What a $500 Billion AI Data Center Actually Costs
On a decommissioned uranium enrichment site in southern Ohio, a consortium is building what would be the largest data center ever attempted — 10 gigawatts of compute with its own 9.2GW power station attached. The construction bill is remarkable. The financing structure is the part worth understanding: roughly two thirds of the money goes to a single supplier, who is also guaranteeing the debt used to buy from them — a guarantee cut from $250B to $105B in August after investors balked.
Thank you for reading this post, don't forget to subscribe!
| Figure | Was | Now | Change | Why |
|---|---|---|---|---|
| Nvidia debt guarantee, Ohio site | ~$250B | up to $105B | ▼ −58% | Scaled back after investors raised concerns over Nvidia’s risk exposure; now covers the first phase (~5GW) only, not the full 10GW. HIGH |
| Status of the guarantee | in talks | agreed | firmed up | Reported as a near-final agreement covering phase one. MEDIUM |
| Chip financing | ~$350B | ~$350B | no change | Separate from the debt guarantee; no revision reported. MEDIUM |
The reduction strengthens rather than weakens this report’s argument. Nvidia pulled back
because investors questioned the risk of a supplier underwriting its own customer —
the exact circular-financing concern set out below. Sources: Wall Street Journal (14 Aug 2026),
Reuters, Fortune (18 Aug 2026). Previous value retained above; see our
corrections policy.
The building is the cheap part
In March 2026 the US Department of Energy announced a public-private partnership to build a 10-gigawatt data center at the former Portsmouth Gaseous Diffusion Plant in Pike County, Ohio — a Cold War uranium enrichment facility, now rebranded the PORTS Technology Campus. SoftBank’s Masayoshi Son has put the full build at roughly $500 billion. The first phase alone — about 800 megawatts, roughly 8% of the eventual capacity — is costed at $30–40 billion and due in early 2028.
Break the build into layers and the intuition most people have about data centers inverts. The concrete-and-steel shell, the cooling plant, the substations, even a purpose-built 9.2GW power station — all of it together is a minority of the bill. The silicon is roughly two thirds. This is not a construction project with computers in it; it is a chip purchase with a building wrapped around it, and the building is the part that will still be standing in forty years.
Which is what makes the financing the real story. Nvidia has agreed to guarantee up to $105 billion of the project’s debt — cut back from the roughly $250 billion reported in July, and now covering only the first phase — alongside financing of up to $350 billion of chip purchases, chips it manufactures. The structure exists because OpenAI does not hold an investment-grade credit rating, so the developer, SoftBank’s SB Energy, would borrow against Nvidia’s balance sheet instead. A supplier is underwriting its customer’s ability to buy from it. Bernstein’s Stacy Rasgon put it plainly: the arrangement “will clearly fuel ‘circular’ concerns.”
- Chips dominate. Roughly two thirds of spend is silicon — the physical plant is the smaller half of a very large number.
- Power had to be built, not bought. No US grid has 10GW of spare capacity to hand, so the project includes its own gas station.
- The supplier is underwriting the buyer. Vendor financing at this scale is the mechanism behind the circular-financing critique.
The cost stack, layer by layer
Read the confidence column carefully. Four figures here are reported; the rest come from a widely-circulated independent teardown of the project and are engineering estimates, not disclosures. They are useful for understanding proportion, not for quoting as fact.
| Layer | Estimate | What it buys | Confidence |
|---|---|---|---|
| Chips | $350B | ~4.5M units: high-bandwidth memory plus logic die | MEDIUM matches reported chip financing |
| Shell & construction | ~$55B | Concrete, steel, labour; storm-rated structure | MODELLED |
| Power generation | ~$37B | 9.2GW gas plant, turbines, transmission | MODELLED |
| Cooling | ~$35B | Closed-loop direct-to-chip liquid system | MODELLED |
| Network | ~$30B | In-rack copper interconnect, fibre, switching | MODELLED |
| Electrical distribution | ~$20B | Substations, backup generation, fuel cells | MODELLED |
| Land | $0.1–0.3B | 3,700 acres, leased federal site | MODELLED |
| Modelled total | ~$527B — against SoftBank’s stated ~$500B HIGH | ||
The chip line is the one to trust most: it independently matches the ~$350B of chip purchases Nvidia is reported to be financing for this site. That two separate routes land on the same number is the strongest signal in the table.
Why power is the hard constraint
You cannot plug ten gigawatts into the American grid. So the project builds what would be among the largest gas-fired stations on earth — 9.2GW — on the same site as the compute.
And the binding constraint isn’t money, it’s turbines. Large gas turbine manufacturing is concentrated in three companies, and their order books tell the story: GE Vernova’s backlog reached 116GW in Q2 2026, with the company targeting 125GW under contract by year-end against manufacturing capacity of roughly 20GW a year — close to five years of output already committed. Siemens Energy’s backlog stands near 69GW. Natural gas investment is hitting a decade high of ~$330B in 2026, but capital cannot buy delivery slots that don’t exist.
The same squeeze applies further down the stack. Backup diesel generation carries multi-year lead times, which is why projects increasingly bridge with fuel cells — deliverable in months rather than years, at a price premium. Our infrastructure index tracks the wider power picture, including the IEA’s projection that data-centre electricity roughly doubles to ~945 TWh by 2030.
Cooling: the water claim, checked
Traditional evaporative cooling consumes roughly 2.6 million gallons per megawatt per year. At 10GW that arithmetic produces the alarming comparisons — a facility drinking like a major city. Closed-loop direct-to-chip cooling, where a cold plate sits on the die and the coolant is filled once and recirculated, can cut that toward zero. Nvidia claims its liquid-cooled GB200 NVL72 architecture delivers 300× better water efficiency and 25× better energy efficiency than air cooling.
But the popular version of this claim overstates it. Direct-to-chip cooling does not inherently eliminate water use, because the heat still has to leave the building somehow. If the final heat-rejection stage uses dry coolers, site water consumption approaches zero. If it uses cooling towers — as many liquid-cooled facilities still do — evaporation and blowdown continue to consume water. The saving is real, and it is conditional on the rejection method, which is rarely specified in announcements. See our energy and water per query index for the per-prompt view.
The financing loop
The sequence matters, because it has already changed once. In September 2025 Nvidia announced an investment of up to $100 billion in OpenAI, tied to OpenAI deploying 10GW of Nvidia systems. That commitment collapsed in February 2026; Nvidia instead took a $30 billion equity stake in OpenAI’s $110 billion round.
What replaced it is structurally different and larger. Nvidia has now agreed a guarantee of up to $105 billion on financing for the Ohio site — reduced from the ~$250 billion reported in July after investors raised concerns about its risk exposure, and limited to the project’s first phase (~5GW of the 10GW total) — plus separate financing of up to $350 billion in chip purchases. The reason for the unusual shape is a credit problem: OpenAI lacks an investment-grade rating, so SB Energy would borrow against Nvidia’s balance sheet rather than OpenAI’s.
Set out plainly: a chip vendor guarantees the loans that a developer uses to buy that vendor’s chips. Vendor financing is not new or inherently improper — but at this scale it makes demand harder to read, because bookings no longer independently confirm that an end customer wanted the capacity. Nvidia’s wider financing commitments across the sector have been estimated near $750 billion and are drawing scrutiny for the same reason. Our AI Bubble Tracker holds the balanced version of this argument, and AI’s Hidden Debt covers the off-balance-sheet dimension. For the contrasting structure — where a third-party asset manager, not the supplier, funds the build — see Off the Books: Who Actually Owns AI’s Data Centers.
FAQ
What does a large AI data center cost to build?
The Ohio project is costed at roughly $500 billion for 10GW, with phase one (~800MW) at $30–40 billion. Chips account for around two thirds; land, shell, power, cooling and networking make up the rest.
Why build a power plant instead of using the grid?
No US grid region has ten gigawatts of spare capacity available on this timeline, so the project includes 9.2GW of on-site gas generation. The constraint is turbine manufacturing, not funding — the two largest suppliers already carry roughly 185GW of combined backlog.
Do AI data centers really use as much water as a city?
Evaporative cooling does consume roughly 2.6M gallons per MW per year, which at this scale is city-sized. Closed-loop direct-to-chip cooling can reduce that to near zero — but only if the final heat rejection uses dry coolers rather than cooling towers.
What is circular financing in AI?
When a supplier funds, guarantees or invests in its own customers, so revenue is partly financed by the vendor booking it. Nvidia guaranteeing project debt used to buy Nvidia chips is the clearest current example.
Is Stargate funded by the US government?
No. The federal role here is the site — a leased, decommissioned DOE facility — and permitting facilitation. The capital is private. See AI Government Spending for why Stargate should never be added to federal AI totals.
Project facts from the US Department of Energy announcement (March 2026) and contemporaneous reporting (Spectrum News, CBS, Barchart) covering the PORTS Technology Campus: 10GW compute, up to 10GW generation including 9.2GW gas, phase one ~800MW at $30–40B. Total project value per SoftBank (Masayoshi Son). Financing details — the collapsed $100B commitment, the $30B equity stake in OpenAI’s $110B round, the guarantee (~$250B reported in July, cut to up to $105B and phase-one only in August) and ~$350B chip financing — from reporting collated by Benzinga, Yahoo Finance, BigGo and Business Standard; analyst comment from Bernstein. Turbine backlogs from GE Vernova and Siemens Energy Q2 2026 disclosures. Cooling figures from Data Center Frontier, EESI and Nvidia’s published GB200 NVL72 claims. Rows marked MODELLED are not disclosed figures — they come from an independent public teardown of the project and are reproduced here for proportion, not as fact; we have corroborated only the chip line against reported financing. Deal terms remain under negotiation and may change. Corrections: see our methodology and corrections policy.