Here's a fun exercise. Draw a box for Nvidia. Draw a box for CoreWeave. Draw a box for Nebius. Now draw the money.
Nvidia invests $2 billion in CoreWeave. CoreWeave takes that money — plus $18.81 billion in debt it's raised since going public — and buys GPUs. From Nvidia. Nvidia books the sale as revenue. Nvidia also personally guarantees, via a $6.3 billion backstop running through April 2032, that if CoreWeave can't rent out that capacity to anyone else, Nvidia will buy it back itself.
Nvidia invests $2 billion in Nebius too. Same shape. Nebius has raised $8.32 billion in debt against $3.92 billion in equity since Q4 2024, and most of that capital finds its way back to Nvidia's balance sheet as chip revenue.
You are not misreading this. The chip vendor is capitalizing its own customers so they can buy its chips, then insuring the trade so it can't lose. On paper that's three companies transacting. In practice it's one balance sheet wearing three name tags.
This isn't a conspiracy theory from a finance subreddit. It's in the 13F filings, the DDTL facility terms, and the earnings calls. I went and read them so you don't have to.
The numbers, because vibes aren't a source
CoreWeave, Q1 2026:
Revenue: $2.08B (+112% YoY)
Operating cash flow: $2.98B
Capex: $7.70B
Free cash flow: -$4.71B
Debt (QoQ): $24.86B (+16.1%)
Interest expense: $536M (25.8% of revenue)
They're spending $7.7 billion to make $2.98 billion in operating cash. The gap gets plugged with debt — six separate Delayed Draw Term Loan facilities, the newest a $3.5 billion senior note raised on June 11. Total debt issuance since IPO: $18.81 billion, against $3.5 billion of equity. That's not a capital structure, that's a lever with no fulcrum.
Nebius is the same shape with smaller numbers: $339M revenue (+684% YoY, which sounds incredible until you remember it's growth off a tiny base), $2.47B capex, -$214.9M free cash flow, $8.45B debt against a $9.37B cash pile that exists because they keep issuing more debt to keep it there.
Both companies have contracted 3.5 GW of power capacity. Both are only running a fraction of it live. The bet is that hyperscaler demand shows up to fill the rest. Microsoft has committed ~$60B across CoreWeave, Nebius, and Nscale. Meta committed $35.2B to CoreWeave and up to $27B to Nebius. Combined, Microsoft and Meta's neocloud commitments run about $122B — roughly 90% of AWS's trailing twelve-month revenue, promised to companies that don't have AWS's balance sheet.
Then Meta broke the loop
On July 1, Bloomberg reported Meta is building its own cloud business to sell AI compute — meaning the company that just handed CoreWeave a $21 billion commitment and Nebius up to $27 billion might become their competitor instead of just their whale customer.
CoreWeave dropped 13.9% to $85.68 in a day. Nebius dropped 17% to $229.18. Nebius alone shed roughly $12 billion in market cap. This is what happens when a market finally does the arithmetic: if your two biggest customers are also your two biggest funders, and one of them decides to compete with you instead of pay you, there is no third customer waiting in the wings to cover a $17.33 billion funding gap CoreWeave already projects for the rest of the year.
Nvidia's actual defense, and why it doesn't land
Nvidia published a memo denying circular financing and rejecting comparisons to prior tech-bubble blowups. Michael Burry — yes, that Michael Burry — responded that the memo dodged his actual argument, which isn't "Nvidia is Enron," it's "Nvidia is Cisco." His claim: hyperscalers have quietly stretched the assumed useful life of GPU hardware, which understates depreciation and inflates near-term earnings by an estimated $176 billion industry-wide between 2026 and 2028. He's also flagged that insurance companies are routing retiree annuity assets into the debt stack behind these data center buildouts — meaning the people holding the bag if this unwinds aren't the people who understood the trade.
Nvidia's rebuttal focused on denying it's circular. It didn't really address the depreciation math or the annuity exposure. That's the tell. When you get accused of three things and only rebut the one that's easiest to rebut, the other two are the ones worth reading.
Why this should matter to you, specifically
If you're building on any of this infrastructure — renting GPUs from a neocloud, deploying inference on rented capacity, planning a roadmap around "compute will keep getting cheaper" — you're building on top of counterparties whose revenue is substantially their vendor's own capital coming back around. That's fine when growth is exponential and nobody's checking the plumbing. It stops being fine the moment one hyperscaler decides to become a competitor instead of a customer, which just happened, live, in front of everyone, on July 1.
Debt-funded capacity that depends on one or two customers not walking away is not a moat. It's a single point of failure with better marketing. If you're picking a compute provider for anything you plan to still be running in 2027, read the 10-K, not the press release. Look at free cash flow, not revenue growth. Revenue growth is the number companies want you to see. Free cash flow is the number that tells you who's actually solvent when the music stops.
Nvidia isn't going anywhere — they make the chips everyone needs and they have the balance sheet to survive any of this unwinding. The neoclouds sitting on $25 billion in debt with negative free cash flow and one whale customer eyeing their business model? That's the part of the loop that breaks first.
Sources: io-fund, Michael Burry's Substack, The Motley Fool, 24/7 Wall St.
Top comments (0)