On the Evening of 6/10, the Market Gave a Contradictory Answer
On 6/10/2026, Oracle reported its fiscal 2026 fourth quarter: remaining performance obligations (RPO — the order book that is signed but not yet delivered) reached $638 billion, up 363% year over year (Oracle press release). No public disclosure from a large software peer shows an order book of that magnitude.
The stock's response was a 10% drop after hours (CNBC), and a decline of about 35% across month 6 as a whole (Motley Fool). By 7/6, Oracle closed at $143.76, down about 58% from its 9/2025 peak of $345.72. That same day, at the other end of the loop, Micron closed at $984.75 and SK Hynix traded back near the all-time high it set early in month 6 — both memory stocks sitting comfortably above 1.8 times the top of our research reports' bull-case valuation bands.
One AI supply chain, one end hugging a one-year low, the other hugging all-time highs. To make sense of that split, you first have to answer the question the market has argued about for the better part of a year: is the ring of cross-investments and cross-purchases among NVIDIA, OpenAI, Microsoft, Oracle, and CoreWeave a stroke of genius in a scarcity market, or a Ponzi that books revenue in a circle?
We laid out every dollar and every public document, and counted.
First, Lay Out Where the Money Flows
The charge of circular financing sounds complicated; drawn on paper it is a single ring: NVIDIA puts money into its customers, the customers pay cloud providers, and the cloud providers pay the money back to NVIDIA for chips. Checked deal by deal against primary documents, the ring actually consists of the following:
| Money flow | Amount and nature | When | Status and source |
|---|---|---|---|
| NVIDIA → OpenAI | Intent to invest "up to" $100 billion, disbursed in tranches per gigawatt deployed | 2025-09 | Letter of intent (LOI); stalled as of 2026-02, see next section (NVIDIA newsroom) |
| OpenAI → Microsoft | $250 billion Azure cloud purchase commitment; Microsoft simultaneously holds 27% of OpenAI | 2025-10-28 | Restructuring agreement in effect (Microsoft official blog) |
| OpenAI → Oracle | Roughly $300 billion, five-year cloud compute contract (2027–2032) | 2025-09 | As reported; Oracle's $638 billion RPO corroborates the magnitude (CNBC) |
| OpenAI → CoreWeave | Cumulative compute leases of roughly $22.4 billion (including a $6.5 billion add-on in 2026-01) | Cumulative since 2023 | As reported (TechCrunch) |
| NVIDIA → CoreWeave | A further $2 billion equity investment; separately, a commitment to purchase its unsold capacity through 2032-04-13 | 2026-01; 2025-09 | Investment as reported; the buyback commitment is in an SEC 8-K (SEC) |
| NVIDIA → cloud providers | Multi-year cloud service purchase commitments totaling $30 billion, scheduled past fiscal 2032 | As of 2026-04-26 | NVIDIA 10-Q (NVIDIA IR) |
Start with what this table confirms: the money genuinely moves, the contracts are genuinely signed, and SEC filings disclose down to contract terms. A Ponzi by definition requires fabricated returns and new money paying off old money — that does not fit here. Calling this a Ponzi is rhetoric, and a misdiagnosis.
Then note what the table exposes: every node's revenue in the ring presupposes some other node's capital spending, and the node spending the most (OpenAI) is itself still losing money. This is vendor financing — the supplier puts up money so the customer can afford to buy. The term has a well-documented rap sheet in history, and we settle that old account below.
Three Numbers Everyone Still Quotes Whose Terms Have Expired
In the fight over this ring, the three numbers cited most often by bulls and bears alike have all changed underneath them. This is also where this piece differs most from the wave of commentary from the fall and winter of 2025.
First: "NVIDIA is putting $100 billion into OpenAI." The original announcement in 9/2025 said NVIDIA "intends to invest up to $100 billion" — an intent capped at $100 billion, disbursed in tranches as OpenAI deploys each gigawatt of NVIDIA systems — a milestone-conditioned intent from day one, signed as a letter of intent. In late 1/2026, The Wall Street Journal reported the plan had stalled; on 2/1, Jensen Huang responded in person in Taipei: "It was never a commitment," adding that "we will invest one step at a time" (Bloomberg, Fortune). Per CNBC, as of late 2/2026 the NVIDIA amount actually under discussion was "up to $30 billion," as part of an OpenAI round exceeding $100 billion at a pre-money valuation of roughly $730 billion (CNBC). From $100 billion to the $30 billion tier — a seventy-percent shrink, and the largest intent-level crack that has already opened anywhere in the loop.
Second: "OpenAI has signed $1.4 trillion in compute commitments." That widely circulated sum comes from media tallies around 11/2025. By 2/2026, the figure OpenAI gave investors had become: roughly $600 billion in total compute spending through 2030 (CNBC). The two numbers measure different things (one is nominal signed total, the other an actual spending target) and cannot simply be netted against each other, but the direction is unambiguous: the company is cooling its own commitment sheet.
Third: "OpenAI lost $38.5 billion in one year." The number is real, but it has to be taken apart. Per audited financial documents that leaked and were verified by the Financial Times: OpenAI's 2025 revenue was $13.07 billion (versus $3.7 billion in 2024 — an increase of roughly 2.5 times in a single year), total costs and expenses were $34 billion, and the operating loss was $20.92 billion; the bulk of the $38.5 billion net loss was a $41.55 billion one-time non-cash loss from fair-value changes in convertible interests and warrant liabilities during the conversion from nonprofit to for-profit (Fortune). Meanwhile, per CNBC, actual cash burn in 2025 was about $8 billion, below the original $9 billion target. Bears quote the $38.5 billion, bulls quote the $8 billion, and both are picking the figure that suits them; the honest operating-loss run rate is the roughly $20.9 billion-a-year tier — revenue is tripling, and the losses are real too.
NVIDIA's 10-Q: Four Roles Played by the Same Company
Commentary can argue about definitions; statutory filings cannot. NVIDIA's fiscal 2027 first-quarter report (10-Q, for the quarter ended 4/26/2026) is the hardest document in this whole debate, and we checked the original text paragraph by paragraph (NVIDIA IR). Four numbers deserve to be laid out on their own:
Revenue concentration: 54%. Total quarterly revenue was $81.6 billion (up 85% year over year), of which the top three direct customers accounted for 21%, 17%, and 16% — 54% combined, roughly $44.1 billion. In the same quarter a year earlier, only two direct customers exceeded 10%, at 30% combined. Concentration nearly doubled in a year. The 10-Q also carries a line disclosing the indirect loop: "We estimate that one AI research and deployment company contributed to a meaningful amount of our revenue by purchasing cloud services from our customers" — no name given, none needed.
Outbound investment in a single quarter: $17.9 billion. Net additions to non-marketable equity investments were $17.9 billion for the quarter; a year earlier the figure was $649 million. The investment book swelled from $3.39 billion in 4/2025 to $22.25 billion at the end of 1/2026, plus this quarter's net additions — more than tenfold in a little over a year. The 10-Q states plainly that investees include "AI model makers," that is, companies that may indirectly buy NVIDIA products.
Lease guarantees: $3.5 billion. Starting in fiscal 2026, NVIDIA began guaranteeing partners' data-center facility lease obligations against default, in exchange for warrants; maximum gross exposure is $3.5 billion over terms of 5 to 7 years, with $712 million placed in escrow by the partners.
Compute buyback: $30 billion. NVIDIA itself has signed multi-year cloud service purchase commitments totaling $30 billion, paid out as $6 billion, $7 billion, $7 billion, $5 billion, $3 billion, and $2 billion through fiscal 2032 and beyond — selling chips to cloud providers, then renting the compute back from them. The 10-Q's own text adds: "Cloud service capacity may be reduced or terminated."
Read the four numbers together: on the same chain, NVIDIA simultaneously plays supplier (selling the chips), shareholder (investing in customers' equity), guarantor (backing customers' leases), and leaseback customer (buying customers' compute). Each role has business logic on its own; stacked, the four mean this: customers' credit risk is moving, batch by batch, from customers' balance sheets onto NVIDIA's.
The Ledger from 1999
Vendor financing last took the stage at scale in the telecom bubble of 1999–2001. Per the tally compiled by investor Tomasz Tunguz: Lucent committed $8.1 billion in customer financing, Nortel extended $3.1 billion (of which $1.4 billion was left uncollected), and Cisco promised $2.4 billion in customer loans; it ended with 47 competitive local exchange carriers (CLECs) going bankrupt between 2000–2003, and 33% to 80% of the vendors' loan portfolios going unrecovered (Tunguz). Peak telecom capital spending in 2000 was about $120 billion (roughly $213 billion in 2023 dollars), with cumulative investment above $500 billion (Fabricated Knowledge).
The similarities are glaring: suppliers levering up their customers, demand figures citing one another, and the order book traded as a leading indicator of revenue. Bloomberg's bear-case verdict, in its map of this round of AI circular deals, points the same way: the interlocking structure skews incentives, and if end demand falls short, losses get amplified along the chain (Bloomberg). The Next Platform used the word outright: "round-tripping."
The differences are just as glaring, and an honest comparison has to lay them out. The CLECs borrowing back then were shells with next to no revenue; the core nodes of today's loop are Microsoft, Oracle, and NVIDIA — giants earning tens of billions to over a hundred billion dollars a year in net profit. Back then the loans largely became bad debt; today the cash genuinely moves and the compute genuinely ships. Back then disclosure was a fog; today the concentration, guarantee exposure, and buyback schedules are all written into the 10-Q. The representative bull argument comes from asset manager Janus Henderson: in a market where advanced chips are scarce and AI buildout is extraordinarily expensive, long-term purchase commitments plus financing are rational supply lock-in — a "virtuous circle" (Bloomberg).
So what the ledger of 1999 leaves for today is one concrete question: the moment end demand is disproven, who is holding the concentrated risk? Back then, end demand meant dial-up internet users; today it means the monetization of the OpenAIs — annualized revenue of roughly $25 billion (as of late 2/2026, per The Information) set against a spending target of roughly $600 billion through 2030. The gap is filled by financing, the financing is held up by the story, and the story gets renewed by the next quarter's deployment milestone.
The Evidence in Our Own Library: the Market Is Already Pricing the Two Ends Apart
This site holds two things few others do, and they add a quantitative corner to this debate.
The first is scorecards. Late in month 5 we scored the loop and its first-tier supply chain — 9 companies (NVIDIA, Microsoft, Oracle, AMD, SK Hynix, Micron, Vertiv, CoreWeave, Super Micro) — on the Baillie Gifford growth-investing Ten Questions, where Q9 ("what expectations are priced into today's share price") and Q10 ("what has the market not yet realized") are built to measure pricing perception. The results line up with the memory piece from the middle of month 6 with cold consistency: across 9 companies and 18 cells, none scored above 3 (out of 10, Baillie growth leaderboard). Our Oracle research report wrote on 5/20 that the market had "fully, even excessively priced" the AI narrative, and that the real open question was "whether customer concentration, low margins, negative free cash flow, and high leverage are being underestimated"; the CoreWeave report's verdict was that "the market has priced the flower as if its future bloom were certain."
The second is valuation bands. Every research report carries a conservative/base/bull three-scenario DCF range. Put the 7/6 closing prices back into the month-5 bands and the split is plain:
| Company | Report snapshot price (month 5) | Price (7/6) | Change | Which band the price sits in now |
|---|---|---|---|---|
| Oracle | 181.46 | 143.76 | −21% | Back inside the base-case range (110–150) |
| CoreWeave | 99.81 | 86.46 | −13% | Below the bull-case range's lower edge (95) |
| Microsoft | 423.54 | 386.74 | −9% | Just below the base-case range's lower edge (390) |
| NVIDIA | 219.98 | 195.55 | −11% | Still inside the bull-case range (180–220) |
| Micron | 698.74 | 984.75 | +41% | About 89% above the bull-case range's top |
| SK Hynix | ₩1,940,000 | ₩2,343,000 | +21% | About 80% above the bull-case range's top |
Same month 5, same method, and two months later the market handed back two opposite sets of readings: it is discounting the levered, borrow-to-build-compute end of the loop — Oracle has given back the entire premium from the day its contract announcement in 9/2025 jumped the stock 42% in a session, falling back into our base-case range — and it keeps paying up for the physical-scarcity end — the two memory names stand at 1.8 times the top of their bull-case ranges. The market did not wait for the debate to settle; it is already using price to separate "credit risk inside the loop" from "physical bottleneck outside the loop."
One more piece of circumstantial evidence comes from our mention network: across the 998 research reports in our library, NVIDIA is mentioned in 55 — the biggest hub in the entire dataset; CoreWeave's mention count is zero. A company this central to the circular-financing narrative has left no footprint in a fundamentals research network spanning 919 tickers — it handles GPUs and money, yet produces almost nothing that anyone else's research needs to cite. A pipe, by itself, is a hard place for value to settle.
Where the Crack Is Most Likely to Start
Making "when does the loop crack" concrete, ranked by the evidence in front of us:
First in line: Oracle's financing dependence. Fiscal 2026 capital spending was $55.7 billion (up 162% year over year), operating cash flow $32 billion, free cash flow −$23.7 billion; the fiscal 2027 financing plan is $40 billion, double the prior expectation (Oracle results, CNBC). Converting a $638 billion order book rests on the bond market staying willing to supply blood at acceptable spreads. As reported, after news of the NVIDIA-OpenAI stall broke in month 2, Oracle publicly stated the matter had "zero impact" on its financial relationship with OpenAI and that it remained "highly confident" in OpenAI's ability to meet its commitments — and the very need to come out and declare "zero impact" is the question the market is asking.
Second in line: CoreWeave's leverage and guarantee triggers. NVIDIA's buyback commitment on CoreWeave's unsold capacity runs to 4/2032, and the lease-guarantee exposure is $3.5 billion. If AI compute rental prices weaken, the guarantees and buybacks turn from endorsement into cash outlay, and the loop's cost turns explicit for the first time.
Third in line: OpenAI's monetization pace. Roughly $25 billion of annualized revenue against a roughly $600 billion spending target through 2030 — the gap in between takes round after round of financing. The first hard test is already on the calendar: per NVIDIA's official announcement, the partnership's first gigawatt of systems is slated to deploy in the second half of 2026 on the Vera Rubin platform. On schedule, the loop rolls over; delayed, and the $30 billion-tier investment talks and every milestone behind them slip with it.
What has already happened should count as having happened. The $100 billion intent shrinking to the $30 billion tier, the "$1.4 trillion in commitments" reined in by the company's own updated figures, Oracle down about 35% in month 6 alone — intent-level cracks arrived one after another in the first half of 2026. Our judgment: before 12/31/2027, at least one contract-level crack appears among the loop's five core parties (NVIDIA, OpenAI, Microsoft, Oracle, CoreWeave) — a signed contract publicly cut back, renegotiated, or defaulted on, or a guarantee/buyback materially triggered, or a large writedown taken as a result — with subjective odds of roughly 65%.
Conclusion: Laying the Judgment Out
The Ponzi label fails. The cash genuinely moves, the compute genuinely ships, and disclosure runs to the contract level. Those calling it a Ponzi underestimate how real these transactions are; those dismissing every skeptic as ignorant of scarcity markets are looking away from the four numbers in the 10-Q.
It is vendor financing with leverage, and the risk sits in concentration. NVIDIA's top three customers account for 54% of a single quarter's revenue (30% a year ago), it invested $17.9 billion outbound in one quarter, and it simultaneously serves as supplier, shareholder, guarantor, and leaseback customer. We put the subjective odds that this concentration keeps trending up over the next four quarters at roughly 70% — and the customer-concentration disclosure in each quarterly 10-Q is the cheapest dashboard for tracking that call.
The market has started giving its own answer. The levered end has fallen back into base-case valuation ranges while the scarcity end stands at 1.8 times the top of bull-case ranges — that split is itself the market's vote on which part of the loop holds real value. We put the subjective odds of a contract-level crack before end-2027 at roughly 65%; if it cracks, first in line is Oracle with the heaviest financing dependence, second is CoreWeave with the most concentrated guarantees.
To track this judgment, watch four public signals: the top-three customer share in NVIDIA's quarterly 10-Q (currently 54%); whether Vera Rubin's first gigawatt deploys on schedule in the second half of 2026; the issuance spreads on Oracle's $40 billion of fiscal 2027 financing; and the valuation of OpenAI's next round along with NVIDIA's actual check size. This site will keep updating all four.
Evidence Boundaries
The load-bearing numbers in this piece fall into four classes with different provenance, stated separately:
Primary disclosures (the concentration/investment/guarantee/buyback commitments in NVIDIA's 10-Q, the buyback terms in CoreWeave's 8-K, the RPO and cash flow in Oracle's results, the Microsoft-OpenAI restructuring terms) are checked directly against SEC filings and official company statements — the hardest layer in this piece.
Reported figures (the roughly $300 billion OpenAI-Oracle contract, OpenAI-CoreWeave cumulative roughly $22.4 billion, the details of OpenAI's audited financials, the $30 billion-tier NVIDIA check under discussion, OpenAI's revenue run rate) come from CNBC, Fortune, Bloomberg, TechCrunch and others reporting on primary documents or informed sources; we cross-checked where possible, but the contracts themselves are not public.
Prices and valuation bands (7/6 closing prices, the three-scenario report ranges, scorecard scores) come from this site's price table and its own scoring system; the valuation bands are late month-5 report snapshots, with methodology in the underlying reports; prices are point-in-time snapshots with a shelf life measured in weeks.
Odds and judgments (the 65% crack odds, the 70% concentration odds) are the subjective judgment of this piece and each does not constitute investment advice.
The falsification conditions for this piece's core claim — "the deals are real, the risk sits in concentration" — are clear: if NVIDIA's customer concentration falls for two consecutive quarters, the Vera Rubin milestone lands on schedule, and Oracle completes its $40 billion financing at normal spreads, the "crack" call should be marked down, and this site will publish the reconciliation when the test dates arrive. The 1999 analogy is a directional reference; the differences in scale and counterparty quality are laid out in the body, and the direct multiple conversions in circulation (the "2.8 times the telecom era" variety) failed our verification and are not used here.
Main Sources
Primary disclosures: NVIDIA 10-Q (FY2027 Q1, ended 2026-04-26) · CoreWeave 8-K: Order Form with NVIDIA (2025-09-09) · NVIDIA-OpenAI strategic partnership announcement (2025-09) · Oracle Q4 FY2026 results (2026-06-10) · Microsoft: the next chapter of the Microsoft-OpenAI partnership (2025-10-28)
Reporting and commentary: Bloomberg: the AI circular-deals map · Bloomberg: Huang's "never a commitment" (2026-02-01) · CNBC: OpenAI resets spending expectations (2026-02-20) · Fortune: OpenAI's 2025 financials leak (2026-06-16) · TechCrunch: NVIDIA's $2 billion top-up in CoreWeave (2026-01-26) · CNBC: Oracle Q4 and the financing plan (2026-06-10) · Motley Fool: Oracle down 35% in month 6 (2026-07-06)
Historical comparison: Tunguz: NVIDIA versus Nortel on vendor financing · Fabricated Knowledge: lessons from the rise and fall of telecom
Our reports and data: NVIDIA investment memo · Oracle long-term owner's study · CoreWeave deep value study · Microsoft deep value study · Micron deep value study · SK Hynix long-term value study · Baillie growth leaderboard
Further reading: Has the memory supercycle topped? · The sector rotation inside the AI trade