On the grounds of a decommissioned Cold War uranium-enrichment plant in Piketon, Ohio, SoftBank's SB Energy is pouring the shell of what will be one of the largest computing campuses ever attempted. OpenAI has signed a twenty-year lease on the site, known as PORTS-Pike, with an initial buildout of 4.25 gigawatts and an option to reach 8 (AIStockWire, Aug 17). The rent is enormous and the tenant has never turned an operating profit. So who signed as guarantor? Nvidia did. On August 17 the chipmaker disclosed, through a securities filing, that it will stand behind up to 105 billion dollars of defined lease and power payments plus a residual-value commitment on the campus, phasing in as capacity comes online between 2028 and 2030 (AIStockWire, Aug 17; TechJournal, Aug 19).
Here is the contradiction that cannot hold. Nvidia books each generation of chips destined for the site at roughly 1.5 million GPUs and 150 billion to 200 billion dollars of revenue, and frames OpenAI's total compute commitments at about 600 billion dollars through 2030 (AIStockWire, Aug 17). Oracle, meanwhile, holds a separate five-year, 300 billion dollar cloud contract with OpenAI that implies roughly 60 billion dollars a year (Finviz, September 2025). Yet the revenue on both sets of books exists only because the same two companies have quietly agreed to carry the credit risk of the customer generating it. Nvidia is not merely selling chips; it is insuring its own sales. Oracle is not merely renting compute; it is holding fifteen-year property obligations against a client that can walk after about five years (Implicator.ai, Dec 13, 2025).
The trigger this week was the Ohio filing. The pressure underneath runs back further. Oracle raised 18 billion dollars in new debt in September 2025, and by the end of November reported owing more than 124 billion dollars including operating-lease liabilities (Wyoming Star News, Dec 18, 2025). Its backlog swelled past half a trillion dollars almost entirely on contracts with one unprofitable buyer. Then the private capital that was supposed to spread the risk started declining the invitation: Blue Owl Capital walked away from funding a planned 10 billion dollar Stargate data center in Saline Township, Michigan, leaving Oracle hunting for replacement financing (Financial Times, Dec 17, 2025). The bond market drew its conclusion before the equity market did. The cost of insuring Oracle's debt against default hit a near eighteen-year high in July (Bloomberg, Jul 20, 2026).
Name the actors and their wants. Larry Ellison's Oracle wants to convert its database cash cow into the landlord of the AI age before its core business ages out; the cost is a balance sheet stretched thin enough that the company cut roughly 30,000 jobs this year while spending on data centers continued (Tech Insider, August 2026). Jensen Huang's Nvidia wants every frontier lab building on Nvidia silicon, and has concluded that the bottleneck is no longer chips but sites, so it now supplies the credit that gets sites built. Sam Altman's OpenAI wants more compute than any balance sheet on earth could fund directly, so it signs leases others must finance. SoftBank's SB Energy wants the fees of owning infrastructure someone else insures. Each party is acting rationally. Together they have built a structure where the same dollars appear as chip revenue at Nvidia, capacity backing revenue at OpenAI, and contracted backlog at Oracle.
The history that fits is Lucent Technologies in the late 1990s. After Washington opened local telephone markets in 1996, a swarm of underfunded telecom carriers ordered switching equipment they could not pay for, and Lucent lent them the money to buy its own gear, booking the sales as revenue. For three years the income statement looked magnificent. When the carriers began failing in 2001, Lucent's receivables and loan book collapsed with them, and a company that had been the most widely held stock in America lost nearly everything. The mechanism there is identical here: vendor financing converts a supplier into an undisclosed creditor of its own customers, and reported growth outruns collected cash until it cannot.
The counterexample argues the other way, and it is Nvidia's own defense. Boeing and Airbus have financed aircraft buyers for decades without ruin because an airplane is a fungible asset: if the airline defaults, the plane flies for someone else next month. Nvidia makes exactly this claim about PORTS-Pike, arguing its support covers only land, power and shell, that its exposure shrinks as OpenAI pays, and that a warehouse of CUDA-running systems is re-rentable across its ecosystem rather than tied to one shaky tenant (AIStockWire, Aug 17). That argument holds if AI demand keeps compounding and the hardware stays general-purpose. It breaks if a whole generation of accelerated-computing campuses hits the market at once, the way empty fiber did in 2002, when even fungible assets find no second tenant because everyone built the same thing in the same years.
Walk the chain forward. First order: Oracle's borrowing costs stay elevated, so every new gigawatt costs it more to finance than the last, which is why private lenders like Blue Owl are becoming choosier about which Stargate sites get funded (Financial Times, Dec 17, 2025). Second order: the burden migrates to the strongest balance sheet in the room, Nvidia, whose guarantees multiply with every deal its critics call circular and which Huang has dismissed by name (AIStockWire, Aug 17). Third order: the risk lands on people nowhere near the deal, including the ratepayers and county around Piketon, where SB Energy and SoftBank have committed to at least 10 gigawatts of new generation and at least 4.2 billion dollars of grid work with American Electric Power's Ohio utility (AIStockWire, Aug 17). If the leases hold, those communities get a twenty-year industrial base. If they break, the region inherits the largest stranded asset in Ohio since the enrichment plant itself closed.
The observable sequence is straightforward. If the read is right, Oracle's credit-insurance costs stay elevated or worsen into its next results, and future Stargate sites keep losing or changing private funders the way Michigan did. If the read is wrong, OpenAI's revenue keeps doubling fast enough that lease payments retire Nvidia's exposure on schedule, and Oracle's insurance costs fall back toward normal levels within two quarters.
What nobody can dispute is where the consequence would concentrate. The profits sit on the books of Nvidia and Oracle today. The obligation sits in twenty-year leases, fifteen-year property terms and a hundred-billion-dollar guarantee that only matters if something goes wrong. Markets price what appears in revenue. They consistently misprice what appears in a footnote.
The profits sit on the books of Nvidia and Oracle today; the obligation sits in a footnote waiting for something to go wrong.
Method. This analysis rests on the sources cited below. ARCANE does not publish a proprietary universe, cohort weighting or exclusion list for this piece — the reading is the desk's, argued from the record, not a screened back-test.