The invoice for a trillion dollars of trust arrives in pieces

Deep Read

Who actually holds the risk

Monday morning, Jensen Huang named an asset class into existence. Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, six institutions managing roughly $4 trillion between them, to build "AI compute infrastructure financing platforms" meant to mobilize more than $500 billion of third-party capital for the purchase of Nvidia's own product. His pitch called Nvidia compute "broadly adopted, flexible across models and workloads, fungible and transferable." The same day, Anthropic ran the customer-side version of the identical trade, twice. Theseus Infrastructure, formed with Macquarie Asset Management and GIC, puts Anthropic's US data centers on somebody else's balance sheet as a long-term tenant. And Bloomberg named Anthropic as the likely unnamed counterparty on Riot Platforms' 191-megawatt, 20-year Rockdale, Texas lease worth $9.1 billion in base contract value, up to $16.1 billion with both extensions exercised.

Read those two announcements together and the shape is exact. Every structure separates the user of compute from the owner of the risk underneath it. Labs preparing IPOs convert capex into rent and keep the balance sheet clean for the prospectus. Nvidia insulates its own demand from its customers' cost of capital by arranging the capital itself. The professional credit investors buying this paper are, in theory, doing real underwriting on genuine risk, a step up from Lucent and Nortel financing their own 1990s customers into a receivable that came back worthless. What none of Monday's announcements contained was any residual-value guarantee, offtake term, or price; the platforms remain "subject to execution of final agreements." Where the risk actually landed was always going to be visible only in documents nobody had published yet.

The rest of the week was the search for those documents, one earnings call at a time. CoreWeave reported Tuesday: revenue up 112% to $2.575 billion, ahead of consensus, backlog at $104.2 billion, essentially the Street's own number, well short of Cantor Fitzgerald's higher $131 billion model. The equity market read that half of the release and sent the stock up as much as 16% after hours. The other half: GAAP net loss widened to $626 million, and net interest expense hit $640 million, larger than the loss itself, for the first time. First-half capex ran $14.1 billion against $5.5 billion of cash on hand. A company can grow its revenue, beat every estimate, and still be spending more to service its debt than it is losing on operations, and the stock rewarded it anyway on the day of the print.

Wednesday sharpened the pattern into a genuine split verdict. Cisco booked $9.3 billion in AI infrastructure orders for its fiscal year, clearing its own $9 billion guide, and the market spent a full day deciding the orders number was not the point: non-GAAP gross margin fell to 66.3% from 68.4%, the fiscal 2027 first-quarter guide signaled further compression, and the stock fell 7% even as five separate firms raised their price targets the same afternoon, calling the guidance "robust" and citing a "networking super-cycle." Applied Materials beat its own consensus on revenue and earnings and slipped anyway, on a stock that had already run roughly 200% into the print. Cerebras beat its own guidance across every reported line and fell hardest of the week, 15.7%, because its hardware segment declined while cloud and services grew 287%, and investors are now pricing wafer-scale hardware on the shape of its transition to services. Lenovo, the same afternoon, jumped 19% on a genuinely clean beat: AI product revenue up 60%, an AI server order pipeline that grew 157% quarter over quarter to $54 billion. Four companies beat consensus in the same week. The market paid for growth with expanding margin and charged for growth that cost margin to get, and it did not care what the headline number said.

Underneath all of it, the actual constraint kept repricing. HBM4 now costs Nvidia roughly $31 to 32 per gigabyte against $17 to 18 for HBM3e, and other GPU and ASIC buyers pay more still. Samsung's and SK Hynix's shares rose on the memory scarcity all week, and by Thursday the reported plan was a combined $144 billion of shareholder returns funded by the same HBM cycle that is starving everyone downstream of memory. Nvidia's own workstation GPU, the RTX Pro 6000, crossed $16,000 this week, nearly double its March 2025 launch price, on the same shortage. None of this is the financing complex directly. All of it is the thing the financing complex is a bet on: that the scarcity holds long enough for the leases, the guarantees and the depreciation schedules to clear before anyone has to test what the collateral is actually worth.

The test arrived Saturday, out of sequence with the rest of the week's news cycle, in a Bloomberg story written for bond desks. Roughly $70 billion of residual-value guarantees, backstops by chipmakers promising to cover losses on hardware financed for their own customers, now sit behind AI infrastructure debt and appear on no balance sheet any of the guarantors publish. The anchor case is Project Big Sky, a $35 billion deal in which Broadcom backstops most of the financing that Apollo Global Management and Blackstone are providing for custom chips leased to Anthropic. Bank of America strategists estimate the broader AI XPV platform Broadcom struck in June could carry $370 billion of senior debt by 2029, with worst-case losses near $42 billion in a full default scenario. Nvidia itself guarantees up to 25% of certain projects. Meta has structured roughly $27 billion and $13 billion of data-center packages on similar mechanics. These guarantees are specifically what lifts the underlying paper to investment grade, which means the credit quality of a meaningful and growing slice of the bond market now rests on chipmakers' private promises rather than public liabilities.

The same weekend, in a much smaller story that reads like a tell, Nvidia quietly cut its own guarantee on OpenAI's planned Ohio campus from an initially discussed $250 billion to under $120 billion, while The Information reported Nvidia is now in talks to invest up to $3 billion directly in SB Energy, the developer building the site, half at signing and half at SB Energy's own planned IPO. Guaranteeing a customer's debt is exposure with no upside attached. Owning a piece of the power company at least converts some of that exposure into an asset with a claim on the thing everyone actually needs, which is electricity, not credit. Read against the week that came before it, that one substitution looks less like an isolated negotiation and more like the company that is best positioned to see the whole financing complex quietly deciding it would rather hold equity than a guarantee.

None of this is a bubble call in the lazy sense, and it would be dishonest to make it one. The demand underneath the buildout is real by every measure the week produced: TSMC's third straight record month, Foxconn's AI servers crossing half of revenue, Supermicro guiding 30% above the Street, Lenovo's order pipeline up 157%. The financing structures being built to fund that demand are also, on their own terms, more sophisticated than the vendor-financing schemes that destroyed capital in prior cycles, because professional credit investors are underwriting them with eyes open, after the BIS named AI capex and circular financing a top systemic risk back in June. What is new, and what this week actually revealed, is the size of the gap between what gets announced on a Monday and what gets guaranteed in the documents nobody reads until a bond desk does the arithmetic on a Saturday. Prediction F4 called this divergence in June: that stress in this cycle would show up in credit markets before it showed up in share prices. Five separate earnings prints and one Bloomberg investigation later, that is exactly the order it is arriving in.