Contents
Option A: the investor's servers at CMU ITSC (colocation)
Research note, October 27, 2026, Alexander Grabowski (B4L). Sources retrieved October 27, 2026; figures are cited and our estimates marked.
In option A the investor buys GPU servers and places them in the data center of Chiang Mai University's Information Technology Service Center (ITSC), at 239 Huay Kaew Road on the main campus.[235] B4L plans and researches the option on the investor's behalf. Money is in THB, with USD in parentheses at 33 THB per USD, which is B4L's planning rate.
In short
- ITSC rents space by the height a server takes up in the rack, from 1,500 THB (USD 45) a month for 1U to 20,000 THB (USD 606) for a full 42U rack. The tariff dates from June 2021, and the ITSC wiki still showed it in February 2026.[14][220]
- Outside organizations need an academic cooperation agreement with CMU. The rules ban private carrier circuits, cap power at a level ITSC does not publish, and limit on-site work to 6 hours a day and 60 hours a month.[14]
- The whole data center runs on backup power systems of 120 kVA and 125 kVA and one 320 kVA generator.[15] A single DGX B200 draws up to about 14.3 kW.[189]
- In our reading, one or two GPU servers can fit if ITSC agrees to supply the power, but a GPU cluster cannot.
- CMU runs its own GPUs, about 24 A100s plus a 4-GPU H100 node, and charges outside users 200 THB (USD 6.06) per GPU-hour.[13][186] It launched a CMU Academy of AI with NVIDIA, NT, AIS, True, and others on February 27, 2026.[79] We found no link between CMU and ThaiLLM.
- We estimate that hosting one 8x B200 server would cost about 38,000 to 84,000 THB (USD 1,150 to 2,530) a month with power at PEA rates. That is small next to the cost of buying the hardware or renting it.
What ITSC offers
The facility
ITSC's annual report, listed as the 2024 report, describes the data center's main systems.[15] Power comes from two power and backup systems of 120 kVA and 125 kVA, described as "Fully Redundancy," plus a 320 kVA generator; we read the two systems as UPS units. Precision air conditioning cools the room. For fire protection it has VESDA (very early smoke detection) and Halon gas suppression.
ITSC monitors the room 24 hours a day and controls entry with face scans and RFID cards.[236]
The data center has held ISO/IEC 27001 (information security) since April 7, 2019.[15] On April 4, 2026, CMU also became the first Thai university certified to ISO/IEC 27701 (privacy).[237] Neither standard says how resilient the power and cooling are. That is what a tier rating measures, as option B explains, and we found no Uptime Institute or TIA-942 tier rating for ITSC. Nor does ITSC publish its floor area, rack count, cooling capacity, or free capacity.
Prices and rules
| Space per server | THB per month | USD per month | What fits (our mapping) |
|---|---|---|---|
| 1U | 1,500 | 45 | |
| 4U | 3,100 | 94 | one 8x RTX PRO 6000 server |
| 10U | 6,000 | 182 | one DGX B200 or B300, or one DGX H200 (8U) |
| 20U | 11,000 | 333 | two 10U servers |
| 42U | 20,000 | 606 | a full rack |
The wiki and the 2021 tariff (no. ร 1/2564, dated June 1, 2021) give the same prices.[14][220] As we read them, neither mentions a separate charge for power. Any discount is at the discretion of the ITSC director.[220]
These are the wiki's rules that matter most for GPU servers:[14]
- Servers must have two power supplies.
- Only named contacts may enter, and only on official working days.
- Visits are capped at 6 hours a day and 60 hours a month.
- No leased lines or other private circuits may be brought in.
- Power use must stay under a limit that ITSC sets.
- ITSC staff reboot equipment only on written request.
- Customers give 30 days' notice to leave.
The sources disagree on who may use the service. The wiki and the 2021 tariff accept outside organizations that cooperate academically with CMU,[14][220] while the public service page mentions internal units only.[238] We give more weight to the formal tariff and to the wiki, which was edited more recently. The investor therefore needs an MOU with CMU first, and B4L can prepare and negotiate it on the investor's behalf.
ITSC lists three contacts. Colocation inquiries go to 053-941779 or naranont.a@cmu.ac.th,[14] operations answers on 053-943870,[236] and the One Stop service is on 053-943800 ext. 1.[238]
The BKNIX internet exchange has one of its two Chiang Mai sites at ITSC.[239]
What the hardware needs from it
| Server (8 GPUs, see hardware) | Height | Power per server | Per rack at 10 kW (est.) | Per rack at 15 kW (est.) | Share of 245 kVA (est.) | Share of 120 kVA (est.) |
|---|---|---|---|---|---|---|
| RTX PRO 6000 (Supermicro) | 4U | 6.8 kW nameplate[20] | 1 | 2 | 2.8% | 5.7% |
| DGX H200 | 8U | 10.2 kW max[188] | 0 | 1 | 4.2% | 8.5% |
| DGX B200 | 10U | 14.3 kW max[189] | 0 | 1 | 5.8% | 11.9% |
| DGX B300 | 10U | 15 kW, 19.7 kW peak[176] | 0 | 0 (1 if power-capped) | 6.1% (8.0% peak) | 12.5% (16.4% peak) |
The 10 and 15 kW rack limits in the table are our assumptions. Most racks consume 5 to 15 kW,[240] and ITSC's real limit is the first question to ask. The shares divide server kW by UPS kVA, taking 1 kVA as roughly 1 kW, which is our simplification. The 120 kVA column applies if the two UPS units back each other up rather than adding their capacities. ITSC's existing load is unknown.
One user measured 6.5 to 7 kW at the wall for an 8x B200 node, an anecdotal figure.[230] Circuits must still be sized for the nameplate rating.
NVIDIA's layout guides assume much denser racks. For the DGX B200 they typically place 2 systems in a rack at 28.6 kW, and 4 per rack (57.2 kW) only in sites built for extreme-density air cooling.[241] For the DGX B300 they place 2 per rack at 30 kW on average with plain air cooling, while 4 per rack (58 kW average, 76 kW peak) need an active rear-door heat exchanger.[176]
NVIDIA's documents disagree on inlet air. The B300 user guide allows 10 to 35 °C,[195] but the planning guide limits it to 30 °C up to 1,000 ft of altitude and 25 °C up to 5,000 ft.[176] Chiang Mai's weather station sits at 313 m, about 1,030 ft,[242] so, read literally, the 25 °C limit applies. We weight the planning guide, which is meant for site design.
Scaling up shows where the room runs out (our estimates). Four DGX B200s draw 57.2 kW, which is 23% of the 245 kVA total, or 48% of 120 kVA. Cooling comes on top: at an assumed PUE of 1.6 (PUE is total facility power divided by IT power), those servers would put about 92 kW on the 320 kVA generator. Eight servers would need 114 kW of IT power (47% or 95% of the UPS capacity) and 183 kW at the generator.
One or two GPU servers might fit if ITSC gives one rack 15 to 30 kW. A cluster does not fit in this room.
CMU's GPUs and AI program
ITSC built its ERAWAN HPC system in 2021 and 2022 with PMU-C funding.[243] At launch it had NVIDIA HGX A100 GPUs with 1,920 GB of GPU memory in total.[185] We read that as about 24 A100 80GB cards, which matches the 16 + 8 GPUs listed in two of the system's job queues. A later GPU-H100 queue has one node with 4 GPUs.[13] Outside users pay 200 THB (USD 6.06) per GPU-hour, and CMU staff pay 5 THB.[186] For comparison, Lambda rents an A100 80GB at USD 2.79 and an H100 at USD 3.99 per GPU-hour on demand.[23] ERAWAN is cheap for CMU staff but not for the investor.
The Faculty of Engineering received a modular data center, Vayu Ready, from Krungthai Computer Services on October 30, 2023.[244] It has one GPU node with 4x A100 40GB, and an A100 add-on costs 3,000 THB (USD 91) a month.[187]
CMU launched its Academy of AI (AAI) on February 27, 2026. The academy rests on four pillars, one of them "Infrastructure Preparation."[79] Its partners include NVIDIA (Thailand), IBM, Google, AWS, Oracle, AIS, True, and National Telecom (NT), and the cooperation covers "AI Infrastructure & Cloud Sovereignty."[79]
On October 16, 2026, ITSC published a draft tender for a computing service "for students in data analytics and AI."[245] The terms of reference describe a 32.892 million THB (USD 997,000) virtual desktop system for 200 clients, with NVIDIA L4-class GPUs.[246] In our reading, that is classroom infrastructure rather than hardware for running LLMs.
NSTDA's list of ThaiLLM development partners does not name CMU,[3] and we found no link from CMU to Siam AI or Typhoon. The only indirect link is NT, which provided the compute for the ThaiLLM Playground[30] and is also an AAI partner (see the ThaiLLM note).
This leaves room for a partnership (our judgment). These sources show no Blackwell-generation GPUs at CMU. If the investor's servers set aside GPU hours for CMU researchers or AAI courses, CMU gains a reason to sign the MOU and to offer workable power terms.
CMU STeP
B4L's office is at CMU STeP on the Mae Hia campus, which makes B4L the investor's local point of contact.[247] Mae Hia is about 5 km south of the main campus, where ITSC sits.[248] STeP houses the Northern Science Park.[249]
BOI Announcement 22/2565 makes the park an investment promotion zone with two incentives.[234] Development in bio, nano, and advanced materials gets 2 extra years of corporate income tax exemption, and R&D, electronic design, engineering design, and laboratories get a 50% tax reduction for 5 years. As we read the list, a data center is not on it. BOI's data-center category needs at least 2 MW of IT load,[250] far above option A (see Thai compute).
A 2022 Vertiv case study reports a 120 kVA UPS for a "campus data center" (vendor claim).[249] We found no offer of colocation or server-room space at STeP.
Network
BKNIX Chiang Mai, the city's internet exchange, has two sites: Symphony on Thipnate Road and CMU on Huay Kaew Road.[239] PeeringDB lists 8 peers and 75G of total capacity, best effort with no SLA. The peers include UniNet (10G), ByteArk (20G), and CMU.[251] AIS, True, and NT are absent (our reading), so most traffic to Thai users probably still passes through Bangkok.
The sources disagree on CMU's port. BKNIX's IXP Manager shows 10 Gbit,[233] while PeeringDB shows 1G.[251] We weight IXP Manager, the system the exchange itself runs; the PeeringDB record dates from December 2024.
ITSC runs CMU's link to UniNet, Thailand's research and education network.[235] The latest bandwidth figures we found are from 2019: 6.6 Gbps domestic and 9.13 Gbps international.[252] Because private circuits are banned, the investor's servers could reach users only through CMU's network. Whether commercial traffic is allowed, and at what transit price, is an open question.
Nobody has measured the latency to Bangkok. We estimate 8 to 15 ms round trip over roughly 700 km of fiber.
Costs at ITSC
The worked example is one DGX B200, which takes 10U.
Space costs 6,000 THB (USD 182) a month, or 20,000 THB (USD 606) if ITSC wants the server alone in a rack.[14]
Power is not in the tariff, and the electricity tariff CMU itself pays is unknown. As a reference we use PEA's time-of-use tariff at 22 to 33 kV from September 2026 bills:[25] a demand charge of 132.93 THB per kW, energy at 4.1839 THB/kWh on-peak (09:00 to 22:00 on weekdays), and energy at 2.6037 THB/kWh off-peak. The Ft surcharge, a fuel adjustment, adds 0.1623 THB/kWh.[175]
For a flat 24/7 load we estimate 3.81 THB/kWh including VAT. On-peak hours are 65 of the 168 hours in a week, so energy averages 3.215 THB/kWh. Ft adds 0.162 THB/kWh, and the demand charge spread over 730 hours adds 0.182 THB/kWh, for 3.56 THB/kWh before VAT. VAT at 7% brings this to 3.81 THB/kWh.
We assume a PUE of 1.6 for an older air-cooled room (est.). The Uptime Institute's 2025 survey average is 1.54.[253]
| Load case (est.) | IT kW | Facility kWh per month | Power, THB per month | Total with 10U space | Total with a full rack |
|---|---|---|---|---|---|
| Light, about 50% of nameplate | 7.15 | 8,350 | 31,800 | 37,800 (USD 1,150) | 51,800 (USD 1,570) |
| Typical, about 70% | 10.0 | 11,680 | 44,500 | 50,500 (USD 1,530) | 64,500 (USD 1,950) |
| Full nameplate | 14.3 | 16,700 | 63,600 | 69,600 (USD 2,110) | 83,600 (USD 2,530) |
All figures in the table are our estimates, computed as kW x 730 h x 1.6 x 3.81 THB plus the space fee.
Power is 60 to 90% of the bill. If ITSC included it in the 10U price, it would be selling about 44,500 THB of electricity for 6,000 THB. In our judgment, the MOU should set a separate price for power.
Three comparisons put these sums in proportion. For colocation in the United States, Lenovo assumes about USD 1,500 (49,500 THB) a month per high-density rack,[196] and a GPU-finance firm assumes USD 280 per kW per month,[21] which by our arithmetic is about USD 4,000 (132,000 THB) for 14.3 kW. Buying the server costs far more: a new HGX B200 server costs USD 360,000 to 470,000,[21] which spread over 5 years is 198,000 to 259,000 THB (USD 6,000 to 7,800) a month (our arithmetic). Renting 8 B200 GPUs instead, at Lambda's USD 6.69 per GPU-hour,[23] costs about 1.29 million THB (USD 39,000) a month at full use.
Hosting is therefore a small share of the cost of owning a server, and utilization decides whether owning beats renting, as the hardware note shows.
Transit and IP fees, remote hands, and insurance are still unknown.
Alternatives in Chiang Mai
| Site | What is published | Status |
|---|---|---|
| NT Data Center Chiang Mai | Tier 3 (operator claim, no certificate found), 25 racks, under renovation[12]; IT power and floor area "N/A"[254] | completion estimated for June 2026; reopening not confirmed as of Oct 27, 2026 |
| AIS Data Center Chiang Mai | near the central business district; no tier or size[183] | operating |
| AwareIDC | colocation, 99.9% uptime SLA (vendor claim)[255] | operating |
| Edge Centres EC61 | solar edge site for 64 quarter racks of 1 kW (2022 news)[256] | unknown |
NT is a carrier, so we infer that its site would not ban private circuits. None of these sites publishes a per-rack power limit.
What this means for the investor
Option A is a fast first step that needs little capital. It needs no building and no grid connection, and it puts the servers in an ISO 27001 room 5 km from B4L's office. The RTX PRO 6000 server had a lead time of 3 to 5 weeks in October 2026,[20] so the timeline depends on the MOU rather than on the hardware.
We advise starting with one server. The 8x RTX PRO 6000, at 6.8 kW, fits a normal rack most easily, while a B200 or B300 needs a rack of 15 to 20 kW.
ITSC serves best as the pilot for option B. A year there gives measured power, utilization, and demand, which are the numbers that size option B. CMU is also the likeliest anchor partner for option B.
The investor should also know the ceiling. Two servers is the practical limit. The access hours, the ban on private circuits, and air-only cooling make ITSC a pilot site and not a home for a cluster.
The investor's MOU with CMU should cover the points below, and B4L can prepare and negotiate it on the investor's behalf.
- Purpose. The agreement rests on academic cooperation, such as GPU hours for CMU researchers and AAI courses.
- Power. ITSC commits a number of kW per rack on two independent (A and B) feeds and prices it, either as metered pass-through or as a flat rate per kW.
- Cooling. ITSC guarantees a range of inlet temperatures.
- Access. The investor gets escorted 24/7 access for incidents, and remote hands (on-site help from ITSC staff).
- Network. The MOU fixes the port speed, the public IP addresses, permission for commercial traffic, and the transit price, and grants an exception for an NT or AIS circuit.
- Term. The term runs 3 to 5 years to match the hardware's life, instead of 30 days' notice, with first refusal on more space.
- Liability. The MOU settles liability, insurance, and data protection under the PDPA (Thailand's personal data law).
- Exit. The MOU sets the terms for removing equipment and for handling disks.
Open questions
These questions go to ITSC (053-941779, naranont.a@cmu.ac.th):
- What is the per-rack power limit, and can one rack get 15 to 30 kW?
- How much UPS, generator, and cooling capacity is free? Are the 120 and 125 kVA units a redundant pair?
- Is power included in the price, passed through, or priced per kW?
- Can the investor (or B4L acting for it) get 24/7 incident access, and remote hands out of hours?
- Which MOU template applies, who signs for CMU, and how long does approval take?
- Is liquid cooling possible now, or is it planned?
- What uplink speed and public IP addresses come with a server? Is commercial traffic allowed, and what does transit cost?
- What insurance covers customer equipment?
- What is the cold-aisle temperature, and the floor load limit per rack?
Four more questions go to others:
- Would shared GPU hours fit the "Infrastructure Preparation" pillar of CMU AAI?
- Does CMU STeP (053 948678, info@step.cmu.ac.th)[247] have a server room with spare power?
- Has NT's Chiang Mai site reopened, and what are its per-rack limit and price?
- What port speed does CMU actually have at BKNIX?