Nvidia guaranteed the building so it could sell everything that goes inside
Posted on: 30 September 2026
On Monday the buyback was read the obvious way, as a board convinced its stock is cheap, and the numbers seem to agree. Nvidia trades at around 16.5 times expected earnings for the next twelve months, the lowest multiple since January 2015 and a long way below its fifteen-year average of 30. Those who think the AI cycle has peaked read that number as confirmation, those who think otherwise see an opportunity and Jensen Huang has just put $150 billion on the second view. Both are opinions about the price of a share, though. What interests me more is what the buyback says about the risk Nvidia took on six weeks earlier, because the cash that will fund it is the same cash standing behind a signature of comparable size.
The campus is called PORTS-Pike and sits in Pike County, southern Ohio, on federal land where uranium was once enriched for the government. SB Energy, SoftBank's energy arm, will build and run it while OpenAI occupies it on a twenty-year lease. The first phase is worth 4.25 gigawatts with an option on another 3.75 and the first 800 megawatts should switch on in 2028. Nvidia will supply the compute and put $1.5 billion straight into SB Energy. Above all it will guarantee up to $105 billion of its customer's obligations, and that customer only starts paying once the capacity is actually available.
The final figure came out of a negotiation that at the end of July, according to the Wall Street Journal, started at $250 billion. On the Monday the story broke the stock lost 4.5 per cent by late morning and credit default swaps on Nvidia's bonds posted their sharpest intraday rise since they began trading actively in November. Bondholders understood at once who would carry the weight of that signature. Three weeks later the guarantee had shrunk to less than half.
To measure the jump you only need the previous quarter. In the 10-Q for the period ending in April, as Tom's Hardware pieced together, Nvidia's maximum exposure across every lease guarantee it had given its partners came to $3.5 billion, taken in exchange for warrants and booked as credit derivatives with a fair value described as immaterial. A single contract is now worth thirty times that total.
The comparison doing the rounds since August is Lucent, and it rests on solid documents. At 30 June 2000 Lucent reported credit commitments to customers of about $7.7 billion, plus a billion in guarantees on their debt. The same 10-Q contains a passage you could copy out today almost untouched, where the company explains that network operators were requiring suppliers to arrange financing as a condition of bidding for their projects. Lucent planned to sell those loans on to banks and investors to free up capacity, and it put in writing that all of this depended on its own credit rating and on being able to keep selling them. On 10 August this year Nvidia launched, with six large financial institutions including BlackRock, platforms meant to channel more than $500 billion of third-party capital into AI infrastructure. The shape is identical.
Nvidia's defenders have a serious argument and it deserves to be taken whole. Lucent's customers were largely upstart carriers built on venture money, with no profitable business underneath the networks they were laying, whereas Nvidia's biggest buyers are Microsoft, Google, Amazon and Meta, which pay for their spending out of operating cash flow. True. Except that the Ohio guarantee concerns none of them. It concerns OpenAI, which has no investment-grade rating and for exactly that reason needed someone else's signature to borrow on acceptable terms. Nvidia's exposure sits in the thinnest layer of its customer base, the labs and the companies that rent out compute, where back in January it had already put $2 billion into CoreWeave at $87.20 a share.
The technology transition I know best is a different one, but it taught me to recognise certain patterns. From 2004 I worked in the European Digital Cinema Forum, with an observer's seat at the DCI table, when cinemas had to give up film and nobody wanted to pay for the projectors. At Microcinema we started operating in 2005, the same year DCI published the first version of its specification and the first screens of the UK Film Council's Digital Screen Network came on. We were already talking to chains such as UCI and to the Italian multiplexes that in 2009 would become The Space. Later we joined the Italian consortium that developed ISIDE with the European Space Agency to deliver films to cinemas by satellite, which together with OpenSky came to serve around 1,000 screens in Italy, but that is another story. The economic knot was eventually untied by the virtual print fee, which made its European debut in 2008 in a deal between Arts Alliance Media and the French chain CGR. Each time a distributor booked a film onto a converted screen it paid a fee, calculated on what it saved by not striking and shipping a 35mm print. The party that paid was the party that saved, and the saving existed whatever the health of the cinema. The scheme also had an exit clause: once the projector was paid off, the fee stopped.
The flaw showed up later. In 2019 Renana Teperberg, who ran the commercial side of Cineworld, told Screen that putting a time limit on the virtual print fee had been a mistake, since a 35mm projector lasted for years with little more than new lamps while a digital one is essentially a computer. The system had paid for the first generation and nobody had worked out who would pay for the second.
Nvidia learned that lesson and turned it upside down. In its announcement of the deal the company says the whole economic point lies here, that the site is long-lived while the compute inside can be upgraded with every generation. The guarantee covers a twenty-year contract and Nvidia releases a new architecture roughly once a year. Whoever guarantees the building buys the position of natural supplier for every replacement that comes through that door until the late 2040s. The most elegant clause of the virtual print fee survives, since OpenAI pays only when the capacity exists just as the distributor paid only when the film was booked. What made the fee healthy is missing, though. Here the one signing is the one selling, and its return on the guarantee is the sale itself.
Huang dismissed the charge of circular financing in a single line: OpenAI pays rent. True. But the rent is precisely what Nvidia is guaranteeing.
At this point Monday's buyback stops being stock market news. The guarantee and the buyback draw on the same cash flow, the one Huang's statement presents as able to fund both investment in the transformation and returns to shareholders. They use it in opposite ways. The guarantee moves no money while the customer keeps paying and barely shows on the balance sheet. The buyback actually sends cash out to whoever chooses to sell, so each share retired leaves the remaining ones a larger slice of an obligation nobody has yet had to honour. Those who leave get paid. Those who stay keep the tail risk. Bondholders had already priced this in July, while on Monday shareholders answered with a 1.2 per cent rise before the open.
For anyone on a board trying to decide how far to trust this chain, cinema and Lucent mostly teach what to avoid. Reading the buyback as a view on the share price is the first mistake, because it is also a decision about who should carry the risk of the guarantee, taken by the very people who signed it. The second is treating the $500 billion platform with the asset managers as proof of strength. In 2000 the same architecture sat in Lucent's 10-Q among the conditions the company said it depended on, and in its 2001 accounts Nortel admitted that placing those loans with third parties had become harder just as the industry went into crisis.
One question remains that none of the filed documents actually addresses. In 2019 British exhibitors were discovering that the projector paid for by the virtual print fee already needed replacing and that nobody had planned who would pay. In Pike County the first building will switch on in 2028 with that year's chips inside, when the contract still has nineteen years to run. This time the supplier of the replacement is already written into the deal. Who will have the money to buy it is far less clear.