Nvidia, OpenAI, and CoreWeave Overlap Contracts, Stakes, and Debt

The Thesis

Nvidia sells chips to CoreWeave. CoreWeave pledges a Meta services contract to borrow from banks, then builds data centers to house Nvidia chips. Nvidia separately invested $30 billion in OpenAI, CNBC reported in March 2026. OpenAI stands to hold roughly $2.6 billion in CoreWeave and Cerebras stock, The Information reported in May 2026, through arrangements tied to buying their chips and cloud time. OpenAI also lent Cerebras money. Customer contracts, equity grants, vendor investments, and contract-backed debt overlap. The records do not trace the same cash around a circle.

That is the financing shape behind the AI capex buildout. Chipmakers, hyperscalers, neoclouds like CoreWeave, and model labs sell to each other, take stakes in each other, and borrow against customer contracts. TEXXR counted more AI data-center capex articles in 2026 Q2 than in 2024 Q1; article volume is not dollars spent or proof that these arrangements tightened every quarter. Bloomberg’s October 2025 report called the deals “circular.” That label describes the overlap. It does not prove a round trip for any dollar.

Nvidia supplies the chips. CoreWeave builds capacity around them. Salesforce is a step removed: its Anthropic stake links its balance sheet to the buildout, while its bond sale shows what creditors charged one software company in the same period. This page tracks the rise in capex coverage, the financing arrangements, and the constraints the record identifies. Whether demand can support the capacity is the arithmetic on the token ceiling. This page does not forecast where any of these stocks go next.

The Evidence

TEXXR’s archive counts articles tagged to AI data-center capex each quarter; it does not measure capital flows. Quarterly volume ran at 85 articles in 2024 Q1. By 2024 Q4 it was 132. By 2025 Q4, 304. By 2026 Q2, 534 — a 6.3x jump in nine quarters, out of 2,879 articles tagged to the topic across 45 observed quarters ending in 2026 Q2. The sharpest single-quarter shift in what that coverage was actually about, measured by semantic drift — the quarter-to-quarter change in the archive’s article-topic mix — lands earlier than the current boom’s timeline suggests: 2020 Q4, when Nvidia’s early supercomputer commitments and pandemic-era cloud spending first pulled the topic away from ordinary data-center announcements. The latest coverage increase is the second inflection, not the first.

The Stargate venture shows how unstable a single mega-deal can be even while the totals climb. OpenAI, SoftBank, and Oracle unveiled the $500 billion project in January 2025. By July, the Wall Street Journal reported it had “struggled to get off the ground” and had sharply scaled back its ambitions. By September, it had expanded again — five new US sites, pushing planned capacity toward 7 gigawatts. By February 2026, The Information was reporting that OpenAI had scrambled for compute as Stargate stalled amid clashes with SoftBank. The headline number survived; the plan under it kept changing.

Nvidia’s OpenAI plans show why a pledge and an investment need separate labels. Nvidia announced an investment of up to $100 billion in September 2025. By January 2026, that plan had stalled in its original form. CNBC reported the March 2026 investment described above; the cited reports do not establish that it replaced the September plan. CNBC reported in May 2026 that Nvidia’s disclosed equity investments across the AI stack had exceeded $40 billion in 2026, including its OpenAI investment.

CoreWeave links a customer contract to an equity grant, then a different customer contract to debt. Its five-year, $11.9 billion OpenAI contract came with $350 million of CoreWeave stock issued to OpenAI at its 2025 IPO. In February 2026, reports said CoreWeave sought bank financing backed by its Meta services contract. Bloomberg reported that the loan closed on March 31, 2026, at $8.5 billion. CoreWeave called it “the largest chip-backed debt deal of its kind.” CoreWeave also uses special-purpose vehicles for construction debt. Those vehicles and the Meta-backed loan are different financing arrangements; the cited loan report does not trace cash between them.

Debt forecasts put a scale on the funding question, not on completed borrowing. Morgan Stanley estimated in August 2025 that hyperscalers would fund $1.4 trillion of a projected $2.9 trillion in AI infrastructure need. The $1.5 trillion remainder is derived by subtraction; it was not all committed debt. In June 2026, Reuters reported that Morgan Stanley had lifted its forecast for global AI-tied debt issuance in 2026 to nearly $570 billion, more than twice its prior forecast for that year. The two estimates cover different measures and periods. Upstream, TSMC raised its 2026 capex guidance in July from $52–56 billion to $60–64 billion. A step removed, Salesforce sold $25 billion in bonds in March 2026. The Financial Times said investors demanded a significant premium to take part; that bond pricing does not isolate AI risk from other risks to Salesforce.

The record also carries the counter-evidence. The Wall Street Journal’s November 2025 accounting of the buildout put 2025’s US data-center capacity — built, underway, planned, or stalled — above 80 gigawatts, running into what it called the “ground truths of physical” limits: power, water, chips, and local opposition that don’t scale as fast as balance sheets do. One economist’s estimate, cited in a January 2026 Financial Times piece, put the whole AI investment boom at roughly 1% of US GDP — comparable in scale to prior investment booms, which is a benchmark, not a verdict on how this one ends.

The Companies

Nvidia sits at both ends of the loop. It supplies the chips every other name here is spending to acquire, and it increasingly funds the customers buying them: the OpenAI equity stake, the CoreWeave stake, the more than $40 billion in 2026 equity commitments across the stack. When Jensen Huang said an investment “might be the last time” Nvidia writes a check that size, TEXXR’s record treated the remark as news, not reassurance.

CoreWeave is the neocloud in the middle — Nvidia-backed since before its 2025 Nasdaq IPO, running 43 active data centers as of April 2026 (the same month it signed a multiyear Nvidia-chip supply deal with Anthropic), and holding a $99.4 billion revenue backlog against contracts with OpenAI, Meta, and Anthropic. It is also the clearest single case of the loop’s financial engineering: debt moved off its balance sheet through special-purpose vehicles, then more debt raised against the very hyperscaler contracts that are supposed to prove its demand is real rather than circular.

Salesforce isn’t building data centers. It’s the second-order name: a roughly $5 billion equity stake in Anthropic ties its balance sheet to the same cycle, its Agentforce pricing is shifting from flat subscriptions toward usage-based AI fees as compute costs eat into software margins, and its March 2026 bond sale — priced at a premium — is the record’s clearest sign that AI-linked credit risk has reached investment-grade enterprise software, not just the chipmakers and clouds.

The Lenses

Benjamin Graham asks whether each deal works on its own cash flows. Security Analysis rejects new-era arguments that suspend ordinary underwriting. Applied here, each contract needs a price, funding cost, useful life, and enforceable customer obligation. Revenue backlog and announced capacity cannot substitute for those inputs. The Intelligent Investor adds the margin-of-safety test: assumptions need room to fail without wiping out the capital structure.

Robert Shiller asks how the story alters the financing environment. Narratives spread through markets and feed back into behavior. Bloomberg put “circular deals” into circulation in October 2025. Banks, regulators, and investors can apply that label to separate transactions before any filing proves that cash literally completed a circle. Bloomberg’s October 2025 report does not show whether the label changed loan terms or capex plans.

What Moved

Actual capex puts the funding question in numbers. The DeadRisk market layer records Applied Digital FY2026 capex, for the year ended May 31, 2026, at $2.9 billion against $89.7 million of operating cash flow — about 32 times, derived by division. FY2025 capex was $681.6 million against negative $115.4 million of operating cash flow. Its reported debt rose from $910.8 million at FY2025 end to $5.1 billion at FY2026 end. Those aggregates show a need for funding beyond operating cash; they do not identify who financed which assets. Separately, Financial Times sources said Amazon had discussed spinning $8 billion of Grace Blackwell chips into a special-purpose vehicle, then leasing them back for its US data centers. Those were talks, not a closed financing deal. Amazon’s FY2025 capex was $131.8 billion versus $83.0 billion in FY2024, while free cash flow fell to $7.7 billion from $32.9 billion, according to the DeadRisk market layer. The proposed vehicle’s $8 billion asset size exceeds that earlier year’s free cash flow; it is not FY2025 borrowing. CoreWeave’s $8.5 billion Meta-backed loan, detailed above, did close. The two records describe different transactions, not a traced cash circuit.

Anthropic’s IPO filing puts a counterparty flow beside those funding arrangements. Reuters reports that Anthropic routed 47% of its 2025 sales, about $2.16 billion, through cloud partners Amazon and Google. Reuters’s analysis puts distribution fees paid back to those partners at about $351 million. That is 16.3% of routed sales, derived by division; Reuters does not split either amount by partner. A distributor receiving fees, a supplier offering an equity right, and a lender taking contract collateral can all sit near the same buildout. These records do not show the same cash completing a circle.

Keep rights and plans separate from cash. Qualcomm’s AWS partnership includes Amazon warrants covering $4 billion of Qualcomm stock; CNBC did not report an exercise or cash outlay. OpenAI’s Samsung chip work has no reported spend or delivery schedule. Samsung and TSMC’s ASML commitments target 2028 and 2030 machine use, not installed 2026 capacity. The Wall Street Journal reported that more than 10 US states have rolled back tech tax breaks; it did not document canceled capex. Amazon says it stopped using county NDAs for data-center projects. GeekWire reports Amazon’s pledge to spend more than $1 billion over five years in hosting communities. A pledge and a permitting constraint are not completed capacity.

Physical supply has its own clock. South Korea’s $919 billion plan is planned AI-infrastructure investment, not outlay. SK Hynix’s Indiana groundbreaking began a $4 billion packaging project in August 2026; Bloomberg places HBM mass production in the second half of 2029. Corning’s 2026 Q2 sales rose 17% against 2025 Q2 to $4.74 billion, while Corning reported production limits on AI connectivity. Microsoft’s Project Zenith would run large models locally; its announcement does not establish displaced data-center demand.

Supplier alternatives do not yet settle the cash-flow test. SemiAnalysis’s Jalapeño tests cover specified open-weight models, not every workload. OpenAI’s Ultrafast preview puts Cerebras in an API tier, while AMD and Cerebras are integrating server racks. Cerebras’s 2026 Q2 revenue was $180 million, up 74% from 2025 Q2 but below the $194.23 million estimate Reuters cited. Reported profit needs its own check: Amazon and Alphabet’s other income, mostly investment gains, totaled about $121 billion after tax in 2026 Q2; the Wall Street Journal put that at 66% and 71% of their respective quarterly profits.

• 2026-07-18Pillar launch. TEXXR counts 534 AI data-center capex articles for 2026 Q2, compared with 85 in 2024 Q1.
• 2026-06-10Morgan Stanley forecasts nearly $570 billion of global AI-linked debt issuance in 2026, more than twice its prior forecast.
• 2026-03-05CNBC reports Nvidia’s recent $30 billion OpenAI investment. That amount is $70 billion below the September 2025 plan’s ceiling, based on subtraction; the cited reports do not call it a replacement.

Sources