Data Center Foundations for the Developer and Investor Seat

Why this page exists

I can read a Thai data center headline and understand every word of it and still not know whether the deal is good. "300 MW campus, THB 72.7 billion, Rayong" tells me nothing until I know whether 300 MW is IT load or grid capacity, whether the power is confirmed or applied for, what the tenant pays per kilowatt per month, and who eats the loss if the substation slips two years. This page builds that vocabulary from the physics up, then attaches the Thai regulatory and commercial facts to it, so that the next headline decomposes into numbers I can underwrite.

The order is deliberate. Power first, because everything downstream is a consequence of it. Then heat, because heat is just the power arriving somewhere you did not want it. Then redundancy, because that is what the tenant is actually paying a premium for. Then the development sequence, the four product shapes, and the deal itself.

Where Thai figures appear, they are sourced. Where the engineering is universal, it is my own and I say so. Where I have done arithmetic on published numbers, I flag it as mine.

Capacity is priced in megawatts because the building is a wrapper around an electrical service

Every other kind of real estate is sold by area because area is what the tenant consumes. A warehouse tenant fills a floor plate. A data center tenant fills a switchboard.

Three things follow from that. The scarce input is electrical capacity, not land. The cost is concentrated in electrical and mechanical plant rather than in the shell, so the concrete is the cheap part of the build. And density keeps moving, which means the same floor area can host several times the load it hosted a decade ago, so square meters have no stable relationship to revenue. A hall built for 5 kW racks and a hall built for 120 kW racks can be the same size and differ by an order of magnitude in what they earn and what they cost.

The utility reinforces this. Grid allocation is granted in megawatts, applications are counted in megawatts, and Thailand's new capacity guarantee is levied per megawatt at THB 4.5 million of reserved capacity (Bangkok Post, 17 April 2026). When the state, the lender and the tenant all count in the same unit, the developer has no choice.

Square meters still matter for exactly one question, which is whether the plant physically fits. Generator yard, fuel tanks, chiller plant, transformer bays, cooling towers, loading, setbacks and fire separation all consume land that earns nothing. On a constrained urban plot the MEP yard is what stops you, not the white space. That is why the interesting Thai sites are 50 to 100 rai of industrial estate land rather than city blocks.

IT load is the number that matters, facility load is the number you pay for, and PUE is the ratio between them

IT load is the power delivered to computing equipment, conventionally metered at the output of the power distribution unit or at the rack inlet. Facility load is everything the site draws at the utility meter, which is IT load plus cooling, plus lighting, plus office and security, plus every conversion loss between the boundary and the server.

PUE is facility load divided by IT load. A PUE of 1.30 means that for every kilowatt reaching the servers, another 0.30 kW is spent getting it there and taking the heat away. It is a ratio, so it flatters nothing on its own. A site with a magnificent PUE and terrible utilization is still a bad business.

There is a third number, contracted capacity, which is what the tenant reserves and pays for, and a fourth, actual draw, which is what the tenant uses. The gap between them is the diversity factor, and it is one of the most important quantities in the industry. Tenants routinely draw well under their contracted kilowatts, so operators oversubscribe their utility connection against that behavior and earn on capacity that never materializes as load.

That same gap is what regulators now treat as a problem rather than a margin. Malaysia's grievance is explicit, with 603 MW of actual demand against 1,276 MW of declared maximum capacity, roughly 47% against an 85% target (TechRepublic). Thailand's guarantee is keyed to exactly this, with half the THB 4.5m per MW returned when the project reaches 50% of proposed consumption within one year and the balance at 70% (Bangkok Post). Read plainly, the Thai state has priced the tenant's diversity factor and handed the bill to the developer.

How PUE is gamed

I have never seen a marketing PUE I trusted without asking six questions.

Where is the measurement point. If IT load is metered at the rack inlet rather than at the UPS output, then the server power supply and voltage regulator losses sit inside the denominator and inflate it, which pushes the ratio down without a single watt being saved. Move the boundary, improve the number.

Design or actual. Design PUE is a calculation at full load and favorable ambient conditions. Annualized measured PUE on a partly filled hall is a different animal, often much worse, because chillers, pumps and fans have substantial fixed draw regardless of how much IT is running.

Which season. Quote the coolest month and the number improves. In Bangkok this matters less than elsewhere, because the wet bulb temperature stays high year round and free cooling hours are close to nil, which is precisely why Thai PUE is structurally harder to achieve than in Tokyo or Sydney.

What is excluded. Site substation losses, office space, the security building and the fuel polishing plant can all be quietly left out of the numerator.

Whether water was traded for power. Evaporative cooling lowers PUE and raises water consumption. Dry coolers do the reverse. A PUE improvement can be a WUE deterioration in disguise, which is the tension I take up below.

What load fraction. A hall at 30% occupancy will not hit its design PUE and nobody quotes the 30% number.

Thailand has now turned PUE into a tax variable, which sharpens all of this. The BOI's higher-incentive data center category carries an eight-year corporate income tax exemption and a requirement of PUE no worse than 1.3, against five years for the standard category (Formichella & Sritawat, Silk Legal). Note the counterintuitive twist that both tiers drop one grade inside the EEC, so a compliant project outside the corridor gets better treatment than the same project inside it. Sources disagree on whether the governing notification is dated 5 June 2025 or 5 June 2026; Silk Legal and LexBangkok give 2025, Mahanakorn Partners gives 2026. Once a ratio determines a tax outcome, the incentive to move the measurement boundary stops being cosmetic.

Every conversion between the transmission line and the chip takes a cut, and the losses become heat you then pay to remove

The chain, in order, and what each piece is for.

Transmission line. EGAT owns and operates Thailand's high voltage network and is its sole transmission licensee; MEA distributes in Bangkok, Nonthaburi and Samut Prakan, and PEA covers the other 74 provinces (Tilleke & Gibbins). The connection voltage class you land at determines your tariff and your wheeling charge, and the ERC's draft third party access table distinguishes 69 kV and above, medium voltage of 22 to 33 kV in PEA territory and 12 to 24 kV in MEA territory, and low voltage (ERC hearing 582).

Utility substation and the customer substation. Step down from transmission to medium voltage, then a customer-owned substation on site. For a large campus this is a serious asset in its own right; Fuji Electric's Thai subsidiary is building the campus substation for NTT's BKK4, its first such project in Thailand (RCR).

Medium voltage switchgear and ring main. Sectionalizing, protection and the ability to isolate a fault without dropping the site.

Unit substation transformers. Step medium voltage down to 400/230 V for distribution inside the building.

Low voltage switchgear.

UPS. Static double conversion is the norm. It rectifies incoming AC to DC, floats the battery or flywheel on that DC bus, then inverts back to clean AC, so the load never sees a transfer. This is where you buy ride-through, typically enough for the generators to start and take load.

Generators. Diesel gensets sized to the full facility load, held in N+1 or 2N, with fuel storage on site. In Thailand this is about to become a regulated question rather than an engineering one, because authorities are considering charging operators when demand spikes force utilities to deploy emergency generation, and considering mandating local energy storage at heavy load facilities (Eco-Business).

PDU and remote power panel. Distributes UPS output to the floor, sometimes with a further step-down transformer.

Busway. Overhead bus running the length of the row, tapped by plug-in units at each rack. Busway replaced the old under-floor cable spaghetti because it lets you change rack density without rewiring.

Rack PDU. The vertical strip in the cabinet with the outlets and the per-outlet metering.

Server power supply unit. Converts AC to a 12 V or increasingly a 48 V DC bus inside the chassis.

Voltage regulator modules. On the board, converting 12 or 48 V down to well under a volt at hundreds of amps for the processor or GPU. This is the last and often the least efficient conversion in the chain.

As an engineering rule of thumb, and this is my own knowledge rather than a sourced figure, transformers run in the high nineties percent, double conversion UPS in the mid nineties, a good server PSU in the mid nineties, and board level regulation somewhat worse than that. Compounded, appreciably less of the power entering the fence reaches the silicon than the headline megawatt number suggests. Two consequences matter commercially. First, every lost watt appears as heat inside the building, so electrical inefficiency is charged twice, once on the meter and again through the chiller. Second, where the industry draws the IT load boundary determines which of these losses count against PUE and which hide inside it.

For Thailand there is a specific reason to care about the conversion stack right now. The proposed Type 9 tariff is a marginal cost pass-through indexed to imported LNG rather than a fixed rate. Energy Minister Ekkanat Prompan said on 28 August 2026 that with spot LNG at USD 23 per MMBtu the fuel component alone exceeds THB 5 per kWh (Thansettakij). The widely quoted THB 5 to 6 per kWh band traces to an unnamed Ministry of Energy source published a day before the NEPC met (Thansettakij, 14 July 2026); the ERC has published no rate, no megawatt threshold and no energy-versus-demand split, and Bangkok Post reports an expected Q4 2026 effect (13 August 2026). For reference, data centers currently pay the general industrial rate of a little over THB 4 per kWh, and the average billed tariff for September to December 2026 was set at roughly THB 3.86 to 3.89 per kWh after the ERC stripped public street lighting out of the base tariff (sources give both 3.86 and 3.89; Thansettakij, Nation Thailand). Underwriting a fixed Thai power price today is underwriting something that does not exist.

Heat rejection is the same energy leaving, and above roughly 120 kW a rack the air simply cannot carry it

Essentially all electrical power delivered to a data hall becomes heat. The cooling system's only job is to move that heat from the chip to the outdoors.

The classical chain runs from the chip through a heat sink into room air, from room air across the coil of a computer room air handler, from that coil into a chilled water loop, from chilled water into the evaporator of a chiller, out of the chiller's condenser into a condenser water loop, and from the cooling tower into the atmosphere by evaporating water. Alternatives exist. Air-cooled chillers reject straight to ambient air and use no water but consume more power. Economizers use outside air or a water side free cooling loop when ambient conditions allow, which is a large saving in temperate climates and close to worthless in Bangkok, where the wet bulb stays high all year. That climate fact, not any regulation, is why Thai PUE is structurally harder than Japanese or Australian PUE.

Why air runs out at around 120 kW a rack

This is arithmetic anyone can do. Air has a specific heat of roughly 1 kJ per kilogram per kelvin. To carry 120 kW at a 12 K temperature rise across the rack you need about 10 kilograms of air per second, which at data hall density is on the order of 8 to 9 cubic meters per second, or roughly 18,000 cubic feet per minute, through a single cabinet. That is a gale. Fan power scales with roughly the cube of flow, so pushing toward that figure means the fans start consuming a meaningful fraction of the load they are cooling, and the acoustics and the pressure differentials become unmanageable. The wall is not a standard, it is physics.

The progression in practice, from my own experience of the equipment rather than any published table, runs from legacy halls at a few kilowatts a rack, to hot and cold aisle containment in the low tens, to rear door heat exchangers taking water to the back of the cabinet in the tens, and then to liquid at the chip.

Direct to chip puts a cold plate on the processor and pumps coolant through it, usually via a coolant distribution unit that isolates the facility water loop from the technology loop. It captures most but not all of the rack heat, so you still need air handling for memory, drives and network gear. It is the mainstream answer for current AI racks and it is what Thai operators are deploying; GULF's GSA01 in Samut Prakan is described as the first in Thailand with liquid cooling for AI chips (Gulf), and DayOne has liquid cooling at Chonburi Tech Park.

Immersion submerges whole boards in dielectric fluid, single phase or two phase. It captures nearly all the heat, allows extreme density, and eliminates server fans. It also breaks warranty and serviceability norms, needs a different floor structure to carry the tank weight, and creates a fluid supply chain dependency. It remains a minority approach.

The underwriting point is that these are not interchangeable retrofits. A hall commissioned for air at 10 kW a rack cannot take 120 kW racks by buying more CRAHs. You need different floor loading, different pipework, different power density per square meter and different commissioning. When I look at an existing Thai asset, the question is not what it cools today but what it can be converted to cool, and at what cost.

Water, and why WUE is the politically dangerous ratio

WUE is water consumed per kilowatt hour of IT energy. It is contested where PUE is not, for three reasons.

First, PUE and WUE trade against each other. Evaporative cooling is thermodynamically efficient and improves PUE while consuming water. Dry cooling saves water and worsens PUE. Thailand has tied its eight-year tax exemption to PUE ≤ 1.3, which is a direct incentive to evaporate, at the same moment its screening regime demands water management plans. The tax code and the permit are pulling in opposite directions.

Second, water is local in a way electricity is not. A grid megawatt is fungible. A cubic meter drawn from Rayong's raw water system is taken from a specific catchment that specific people also use. EEC water demand was 658 million cubic meters in 2022 and is projected at 800 million by 2027 and 1 billion by 2036, with Rayong's waterworks already running at about 120,000 cubic meters a day against 151,200 of capacity (Mongabay).

Third, the comparisons are devastating and easy to make. TDRI estimates a 100 MW data center consumes water equivalent to the annual use of 1.3 million people and electricity equivalent to 13 million, while creating about 50 direct jobs (Thailand Construction). WHAUP puts data center water use at 12 to 16 times that of a typical industrial user (Nation Thailand).

The single most useful Thai datapoint is Bridge Data Centres' QHI01 at Khlong Tamru, Chonburi, a 0.2 GW site with a water agreement for 3.3 million cubic meters a year, or 9,000 a day, equivalent to the annual use of 36,900 residents, while contractors cite 4.38 million cubic meters (Mongabay). The two figures are 33% apart on the same facility, which tells you how soft this disclosure is. My arithmetic on the lower figure, assuming the 200 MW runs near continuously, gives roughly 1.9 litres per kWh, which is in the range I would expect for an evaporatively cooled site and is a reasonable sanity check on any Thai proponent's water plan.

Thai industry itself ranks water above power. In a Federation of Thai Industries survey of 160 executives, 68.8% flagged water supply as the leading concern against 61.3% for electricity demand (Thailand Construction). Water management is now a formal BOI screening dimension, and the NEPC's July 2026 package requires large data centers to submit water management plans before the state expands the grid for them (EPPO). I would treat water rights in Chonburi and Rayong as underpriced diligence relative to the attention power gets.

Redundancy grammar describes topology, and what the tenant buys is a failure domain, not a guarantee

N is exactly the capacity the load requires. N+1 adds one spare unit per system, so any single unit can fail or be serviced without loss. 2N is two complete, independent systems, each capable of the whole load. 2(N+1) is two systems each with its own spare. Distributed redundancy shares a common spare across several systems rather than duplicating each one, which is cheaper per MW and harder to operate.

The Uptime Institute tiers layer intent on top of that topology. Tier I is basic capacity with no redundancy. Tier II adds redundant capacity components but a single distribution path. Tier III is concurrently maintainable, meaning every capacity component and every distribution path can be removed from service for planned work without dropping the load; a single unplanned fault can still take the site down. Tier IV is fault tolerant, meaning the site survives a single unplanned failure anywhere, with physically compartmentalized systems and continuous cooling.

Three things are commonly misunderstood. Tier is a property of the topology and, in the certified case, of what was actually built, not of the operator's competence. Tier is not an availability percentage; the old table of nines was withdrawn for good reason. And nobody outside a certification process should say "Tier III" when they mean "we think it is roughly Tier III."

What the tenant is actually buying, then, is three things. A topology that determines what maintenance can happen without them noticing. A contractual SLA whose remedy is service credits, normally capped at a fraction of the monthly fee, which is compensation and not insurance against their own outage costs. And a set of failure domains, meaning the A and B feeds they dual cord into, the separate rooms, the separate risers.

The commercial insight follows. A tenant whose application already replicates across sites is paying twice if it also buys 2N inside one building. That is exactly why hyperscale self-builds frequently commission to N or N+1 and put resilience in software across availability zones, and why Tier III and IV topologies are priced for the enterprise and government tenants who cannot fail over. When someone tells me a Thai site is Tier III certified, my question is who the tenant is, because that determines whether the premium is a feature or a cost.

Thailand has quietly written this into its tax law. The BOI's promoted data center conditions include concurrently maintainable systems, continuous rated generators, UPS and cooling redundancy, independent distribution paths, fire protection, 24 hour security and ISO/IEC 27001 certification (LexBangkok, Silk Legal). ISO 27001 is a condition precedent to exercising the exemption, and revenue earned while it is pending is taxed at 20%, with non-compliance exposing the promotee to withdrawal and retroactive tax (Fangda). A concurrently maintainable topology is no longer just an engineering choice in Thailand; it is a tax condition.

The development sequence is a power sequence with land, fiber and permits attached

Written as the calendar actually runs, rather than as the brochure presents it.

Land under option. You option, you do not buy. Zoning and town planning, flood exposure, soil bearing, distance to residential, and enough area for the MEP yard. Thai brokers report investors seeking 50 to 100 rai for projects up to 500 MW, with site selection due diligence running more than a year and power availability outweighing land cost as the primary obstacle (Bangkok Post). Provincial caveats matter. Savills reports Chachoengsao constrained on both electricity and water, Chonburi similarly utility-constrained despite being preferred, and parts of Ayutthaya designated flood retention areas. The Ministry of Energy has separately named Bangkok, Samut Prakan, Prachinburi, Ayutthaya and Saraburi as having transmission headroom outside the EEC (Nation Thailand), which sits awkwardly with the broker view on Ayutthaya. Transmission headroom and site developability are different tests and both must pass.

Power. This is the gate, and Thailand has recently inverted its order. From 30 March 2026 a new data center project must obtain written ERC confirmation that national supply is sufficient for its proposed load, and BOI Notification Sor. 2/2569 makes a power-readiness confirmation letter from the ERC Office a prerequisite for investment promotion (Eco-Business, Formichella & Sritawat). Note the circularity that Direct PPA eligibility in turn requires BOI promotion. Then comes the guarantee, THB 4.5 million per MW of reserved capacity, refunded half at 50% consumption within a year and fully at 70%, with allocated capacity revocable if commercial operation is not demonstrated within five to seven years. On a 100 MW project that is THB 450 million. It is a bank guarantee rather than cash, so it is a fee-bearing contingent liability that consumes bank lines rather than an outright outflow, but no implementing regulation carrying the figure had been published as of 30 August 2026 (Bangkok Post, Mahanakorn Partners).

Fiber. Dual diverse routes, distance to a carrier hotel and to the internet exchanges, and access to subsea capacity. Thailand's landings sit at Si Racha, Petchaburi, Songkhla, Satun and Chumphon, and only the Malaysia-Cambodia-Thailand system lands in Rayong (Submarine Networks). The EEC power cluster is therefore not itself a major landing point and depends on terrestrial backhaul. The NBTC approved the Google-backed TalayLink conduit on 27 April 2026, a 300 km conduit able to house eleven cables landing in the Songkhla and Satun area, with the stated rationale of reducing reliance on Singapore, since roughly 70% of Thai international traffic currently routes terrestrially through Malaysia (Submarine Networks). Latency diversity and power headroom therefore sit at opposite ends of the country.

Water allocation, including a dry season plan, now a screening criterion in its own right.

Permits. EIA where required, factory operation licence, building permit, and if the site sits inside an IEAT estate then the land use permit and business commencement notification. If you self-generate at 1,000 kVA or more you need an ERC generation licence, and the ERC must test the application against the Power Development Plan (JETRO/ERC licensing regulation). That last point is a live risk, because PDP 2026 remains a draft, with public hearing scheduled for 8 September 2026 and proclamation targeted within 2026 (Thansettakij).

Anchor tenant. A signed lease, or at minimum a term sheet with a ramp schedule, before notice to proceed. In practice power and anchor race each other, because the tenant will not commit without a power date and the utility will not confirm without a committed customer. Thailand's guarantee is best understood as an attempt to break that standoff by pricing the option to hold capacity.

Notice to proceed and long lead procurement. Transformers, medium voltage switchgear, chillers and generators now carry procurement lead times that can determine the schedule more than construction does.

Financial close. Which is where the lender's security package meets the regulator, and in Thailand that collision is not benign, as set out below.

Four products, and the difference between them is which risks the landlord agreed to eat

Powered shell. The developer delivers land, entitlements, the building envelope and power to the boundary or to a switchroom. The tenant fits out its own mechanical and electrical plant. The landlord carries entitlement risk, power delivery date risk and shell construction cost. The tenant carries MEP capex, technology choice and PUE performance. Returns are the lowest of the four and the risk profile is closest to industrial real estate.

Turnkey wholesale colocation. The landlord delivers commissioned white space with power and cooling to the rack position and sells committed kilowatts of critical capacity. The landlord now owns MEP capex, the availability SLA, and typically a PUE commitment; power is usually passed through at or near cost. The tenant pays base rent on reserved capacity whether or not it draws it. This is the dominant shape for hyperscale demand in Southeast Asia and it is what most announced Thai projects are.

Retail colocation. Cabinets, cages and cross-connects on shorter terms at a much higher price per kilowatt. The landlord carries churn, sales cost and the burden of keeping an interconnection ecosystem alive, which is also the moat. Telehouse Bangkok, with 9.5 MVA at Rama 9 and an AWS Direct Connect location, is a connectivity play rather than a hyperscale one (Telehouse).

Hyperscale self-build. The operator buys land and power directly, builds to its own reference design, commissions at lower physical redundancy because it replicates in software, and funds on balance sheet or through joint ventures. Cheapest per MW at scale, and it converts a lease obligation into an owned asset, but it consumes the operator's capital and imports local execution risk.

The Thai reading is that the market is overwhelmingly wholesale. Microsoft is explicit that it is relying on partnerships with Thai and global companies to develop and operate its new cloud region, with True IDC the named landlord (Light Reading). AWS committed more than USD 5 billion and Google USD 1 billion without disclosing owned megawatts. So the developer carries construction and power risk in exchange for long contracted cash flow, if and only if it lands the anchor.

That "if" is undifferentiated in Thailand on public information, because outside Microsoft to True IDC no anchor tenant is disclosed for DayOne, Bridge, Beijing Haoyang, Digital Edge, Evolution, Empyrion, GSA, Skyline or NTT BKK4. Evolution's disclosure that Phase 1 of TH01 is sold is the only other public demand signal I found.

The Thai field, and why the megawatt numbers are not comparable

The table below exists because the capacity label, not the capacity number, is the information.

Operator	Site	Stated capacity	IT or gross?	Owner / backer
DayOne	Chonburi Tech Park, Amata City Chonburi	180 MW grid, over 100 MW IT, 300 MW after CTP1 expansion	Both stated separately	Coatue and Hillhouse largest holders; INA joined Series C
True IDC	EEC hyperscale campus, Chonburi	250 MW, THB 77bn	Not stated	CP Group; GIP/BlackRock partnership, stake undisclosed
GSA Data Center	GSA01 Bang Phli, Samut Prakan	25.6 MW, operational	Not stated	GULF 40%, Singtel 35%, AIS 25%
NTT GDC	BKK4, Chonburi	100 MW supply agreement, energization Q2 2027	Grid supply	NTT; power from B.Grimm Power
STT GDC Thailand	Bangkok 1 and 2	22 MW live plus 24 MW, 46 MW campus	IT load, stated explicitly	JV with Frasers Property; platform now KKR 75% / Singtel 25%
Bridge Data Centres	QHI01, Khlong Tamru, Chonburi	up to 200 MW at full build; BOI tranche 134 MW	Not stated	Bain Capital, stake sale process live
Beijing Haoyang	WHA Eastern Seaboard IE 4, Rayong	300 MW, THB 72.67bn	Not stated	Beijing Haoyang Cloud & Data
Digital Edge B.Grimm	BKK1, Chonburi EEC	100 MW, RFS Q4 2026	IT load, stated	Digital Edge JV with B.Grimm Power, BGRIM reported at 40%
Empyrion Digital	TH1, Bang Na, Bangkok	20 MW, go-live Q3 2027	IT load, stated	Seraya Partners
Evolution Data Centres	TH01 Bang Na, TH02 Samut Prakan	12 MW sold plus 40 MW; 200 MW	Not stated	Central Pattana partner on TH01

Sources for the rows above are the operators' own releases and BOI announcements: DayOne, True IDC, Gulf, B.Grimm Power, STT GDC, BOI, WHA, Digital Edge, Empyrion, Evolution.

Only four of the ten label their megawatts. Any capacity stack built by adding this column is not comparing like with like, and the difference is not small. DayOne's own disclosure of 180 MW grid against over 100 MW IT implies a ratio near 1.7, which is what you get once you add PUE and redundancy headroom on top of IT load. Compute cost per MW on a gross number and you will flatter your own deal by a wide margin.

The market-level capacity numbers are equally unsettled. BMI puts Thai live IT capacity at 216 MW in Q2 2026 with 944 MW under construction or planned; Fangda gives roughly 350 MW operational as of 2024 against a 2.87 GW pipeline; Knight Frank gives Bangkok alone at 122 MW at end-2025 rising to a forecast 402 MW by 2027 (TechNode on BMI, Fangda, TechNode on Knight Frank). They are probably measuring different things, carrier-neutral colocation against everything including enterprise self-build, but no source reconciles them and I would not underwrite off any one of them.

One more market fact deserves emphasis because it cuts against the scarcity narrative. Bangkok colocation vacancy is 23.3%, against Johor at 0.7%, Singapore at 4.9% and Jakarta at 20.5% (Knight Frank Data Centre Atlas 2026). Thailand is simultaneously power-rationed at the hyperscale end and oversupplied at the leasing end. Those are two markets, and conflating them is how speculative colocation gets built into a vacancy.

The deal

Cost per MW, and getting the denominator right

Cushman & Wakefield's 2026 construction cost guide puts Thailand at USD 7.0 million per MW low, USD 8.8 million mid and USD 10.5 million high, up 14.7% year on year on a mid-specification build, eighth of fourteen Asia Pacific markets (C&W). The peer set on the mid figure runs Vietnam at 7.2, Indonesia 8.3, Philippines 8.2, Malaysia 9.6, Australia 10.0, Korea 10.6, Singapore 14.4 and Japan 16.0. Thailand is mid-tier and no longer a low-cost outlier; its 14.7% inflation is well above the regional average of about 10%.

Two cautions. That is a mid-specification build, so an AI-density facility with liquid cooling should not be underwritten at the mid. And it is a construction index, with land reported separately, so it is not an all-in development number.

Thai BOI-registered project values bracket the C&W anchor rather than contradicting it. My arithmetic on the January 2026 approvals gives True IDC at THB 45.3bn for 223 MW, about USD 6.5 million per MW; GSA Data Center 05 at USD 1.18bn for 120 MW of IT load, about USD 9.8 million; and Stellar DC at USD 300m for 25 MW, about USD 12 million (TechNode). BOI registered capital is not project cost and its treatment of land and IT equipment is unclear, so use these as a sanity range and not as comps.

Leases, tenor and the rent nobody publishes

The hyperscale norm, and every published data point I could find on it is American, runs to a 15-year base term with two or three five-year extensions, escalators of 2 to 4% fixed or CPI-linked, take-or-pay for AI anchors, revenue split roughly 70 to 80% base rent on reserved capacity and 15 to 20% metered power pass-through, PUE caps typically 1.3 to 1.4, ramp tranches with hard dates, and letters of credit or parent guarantees for non-investment-grade tenants (Build.inc). Indicative US rents are USD 100 to 150 per kW per month hyperscale, 150 to 250 wholesale, 200 to 400 retail.

No Thai lease terms are disclosed anywhere. No tenor, no escalator, no security package, no anchor credit for True IDC Rayong, Digital Edge BKK or DayOne Chonburi. There is also no public Thai colocation rent benchmark. Both Cushman & Wakefield and CBRE cover Bangkok only inside paywalled report bodies. Anyone quoting a Thai yield on cost is quoting an assumption dressed as an observation, and should say so.

Contracted versus merchant, and what the market actually pays for

Contracted means signed leases with creditworthy counterparties at notice to proceed. Merchant means building against a view of demand and leasing later. The valuation gap between them is the whole game. ARC Group reports platform assets combining secured power, multi-market execution and a delivery record trading at 25 to 35 times EBITDA, against regional telcos at roughly 7 times, and cites AirTrunk at 21 times contracted EBITDA (ARC). I could not open the PwC page those multiple figures are sourced to, so treat them as reported rather than verified.

The largest regional platform mark is the KKR and Singtel acquisition of the remaining 82% of STT GDC for S$6.6bn at an enterprise value of S$13.8bn across 2.3 GW of design capacity, announced 4 February 2026 (STT GDC). My arithmetic on those disclosures gives roughly S$6.0 million per MW of design capacity and implies about 42% net debt in the enterprise value. Design capacity is not operational capacity, so that is not a stabilized asset comp and should not be used as one.

The premium is paid for secured power and contracted revenue, not for concrete. Which is exactly why Thailand's power-confirmation gap is where value sits.

Yield on cost

Stabilized net operating income over total development cost. Cushman & Wakefield estimated the Asia Pacific pipeline at roughly 13% gross yield on cost, rising to about 14% in Tokyo, Mumbai, Johor, Sydney and Beijing, from a 12,452 MW pipeline requiring USD 116.2bn and generating over USD 14.9bn of annual colocation rent (C&W). My arithmetic on those two numbers gives about USD 9.3 million per MW of all-in mid-spec cost, consistent with the 2026 cost guide. No Thailand-specific yield on cost is published, and without a Thai rent benchmark one cannot be constructed from observed data.

Project finance versus corporate finance, and the Thai trap

Project finance is non-recourse to the sponsor, sized on contracted cash flow with debt service coverage covenants, and secured on the SPV's shares, its contracts, its accounts and its assets, with direct agreements giving lenders step-in rights against the offtaker and the utility. Corporate finance is faster and cheaper to arrange, has no ring-fence, and consumes the sponsor's balance sheet.

Regional norms are five to seven year tenors with interest only during construction, at USD effective rates of 6.3 to 8% as observed in 2024 (Octus). Tenor and margin are undisclosed on every Thai financing I found.

Thai precedents, such as they are. Digital Edge and B.Grimm Power signed an USD 880 million green loan on 6 May 2026 for their 100 MW Chonburi campus across a nine-bank syndicate, described as Thailand's largest data center financing (Digital Edge). Against a reported total JV investment plan of about USD 1 billion that would be roughly 88% loan to cost, which is implausible for project finance, so either the facility covers more than the campus or the cost figure understates it. The True IDC and GIP Rayong financing for a 102.6 MW facility is reported at USD 550 million by Linklaters, USD 530 million by Crédit Agricole CIB and USD 560 million by Octus; the three conflict and none is authoritative. Taking the range, that is roughly USD 5.2 to 5.5 million of debt per MW, which against an USD 8.8 million per MW cost anchor implies something near 60% leverage. DayOne's Thai project took a THB 28 billion green loan from a UOB-led six-bank syndicate in March 2026 (Baker McKenzie), and True IDC was reported in July 2026 to be seeking about USD 2 billion in three tranches, roughly USD 1.8 billion of it in dollars with two smaller baht tranches (Bloomberg). That tranche design is the market telling you how thin baht funding depth is for this asset class.

Now the trap, and it is the most under-priced legal fact in the Thai file. A PEA notification published in the Royal Gazette on 4 June 2026 and effective 5 June 2026 imposes a three-year lock-up running from the actual electricity usage date, during which the original shareholders named in the power usage application must retain more than 50% through any transfer, with carve-outs only for common-control transfers, conversion to a public company, and statutory amalgamation (Baker McKenzie). Baker McKenzie's own emphasis is not on M&A but on financing, flagging uncertainty over whether lenders can enforce security over the PPA or exercise step-in rights during the lock-up, and adverse exposure if a share pledge is enforced and produces a prohibited change of control. In plain terms, the standard project finance security package may be unenforceable in the first three years of operation, which is precisely when it matters. The rule applies in PEA territory only, not MEA, and it catches large industrial users as well as data centers. It does not block a minority sell-down, since the originals need only stay above 50%, so the sponsor-recycling story is constrained rather than dead.

What actually kills a data center deal

Power that never arrives. This is the first and largest killer everywhere, and in Thailand it is quantified. PEA data as of June 2026, reported by Thansettakij on 25 July 2026 as a single-outlet exclusive, showed 95 projects applying for 26,045.2 MW against 4,882.7 MW that PEA could confirm, and of that only 3,240.7 MW firm with 1,642 MW conditional. Signed power purchase agreements covered just 2,768.2 MW across 18 projects (Thansettakij). That last figure is the honest one. Roughly one tenth of the applied queue is contractually real. The June snapshot is already stale: on 28 August 2026 PEA Deputy Governor Pongsakorn Yutthakowit put applications above 30,000 MW per Thansettakij, or at 20,000 to 30,000 MW per Bangkokbiznews, with no updated confirmed-supply split (Thansettakij, Bangkokbiznews). These are PEA-only figures excluding MEA territory, and gross applications are non-binding, so the queue overstates real demand by design. EPPO's own planning scenarios for PDP 2026, dated 24 March 2026, model data center demand to 2050 at 6,799 MW low, 8,811 MW mid and 19,808 MW high, the last built to encompass all MEA and PEA inquiries (Thansettakij). Confirmed supply sits below the plan's own low case. Anyone underwriting scarcity value in that queue is underwriting a number the regulator is actively discounting.

Relief that arrives after you need it. EGAT's near-term EEC package adds 1,150 MW across five substation works, from Rayong 2 already complete through Phanthong and Phanthong 2 to Pluak Daeng in Q1 2027 (EGAT). The larger THB 31,050 million transmission program adding about 2,667.6 MW is phased from 2028 with completion in 2033, and as of 14 August 2026 the ERC had only opened public consultation on it, not approved it. Bangkok Post has presented it as an EGAT commitment; the Thai-language and law-firm accounts present it as a Ministry proposal awaiting Cabinet. Treat it as proposed.

The anchor walks or shrinks. Because Thai anchor tenants are almost entirely undisclosed, lease-up risk cannot be differentiated across the field on public information. That is itself a diligence finding.

Cost overrun and long-lead equipment, against a Thai construction cost inflating at 14.7% a year.

Water refusal and community opposition. Documented and named at Khlong Tamru in Chonburi and Ban Chang in Rayong, including a subdistrict chief reporting no visible EIA for the QHI01 project, and a People's Party MP calling on 27 July 2026 for a parliamentary committee to scrutinize data center approval (Mongabay). Of twelve companies approached by the Business & Human Rights Resource Centre in March 2026, only two responded. Silence is a governance signal.

Regulatory reclassification. The NBTC is examining moving data center services from a Type 1 to a Type 3 telecommunications licence, targeted for end-2026 after consultation. Type 3 requires Thai-majority shareholding. That is a binary that voids a foreign-controlled thesis outright, and it has no equivalent in Malaysia, Indonesia, Singapore or Vietnam (Silk Legal). Alongside it, a Data Centre Business Policy Committee was created by Cabinet decision of 5 August 2026, gazetted 13 August and effective 14 August 2026, adding a national approval layer above the BOI and the utilities.

A financing structure defeated by the lock-up, as above.

Power price you cannot fix. Type 9 is being designed as a marginal cost pass-through, not a level, which means the tariff is not hedgeable at underwriting. The hedge is Direct PPA, and Direct PPA is not available. The pilot was approved on 25 June 2024 with a 2,000 MW cap and a requirement of at least 50 MW of IT base load per building; the ERC's draft criteria went out on 3 October 2025 with a seven-day comment window; the NEPC removed the cap and opened the scheme to all industry on 15 July 2026; EPAC approved the principle on 3 August 2026 and the ERC entered consultation. As of late August 2026 the ERC's deputy secretary-general said the Grid Code and the wheeling charge were targeted for completion within the year, and the Energy Minister said the TPA charge criteria were still to be announced (Bangkokbiznews, Thansettakij). No Direct PPA has been reported signed.

The draft wheeling economics matter, because market commentary is anchored on the wrong number. The ERC's own consultation document, approved 22 October 2025 and consulted from 27 October to 10 November 2025, contains a voltage-differentiated table, not a postage stamp rate.

Component (THB/kWh)	69 kV and above	Medium voltage	Low voltage
Transmission	0.6868	0.7105	0.7528
Distribution	0.4350	1.2005	1.7707
Wheeling subtotal	1.1218	1.9110	2.5235
System security / ancillary	0.4978	0.4978	0.4978
Policy expense (Sep–Dec 2025 Ft)	0.1447	0.1447	0.1447
My arithmetic, energy-based total	1.7643	2.5535	3.1660
Peak demand charge (THB/kW/month)	74.14	132.93	210.00

Source: ERC public hearing 582. The totals row is my own addition and is not stated in the document. The widely circulated figure of roughly 1.07 THB/kWh on a postage-stamp basis comes from a Hunton client alert on the September 2024 draft and is repeated by vendors; like for like against wheeling only it is close to the 1.1218 figure, but comparing it to the full grid stack understates the cost by a wide margin. Note also that the 0.4978 ancillary charge is built mostly from surplus reserve capacity rather than from any service the buyer consumes, that positive imbalance energy is forfeited to the utility without compensation on fifteen-minute settlement while negative imbalance above 2% is charged at up to twice the retail tariff, and that the policy expense is Ft-linked and resets every four months, so it is not contractible. None of these rates were adopted, the register shows no TPA tariff consultation in 2026, and the Ministry has since described a TPA charge that will also include a system loss component absent from this table. Treat the whole stack as a superseded proposal, with a plausible range of roughly 1.1 to 2.5 THB/kWh of wheeling depending on connection voltage.

Obsolescence. A hall commissioned for 10 kW racks cannot host 120 kW racks without replacing the cooling and often the floor. When I underwrite an existing Thai asset I underwrite the retrofit, not the build.

What I would take away

The unit is the megawatt because the scarce input is electrical service, and the only megawatt worth quoting is IT load. PUE is a ratio whose measurement boundary is negotiable, which is why tying tax to it changes behavior rather than outcomes. Heat is the same energy leaving, and at roughly 120 kW a rack the air physically cannot carry it, so liquid at the chip is a design decision made years before the tenant arrives. Redundancy grammar describes topology, and a tenant that already replicates in software is buying the same insurance twice when it pays for 2N. The development sequence is a power sequence, and Thailand has just inverted it so that power confirmation precedes investment promotion. And the deal is the spread between an USD 8.8 million per MW mid-spec build and a contracted rent that nobody in Thailand publishes.

The two facts I would put in front of an investment committee before anything else are these. Only about 2,768 MW of Thai data center load had signed power purchase agreements against a queue that has since passed 30,000 MW, so the scarcity is real but the queue is not evidence of it. And a PEA notification effective 5 June 2026 may make the standard project finance security package unenforceable for the first three years of operation, which is a legal fact that changes the capital structure, not just the diligence list.

Topics

th-dc-landscape · th-power-sector · ai-infrastructure