Where AI runs in Thailand

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.

This note maps Thailand's AI compute and the rules that shape a new GPU site. It is the background for option A, option B and the renting baseline, and it pairs with the notes on ThaiLLM and hardware. Money is converted at 33 THB per USD, B4L's planning rate.

In short

• The national supercomputer LANTA has 704 NVIDIA A100 40GB GPUs from 2022, trained ThaiLLM and sells time to Thai companies at 135 THB (USD 4.1) per 4-GPU node-hour.[57][55][58] We found no successor.
• Commercial GPU capacity sits in Bangkok and the Eastern Economic Corridor (EEC). Siam AI lists the H100 at USD 2.69 and the H200 at USD 2.99 per GPU-hour, and Float16 lists the H100 at USD 4.32.[61][129]
• AWS and Google Cloud run Thai regions but list no GPU instances there in their own documentation (read October 27, 2026), and Microsoft's region is not live.[10][11][133]
• The Board of Investment (BOI) approved 498.7 billion THB (USD 15.1 billion) of data-center projects in 2025.[134] Its best incentive, 8 years of exemption from corporate income tax (CIT) outside the EEC, requires a power usage effectiveness (PUE) no higher than 1.3 and at least 2 MW of IT load.[18][135]
• In July 2026 the energy policy council approved a separate electricity tariff for data centers (5 to 6 THB/kWh is under consideration, against 3.95 THB) and a bond of 4.5 million THB per reserved MW.[134][136]
• US rules require a license for advanced AI chips only when the buyer is headquartered in, or owned from, Country Group D:5 (which includes China) or Macau; a rule on remote access is reportedly being drafted.[137][138]
• We found no GPU cloud in Chiang Mai. NT's site there has 25 racks, and the only sizeable GPUs belong to the ERAWAN cluster of Chiang Mai University (CMU), with 28 GPUs.[12][13]

LANTA, the national supercomputer

LANTA belongs to NSTDA's National Supercomputer Center (ThaiSC). It is an HPE Cray EX cluster of 346 nodes, 176 of them GPU nodes holding 704 NVIDIA A100 GPUs (4 per node), and it has been in service since November 2022.[56] TOP500 lists the GPUs as the 40 GB version.[57] Storage is a Cray ClusterStor E1000 with 10 PB of disk plus 945 TB of flash, and the network is 200 Gbps HPE Slingshot.[56]

ThaiSC calls 8.15 PFlop/s the "peak performance," but TOP500 shows that 8.15 PF is the measured Linpack result (Rmax) and that the theoretical peak (Rpeak) is 13.77 PF, at 310.46 kW.[56][57] We use the TOP500 figures. The system is liquid-cooled, which NSTDA says cuts cooling energy by 33% against air cooling.[55] ThaiSC is based in the INC2 building, Thailand Science Park, Pathum Thani.[56]

LANTA ranked 70th on the TOP500 list of November 2022 and 223rd in June 2026, with unchanged Rmax and Rpeak, so no hardware has been added.[57] Neither the ThaiSC director's slides of May 2026 nor NSTDA's report for the third quarter of FY2026 describes a successor.[139][59]

Use. In the report period, October 1, 2025, to June 30, 2026, ThaiSC delivered 3.18 million GPU-hours to 730 users in 69 organizations.[59] If a GPU-hour is one A100 for one hour, that is about 69% of the 4.61 million available hours (704 x 24 x 273 days; our estimate). Education accounts for 62% of users and enterprise for 21%.[139]

Price. Public and private legal entities can buy time, with a minimum project of 5,000 THB valid for one year; a GPU node-hour is charged as 3 Service Node-Hours (SHr).[58] Free proof-of-concept support and research vouchers are offered from time to time.[55]

Thai GPU clouds

The GPU counts are the providers' own claims, and the prices are list prices read on October 27, 2026.

AIS announced GPU-as-a-service with Singtel and Bridge Alliance in August 2024, without a price.[148]

Siam AI

Siam AI Corporation was registered on January 15, 2024, with 695 million THB of capital, and OBON Corporation holds over 99.86% of it.[68] DDN, its storage vendor, calls it Thailand's first and only NVIDIA Cloud Partner, and NVIDIA's "sovereign AI" event in Thailand in December 2024 was held with Siam AI, a private company.[142][149]

CMKL University said in September 2024 that Siam AI operates 1,024 H100s, though the same page also mentions an "HGX H200 1,024-GPU cluster"; we read H100.[140] The CEO said phase 1 ordered 9 DGX B200 systems, and Kaohoon reported that 8 were bought at about 11 million THB each.[68][141] Siam AI said 7 billion of the 10 billion THB it has invested went into a GB200 NVL72 module at STT Bangkok 1.[66] We found no evidence that all 48 planned racks (about 3,456 GPUs, our estimate) arrived.

In December 2024 Siam AI signed a data-center contract with GSA Data Center, a Gulf joint venture, and an MOU with True IDC.[150][151] BOI approved a Siam AI project worth 3.2 billion THB (USD 97 million) in Chonburi and Pathum Thani.[66] In January 2025 the CEO cited about USD 2 per card-hour, with most customers on contracts of 3 to 4 years.[68]

BDI says the ThaiLLM Playground gets compute from NT and SIAM.AI CLOUD, while the ThaiLLM site credits only NT; we give more weight to BDI, the operator.[5][30] The OpenThaiGPT-ThaiLLM model card credits Siam AI for its post-training on 8 x H100.[48]

The export case. On May 8, 2026, Bloomberg, as reported by the Bangkok Post, identified Bangkok-based OBON Corporation as "Company-1" in a US case over at least USD 2.5 billion of NVIDIA-based servers allegedly diverted to China.[69] The US indictment of March 2026 does not name Company-1, and its charges are accusations.[152] Siam AI says it has never exported AI servers to China and is not involved in the proceedings.[153][70] No court has made a finding.

Hyperscalers: regions without GPUs

Alibaba Cloud (two data centers) and Huawei Cloud (three zones) also operate in Thailand; we found no GPU offer from either.[66][159] In our reading, hyperscaler GPUs with the data kept in Thailand are not available today.

Data-center build-out and BOI

About 70% of data-center investment is in the EEC, and a "Power and Water Map" is being prepared to steer projects elsewhere.[134] True IDC broke ground on March 31, 2026, on an AI site in the EEC of up to 250 MW, with phase 1 due in 2027.[162] STT Bangkok 1 was Thailand's first NVIDIA DGX-Ready data center and runs AI clusters with direct-to-chip liquid cooling.[163]

The BOI rules. Announcement Sor. 9/2568 of November 14, 2025, creates a "high energy efficiency" category of data center (8.2.1.1), which must have a PUE, the ratio of total facility power to IT power, no higher than 1.3 before the tax exemption applies.[18] Both categories need at least 2 MW of IT load; 4 or more telecom circuits totaling 60 Gbps; concurrent maintainability with redundant generators (a design like Tier III, in our reading); ISO/IEC 27001; a water plan; a plan for benefits to Thailand; and Thais in at least 50% of executive and specialist posts within 3 years.[18][135] The incentive groups are as follows:[135]

GPU hosting qualifies. In May 2025 BOI tied the 8-year tier to PUE and to "data hosting using advanced computing capabilities such as graphic processing units (GPU)."[160] The text of November 2025 no longer mentions GPUs; it lists services such as colocation and data hosting.[18] We read the later text as binding, so the tier turns on PUE and not on GPUs. BOI's statistics for 2025 show one application for GPU data hosting worth 8,110 million THB.[164] Since Sor. 2/2569 (March 30, 2026), applicants have also needed a power-supply letter from the Energy Regulatory Commission (ERC) before applying. A law-firm note that we could not read says a later notice dropped this requirement, so it should be confirmed with BOI.[165][135]

Government cloud and national AI compute

The government's central cloud, GDCC, gets 4,980 million THB (USD 151 million) in FY2027, up from 2,264 million THB in FY2026, when it served 1,212 agencies.[76] The money pays NT to run it.[77] In August 2026 the digital economy minister said GDCC would become a "National Cloud" with strong GPU infrastructure, but gave no GPU count or model.[78]

The National AI Strategy (June 2026) lists ThaiLLM as center of excellence #7 and a "Government AI Computing Center" as #8, and it proposes at least 2,000 million THB for open-source AI infrastructure and a National Data Bank.[75] We found no site, budget line or GPU count for center #8. In the FY2027 budget bill, BDI's ThaiLLM center gets 52 million THB (USD 1.6 million), according to an opposition MP's analysis.[39]

Electricity for data centers

Green tariffs. UGT1 (unspecified green power) started in January 2025 at 4.21 THB/kWh, 0.06 THB above the standard rate; its premium is now 0.0375 THB/kWh.[166][167] UGT2 (power from named new plants) was set by an ERC resolution on April 1, 2026.[168] Its rate is 4.5475 THB/kWh in a BOI briefing from 2024 and about 4.56 THB in trade press from March 2026.[169][170] The trade press gives fixed 10-year contracts, while Thansettakij says 10 to 25 years, aimed at large users and data centers.[170][171] We rely on neither figure until the Provincial Electricity Authority (PEA) publishes its rate.

Direct PPA. ERC's pilot for direct power purchase agreements (PPA) lets solar and wind plants sell 2,000 MW directly to data centers over the grid.[172] The draft of October 2025 admits only new BOI-promoted data centers with at least 50 MW of IT base load per building.[173]

New data-center rules. In June 2026 ERC said it will license electricity sales to data centers and rank projects by national benefit.[174] On July 15, 2026, the National Energy Policy Council approved in principle a separate tariff class for data centers.[136] A rate of 5 to 6 THB/kWh is under consideration, against 3.95 THB for September to December 2026 (27 to 52% higher, by our arithmetic), and large data centers must post 4.5 million THB (USD 136,000) per reserved MW, with half refunded at 50% use and the rest at 70%.[134] The size that counts as "large" is not published.

Chiang Mai power today. PEA's time-of-use tariff from September 2026, at 22 to 33 kV, charges 132.93 THB per kW of peak demand plus 4.1839 THB/kWh on peak (09:00 to 22:00 on weekdays) and 2.6037 off peak, the same from 30 kW (type 3.2) as from 1,000 kW (type 4.2).[25] The fuel adjustment (Ft) adds 0.1623 THB/kWh.[175] A flat 24-hour load pays about 3.56 THB/kWh before VAT and 3.81 with VAT, or about 33,000 THB (USD 1,000) per continuous kW per year (our estimate). NVIDIA puts a DGX B300 at about 15 kW, so one server costs about 0.5 million THB a year in IT power, or 0.65 million at PUE 1.3 (our estimate).[176]

US export rules for AI chips

The AI Diffusion Rule of January 2025 put Thailand in Tier 2.[177] On May 13, 2025, the US Bureau of Industry and Security (BIS) rescinded it, stopped enforcing it and promised a replacement.[178] In July 2025 Bloomberg reported a draft rule to restrict shipments of AI chips to Malaysia and Thailand; we found none published by October 27, 2026.[179]

BIS guidance of May 31, 2026, says a license is required worldwide for advanced computing items going to entities headquartered in Country Group D:5 or Macau, or owned from there; the non-enforcement of destination licenses applies only to other buyers, and existing data centers may keep running the chips already installed.[137]

In July 2026 a White House official alleged that the Chinese firm Moonshot "accessed GB300s in Thailand."[180] The Commerce Department is reportedly drafting a rule to stop Chinese AI firms from renting GPUs through data centers in countries such as Thailand.[138] The Remote Access Security Act passed the US House in January 2026 but, as of August, not the Senate.[181]

Our reading. A buyer that is not headquartered in, or owned from, Country Group D:5 or Macau can import H200 or B200-class servers into Thailand today without a US license. We recommend that the investor check its ownership chain against that test, expect end-user statements and possibly a site check from the vendor, and plan know-your-customer checks if it rents GPU time to others. US policy has changed several times in 18 months, so the position should be verified with the vendor before ordering.

Chiang Mai today

NT lists its Chiang Mai data center as Tier 3 (an operator claim) with 25 racks, under renovation with completion estimated for June 2026; its own site page gives IT power and floor area as N/A.[12][182] We could not confirm that it has reopened. AIS runs a Chiang Mai data center with no published size or tier.[183] None of the GPU clouds above names a location in Chiang Mai.

The only sizeable GPUs we found are in ERAWAN, the HPC system of CMU's Information Technology Service Center (ITSC), in service since December 2022, whose scheduler lists 28 GPUs, including one 4-GPU H100 node.[184][13] Its total memory at launch, 1,920 GB, fits 24 A100 80GB cards (our estimate).[185] Outside users pay 200 THB (USD 6.06) per GPU-card-hour, more than twice Siam AI's list price for an H100.[186] The Faculty of Engineering's modular data center adds 4 A100 40GB.[187] The note on option A gives the colocation terms.

What this means for the investor

This section gives our judgment and recommendations.

Renting is the bar to beat. Eight H100s at Siam AI's list price cost about USD 188,000 (6.2 million THB) a year (2.69 x 8 x 8,760 hours), more than half the roughly 11 million THB that Siam AI reportedly paid for a DGX B200. We recommend that the investor price every purchase against this figure. LANTA (about 34 THB per A100-hour) suits fine-tuning but not round-the-clock serving.

The north has no GPU capacity. All the commercial GPU offers we found are in Bangkok or the EEC, and the hyperscalers have no GPUs in Thailand. A Chiang Mai site would be the region's first, which is a selling point to CMU and to firms that must keep data in Thailand.

BOI's thresholds set the size of option B. The 8-year exemption needs 2 MW of IT load (about 130 servers of 15 kW), a PUE of 1.3 and ISO 27001, and Direct PPA needs 50 MW. We recommend that the investor decide early whether option B is planned for 2 MW or gives up the incentive; option A avoids the question.

Power prices carry policy risk. We recommend budgeting about 3.6 THB/kWh, with a sensitivity case at 5 to 6 THB in case the data-center tariff class reaches a site of this size.

Export rules allow a purchase now, subject to the ownership test above, but renting capacity to others needs customer screening from day one.

NT is the natural public partner. It hosts ThaiLLM, runs the growing government cloud and has a Chiang Mai site. We recommend that the investor complete due diligence on the OBON allegations before any tie-up with Siam AI.

Open questions

• ThaiSC (NSTDA): Is a successor to LANTA budgeted? Does the reported "GPU-hour" mean one A100-hour? Can companies reserve dedicated nodes?
• Siam AI: How many GPUs are in service in 2026, and where is the H100 cluster? What has its NVIDIA Cloud Partner status been since May 2026, and what does a committed contract cost?
• NT's data-center unit: Has NT Chiang Mai reopened, and with how many MW and how many kW per rack? How many GPUs will the "National Cloud" get?
• BOI: Is the ERC letter still required? Is any incentive open to an AI facility below 2 MW at CMU STeP?
• ERC and PEA Region 1 (Chiang Mai): What size threshold applies to the data-center tariff and the bond, and what is PEA's UGT2 rate for 2026?
• The export-compliance teams of server vendors: Is there any new rule on remote access or on Thailand, and what paperwork is required?
• Department of Business Development: Are Thai AI Cloud and TATC registered Thai companies?
• MDES: What are the site, budget and timing of the Government AI Computing Center?