Nvidia has reduced its commitment for OpenAI’s planned data center in Ohio from $250 billion to less than $120 billion. In doing so, it yielded to its own shareholders, who had grown uncomfortable with how much risk the chipmaker was taking on for a single customer. According to the Wall Street Journal, the two companies were already finalizing the agreement last week and could have signed it within days. At the same time, rival Anthropic released figures that appear to be the exact opposite of a cooling market.
Original terms of the agreement
At the end of July this year, the Wall Street Journal reported that Nvidia was negotiating a financial guarantee of approximately $250 billion for a data center project in southern Ohio. The total cost of building the campus would exceed half a trillion dollars once the cost of chips was added to the land, construction, and energy infrastructure.
However, the guarantee was never intended to finance the processors themselves. Nvidia was committing to guarantee the lease of the campus and the debt associated with its construction, meaning it would sign on as a co-borrower for loans that banks would otherwise be unwilling to provide to OpenAI. The reason is that OpenAI does not have an investment-grade credit rating, so it cannot secure inexpensive financing for such a large project on its own. Moreover, the data center will not earn a single dollar until someone completes it and fills it with graphics cards.
Alongside this, a second, separate part of the financing was also under discussion. It was intended to cover the purchase of the Nvidia chips themselves, with as much as another $350 billion being considered. It was precisely this combination that sparked the greatest distrust in parts of the market. The supplier would effectively be guaranteeing the money its customer used to buy its own products.
What forced Nvidia to back down
As soon as the $250 billion figure became public in the final days of July, Nvidia shares lost around five percent in a single trading day. Shareholders simply concluded that a contingent liability of nearly a quarter of a trillion dollars, tied to one company’s ability to continue raising capital, had a completely different risk profile from Nvidia’s existing business model of selling chips for cash.
More vocal critics also spoke out. Michael Burry, who built his reputation by betting against the mortgage market in 2008, posted a comment on X saying that everything had come full circle because Nvidia was guaranteeing purchases of its own chips. Shortly before that, he had increased his short position against Nvidia shares. Commentator Ed Zitron questioned both the financing structure itself and the fact that the campus was being built by SoftBank, asking where the money was supposed to come from.
The pressure worked, and according to the WSJ, Nvidia and OpenAI are finalizing an arrangement in which the chipmaker guarantees only the first phase of construction. The guarantee has therefore fallen below $120 billion, less than half the amount originally discussed. Reuters notes that the change came after investors raised concerns about Nvidia’s overall exposure to major financial commitments.
The planned data center in Ohio
Once completed, the project is expected to have a capacity of ten gigawatts, which would make it the largest announced data center in the world. It is being built by SB Energy and is part of a broader wave of construction launched by OpenAI under the Stargate banner.
According to the July report, the first phase is expected to deliver approximately 800 megawatts of capacity by 2028. Technology website The Decoder adds that under the new agreement, Nvidia’s guarantee covers the first construction phase, with a capacity of approximately five gigawatts. The remaining costs therefore fall to OpenAI, SoftBank, and other financing partners, as the construction itself is not being scaled back. The only thing changing is who will bear the impact of a potential collapse.
The lease agreement for the entire ten-gigawatt campus is also still being negotiated, and the final signature is still pending. OpenAI wants to secure its own computing capacity rather than merely leasing it from Microsoft or Oracle. However, the company’s ability to support such a large investment remains under scrutiny because it is still not profitable. According to available estimates, it will lose around $14 billion this year despite growing revenue.
According to information from July, Microsoft, Google, and Anthropic were also interested in the Ohio site.
Meanwhile, Nvidia is looking for money elsewhere
A few days before reducing the guarantee, Nvidia announced a partnership with six major financial institutions and launched a platform for financing computing capacity. The goal is to attract more than half a trillion dollars in external capital for the construction of artificial intelligence infrastructure.
Instead of a single company carrying the risk on its balance sheet, it will be spread among banks, funds, and investors that will be compensated for taking it on. For smaller computing capacity providers, however, this means tougher terms. Over the past year, they have borrowed against multi-year computing capacity contracts on the assumption that one of the technology giants would always have their backs. Guarantees will now be shorter, divided into phases, and lenders will require actual revenue to cover these commitments.
Anthropic counters with record revenue
While Nvidia was scaling back, Anthropic released figures that dispel concerns about a cooling market. According to reports from Bloomberg, its second-quarter revenue jumped from just under $800 million to more than $11.5 billion, representing a fourteenfold year-over-year increase. According to available information, the company also reported a positive adjusted operating profit, the first operating profit in its history.
There is one catch, however. A significant portion of this increase comes from a massive contract to purchase computing capacity from SpaceX, which costs Anthropic approximately $1.25 billion per month and whose profit margins may not be repeated in subsequent quarters. According to sources close to investors, Anthropic itself warned that no one should count on sustained profitability in the immediately following quarters because it is preparing its own massive infrastructure investments.
People familiar with Anthropic’s finances say it is targeting revenue of between $190 billion and $200 billion for 2028. The company wants to go public at a valuation approaching $2 trillion, as early as the end of September or the first few days of October. OpenAI is moving toward a similar milestone at a similar pace, with its market debut also being discussed for as early as September.
