On August 17, 2026, NVIDIA filed a disclosure with the SEC that quietly redrew the map of AI infrastructure finance. The company guaranteed up to $105 billion in lease and power payments for a single data center campus in Pike County, Ohio, to be leased exclusively by OpenAI for 20 years. The facility, developed by SoftBank-owned SB Energy and called the PORTS-Pike Technology Campus, will scale to 8 gigawatts of IT capacity — enough electricity to power roughly six million American homes. The first 800 megawatts are scheduled to come online in 2028, with full buildout extending through 2032.
The same week, Reuters reported that Anthropic — OpenAI's closest competitor — told IPO bankers it projects $190 billion to $200 billion in 2028 revenue, a two-years-forward forecast underpinning a target valuation approaching $2 trillion. Anthropic is simultaneously in advanced talks to acquire Israeli AI infrastructure startup Decart for roughly $6 billion, a deal aimed at squeezing more performance out of every chip it already buys.
Two stories, one signal. The AI industry has stopped pretending compute is a cost line. Compute is now the asset, the collateral, the product, and the debt — all at once. For anyone building, buying, or financing AI infrastructure, the Ohio deal and the Anthropic IPO forecast are not background noise. They are the new baseline.
The Ohio Deal: What NVIDIA Actually Guaranteed
The SEC filing corrected earlier reporting that had placed NVIDIA's Ohio commitment as high as $250 billion. The actual figure is a guarantee ceiling of $105 billion. NVIDIA also made a direct $1.5 billion equity investment in SB Energy, the SoftBank subsidiary developing the site. OpenAI signed a 20-year lease for the full campus.
The structure matters. NVIDIA is not writing a check for $105 billion. It is guaranteeing that OpenAI's lease and power payments to SB Energy will be met. If OpenAI pays its rent from its own revenue — as Jensen Huang insists it will — the guarantee never activates. If OpenAI defaults, NVIDIA is on the hook for up to $105 billion. The deal also makes NVIDIA the exclusive chip provider for the facility. Every rack in Ohio will run on NVIDIA silicon. The guarantee is, in effect, a vendor financing arrangement: NVIDIA is backing its largest customer's lease on a building that will be filled with NVIDIA's own product.
The scale is unprecedented. Eight gigawatts is roughly four times the capacity of the largest existing data center campuses in the world. The project is expected to create 35,000 construction jobs during its six-year buildout and 2,500 long-term operating jobs. The land and power agreements are already secured. SB Energy has the site. NVIDIA has the chips. OpenAI has the lease. The only thing none of them have yet is the revenue to pay for it.
The Circular Financing Critique — and Huang's Rebuttal
The deal immediately drew comparisons to the $500 billion Wall Street AI compute financing arrangement NVIDIA announced on August 10 — the deal we covered last week, the one Michael Burry called “shades of Enron.” The critique is straightforward: NVIDIA is financing its customers' purchases of NVIDIA products, effectively creating demand for its own chips with its own balance sheet. The customer pays NVIDIA back from revenue that depends, in part, on running workloads on the chips NVIDIA sold them. If the AI revenue does not arrive, the whole stack unwinds.
Jensen Huang addressed the circular financing charge directly in an August 17 blog post titled “Securing the Infrastructure of Intelligence.” His argument rests on a single load-bearing assumption: OpenAI will pay the lease from its own revenue. The guarantee only activates on default. NVIDIA is not lending money to OpenAI to buy chips. NVIDIA is guaranteeing OpenAI's lease payments to a third-party developer (SB Energy). The chips are purchased separately. The revenue, Huang argues, is real and growing — OpenAI's run rate has grown more than 10-fold annually for three consecutive years.
The rebuttal is stronger than critics allow, but weaker than NVIDIA would prefer. The guarantee structure is genuinely different from a direct vendor loan. But the dependency is still circular in aggregate: OpenAI's ability to pay the lease depends on AI inference and training revenue that is not yet proven at the scale required to service an 8-gigawatt campus. NVIDIA's $105 billion exposure is a bet that AI revenue in 2028 and beyond will be large enough to cover the operating cost of one of the largest power consumers on Earth. That bet is not irrational — but it is a bet, not a certainty.
The Anthropic Counter-Bet: $200 Billion in 2028
The same week NVIDIA was filing its Ohio guarantee, Anthropic was telling IPO bankers to value it on 2028 revenue, not 2026 revenue. The Reuters report put Anthropic's internal forecast at $190 billion to $200 billion in 2028, supporting a target IPO valuation of approximately $2 trillion. Anthropic filed confidentially with the SEC on June 1 and is targeting a September or October 2026 public listing.
The valuation method is unusual. Bankers and investors are applying an enterprise value-to-revenue multiple to 2028 forecasts — two years forward — rather than current or trailing revenue. That only works if you believe the 2028 number is credible. Anthropic's argument: its revenue run rate has grown more than 10-fold annually for three years through early 2026, the enterprise AI market is accelerating, and Claude's deployment footprint in coding, enterprise automation, and agentic workflows is expanding into long-term contracts that compound.
The Decart acquisition talks are the operational complement to the financial story. Decart, an Israeli startup founded in 2023 by Dean Leitersdorf and Moshe Shalev, does two things: it builds AI “world models” trained on video and physical-world data, and it builds a chip optimization stack that helps AI developers “squeeze every ounce of performance from every chip.” A $6 billion acquisition would be Anthropic's largest ever, and it would arrive precisely as the company needs to show public investors a credible path to 77% gross margins. If you can cut inference cost per token by optimizing how your workloads run on existing silicon, your margin expands without buying more GPUs. That is the Decart thesis.
Why Compute Is Now Collateral
The Ohio deal and the Anthropic IPO forecast are two sides of the same coin. For the past two years, AI infrastructure has been financed on the assumption that compute capacity will generate revenue at a multiple that justifies the upfront cost. In 2026, that assumption became the dominant financial instrument in the technology sector.
The pattern is now visible across the stack. NVIDIA guarantees leases and provides vendor financing for the buildings that house its chips. OpenAI leases the buildings and commits to 20-year contracts backed by its inference revenue. Anthropic projects $200 billion in 2028 revenue to justify a $2 trillion valuation and buys a chip optimization company to protect the margins that make that revenue worth more. SoftBank, through SB Energy, provides the land and power, and takes the lease payments. Every party is making a bet on the same variable: that AI inference revenue in 2028 and beyond will be large enough to service the debt.
This is the infrastructure debt problem. It is not that the debt is necessarily bad — infrastructure has always been debt-financed. Railroads, telecom networks, power plants, and cloud data centers were all built on long-duration financing against future revenue. The difference is the certainty of the revenue. A railroad in 1880 carried freight that had an existing market. A cloud data center in 2012 served an enterprise IT market that was already enormous. An AI data center in 2028 will serve a market that does not yet exist at the scale being financed. The revenue is plausible, not proven.
For builders and operators, this matters in a specific way. If the 2028 revenue arrives, the current capacity shortage will look like a massive under-build and GPU prices will stay high. If it does not arrive at the projected scale, the over-build will collapse chip prices, strand assets, and — in the worst case — trigger defaults on guarantees like NVIDIA's $105 billion Ohio commitment. Your hardware procurement strategy in 2026 and 2027 should account for both outcomes.
What Every AI Infrastructure Team Should Do Now
Whether you are running a homelab inference cluster or procuring enterprise GPU capacity, the Ohio deal and the Anthropic forecast change the planning calculus. Here is what to do.
1. Do not bet on GPU prices falling in 2026. The Ohio facility's first 800 megawatts do not arrive until 2028. The Stargate project and the SoftBank-backed campuses are still in buildout. Through 2026 and 2027, enterprise GPU capacity remains supply-constrained. If you need compute now or in the next 12 months, lock it in. Waiting for the Ohio capacity to flood the market is a 2028 strategy, not a 2026 one. For local inference today, the consumer NVIDIA RTX 5080-class cards remain the practical entry point for 70B-parameter-class models; the data center capacity being financed this week will not lower your local hardware costs for at least two years.
2. Treat inference cost as the variable that determines whether the debt gets serviced. The entire $105 billion Ohio guarantee, the Anthropic $2 trillion valuation, and the Decart acquisition thesis all reduce to one number: cost per inference. If inference costs fall faster than expected (through better optimization, quantization, or dedicated inference silicon like the Taalas and Fractile chips we covered earlier this month), the 2028 revenue forecasts become easier to hit and the debt is serviceable. If they do not, the revenue forecasts are at risk. Track inference cost per token for your own workloads. It is the leading indicator for the entire AI infrastructure market.
3. Plan for the over-build scenario. The aggressive buildout — 8 GW in Ohio, 10 GW in Stargate, multiple SoftBank campuses — is a bet that 2028 demand will absorb the capacity. If you are making long-term hardware commitments now, structure them with flexibility. Avoid 20-year leases on capacity you cannot sublease. Prefer cloud GPU commitments with exit clauses over rigid capacity reservations. The homelab equivalent: buy the GPU that runs the models you have today, not the one you think you will need in 2028. The 2028 GPU market is the one this financing is designed to reshape.
4. Watch the Anthropic Decart deal as an inference-cost signal. If Anthropic closes the $6 billion Decart acquisition and integrates the chip optimization stack into its Claude inference pipeline, it is a signal that the largest AI labs believe the path to margin runs through inference efficiency, not just more GPUs. That would put downward pressure on inference pricing across the market — good for consumers, challenging for the revenue forecasts underpinning the Ohio guarantee. If the Decart deal falls through, read it as a signal that Anthropic sees the 2028 revenue arriving through raw capacity expansion rather than efficiency gains. Either way, the deal is a leading indicator for the inference-cost curve that determines whether the $105 billion bet pays off.
The Bottom Line
The week of August 17, 2026 may be remembered as the moment AI infrastructure finance matured — or the moment it overextended. NVIDIA guaranteed $105 billion against an 8-gigawatt campus in Ohio that will not fully come online until 2032. Anthropic told bankers to value it on 2028 revenue that has not been earned yet. SoftBank, SB Energy, OpenAI, and NVIDIA are now a single financial stack where each party's solvency depends on the same variable: AI inference revenue in 2028 and beyond.
None of this is inherently reckless. Infrastructure has always been financed against future demand. But the gap between the financing and the revenue has never been this wide for a market this new. The $105 billion Ohio guarantee is a rational bet if the 2028 AI revenue arrives at projected scale. It is a catastrophic one if it does not. For everyone building in AI — from the homelab operator running local LLMs to the enterprise team procuring GPU capacity — the implication is the same: the next 24 months will determine whether the AI infrastructure debt is an asset or a liability. Track inference cost. Lock in capacity you need now. And keep your 2028 options open.
Get weekly AI & security infrastructure guides
Join the GeniusTechLab newsletter for AI infrastructure breakdowns, security analysis, and hardware recommendations — one email a week, no spam.
Subscribe to the newsletter →