Summary: The Wall Street Journal reported on 14 August that Nvidia now expects the first phase of OpenAI data-center loan guarantees to land below $120 billion, below the $250 billion ceiling discussed in late July. No contract had been executed; the arrangement remained a memorandum of understanding, with the Ohio project among the affected work. The article reads the change as evidence that circular-financing concerns and power constraints are now part of AI capacity planning.
The Wall Street Journal reported on August 14 that Nvidia has pulled back on its plan to guarantee loans for OpenAI’s data-center construction: the first phase of the guarantee is now expected to land below $120 billion, a clear step down from the $250 billion ceiling that was on the table in late July. Reuters came in behind with an account of its own. What stands out is that no contract was ever executed — the whole thing still exists purely as a memorandum of understanding. A guarantee that shrinks like this amounts to the market placing a question mark beside the most aggressive version of the AI infrastructure financing loop.
The July Plan: a $250B Guarantee
Late in July, the Journal and Reuters reported in turn that Nvidia was negotiating to guarantee as much as $250 billion of OpenAI’s data-center debt. Among the targets was a large Ohio campus that OpenAI was working to lease from SB Energy — around 10 GW under lease, with the entire project carrying a possible price tag of $500 billion. Nvidia also put memoranda in place with six institutions — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — with the aim of marshaling $500 billion in third-party capital. The structure itself was plain: private credit and institutional money would pay for the buildings and the power, Nvidia would absorb the debt risk, and OpenAI would hold the capacity for years. A guarantee is credit support rather than an equity stake — Nvidia’s exposure stayed contingent, coming due only when borrowers failed to repay.
Why a Chipmaker Co-Signs Loans
The reasoning is not hard to follow: a lender would sooner extend credit to a profitable borrower (Nvidia) than to one losing money (OpenAI). The guarantee also answers a timing mismatch: a campus takes years to finish, OpenAI’s revenue is climbing fast off a small base, and something has to span the distance between construction spend and usage revenue. Put Nvidia in the guarantor seat and both the interest rate and the feasibility of data-center project debt improve; projects break ground, GPUs ship, and the loop closes — sell the chips, then co-sign the buyer’s loan. Critics have likened the arrangement to “a savings-and-loan that happens to design computer chips,” and others name what they see as the real objective: turning GPUs into a securitizable asset class, never mind that they depreciate far faster than real estate. Vendor financing is not without precedent in the semiconductor industry’s history; the novelty lies in the size — a guarantee counted in hundreds of billions, hitched to depreciation schedules counted in a few years.
What Got Cut, and What It Signals
Three details matter. The first: no signature ever happened. That $250 billion figure was consensus at the memorandum level, and conditionals carry the reporting this week — “initially,” “expected to be less than,” “may guarantee.” “Initially” does leave the door open to phased guarantees that expand later, though treating that as a commitment would be a mistake. The second: the reduction lands on the Ohio project, the single largest component of the blueprint. The third: the market’s read tilts positive — shareholders see this less as a loss than as confirmation. Commitments that were never going to be fulfilled in full are now formally not going to be fulfilled. Attention has moved instead to the other links in OpenAI’s capital chain, above all SoftBank and Oracle.
What It Means for the Industry
Three takeaways. First, the circular-financing worry around AI infrastructure is out in the open now; the contest ahead is over whose capital structure holds up under scrutiny, not whose blueprint is larger. Keep an eye on whether the remaining guarantee ever becomes a signed credit agreement — that figure, not the MOU total, is the one that counts. Second, power reality is biting back: ten large nuclear units roughly match the installed capacity of one 10 GW campus, so generation and grid buildout must proceed in lockstep with construction, and interconnection queues do not run on AI timelines. Third, for teams buying compute, a supplier’s financial engineering now belongs in supply-chain risk assessment alongside its chip roadmap. When a vendor’s growth plan rests on guaranteeing its customers’ debts, capacity commitments deserve the same diligence as pricing.