Thai AI

This study asks what ThaiLLM, Thailand's national language-model effort, builds on, what it runs on and where it is hosted. It then asks what an investor should put money into to host AI hardware and open-weight models in Chiang Mai: GPU servers in Chiang Mai University's data center, or a new data center built to the highest standards and for the best energy efficiency. B4L (buildfor.life) researches and plans this for the investor as its consultant; B4L is not the investor. The study is shared by link only and is not indexed. Every figure carries its source, and our estimates are marked.

Start here

• Findings and recommendation The short answer: what ThaiLLM runs on and where, what that means for a host in Chiang Mai, how colocation at CMU ITSC, the investor's own data center and renting compare, and the staged path we recommend.
• Study brief The roles, the investor's goal, the six questions (ThaiLLM, Thai models, where Thai AI runs, colocation at CMU ITSC, the investor's own data center, and renting as the baseline) and the criteria that decide between them.

The design

• The first data center: a design A facility for 2 MW of IT load on 150 rai in the Sikhio corridor. The design covers the GPU and hosting halls, 2N power with continuous-rated generators, solar and battery, warm-water cooling, two fiber routes to the investor's own point of presence in Bangkok, the open-source stack, telemetry and staffing, and it gives the costs by phase, the schedule and the risks.
• 3D model of the design An interactive isometric site plan and a cutaway of the building, made with isokit. You can rotate and zoom, and hovering over any part shows its key figures; phase-2 items are drawn faded.
• Sourcing: who sells each component in Thailand For every main part of the design, from the PEA line and the transformers to the batteries, chillers, racks, GPU servers and the builders, the note names the makers and their Thai sellers, says what is made in Thailand, and gives Thai reference projects, the rules to write into the tenders and the long-lead items.

Investor deck

• The study as a slide deck Twenty slides that take the investor from ThaiLLM and the gap in the market through the staged path, the site, the first data center and what it can sell, to cost, earnings, what VPS and dedicated servers add, risks and next steps. It has speaker notes, a presenter view and light and dark PDFs, and every slide names the notes its figures come from.
• CCTV platform pitch Seventeen slides on whether the investor's company can become a provider of state CCTV in Thailand: what the state has decided, Bangkok's cameras today, whether the cameras call home, one platform with an agent per camera and a tamper-evident log, storage and AI at DC1, the research budget, the routes in, the numbers, risks and the first decisions. With speaker notes, a presenter view and light and dark PDFs.

Demos

• Demos: console, connections, assistants, hosting business, video agent One page with tabs, each simulated from the study's figures. The operations console shows the designed data center's power, cooling, racks, outside feeds, network and services, with scenarios such as a grid outage or a haze day and a drill in which everything fails at once. The connections map follows traffic from the investor's router and AS in Bangkok over two fiber routes, and power over the two feeders. An assistant that runs inside the data center answers its operators, and a product mockup shows the agent server the investor would sell to companies. The hosting business has three linked mockups: the public website with every server, price and the network and data-center pages; the customer portal, where a company orders, restarts, rescues and reinstalls its servers; and the staff backend, where website inquiries, sign-ups and status updates arrive. The video agent shows Bangkok's camera footage stored at DC1 with an index on top: an operator opens a case and asks where a car is, and the agent draws its path and last sighting on the map.

Tools

• GPU fleet and business-model calculator You choose where the GPUs run (CMU ITSC, the investor's own data center, a cloud, or API tokens only), the servers, from one 8-GPU box to liquid-cooled racks, and how the GPU time is sold (token API, GPU rental, dedicated hosting or own use), and you can add VPS hosts and dedicated servers for rent at the server note's prices. It then works out the upfront cost, power at the PEA tariff, running costs, revenue, payback, ROI, NPV and IRR with a ten-year cash view, and compares nine business models side by side. Every figure is sourced or marked as an estimate and can be edited, and a scenario can be shared by link.

ThaiLLM and the Thai AI landscape

• ThaiLLM: what it is, what it runs on, where it is hosted Who runs the national program, what its models are built on, the supercomputer it was trained on and the clouds that serve it, and the program's terms, budget and roadmap.
• Thai-capable open models Typhoon, OpenThai, Pathumma, THaLLE, WangchanLION, SEA-LION and the general models that speak Thai, compared on their sizes, their licenses for commercial hosting, how they were trained, how good they are and what hardware serves them.
• Where AI runs in Thailand A survey of the LANTA supercomputer, the Thai GPU clouds and their prices, the cloud regions that offer no GPUs, the data-center boom and the BOI rules, the government cloud, electricity for data centers, US chip export rules, and what already exists in Chiang Mai.

Options for the investor

• Hardware: what to buy, what it serves, what it costs GPU servers from an 8-card workstation-class server to a liquid-cooled rack: their prices and supply, which models fit, their throughput, and how buying them compares with renting GPU hours or buying tokens.
• Option A: the investor's servers at CMU ITSC Colocation in Chiang Mai University's data center: what it offers, its prices and rules, how much GPU power it can take, CMU's own GPUs and AI program, and the questions to put to ITSC.
• Option B: a new data center in Chiang Mai What the highest standards of reliability and efficiency mean in practice, cooling in Chiang Mai's climate, water and heat reuse, power and solar, earthquake and flood risk, the cost per megawatt, the time to build, and the BOI thresholds.
• Where to build: a better site than Chiang Mai Ten candidate sites compared on land, solar, wind, grid, hazards, water and BOI incentives. The note works out how much of the power on-site solar, batteries and wind can cover hour by hour and at what cost, sets out the rules for generating power on site and buying green power, and ends with a ranked shortlist led by the Korat and Chaiyaphum wind belt.
• Hosting and agent servers What else the building can sell: colocation, virtual and managed servers, storage, and agent servers, on which a company's data, its AI agents and the models run in the same data center. The note covers Thai rules and prices, an open-source reference design, costed packages and certifications, and it uses Hetzner as a benchmark for scale.
• VPS and dedicated servers Ordinary servers the building could rent out next to colocation: shared and dedicated-vCPU virtual servers and whole machines, with full specifications, prices, add-ons and a costing model. The note compares them with Thai and regional hosts, shows what the 2026 memory price surge does to the margins, and covers the Thai rules on logs and customer identity.
• Rescue and reinstall How customers would repair or reinstall a server that no longer boots, the way Hetzner's rescue system works: a NixOS rescue image booted over a closed provisioning network for one boot, reinstalls, a browser console that never exposes the management controllers, rescue for VPS through Incus, disk erasure between customers, and the cost, phasing and Thai logging rules.
• CCTV footage: storage on disk and tape, and video AI What Thailand has decided and what is only proposed about keeping CCTV footage in the country, how many state cameras there are, who stores footage today and at what price, a design for footage at DC1 on Ceph and LTO tape, Bangkok as the lead case, recognition and training within the PDPA, costs and a staged plan.
• One system for Thailand's state cameras A platform that registers every camera, bundles all streams, gives each camera a health agent, measures the maintenance contractors and logs every view; a costed research program on whether the cameras call home, from a contained network and a test lab to open firmware and a camera made in Thailand; and how the investor could become a provider of state CCTV.
• The open stack: from firmware to tape The most open choice for each layer: OpenBMC and OCP servers, NixOS, Incus or Proxmox, SONiC on white-box switches, open firewalls, and Ceph for S3 with LTO tape run by open software. The note prices examples of the network and the storage, covers backup and a second site, and names what stays closed (CUDA, firmware).
• Funding the open-source projects DC1 runs on A rule for giving part of the earnings back: each year the larger of USD 2,000 per developer and 0.5% of revenue. The note names the projects and how each is funded, splits the money among NixOS, Incus, FRR, vLLM, the foundations and Thai open source, values a NixOS mirror for Thailand, build servers, GPU time and engineering time, and explains the Thai tax rules for paying projects abroad.
• Facility systems: generator and telemetry A continuous-rated generator plant sized for the full load, with fuel permits and storage, the alternatives, grid reliability and testing. The second half covers telemetry of everything: ASHRAE limits, corrosion from haze and sulfur, free Thai feeds for air, weather, rivers, fires and earthquakes, and a costed sensor plan.
• Noise: what the neighbors and the staff would hear The first data center's noise at the fence, 300 m out and at the nearest mapped homes in six cases, from a quiet night to a grid outage on generators, computed by ISO 9613-2 from the design's layout and checked against Thailand's annoyance rule and the IFC and WHO benchmarks, with noise maps, hearing protection in the halls, a costed set of fixes and a background survey to run before the land is bought.
• Connectivity: bandwidth, own AS, peering, lines What bandwidth a new site can get; how the investor gets its own AS and IP addresses through APNIC; the NBTC licenses needed to sell connectivity; where to peer (Telehouse Bangkok, TH-IX, BKNIX, BBIX); what transit, exchange ports and dark fiber cost; and a costed plan for each stage.

Reference

• Sources Every cited source, numbered. Each [n] in the documents links here.