The first data center: a design

Research note by Alexander Grabowski (B4L), October 27, 2026. Sources were retrieved on October 27, 2026. Figures are cited, and our own estimates are marked.

This note turns the study into one design: the investor's first data center, which the investor would build, own and operate. B4L (buildfor.life) prepared it as consultant. The design decisions are the lead consultant's judgment. Where they rest on evidence, the evidence is cited here or in the sibling notes (summary, option B, site selection, facility systems, open stack, hosting and agent servers, connectivity). The arithmetic is in model/first_dc.py in the repository, with results in model/first_dc_results.md; it reuses the hourly solar series and the battery dispatch of the site model. Money is in THB, with USD in parentheses at 33 THB per USD, the B4L planning rate.

In short

• The site is C1, the Sikhio and Lam Takhong corridor in Nakhon Ratchasima (Korat): a 150-rai plot (24 ha) with an option on 100 rai more. Korat is in no seismic zone,[314] EGAT is tendering a new 500/230 kV substation there, Korat 4,[342] and both the railway and EGAT's fiber pass through the district.[784][761] Land listed at about 300,000 THB per rai puts the plot at about 45 million THB (USD 1.4 million).[340] Thep Sathit (Chaiyaphum) is the fallback.
• The building is designed and wired for 2.0 MW of IT, which is BOI's floor for data-center promotion, at an annual PUE (total facility power divided by IT power) near 1.2, against BOI's limit of 1.3.[389] Hall A takes 1.4 MW of liquid-cooled GPUs in two pods, hall B 0.4 MW of hosting and agent servers, and the core room 0.2 MW of storage and network. Phase 1 fits out about 1 MW.
• Power arrives over a PEA 22 kV connection of about 3.5 MVA and passes through two transformers, each able to carry the whole site, with UPS on both the A and B paths (2N). Three continuous-rated generators of about 1,600 kVA stand in a 2+1 arrangement, two of them in phase 1, as BOI's conditions require.[389] They are split between two yards, one at each end of the building and one for each power path, and run in parallel through the bus tie between the two switchboards. Fuel stays within the 15,000 L notification limit, split between two bunded farms, plus 2,500 L built into each set. That lasts about 64 hours at phase 1's full load and 36 hours at 2 MW (our model).[669][671]
• Solar comes as 10 MWp of ground-mount panels with a 20 MWh battery, which cover about 96% of phase 1's power hour by hour. At 2 MW the share falls to about 64% until phase 2 adds the same again on the option land (our model). UGT2 green power covers the rest.[312] At full use the plant costs about the grid price, and it hedges the 5 to 6 THB/kWh data-center tariff under discussion.[134] The solar and battery are bought in tranches matched to committed load. At a reference opening load of about 0.4 to 0.5 MW, the full 10 MWp would lose 69% of its output, while 2.5 MWp and 5 MWh cover about 73% at 3,270 to 3,763 THB/MWh. The field's civil works and connection are built for 10 MWp from the start, and the tranches defer about 0.23 to 0.29 billion THB of phase-1 capex (our model and arithmetic).
• Cooling for the GPUs is warm water direct to the chip, with dry coolers and small trim chillers (the air at C1 is above the dry-cooler limit for about 380 hours a year). Hall B is cooled by air on chilled water. No outside air enters the halls, and gas-phase filtration and corrosion monitoring protect the equipment. We estimate a PUE of 1.20 for phase 1 at full use, 1.19 at full build and 1.27 at the opening load, with near-zero water use (our model).
• The network has two meet-me rooms and two 100G backhauls on separate rights of way to a small point of presence at Telehouse Bangkok, where six Thai exchanges meet.[756] The investor holds its own AS and address block from APNIC and buys transit from two carriers, and SONiC switches run the network inside.
• The open stack is NixOS, Incus, SONiC, FRR, Ceph with LTO tape, vLLM and SGLang behind LiteLLM, Kanidm, Keycloak, OpenBao, NetBox and Prometheus. Closed software is used only where nothing open works: CUDA, firmware and safety-certified controllers.
• The facility costs 0.86 to 1.27 billion THB (USD 26 to 38 million) for phase 1, about a third of it solar and battery, and 1.42 to 2.03 billion THB (USD 43 to 61 million) at full build (our model). An example IT fit-out for phase 1 adds 0.27 to 0.48 billion THB. Opex is 44 to 91 million THB a year at phase 1's full use and 74 to 146 million THB at full build.
• The schedule, which is our estimate, starts with a land option, a grid inquiry and a corrosion survey in the first quarter of 2027, and applications go in when Thailand reopens them in early 2027.[497] Design and construction take 18 to 24 months, so the first IT load arrives between the second half of 2028 and the first half of 2029. The pilot at CMU ITSC runs in the meantime.
• The top risks are the national pause and new rules for data centers,[315][497] the suspension of fuel-tank permits,[393] water at Lam Takhong,[316] the grid connection, and GPU prices and demand.

Design basis

Six principles carry the design, and all of them come from the study:

1. Qualify for BOI's data-center promotion. BOI asks for at least 2 MW of IT, a PUE of 1.3 or lower, concurrent maintainability, continuous-rated generators for the whole load with a backup set, fire protection over the whole area, 24-hour security, ISO/IEC 27001 before the tax exemption, a water plan, at least four telecom systems, and colocation and managed services.[389] Outside the EEC that is group A2, with eight tax-free years.[319]
2. Have the top standard certified rather than claim it. Option B defined the top standard as Uptime Tier III, certified for both the design and the constructed facility, with selected Tier IV features. NVIDIA's GB200 reference architecture and Schneider's AI reference designs ask for Tier 3,[16][272] and a liquid-cooled GPU tray has one feed and one return, so even a Tier IV hall would still be single-path at the rack.[240]
3. Sell more than GPUs. Hosting and agent servers put companies' data and agents in the same building as the models (hosting and agent servers).
4. Use open source by default, where it is mature and can be supported. Closed products come in only where no open option works or a safety certificate requires one (open stack).
5. Measure everything, inside and outside, from the first day (facility systems).
6. Stage the money. The building is sized for 2 MW, and the IT is bought as demand arrives. The pilot at CMU ITSC comes first (summary).

Site and plot

Why C1. Sikhio ranks first on the shortlist in site selection. It lies in no seismic zone,[314] inside the Korat and Chaiyaphum wind belt, and next to EGAT's 1,000 MW Lam Takhong pumped-storage plant.[341] EGAT's June 2026 tender includes a new 500/230 kV substation in the province.[342] The Northeastern railway stops at Sikhio, 223.79 km from Bangkok, and EGAT's fiber network has nodes at Lam Takhong and Nakhon Ratchasima, which gives two separate rights of way for two fiber routes.[784][761] PTT's provincial gas pipeline reaches Korat along Highway 2.[702] Suranaree University of Technology is in the province,[363] and Korat's chamber of commerce is pitching the area to data-center investors.[362]

The alternative is C4, Thep Sathit (Chaiyaphum). It has the highest solar yield of the candidates, the densest wind farms and land listed at about 129,000 THB per rai,[345] and Chaiyaphum is among the lowest-income provinces that earn BOI's extra benefits.[319] The evidence on its fiber is the thinnest. The design does not depend on the site: the same building, plant and solar field fit C4, where 150 rai would cost about 19 million THB instead of 45 million (our arithmetic).

The plot. The 150 rai measure about 600 m from east to west and 400 m from north to south. The building sits east of center, with the substation yard to its north. Generator yard A stands at the building's west end, next to electrical room A, and yard B at its east end, next to room B. Each yard has its own bunded fuel farm: farm A lies about 12 m south of its yard beside the access road, and farm B about 30 m east of its yard, across the site road. Staff and visitors park beside the office block on the south side of the building, off the site road, where the cars stay clear of both yards, the tanker route and the fire-access roads. The 10 MWp solar field takes about 101 rai in the west, at the 10.1 rai per MWp of a filing with the Energy Regulatory Commission (ERC),[317] and the inverter station and five battery containers stand between the field and the building; the containers are about 40 m from yard A. Phase-2 solar goes on the 100-rai option beyond the west fence. Land costs about 45 million THB now and 30 million for the option, at the 300,000 THB per rai of a 30-rai listing in Sikhio.[340]

Site plan of the 150-rai plot, a concept render with approximate dimensions, seen from the southeast (north is to the upper right). Open the interactive 3D version to rotate it and read every part.

Hazards. Korat flooded in October 2010,[344] and the Sikhio listing puts its plot at about 260 m above sea level.[340] Water is the main local risk: the Lam Takhong reservoir held 16% in May 2025.[316] The Northeast had 113 days above the PM2.5 standard in 2024, fewer than the North's 129.[321] We found no lightning assessment for the area.

What to check on site before purchase (our list). The title type and zoning color come first; the listing is green-zone farmland.[340] Then come flood levels and drainage; the soil, through a geotechnical survey; and corrosion, through a three-month copper and silver coupon survey that ideally spans the haze season. The investor also needs spare PEA capacity and a second feeder, the actual fiber routes to the fence, and water rights. The distance from schools, temples and hospitals matters, because a ground-mounted array counts as a factory and must keep 100 m from them.[375][376][377] If bi-fuel generators stay an option, the distance to the PTT pipeline matters too.

Building and halls

The building is a single story of 80 by 40 m, with its long axis east to west. A two-story office and operations block of 20 by 15 m stands on its south face by the entrance, with the staff and visitor parking beside it.

The layout is ours, drawn from the design decisions. The loading dock and staging area are at the east end.

Pod sizes follow the hardware on the market. A GB300 NVL72 rack draws 135 kW, up to 155 kW at peak, and sends about 90% of its heat to water.[190] An 8-GPU B300 server draws about 14 to 14.5 kW; NVIDIA puts two in a rack at 30 kW on plain air, or four at 58 kW with active rear-door coolers.[195][176] An air-cooled 8x RTX PRO 6000 server draws 6.8 kW,[20] so two or three fill a 15 to 20 kW rack in hall B. Thai colocation sells racks at 3 kW, with 5 kW as the next step,[390] so hall B is sized for ordinary racks and a few dense ones.

Why one story. Land at C1 is cheap, while racks of 140 kW and their water are heavy. Pipes, busways and cable trays are simpler on one level, and the dock, the halls and the yards meet at grade. Growth means a second building on the plot rather than a floor on top (our judgment). Hall A sits on a slab, with water in floor trenches under leak-sensing cable and with power and fiber overhead, which keeps water and electricity apart (our judgment).

One standard for both halls. The note on hosting and agent servers raised the idea of a Tier III hall for enterprise and GPU customers beside a lean hall in the style of Hetzner, whose units have N+1 UPS on one busbar only and one 2.5 MVA generator each.[487][488] This design instead puts hall B inside the Tier III building on the same 2N power. Customers who bring regulated data and BOI's concurrent-maintainability condition both favor a single grade, and one certificate scope is cheaper to audit (our judgment). A lean hall remains possible later, as a separate building on the plot, if mass hosting takes off.

Building cutaway, concept render: halls A, core and B from west to east, with the electrical rooms along the north wall.

Power

The chain. Power comes from PEA at 22 kV (33 kV where offered), on two feeders from different substations if PEA can supply them, and enters medium-voltage (MV) switchgear. Two 3,150 kVA transformers follow, each able to carry the whole site, and feed low-voltage main switchboards A and B in fire-separated rooms. Each board has a generator section of its own: generator yard A feeds board A and yard B feeds board B. The two sections run in parallel through the A/B bus tie, so the sets still share the load as one 2+1 plant. Lithium-ion UPS sits on both paths (2N), and busways and rack PDUs feed every rack from A and B. CDU pumps and controls are on UPS, so the water keeps moving through a transfer. Solar and the battery connect behind the meter with zero export. A 3.5 MVA site stays well below the 10,000 kVA at which PEA normally connects at 115 kV.[290]

Generators. BOI asks for "an engine generator system with a continuous rating that supports the total power load of the data center" plus a backup set.[389] Uptime accepts only continuous, derated prime or unlimited-runtime standby ratings for Tier III,[682] and Cummins warns that the ISO 8528-1 data-center power (DCP) rating is not appropriate for Tier III because it assumes a reliable grid.[683] The contract therefore asks for the maker's letter on unlimited hours at site conditions (our judgment). The reference set is the MTU 12V4000 DS1650: 1,600 kVA and 1,280 kWe at 400 V, accepting its rated load in one step and compliant with ISO 8528-5 class G3.[684] Phase 1 installs two, one in each yard, and either one carries phase 1's 1.2 MW alone at about 94%. The full build adds a third in yard A (2+1, 2,560 kW for 2.37 MW), and yard B keeps a pad and cabling for a fourth. At the opening load each of two sets would run at about 19%, below the 30% minimum that Cummins advises against wet stacking,[683] so tests run one set on the real load or on load banks (our model and judgment). The engines are specified HVO-ready, as MTU and Kohler approve.[684][690] Bi-fuel gas makes sense only if PTT's pipeline passes near the plot.[702][703]

Two generator yards. The sets stand in two yards, one for each power path. Yard A, at the west end of the building next to electrical room A, holds G1 and the pad for G3; yard B, at the east end next to room B, holds G2 and the reserve pad. With the plant split this way, no single fire, fuel spill, local flood, vehicle accident or exhaust drawn back into the radiators can take out every set, and the separation of the A and B paths that the building already has inside carries on outside. It is one of the Tier IV features that the design takes where they are cheap: Uptime's Tier Standard asks a Tier IV site to isolate complementary systems and their distribution paths physically, so that no single event can affect both.[716] The yards stand at the two ends of the building rather than at the far ends of the 600 m plot, because hundreds of meters of extra cable and longer fuel runs would buy little more protection than the length of the building and a fire wall already give (our judgment). Yard A keeps about 40 m from the battery containers and the nearest solar rows. In the dry-season northeast wind its exhaust drifts toward the field, but the sets run mostly for tests, so soot on the nearest panels is a matter for the cleaning schedule (our judgment). The split adds a second pad, enclosure and bund and a second paralleling section. We have no price for these yet, so the cost model below is unchanged, and the price is among the open questions. What the yards and the dry coolers would sound like at the fence and at the nearest homes, by day, at night and in an outage, is in noise.

Fuel. Up to 15,000 L of diesel needs only a notification to the Department of Energy Business (DOEB), and a tank of up to 2,500 L built into each emergency set does not count.[669][671] Phase 1 holds 20,000 L (two built-in tanks plus 15,000 L of bulk fuel, 7,500 L in each yard's bunded farm), and the full build 22,500 L. That lasts about 64 hours at phase 1's full load, 156 hours at the opening load and 36 hours at 2 MW (our model, at 0.26 L per kWh, a little above the 0.25 L on MTU's sheet). Uptime's minimum is 12 hours at full load.[680] We assume that DOEB counts both farms together as one premises, so the split keeps the bulk total at 15,000 L; whether it does is one of the open questions. DOEB was reported in September 2026 to have suspended new fuel-tank permits for data centers,[393] so both tank farms are laid out for a later license, and a refueling contract is signed before opening (our judgment).

Solar and battery. The plant is 10 MWp of ground-mount panels and a 20 MWh, 4-hour battery behind the meter. Net billing pays only up to 5 kW per meter,[310] so surplus is lost. The site model's hourly dispatch at C1 gives these results (our model):

The rest is bought as UGT2 green power at 4.04 to 4.40 THB/kWh before VAT.[312] Above 1,000 kVA a generating source needs an ERC license, and from 200 kVA a controlled-energy license.[373][374][694] Holding 2 MWh of the battery in reserve gives about 100 minutes of ride-through at 1.2 MW and 50 at 2.4 MW, if its inverters can island the site. That reserve adds to the generators and does not replace them (facility systems).

The ratings are our design and model, with sizes taken from facility systems.

Power quality. PEA's 2024 averages in the Northeast were 1.22 interruptions and 21 minutes per customer,[672] yet power still causes 45% of impactful outages, most often through the UPS.[253] GPU training makes this harder. Uptime reports GB200 NVL72 racks swinging from about 60 to 70 kW to over 150 kW and calls generators the most stressed equipment.[673] In a Vertiv test with STT GDC, 20% input smoothing at the UPS cut generator voltage and frequency swings from about ±10% to ±0.5% (vendor test),[674] and NVIDIA says the GB300 power shelf trims peak demand by up to 30% (vendor claim).[689] The design therefore specifies UPS input smoothing, GPU power smoothing switched on, a 100% load step shown at the factory test, and Class A power-quality meters on each incomer and on the generator bus.[744]

Cooling and air

Hall A: warm water to the chip. Each pod has its own coolant distribution units (CDUs, N+1) feeding cold plates. The facility loop runs to dry coolers on a pad southeast of the building, with trim chillers of about 30% of peak for the hottest hours. The 10 to 15% of heat that still reaches the air goes to rear-door or in-row coils on the chilled-water loop. ASHRAE's W32 and W40 classes are "typically operated without chillers in most locations".[636] A dry cooler leaves water about 5.5 K above the outside air,[286] so with 40 °C facility water the dry coolers work alone while the air is below 34.5 °C. At C1 that limit is passed for about 380 hours a year (4%), against about 3,900 hours (44%) at W32 (our model, from MERRA-2 hourly temperatures, which smooth peaks). The design runs at W40, and accepting W40 water or warmer becomes a purchase condition for GPU systems; the GB300 NVL72 accepts up to 45 °C.[190] NVIDIA's claim that 45 °C water brings warm sites to chillers "just a few days a year" points the same way (vendor claim).[287]

Hall B: air on chilled water. Hall B uses hot-aisle containment with fan walls or in-row coils on chilled water of about 18 to 20 °C, which keeps inlet air in ASHRAE's recommended band of 18 to 27 °C for class A1 equipment.[636]

Air quality. The design uses no outside-air economization, because Korat is hot and humid, has haze, and has no station that measures SO2. Air recirculates, and make-up air passes particle and gas-phase (chemical) filters and keeps the halls at positive pressure. ASHRAE asks for corrosion below 300 Å a month on copper and 200 on silver (level G1), with gas-phase filtration where a site misses that level,[717] and allows 70% humidity only if coupons stay below those rates, otherwise 50%.[636] Haze sampled in northern Thailand was dominated by sulfate and nitrate salts that take up water,[675] and the Northeast has a smoke season of its own,[321] so real-time corrosion monitors watch the intake and the halls,[723] and the coupon survey on the plot comes before the final air-side design.

PUE by phase (our model). For each kW of liquid-cooled IT, the overhead is 0.85 x (0.03 for pumps and fans + trim chillers for 4.3% of the year at half load and COP 5) + 0.15 x (chilled-water air at COP 4.5 + 0.04 for fans) + 0.04 electrical losses + 0.005 controls = 0.114. For each kW of air-cooled IT it is 1/4.5 + 0.04 + 0.04 + 0.005 = 0.307. A fixed 30 kW covers offices, lighting and security. The results:

Every case is below BOI's 1.3, but the opening year comes close, because a half-empty hall carries the fixed load and the air-cooled share is high. PUE is reported to ISO/IEC 30134-2[274] and measured at the IT input every 15 minutes or less, which is The Green Grid's Level 3.[749]

Water. Closed loops and dry coolers evaporate nothing. NVIDIA puts cooling towers at roughly 2.6 million US gallons per MW a year (vendor figure),[287] which this design avoids. Rainwater collected in a pond feeds the fire tank and domestic use, which also answers BOI's water-plan condition.[389] The return water leaves at about 45 °C. Turning it into cooling, as LRZ did with adsorption chillers,[289] or finding a use for it on or near the plot, such as drying, is an idea to test rather than a plan.

Fire, physical security and safety

BOI asks for fire protection over the entire area and a 24-hour security system.[389] The design below is our judgment, after the practice found in the evidence.

• Detection. Aspirating smoke detection covers every space, with each VESDA-E unit serving up to 2,000 m².[741] ITSC and Hetzner both use aspirating detection.[15][489]
• Suppression. The halls get pre-action (double-interlock) sprinklers or water mist, so a broken head alone releases no water. Clean-agent gas protects the electrical, UPS-battery and tape rooms. Electrical rooms A and B, the tape room and each meet-me room are separate fire compartments, and rainwater feeds a fire-water tank.
• Batteries. The lithium containers stand apart from each other and from the building, each with its own detection and suppression.
• Physical security. A perimeter fence and gatehouse and a vehicle barrier guard the site. CCTV covers the fence line, the yards and every door. Card and biometric access, with a mantrap, leads into the halls, and colocation customers in hall B get locked cages. Visitors are escorted, and a two-person rule applies to the tape vault.
• Safety. High-voltage switching needs permits, every isolation is locked out and tagged, switchboards carry arc-flash labels, and written procedures cover the fuel farms and work on live water loops.

EN 50600 grades physical protection by subject (access, fire, environment).[263] The target for the white space, taken from option B, is Protection Class 3 to 4.

Network and connectivity

Inside. Four separate networks run inside the building (our design, after open stack):

• The front end carries customer traffic, APIs, hosting and agents on a 100/400G leaf-spine of white-box switches with SONiC, which runs on about 156 switch models from about 27 vendors.[391] Each tenant gets its own routing table (a VRF) from leaf to hypervisor with EVPN-VXLAN.[581][582][583]
• The storage network uses 25/100G leaves for Ceph.
• The GPU back end is 800G Ethernet with RoCE on SONiC. A 64x800G white box lists at USD 46,800 (1.54 million THB) and NVIDIA's SN5600 at USD 74,457 (2.46 million THB); the SN5600 stays the alternative.[509][510]
• The out-of-band and operational-technology (OT) networks are a 1G management network for BMCs and PDUs and a physically separate network for facility controllers and sensors.

Edge. Two routers at the site and two at the point of presence run NixOS with FRRouting for BGP and nftables for filtering. FastNetMon Community detects floods within seconds and diverts or blackholes them over BGP;[514] floods larger than the links go to upstream scrubbing, which True IDC sells in Thailand.[515] Received routes are checked against RPKI, and the investor's own ROAs are signed in MyAPNIC.[797] OPNsense, which moves about 21 Gbps port to port on Deciso's largest 1U box, guards only the admin VPN and bastions.[513]

To Bangkok. Two 100G backhauls on separate rights of way enter through the two meet-me rooms on opposite sides of the building. One could be an EGAT wavelength (EGAT sells 10 to 400 Gbps at more than 50 substation access points) and the other a railway or road carrier (Interlink has fiber on 3,244 km of State Railway poles).[761][778] Both end at a one-rack point of presence at Telehouse Bangkok, home to TH-IX, BKNIX, BBIX, JTIX, TRUE-IX and AMS-IX Bangkok.[756] Fiber adds about 4.9 µs per km,[826] so Sikhio is about 2.2 ms round trip from Bangkok over 224 rail km (our arithmetic). Two transit carriers with international gateways bring at least 10 Gbps each to the site, and together with the backhauls they meet BOI's condition of four systems and 60 Gbps.[389] CMU pays NT about 133,000 THB a month for 20 Gbps, a useful anchor of about 6,650 THB per Gbps.[758][772]

The investor's own AS and peering. The investor joins APNIC directly: a /23 of IPv4 with a /32 of IPv6 costs AUD 500 once and AUD 2,256 a year.[755] Peering starts at TH-IX, where Google, Microsoft and Cloudflare peer, and at BKNIX, which adds Meta, Amazon and ByteDance; BBIX comes third.[769] A 100G port at BKNIX costs 150,000 THB a month, and a 10G port 30,000.[757] Selling connectivity to tenants needs an NBTC Type 1 license, which law firms read as no longer needing a foreign-business permit since August 2026.[767][802]

A link to ITSC. BOI's cloud-service promotion needs two ISO 27001 data centers in Thailand linked at 10 Gbps or more,[250] and Germany's BSI wants backup sites at least 100 km apart;[520] Chiang Mai is about 500 km away. ITSC bans private circuits in its room,[14] so the link needs an exception from ITSC, or the second site moves to a colocation facility in Bangkok (see connectivity and option A).

The software stack

The selection is ours, drawn from open stack and hosting and agent servers.

Some parts stay closed: CUDA and NVIDIA's user-space driver, the firmware of GPUs, drives and tape, and the certified controllers. The open monitoring stack observes the controllers but never closes their loops. NVIDIA AI Enterprise is optional at USD 4,500 per GPU a year;[464] for the 192 GPUs of the reference fit-out it would cost about 28.5 million THB a year, which the open stack avoids (our arithmetic). Agents reach tools over MCP inside each tenant's zone and reach models only through the gateway, so by default nothing leaves the building (hosting and agent servers).

Telemetry

The sensor plan and its bill of materials are in facility systems. The priced items come to about 0.75 million THB per MW-sized hall, and integration will cost more. Two collectors, A and B, on the OT network poll everything into one time-series store, so monitoring is itself concurrently maintainable, as Tier III requires of controls.[716]

B4L's river-monitoring service taught one rule worth copying: when a feed goes silent, the silence is an alarm (facility systems).

Operations and staffing

Shifts. Two seats are staffed around the clock, one for the facility and one for IT and network. One 24/7 seat needs 4.7 to 5.2 people once leave and training are counted,[483] so each seat runs on five crews. Day staff add engineering, platform and customer work. BOI asks for Thai nationals in at least half of executive and professional posts within three years.[304]

The loaded monthly costs are our estimates. For comparison, a Thai managed-services firm puts the cost of hiring an administrator at 50,000 to 200,000 THB a month.[412]

Runbooks and change control. Nearly 40% of organizations had a major outage caused by human error in three years, and 85% of those came from procedures that were flawed or not followed.[484] Every planned task gets a written method of procedure, reviewed by a change board and carried out with a second person. Standard procedures cover routine work, and emergency procedures cover each P1 alarm. All three kinds live in the same repository as the NixOS configuration (our design).

The path to certification. Uptime's Tier III certification comes in steps: the design documents before construction, then the constructed facility, then operational sustainability once the team has a year of records.[260] ISO/IEC 27001 comes first among the ISO standards, because BOI requires it before the tax exemption.[389] ISO/IEC 20000-1 follows for BOI's cloud activity,[250] CSA STAR and ISO/IEC 27701 for government and personal-data customers,[476][425] ISO 50001 and 14001 for energy and environment,[267][268] and ISO/IEC 42001 for AI management later.[477]

Cost

All figures come from model/first_dc.py and are our estimates; model/first_dc_results.md lists the inputs and their sources. Building and MEP (mechanical, electrical and plumbing) use option B's 321 to 508 million THB per MW of IT, which runs from Thai land and construction plus 7% for liquid cooling up to a US contractor's Tier III top plus 10%.[17][299][26] That rate is applied to 2 MW, less the generator plant, which is priced on its own. Phase 1 carries 70% of it, because it builds the whole shell and the electrical backbone (our split).

The table is in M THB. Without land, solar and battery, the full build comes to 359 to 595 million THB per MW of IT. That is a little above option B's range, because it adds the grid connection, design, certification, network and storage.

Reference IT fit-out for phase 1. The table below is an example only. The real fit-out will be sized to demand later, and nothing in it is a purchase decision.

Twenty colocation racks in hall B hold customers' own servers. The study's staged path still holds: the ITSC pilot comes first, and B300-class servers are bought only when the evidence supports them (hardware).

Opex per year (M THB).

What the numbers say. The low grid bill does not mean free power, because the capital for the solar and battery sits in capex. Counting their capital charges at 8%, as the site model does, energy costs 3,835 to 4,642 THB per MWh at phase 1's full use and 3,875 to 4,691 at full build, against 3,810 from the grid alone. At today's tariff, then, solar is a hedge and a green claim rather than a saving. If the separate data-center tariff of 5 to 6 THB/kWh is adopted,[134] the grid share costs only 2.2 to 2.6 million THB a year at phase 1, where an all-grid site would pay 52.6 to 63.2 million. At the opening load the 10 MWp plant is about three times too big: 69% of its output is lost, and energy costs 9,000 to 11,000 THB per MWh, while 2.5 MWp with 5 MWh would cover 73% at 3,270 to 3,763 (our model). Staff and connectivity dominate opex, and power is a small line. The calculator at ../calculator/ on the study site turns these costs into revenue, ROI and payback for each business mix.

Schedule

Critical path. The critical path runs through the grid connection and the long-lead electrical plant: transformers, MV switchgear and generators. Large generators take 18 to 24 months for units above 2.5 MW,[668] and 48% of Turner & Townsend's respondents name power availability as the main obstacle to on-time delivery.[300] A US contractor puts a Tier III build at 16 to 24 months.[26] Permits run in parallel but can stop the job: the new national rules, the fuel notification and the ERC licenses.

Risks and how the design answers them

What this means for the investor

What follows is our reading, and it is judgment:

• The design qualifies for BOI and can be certified. A 2 MW Tier III building with a PUE of 1.2 on a Korat site meets every condition we found, with margin on PUE except in the first, half-empty year.
• The building is the big expense, and the IT can wait. Phase 1 costs 0.86 to 1.27 billion THB before any GPU, and solar and battery are about a third of that. At today's tariff that plant buys price certainty and green power rather than cheaper power.
• Consider buying solar in tranches. At the reference opening load most of phase 1's 10 MWp is wasted. Matching the solar and battery to committed load keeps the design intact and saves capital early.
• Staff and connectivity are the running costs to manage. Together they are about 60% of phase-1 opex at full use, while grid power after solar is under a tenth.
• Commit to C1 only after the checks. Water rights, a PEA answer on capacity, two fiber routes and a clean corrosion survey decide the plot, and C4 remains a ready fallback.
• Move in the first application window. The 2027 reopening and the fuel rules set the calendar more than construction does.

Open questions

The sourcing note names the companies in Thailand that sell, build or service each main component, with Thai factories and references.