What Happened
Nvidia disclosed in an August 17 filing with the U.S. Securities and Exchange Commission that it will provide up to $105 billion in residual-value guarantees tied to OpenAI's planned data center campus in Pike County, Ohio. The number is large enough to be misread, so the structure matters more than the headline. Nvidia is not writing a $105 billion check, and it is not guaranteeing OpenAI's rent. It is guaranteeing what the finished buildings and the infrastructure inside them will be worth if the whole arrangement falls apart.
The project itself is the largest single data center lease announced so far. OpenAI signed a 20-year lease with SB Energy, a SoftBank subsidiary, for the PORTS-Pike Technology Campus. SB Energy will build, own, and operate the site. The first construction phase covers 4.25 gigawatts of IT capacity, with an option for another 3.75 gigawatts that would take the campus to roughly 8 gigawatts. The Wall Street Journal put the gross figure closer to 10 gigawatts once cooling and supporting infrastructure are counted.
Nvidia's obligation applies only to that first 4.25-gigawatt phase, and it sits at the end of a waterfall. If OpenAI defaults or becomes insolvent, SB Energy has to try to find a replacement tenant first. Failing that, it has to try to sell the facilities. Only after those steps does Nvidia cover the shortfall between the guaranteed minimum value and whatever SB Energy actually recovers — and OpenAI has separately agreed to reimburse and indemnify Nvidia for anything it pays out.
Alongside the guarantee, Nvidia is making a straightforward $1.5 billion equity investment in SB Energy and becomes the exclusive AI compute supplier for the first half of the campus. The actual guarantee agreements have not been filed yet. Nvidia said the exhibits will appear in its quarterly report for the period ended July 26, which should finally reveal the guaranteed minimum values. The details are laid out in Data Center Knowledge's breakdown of the 8-K.
Why It Matters
The most interesting thing about this deal is not the size. It is that a chipmaker has decided its silicon is collateral.
Steven Dickens, CEO and principal analyst at HyperFrame Research, compared the move to captive auto financing. Nvidia, he argued, is recognizing that GPUs hold value long after their first three years of service, much the way a lender treats a leased vehicle. AI accelerators can stay useful for six or seven years, and if that residual value can be underwritten, it becomes something lenders will price. Dickens described it as a way to unlock capital for the buildout — turning the GPU into an investable asset class rather than a line item that only depreciates.
That reframing is a response to a real constraint. Jensen Huang has taken to describing the bottleneck with a new acronym: LPS, for land, power, and shell. Chips and networking gear, in his telling, are no longer the hard part. Sites with grid connections and finished buildings are. And AI labs are growing faster than their balance sheets can support twenty-year infrastructure commitments, which leaves a gap that somebody with a stronger credit profile has to fill.
SoftBank's role fills the other half of that gap. SB Energy owns and operates the campus, and SoftBank has said it plans to develop at least 10 gigawatts of new generation to support projects of this kind. The Ohio site draws power from a 9.2-gigawatt gas plant that the U.S. government owns and Japan is helping finance under a trade agreement, a reminder that AI infrastructure has quietly become an instrument of industrial policy. Huang has framed the economics bluntly: roughly 1.5 million GPUs per system generation, $150 billion to $200 billion in revenue per generation, and about 12 gigawatts of Nvidia compute committed by OpenAI through 2030. Exercise the Ohio option and the combined package reaches roughly 16 gigawatts, worth something near $600 billion. The Decoder walked through the lease terms in detail.
Reaction
The $105 billion figure is notable partly for how much smaller it is than what came before it.
Nvidia originally considered backstopping the project at around $250 billion, then pulled back after its stock fell about 5 percent and investors pushed back on the exposure. A $120 billion figure circulated after that. The final number landed at $105 billion, and it covers asset value rather than the full lease, a materially narrower commitment than the early reporting implied. Fortune read the gap as a signal that chip demand may be less durable than the headline numbers suggest.
The broader unease is about visibility. A Wall Street Journal analysis found that nine large technology companies, Nvidia among them, now carry roughly $3 trillion in mostly AI-related obligations that never appear on their balance sheets. Leases are recorded only once payments begin; purchase commitments only once goods are delivered. Leases that have not yet started total about $1.2 trillion, four times the figure from a year earlier. Alphabet's purchase commitments jumped from $332 billion to $811 billion inside a single quarter. Morgan Stanley analysts warned that investors can barely assess these companies' real leverage. Alphabet and Amazon both recently reported negative free cash flow, and contracts of this shape are close to impossible to cancel.
Not everyone is convinced the framing is even coherent. Dickens objected to measuring an IT buildout in units of electricity, saying he would rather the industry count cores, GPUs, racks, or servers — things that are easier to quantify and easier to picture than a gigawatt.
What's Next
The immediate thing to watch is a filing. Nvidia's quarterly report for the period ended July 26 will contain the guarantee agreements themselves, and with them the guaranteed minimum values. Those numbers determine how much financing capacity the structure actually creates, which is precisely the question the 8-K left open.
Construction runs on a longer clock. The first 800 megawatts are slated to come online in 2028, with the rest of the 4.25-gigawatt phase following. The campus sits partly on a former Department of Energy uranium enrichment facility, an unusually literal handoff of one era's strategic infrastructure to the next. Whether Nvidia exercises its option on the remaining 3.75 gigawatts will be the clearest read on how the first phase is going. OpenAI's own description of the PORTS-Pike project covers the site plan.
There is also a template question. If residual-value guarantees work here, other chip suppliers and other labs will try the same structure, and financing compute against the resale value of accelerators becomes ordinary practice. If GPU resale values disappoint — because a new generation lands harder than expected, or because demand softens — the guarantees get tested in a way nobody has modeled yet.
Closing Thoughts
What Nvidia is doing in Ohio looks less like a chip sale than like a bank underwriting a mortgage on hardware that has not been built yet.
That is a strange position for a company whose advantage was supposed to be architectural. Competitors have narrowed the technology gap over the past two years, and Nvidia's answer has been to lean on the asset it has that most of them do not: a balance sheet deep enough to make other people's projects financeable. Its marketable equity securities rose from $12.9 billion to $30.2 billion in a year. The Ohio guarantee is the same instinct expressed at a much larger scale.
It is a rational move, and it is also a bet that the residual value of a GPU in 2032 is knowable today. Nobody has a long enough record of AI accelerator resale prices to say with confidence how that curve behaves, and the guarantee only gets exercised in the scenario where everyone was wrong about demand in the first place. That is the awkward part: the collateral and the borrower are exposed to the same weather.
Pike County will get the construction jobs and the tax base either way, on ground that spent decades enriching uranium for a different kind of national project. The rest of us will find out, sometime around the end of the decade, whether an accelerator ages like a data center or like a phone.
한글 요약
엔비디아가 8월 17일 미국 증권거래위원회 공시를 통해, 오하이오주 파이크 카운티에 들어설 오픈AI 데이터센터 캠퍼스에 최대 1,050억 달러 규모의 잔존가치 보증을 제공한다고 밝혔습니다. 다만 이는 현금 투자가 아니라 완공된 시설과 설비의 가치를 보증하는 구조입니다. 오픈AI는 소프트뱅크 자회사 SB에너지와 20년 임대 계약을 맺었고, 1단계 IT 용량은 4.25기가와트, 옵션을 포함하면 약 8기가와트 규모입니다. 오픈AI가 이탈할 경우 SB에너지가 대체 임차인을 찾고 시설 매각을 시도한 뒤에야 엔비디아가 차액을 부담하며, 그 금액도 오픈AI가 다시 보전하기로 했습니다. 엔비디아는 별도로 SB에너지에 15억 달러를 투자하고 캠퍼스 전반부의 독점 AI 컴퓨팅 공급자가 됩니다.
핵심은 규모보다 성격입니다. 엔비디아가 자사 GPU를 담보 자산으로 취급하기 시작했다는 뜻이기 때문입니다. 하이퍼프레임 리서치의 스티븐 디킨스는 이를 자동차 할부금융에 비유하며, 6~7년간 쓰이는 AI 가속기의 잔존가치를 인정하면 GPU가 감가상각 항목이 아니라 투자 가능한 자산군이 된다고 평가했습니다. 젠슨 황 CEO는 이제 병목이 반도체가 아니라 부지·전력·건물(LPS)이라고 말합니다. AI 기업의 성장 속도가 재무제표를 앞지르면서 생긴 신용 공백을 엔비디아의 대차대조표가 메우는 셈입니다. 세대당 GPU 150만 장, 매출 1,500억~2,000억 달러, 2030년까지 약 12기가와트라는 숫자가 이 구조를 떠받치고 있습니다.
시장 반응은 안도와 불안이 섞여 있습니다. 당초 2,500억 달러 수준까지 검토됐던 보증은 주가가 약 5% 하락하고 투자자 반발이 이어지자 1,200억 달러를 거쳐 1,050억 달러로 축소됐습니다. 월스트리트저널 분석에 따르면 알파벳·메타·마이크로소프트·엔비디아 등 9개 기술기업이 재무제표에 잡히지 않는 AI 관련 약정을 약 3조 달러 보유하고 있으며, 아직 시작되지 않은 리스만 1조 2,000억 달러로 1년 전의 네 배입니다. 모건스탠리는 투자자가 실제 부채 수준을 가늠하기 어려워졌다고 경고했습니다. 다음 확인 지점은 7월 26일 종료 분기 보고서에 첨부될 보증 계약서이고, 첫 800메가와트는 2028년 가동 예정입니다.