The numbers sound impressive. 10 gigawatts of compute. $500 billion in total investment. $350 billion earmarked for AI chips alone. A 15-year build-out on federal land in Ohio, supported by Japanese capital and US government blessing. But in the absence of data, opinion is just noise.
Let's run the audit.
I have spent fifteen years modeling financial risk in environments where hype consistently outpaces arithmetic. From the 2017 ICO audits that flagged unvested token dumps to the 2020 compound finance rounding error that could have drained $2 million, I have learned one thing: when numbers look too clean, the dirt is in the assumptions.
Context: The Project on Paper
The proposal, as reported, involves a partnership between OpenAI and Nvidia, with SoftBank's SB Energy acting as site developer. The plan is to build a series of data centers totaling 10GW of power capacity, with the first 800MW phase targeted for completion by 2028. Nvidia would provide roughly $250 billion in "lease-develop" financing, essentially pre-purchasing GPU capacity. SoftBank commits $330 billion for energy infrastructure upgrades, linked to tariff relief for Japanese goods. The rest of the $500 billion total is assumed to come from debt, government subsidies, and OpenAI's existing war chest.
The narrative is seductive: unlimited compute for the next GPT generation, a fortress of AGI built in the American heartland. But as a risk consultant, I do not evaluate narratives. I evaluate balance sheets, supply chains, and engineering constraints.
Core: Systematic Tear Down
1. Physics Doesn't Negotiate
10GW is not a number; it is a force of nature. To put it in perspective: the entire state of Ohio currently consumes roughly 150 GW at peak. This single project would consume 6-7% of that state's entire electrical grid. The first 800MW phase alone equals ten of the largest existing data center clusters (like the 80MW Google data center in Council Bluffs).
Electrical infrastructure of this scale requires dedicated high-voltage substations, multiple 345kV transmission lines, and typically a 5-10 year permitting cycle. The plan to have 800MW operational by 2028 implies that environmental impact assessments, grid interconnection studies, and construction must begin within months. That is not aggressive; it is unrealistically optimistic.
Cooling is another hard constraint. 10GW of GPU compute (assuming H100-class chips at 700W per GPU) would generate approximately 7-8 GW of heat. The only viable solution is liquid cooling, and the global supply chain for liquid cooling is currently capable of delivering roughly 500MW of fully deployed cooling infrastructure per year. Scaling to 10GW would require the entire industry to expand capacity 20x within a decade. Even if that is possible, the metallurgical supply chains for copper and aluminum coolant loops would face severe bottlenecks.
2. The Financial Engineering Trap
The financing structure is the most dangerous part. Nvidia is offering $250 billion in "lease-develop" financing. This means Nvidia pre-purchases GPU orders for OpenAI's data center, effectively acting as a bank. This looks like genius financial engineering: Nvidia locks in decades of demand, OpenAI secures hardware without immediate cash outlay.
But consider the risk transfer. If OpenAI's model revenue does not materialize, Nvidia is left holding a special purpose vehicle filled with $250 billion worth of GPUs that have a useful life of roughly 3-5 years. The moment Nvidia starts booking this as revenue (which they likely will under GAAP via vendor financing accounting), they recognize profit now but carry the liability of future hardware depreciation. If OpenAI defaults, Nvidia will have to write down billions of dollars in assets that cannot be deployed elsewhere easily — because the data center is purpose-built for one customer.
Furthermore, the $500 billion total investment implies a capital cost of roughly $50 per watt. Standard hyperscale data centers cost around $10-15 per watt. The premium likely covers Nvidia's financing margins, SoftBank's developer fees, and the premium for federal land. That means the cost of compute here is 3-5x more expensive than building conventional capacity. That premium must be recovered through OpenAI's API pricing. At current rates, OpenAI would need to generate $200-300 billion in annual revenue just to cover depreciation and electricity (assuming a 10-year depreciation and $0.05/kWh). Today, OpenAI's annualized revenue is estimated at $5-10 billion. The gap is a factor of 20-30x.
3. The Single Point of Failure
This project concentrates risk in a way that makes the 2008 mortgage-backed security crisis look diversified. Nvidia becomes both the hardware vendor and the financier. OpenAI becomes both the tenant and the sole customer for the compute. SoftBank's SB Energy is both the developer and the power infrastructure provider. There is no external hedge, no alternate use case for the facility if the AI demand cycle turns.
In my 2022 audit of the Terra/Luna collapse, I showed how a supposedly decentralized stablecoin was entirely dependent on a single yield source. The same principle applies here: if OpenAI's model training hits a fundamental barrier (e.g., no scaling laws benefits beyond GPT-6), the entire capital structure unwinds.
Contrarian: Where the Bulls Got It Right
I am not naive to the opportunity. The bulls would argue that this project is not about 2025 revenue — it is about 2035. If AGI emerges, the compute needed for inference could dwarf training. A 10GW cluster might become the equivalent of the Hoover Dam for the AI age: a long-term asset that generates wealth for decades. The US government's involvement suggests that national security interests may override standard economic feasibility. If China accelerates its own AI compute build-out, the US may subsidize this project to maintain strategic advantage.
Additionally, Nvidia's financing structure could be structured as a sale-leaseback, where OpenAI buys the hardware and Nvidia provides a loan secured by the assets. In that case, Nvidia's balance sheet risk is limited. The Japanese energy investment also provides a geopolitical cushion — Japan secures a foothold in US infrastructure, and the US secures access to Japanese capital and nuclear technology for future power generation.
But these are political and strategic justifications, not financial ones. Politics can change with an election. Strategic advantage cannot be spent on API credits.
Takeaway: The Arithmetic of Accountability
Will the 10GW cluster become the Hoover Dam of AI or the Solyndra of the 2020s? The answer will be written in code, not press releases. Every line of this project's financing agreement, every GPU purchase order, every power purchase agreement must be audited with the same forensic skepticism I applied to the Compound smart contract bug. Until then, this is a massive bet on a single narrative — a bet that ignores the cold reality of physics, supply chains, and marginal cost of capital.
In the absence of verifiable data, this is not an investment thesis. It is a hypothesis, and a highly leveraged one at that. The market always gets its arithmetic right eventually.