Option B: a new data center in Chiang Mai to the highest standards

Research note, October 27, 2026, Alexander Grabowski (B4L). Sources retrieved October 27, 2026; figures are cited and our estimates marked.

In short

• The top reliability grades are Uptime Tier IV and TIA-942 Rated-4, but NVIDIA's GB200 SuperPOD design asks only for Tier 3, TIA-942 Rated 3, or EN 50600 Class 3.[16] We recommend a certified Tier III facility with selected Tier IV features, plus ISO 27001, 50001, and 14001 (our judgment).
• The best published PUE figures run from 1.036 (NREL's HPC data center) to 1.12 (Google Singapore, in the tropics), against an industry average of 1.54.[257][258][253] Our target is an annual PUE of 1.15 to 1.2 with near-zero water use (our judgment).
• Chiang Mai is hot (the station record is 42.5 °C) and smoky in spring, so outside-air cooling would do little; warm-water direct-to-chip cooling, with inlet water up to 45 °C, is the efficient path.[259][190]
• Land and construction for a Thai data center cost about 300 million THB (USD 9.1 million) per MW, according to BCG citing Cushman & Wakefield.[17] We estimate 320 to 510 million THB (USD 9.7 to 15.4 million) for a 1 MW liquid-cooled facility before any GPU is bought (our estimate). Throughout this note USD converts at 33 THB, the B4L planning rate.
• BOI's data-center promotion needs at least 2 MW of IT load, and its best tier a PUE of 1.3 or lower, so a smaller facility gets no data-center incentive.[250]
• Chiang Mai is in seismic Zone 3, the highest of three, and in October 2024 the Ping River rose about 1.08 m above the city flood wall.[28][29]

Four later notes build on this one. Where to build looks for a better site than Chiang Mai, with room for solar and wind; facility systems covers the generator plant and telemetry; open stack covers open hardware, the network, and storage; and hosting and agent servers covers what else the building can sell.

What "highest standard" means

Three families of standards grade a site's physical reliability, and three ISO standards certify how it is run.

Uptime itself warns that its progression "does not mean that a Tier IV data center is better than a Tier II"; the Tiers fit different business needs.[269] Tier III is concurrently maintainable: any component can be taken out for planned maintenance, but the site is still exposed to an equipment failure or an operator's error.[260]

In Thailand, Uptime lists two Tier IV awards, both in Chonburi: PTT Data Center Phase 1 (design and constructed facility) and SUPERNAP Sector 1 (design only).[270] No Chiang Mai facility appears on that list.[270] The Thai entries our research found in TIA's directory are Rated-3, for example True IDC North Muangthong, certified as a TIA-942-C constructed facility in February 2026.[271] In Chiang Mai, NT lists a 25-rack "Tier 3" site under renovation, for which we found no certificate.[12]

AI halls change the question. An article published by TIA notes that each liquid-cooled GPU tray and rack manifold has one feed and one return: "There is no redundancy possible."[240] Rating 4 cooling therefore reaches only the rack manifold. Beyond it the hall is in effect Rating 3, and the standard does not yet address AI architectures.[240] NVIDIA's GB200 SuperPOD reference architecture asks for Uptime Tier 3, TIA-942 Rated 3, or EN 50600 Class 3.[16] Schneider Electric's AI-factory reference designs for GB200, GB300, and Vera Rubin are all specified at Tier 3.[272]

Our recommendation, as a judgment, is Uptime Tier III certification of the design and of the constructed facility, or a TIA-942-C Rated-3 facility certificate. Tier IV features should be added where they are cheap: 2N UPS for network and storage, a second utility feed if PEA can supply one, separate fire compartments, and N+1 coolant distribution units and dry coolers. The white space should meet EN 50600-style Protection Class 3 to 4, and operations should hold ISO 27001, 50001, and 14001. That is the level of STT Bangkok 1, the first Thai site with both TIA-942 Rated-3 and Uptime Tier III design and construction certificates.[273] Full Tier IV adds cost for little gain when the GPU racks themselves are single-path.

What "most energy efficient" means

ISO/IEC 30134 defines the metrics. PUE, the ratio of total facility energy to IT energy, is the headline figure; a PUE of 1.0 would mean zero overhead.

LUMI's two figures need care. Heat pumps that lift its waste heat for district heating consume power, so heat reuse raises its PUE from 1.04 to 1.24.[281] The higher figure is the cost of exporting heat and does not mean the facility is worse. That trade-off is the reason ERF exists.

Singapore has the nearest tropical rulebook. Its second data-center call for capacity (DC-CFA2, December 2025) requires a PUE of 1.25 or better at 100% IT load, Green Mark for Data Centres Platinum, IT equipment meeting SS 715:2025, and at least 50% of new capacity on eligible green energy.[282] SS 726:2026 (August 27, 2026), which IMDA calls the first liquid-cooling standard for tropical data centers, addresses humidity, corrosion, and water use, and IMDA says liquid cooling can cut data-center energy by more than 30%.[283] Thailand's bar is BOI's PUE of 1.3 for its "high energy efficiency" category.[18]

We propose these targets (our judgment): a design PUE at or below 1.2 at full load; an annual PUE of 1.15 to 1.2, measured to ISO/IEC 30134-2; WUE near zero; REF, ERF, and CUE reported; and ISO 50001 certification. Each 0.1 of PUE on 1 MW of IT is 876 MWh a year, about 3.3 million THB (USD 101,000) at the PEA tariff (our estimate).

Cooling in Chiang Mai's climate

ERA5 gives a lower April mean daily maximum (34.6 °C) than the station (36.1 °C).[284][242] We weight the station for peaks, because a reanalysis grid cell smooths out extremes. With few cool hours and smoky spring air, outside-air economizers would do little (our judgment).

Liquid cooling changes the picture. A GB300 NVL72 rack draws 135 kW (up to 155 kW at peak), sends about 90% of its heat to liquid, and accepts inlet water up to 45 °C, which Lenovo says lets dry coolers bypass chillers for most of the year in most climates.[190] LBNL gives an approach of 5.5 K for a dry cooler (from outside dry bulb to leaving liquid) and 2.77 K for a coolant distribution unit.[286] By our arithmetic, dry coolers alone hold a 45 °C supply while the outside air stays below about 36.7 °C (45 - 5.5 - 2.77). On ERA5 figures that limit is passed for only some tens of hours a year (37.5 °C is exceeded about 35 hours a year, 0.4% of 8,760). On station data, though, many afternoons from March to May exceed it, and the 42.5 °C record would give about 51 °C. We infer that the hall can run without chillers for most of the year, with adiabatic spray on the dry coolers or a small trim chiller for hot-season afternoons. NVIDIA says 45 °C coolant moves even warm-climate sites close to an ideal where chillers run "just a few days a year" (vendor claim).[287]

The remaining 10% of the heat, 13.5 to 15.5 kW per GB300 rack, still needs filtered, humidity-controlled air cooling (our estimate). By comparison, an air-cooled DGX B200 draws up to 14.3 kW per 8-GPU system, and NVIDIA says rear-door heat exchangers are not typically suitable for it.[241] An air-cooled GPU hall here would need compressor cooling most of the year and would miss the targets above (our judgment); the hardware note has more on the servers.

Water and heat reuse

Microsoft says that all its new data-center designs since August 2024 cool chips in a closed loop with zero water evaporated, avoiding more than 125 million liters a year per site; the loop still uses chillers.[288] NVIDIA puts cooling-tower use at roughly 2.6 million US gallons (about 9.8 million liters, our conversion) per MW per year, and says dry-cooler designs cut this to near zero (vendor claim).[287] BOI asks promoted data centers for water plans, and the National Energy Policy Council asks the same of large ones.[18][134] We propose a closed loop with dry coolers, with spray water only on peak afternoons (our judgment).

Chiang Mai has no district heating network that we know of, so waste heat is most useful when it is turned into cooling (our reading). At the Leibniz Supercomputing Centre (LRZ) in Germany, adsorption chillers turned an average 120 kW of 45 °C waste heat into about 50 kW of 21 °C cooling, with an overall COP of 12.[289] At that ratio, the roughly 900 kW that 1 MW of liquid-cooled IT puts into water could give about 375 kW of cooling (90% of 1 MW to water, then 900 x 50/120; our estimate). That is more than the hall's air-side load of about 100 kW. Neighboring CMU and STeP buildings at Mae Hia could use the surplus; this is our idea and has not yet been discussed with CMU.

Power and renewables

A facility of 0.5 to 2 MW of IT (0.6 to 2.4 MW in total at PUE 1.2) stays far below 10,000 kVA, so in our reading it should connect at 22 or 33 kV. Our tariff arithmetic for a flat 24-hour load, with 39% of hours on-peak, gives about 3.56 THB/kWh before VAT and 3.81 with VAT. One MW of IT at PUE 1.2 uses 10,512 MWh a year (1.2 x 8,760), which costs about 40 million THB (USD 1.2 million) at full load; at 5 to 6 THB/kWh it would cost 53 to 63 million THB (our estimates). We did not find what size counts as a "large" data center for the new tariff. The ERC also said in June 2026 that it will license power sales to data centers, ranking projects by national benefit.[174] A facility of this size would not qualify for the Direct PPA pilot (our reading of the 50 MW rule).

Solar helps less than it seems. By our arithmetic, matching 1 MW of IT at PUE 1.2 over a year needs 10,512 MWh / 1,387 kWh per kWp, about 7.6 MWp (7.3 MWp at PUE 1.15). That is roughly 43,000 m² at the panel density of CMU's rooftop array, far more than any data-center roof. Annual matching is not hourly matching either: Google's 24/7 carbon-free goal counts every hour, which at night needs storage or other sources.[293] Realistically, rooftop PV covers a small share, UGT covers the rest, and batteries serve for peak shaving (our judgment).

Site hazards

Cost and time

Our estimate takes the Thai figure plus 7% for liquid cooling as the low case (9.1 x 1.07 = USD 9.7 million per MW) and the contractor's Tier III top figure plus 10% as the high case (14 x 1.10 = USD 15.4 million per MW). Small sites lose economies of scale, so a 0.5 MW site will sit nearer the high case (our judgment).

The table excludes IT hardware, the grid connection, certification, and design fees. The IT can cost more than the building. A US reseller lists one GB300 NVL72 rack at about 150 to 190 million THB (USD 4.5 to 5.8 million) for the rack alone, and another US source estimates it at about 265 to 285 million THB (USD 8.0 to 8.6 million) installed (prices read in October 2026).[198][199]

The grid connection sets the schedule. In Turner & Townsend's survey, 48% of respondents name power availability as the top obstacle to delivering on time.[300] We estimate 18 to 24 months in Chiang Mai from the start of design to the first IT load, with the PEA connection and the long-lead electrical plant on the critical path. Prefabrication shortens the work on the hall but does nothing for the grid work. NVIDIA says its DSX POD (October 21, 2026) cuts deployment at a prepared site from six months to as little as 14 days (vendor claim).[301] Schneider's Reference Design 48 is a 1,000 kW, 12-rack modular AI design,[272] and Schneider says prefabricated pods cut on-site installation by 80% (vendor claim).[302] Vertiv claims its MegaMod CoolChip deploys up to 50% faster than an on-site build.[303] None of the three publishes prices.

BOI and scale

A facility below 2 MW therefore gets no data-center promotion. The network rule is demanding here too: BKNIX Chiang Mai, the local internet exchange, lists 8 peers and 75 Gbps of total capacity, best effort.[251] A different BOI route applies to tenants of the Northern Science Park (CMU STeP), where R&D and engineering-design activities get a 50% corporate income tax reduction for five years after the exemption period.[234] Data centers are not on that list (our reading).

There are two ways to reach 2 MW (our judgment). The investor can build with partners who share the cost and the capacity, or build in phases, with land, grid connection, and plant designed for 2 MW and a first module of 0.5 to 1 MW. Whether BOI accepts a phased approach is unconfirmed.

What this means for the investor

What follows is our judgment.

Scale is the hard part; the standards can be met. A certified Tier III, liquid-cooled hall at PUE 1.15 to 1.2 is buildable in Chiang Mai. A single-tenant hall of 0.5 to 1 MW, however, costs 160 to 510 million THB before any GPU, gets no BOI incentive, and pays off only at a high utilization that the investor has not yet shown.

We therefore suggest a phased path. The investor starts with option A or with renting, and measures the real load and the power drawn per server. In parallel, a site study of 6 to 12 months prepares a facility of 1 to 2 MW, certified to Tier III and ready for liquid cooling, to be built with partners. Construction starts when anchor tenants commit enough load to approach 2 MW.

Four decisions come first: liquid-cooled rack-scale systems or air-cooled 8-GPU servers, which fixes the design of the building; the target size, and whether to aim for BOI promotion; the site; and Uptime or TIA-942 certification.

A certificate is worth more than a claim. It is evidence that a partner or investor can check, while "built to Tier IV standards" without a certificate is a marketing phrase.

The summary compares option B with option A and with renting, and the Thai compute note shows where AI runs in Thailand today.

Open questions