Analysis
The same chip, three businesses
Inference can earn fat margins while a GPU landlord loses money on the same chip. In Q2 2026 CoreWeave's data centers lost $49M before interest. The margin had gone to the chipmaker and to whoever sells the answer.
Two claims about AI money are both true. Inference, the work of running a trained model to answer a request, is said to earn high margins. A data center is said to pay back fast for a company like Google. Yet CoreWeave, which owns about $47B of AI data-center property, lost $49M at the operating line in April–June 2026, before a cent of its $640M interest bill.
The claims fit together once you ask who sells what. One Nvidia chip supports three businesses, and they earn very different returns on it.
Three businesses on one chip
| Nvidia | CoreWeave | ||
|---|---|---|---|
| Sells | the chip | hours of GPU time | answers: search, Gemini, cloud services |
| Is paid | in full when the chip ships | by the hour, or by multi-year contract | per query, seat or API call |
| Fills its machines with | — | a few large tenants | mostly its own products |
| Pays for the build with | cash | debt; its August loan priced at SOFR + 550bp | operating cash, and now debt |
Nvidia kept $120B of net income on $216B of revenue in the year to January 2026, or 56 cents of each dollar. Everyone downstream pays that first.
The landlord’s arithmetic
Take a GPU that costs about $45,000 installed. Rent it at $2.50 an hour and keep it busy 80% of the time, and it brings in about $17,500 a year. In CoreWeave’s Q2 accounts, costs other than depreciation came to about 48% of revenue, which leaves some $9,100 a year. Over a six-year life that is a return of about 5–6% a year before tax, against debt that costs about 10%. The figures are an illustration, not CoreWeave’s books, but the quarter says the same thing more plainly. Depreciation was $1.39B and operating income was −$49M: the fleet earned what it cost to wear out, and nothing toward its interest.
Now put the same chip to work for whoever owns the model and the customer. If the answers it serves bring in several times the rental rate per hour, the chip pays for itself in a year or two. That is the high inference margin people talk about. It belongs to the seller of the answer, not to the landlord renting out the hours.
Why Google is different
- It mostly makes its own chips. Google’s TPUs skip Nvidia’s margin on most of its fleet.
- It fills its own machines. Search, YouTube, Gemini and Workspace keep the chips busy, and no tenant can walk away.
- It has cash, and borrows cheaply. Google made $186B of operating cash in the year to June 2026. A project that returns 12% creates value at Google’s cost of money and destroys it at CoreWeave’s.
- Old chips find second jobs. Last year’s hardware moves to lighter internal work, so a six-year life is more believable than it is for a landlord whose old chips sit unrented.
In Hamilton Helmer’s terms, Google holds powers the landlord lacks: scale, its own chips, and customers it already owns. The landlord’s edge is getting chips early, and that edge runs out.
Why the exemption is being tested
Google’s capital spending took 71% of its operating cash in the year to June 2026, up from 32% in 2023. In April–June alone it spent $44.9B against $39.1B of operating cash. Its debt rose from $14.8B at the end of 2024 to $100.2B at June 30, 2026. Its famous returns came from Search, which needed little capital; nobody has yet shown what the AI spending returns. And the price of a token keeps falling: Anthropic’s top model went from $15 to $5 per million input tokens between generations. The seller of answers has margin today. Whether it keeps it depends on how fast prices fall against how fast the chips get cheaper to run.
What the coverage record shows
The TEXXR record traces how the landlord model was built. In November 2024 CoreWeave told investors its signed contracts were worth $17B, about $10B of it from Microsoft. By its Q1 2026 report the backlog was $99.4B, and the 10-Q put it at $103.7B at June 30: six times the 2024 figure in nineteen months. The customer list stayed narrow. The S-1 showed its top two customers at about 77% of 2024 revenue.
The funding moved too. CoreWeave’s first loan secured on Nvidia chips, $2.3B, came in August 2023. By February 2026 the Financial Times described GPU-backed debt held in separate vehicles as a model “pioneered by CoreWeave” and spreading through the industry. In June 2026 a CoreWeave-tied data center sold $900M of five-year junk bonds at 7.5%. Nvidia, meanwhile, has agreed to buy any capacity CoreWeave cannot sell, through April 2032. The record holds 114 CoreWeave stories from 17 outlets; Q2 2026 brought 16, against an average of 10.5 over the previous eight quarters.
The Bezos test
In his 2004 letter, Jeff Bezos built a machine that made a 10% margin, tested it at a 12% cost of capital, and concluded that “there is no growth rate at which it makes sense” to run it. CoreWeave’s margin at that line was −2% last quarter. The landlord fails the test on today’s numbers. The seller of answers can pass it. Whether Google passes it on the AI spending is the open question of the build-out.
What would settle it
- CoreWeave’s Q3 operating line, due around November 10. Positive would mean new clusters earn more than they cost to wear out.
- What three-year-old H100s rent for. If the rate holds, six-year lives hold, for landlords too.
- Google’s capital spending against its operating cash in Q3.
- Token prices for the top models at the next generation.