years of BOI tax exemption.[319] Our screening model finds that, per MW of IT, 5 MWp of solar with a 10 MWh battery covers about two thirds of the power hour by hour below the grid price, and that 10 MWp with 20 MWh covers about 95% at roughly grid cost on about 100 rai. Two risks remain: water, and the pause in data-center approvals of September 2026.[315] The note on site selection compares the candidates.

Hosting and agent servers can carry part of the cost, and BOI requires them in any case. A BOI-promoted data center must offer colocation, managed services, backup and data hosting,[389] and Thai colocation sells at about 38,800 THB a month for a 3 kW rack.[390] The new line of business is agent servers: a company's data, its AI agents and the models sit in one building, and by default nothing leaves it. Our indicative packages run from about 13,000 THB a month (shared models, up to 25 users) to 467,000 THB for a dedicated 8x RTX PRO 6000 server, and one RTX server sold this way breaks even at about 71% sold (our estimates). Keeping data in Thailand does not protect the business on its own, since AWS has run a Thai region since January 2025.[154] The note on hosting and agent servers has the details, including Hetzner as a benchmark.

The same hall can rent out ordinary servers, but memory prices make the margins thin. Virtual servers with shared CPUs could start at 490 THB a month, servers with dedicated CPUs at 1,650 THB, and whole machines at 9,900 THB, priced against NIPA, Hetzner's Singapore cloud and the other regional hosts (our estimates). Memory contract prices rose 93% to 98% in the first quarter of 2026,[829] and memory is now about half the cost of a host. Priced at 1.5 times cost, the virtual servers break even at 40% to 68% of capacity sold, depending on what the hardware costs, and the whole machines at 56% to 94% (our model). A host must also keep traffic data for 90 days and verify every customer's identity.[837][838] The note on VPS and dedicated servers has the packages.

Almost the whole stack can be open source. SONiC runs on white-box switches from many vendors,[391] and NixOS, Incus, Ceph, LTFS tape and Prometheus cover the rest, except NVIDIA's CUDA and GPU firmware, which stay closed. A 1 PB Ceph object store with two tape copies costs about 144 THB per TB a month in hardware and power (our estimate), against about 743 THB for S3 Standard in AWS's Bangkok region.[392] The note on the open stack lists the parts.

The facility needs a continuous-rated generator plant and telemetry of everything. BOI requires continuous-rated generators for the whole load plus a backup set;[389] at full build that means three sets of about 1,600 kVA, costing about 57 to 89 million THB installed (our estimate). New fuel-tank permits for data centers were reported suspended in September 2026,[393] so the plan keeps fuel under the 15,000 L license threshold, which is enough for about 36 hours at 2.4 MW (our calculation). Haze and sulfur make corrosion the risk peculiar to Thailand. The halls therefore need corrosion monitoring next to sensors for temperature, humidity, leaks, smoke, power and coolant, and the operators need the free Thai feeds for air quality, weather, rivers, fires and earthquakes. The note on facility systems has the details.

Bandwidth is easy to buy once fiber reaches the plot; the hard part is getting two separate routes to it. Carriers and the Electricity Generating Authority of Thailand (EGAT) sell links of 10 to 400G. An autonomous system (AS, a network's own routing identity) costs little: APNIC membership with a /23 of IPv4 and a /32 of IPv6 is AUD 500 once plus AUD 2,256 a year.[755] The nearest place to peer is Telehouse Bangkok, where all six established Thai exchanges meet,[756] and a 100G port on BKNIX costs 150,000 THB a month.[757] CMU pays NT about 6,650 THB per Gbps a month for internet,[758] while dark fiber lists at 1,200 THB per core-km a month or more,[759] so a pair from Korat to Bangkok costs about 0.6 to 1.25 million THB a month (our arithmetic). The note on connectivity has the details.

A first design is ready for discussion. Building on all of the above, the note on the first data center sets out a 2 MW (IT) facility on 150 rai in the Sikhio corridor, built in two phases. It has liquid-cooled GPU pods and an air-cooled hosting hall, 2N transformers and UPS, continuous-rated generators, solar and battery bought in tranches as load arrives, two fiber routes to the investor's own point of presence in Bangkok, and an open-source stack. Our model puts phase 1 at about 0.86 to 1.27 billion THB of facility capital expenditure before IT, with the first IT load in the second half of 2028 at the earliest. Renders and an interactive 3D model come with it.

Most of the plant can be bought in Thailand, and much of it is made here. Thai factories make the cables, transformers, switchboards, dry coolers, chillers, coolant distribution units and racks the design needs, while the GPU servers, battery containers, generator sets and large UPS come in through Thai dealers. Service near Korat is thin, and the cooling plant, like the electrical plant, takes most of a year to deliver. Ground-mount solar was built in Thailand in 2026 for about 14 to 18 million THB per MW, below the study's figure,[924][925] while the battery price will only be settled by bids. The sourcing note names the sellers for every main component.

Renting sets a hard baseline. OpenRouter sells gpt-oss-120b at USD 0.03 per million input tokens and USD 0.15 per million output tokens.[19] Owned hardware wins on data control, on Thai models that nobody sells as a service, and at high, steady load. The note on hardware sets buying against renting.

The options side by side

The table draws on these sources: server prices and lead time,[20][21][22] rent,[23][24] the electricity tariff,[25] ITSC,[14] build time,[26] ThaiLLM's terms[27] and hazards.[28][29] The 5-year cost model behind the hardware figures is model/rent_vs_buy.py in the repository, and its results are in model/results.md.

Our recommendation: a staged path

What follows is our advice to the investor, and it rests on our judgment. Each stage is a separate decision, taken on numbers measured in the stage before.

1. Now (months 0 to 3), prepare without buying. Define the workloads and the customers (Thai documents, chat, speech, private company data), test the shortlisted models on them through free APIs and rented GPUs, and have a Thai lawyer read the Typhoon terms. Open talks with CMU ITSC about an agreement for colocation, with BDI about a Chiang Mai role in ThaiLLM, and with NT, which serves ThaiLLM and runs a Chiang Mai site. B4L can prepare and lead these talks for the investor.
2. In stage 1 (months 3 to 12), place one server at CMU ITSC. One 8x RTX PRO 6000 server (5.4 to 6.5 million THB) serves the Thai model portfolio (a strong general model plus Thai specialists) with data kept in Chiang Mai, and it measures real power, load and demand. It is the pilot for option B. If ITSC allows third-party services, two or three design-partner companies can test agent servers on it (see hosting and agent servers).
3. In stage 2 (from month 12), add a second server only on evidence. Add an 8x B200 or B300 server when demand passes the API break-even (7 to 13 billion tokens a day for a B200 against gpt-oss-120b tokens, in our model), when a partner such as CMU commits to capacity, or when a frontier-size model must run in Thailand.
4. In parallel (from month 6), study sites for option B. The study covers a 1 to 2 MW facility on the shortlisted sites in the Korat and Chaiyaphum belt (see site selection): certified Tier III, ready for liquid cooling, at PUE 1.15 to 1.2, with on-site solar and batteries, and with land, grid connection and plant planned for 2 MW so that it can qualify for BOI. Design it with a hosting zone next to the GPU zone, an N+1 continuous-rated generator plant, full telemetry and an open stack (see facility systems and open stack). Build when anchor tenants (CMU, ThaiLLM partners, Thai companies) commit enough load to approach 2 MW.

Two developments would change this advice. A committed anchor tenant, or a public partner (NT, BDI, CMU) that takes a large share of capacity, would justify starting option B earlier. A fall in API prices, or a Thai production API for the Thai models, would weaken the case for owning hardware at all.

Next steps for B4L (on the investor's behalf)

• Ask CMU ITSC about power per rack and spare backup capacity, the price of power, 24/7 access during incidents, network and transit, and the terms of the agreement. The full list is in option A.
• Ask BDI and NT who hosts ThaiLLM and where, whether they want an inference partner in Chiang Mai, and how much capacity NT Chiang Mai will have after its renovation.
• Get quotes for the first server from a Thai distributor and by import, with duty, VAT and on-site warranty in Chiang Mai.
• For hosting and the facility, ask ITSC about serving third-party companies, the Department of Energy Business (DOEB) about the suspension of fuel-tank permits, and the Provincial Electricity Authority (PEA) about a second feed from a different substation, and get quotes for continuous-rated generators and SONiC switches. The lists are in facility systems and open stack.
• For option B, make land and grid inquiries on the shortlisted sites (EGAT and PEA in Nakhon Ratchasima and Chaiyaphum), look into water rights, and measure the wind if wind stays in the plan.