$37.6 billion, six and a half years, and 57 cents of every dollar that never reaches a chip company. What the AI buildout is physically made of, and whether it earns its capital back.
This is the most complete grid-to-GPU construction I've seen in the open, and it meets our reading of the filings at exactly the right place. We asked the same question from the other end: what does a unit of compute earn, read from CoreWeave's own statements (the workings are here: capefearadvisors.substack.com/p/coreweave-twenty-seven-years). Your $8.05M per MW at 55% realisation brackets their filed $8.31M. The revenue sides of the two models agree, which means the whole distance between your 13.4% and our result has to live on the cost side. It does, and it's worth walking.
Your campus runs on roughly $237M of cash opex against $2.0B of revenue, an implied operating margin near 88%. The company you calibrate against files 48 to 56. At your own replacement cadence, the 250MW campus consumes $1.36B of capital a year; at the filed margin it generates $0.97B to $1.15B, and the surplus inverts. Granting ownership fully, so rent leaves the filed number, your campus breaks even at a cash margin between 57 and 68 percent, inside the corridor between the filed figure and your model. At your most generous corner the surplus is approximately zero. Our grid found the same shape from the ledger: ninety-nine cents on the dollar over the schedule, never a dollar.
One assumption worth updating, and it runs your way: the AI-native you calibrate against books six years on GPUs, extended from five in January 2023, not three. And since book schedules touch no cash line, your construction survives any schedule anyone chooses. The debate over lives is about what readers are shown. The cash math never consulted it.
One number for your hurdle: their filed average cash cost of debt ran 9.1% in Q1, their newest unsecured coupon priced at 9.625% at par in June, and those notes last printed at a 10.47% yield on July 10. An 8% unlevered bar sits below the coupon.
Your decay data is the first empirical read on the residual their schedule requires at year six, and "the same exposure, observed twice" deserves to travel. So does your split: fifty-seven cents of every capital dollar never reaches a semiconductor company. Set that beside our most generous corner, which only approaches a dollar by assuming a margin over cash operating costs near 97 percent, power included. Your model prices the plant. Ours priced the concession. The next question belongs to both: what does it cost to run the fifty-seven cents?
Really thoughtful analysis — and I like the framing of pricing the plant versus pricing the concession. I agree the cost side is where the remaining debate sits, particularly around the ongoing cost of operating and replacing the non-GPU infrastructure. The six-year GPU depreciation update is also a helpful correction. Thanks for digging into the numbers and sharing this.
The AI buildout is really a duration mismatch disguised as a capex boom.
Power arrives on a 4–7 year clock. Buildings take years. GPUs can lose most of their economic value in 3–4.
That means utilisation matters more than almost any headline capex number. If fleet utilisation slips from 90% to 70%, the model here takes project IRR from 13.4% to roughly 1%. The infrastructure is still there. The debt is still there. The silicon replacement cycle is still coming.
That is why the cleanest trade may sit one layer upstream: own the suppliers selling into a 15-year physical asset, not the operators carrying the residual risk of rapidly depreciating compute.
I’d love to connect. I’m an advisor at an AI infrastructure company and we are doing some pretty cool things.
Thanks for the detailed explanation!
This is the most complete grid-to-GPU construction I've seen in the open, and it meets our reading of the filings at exactly the right place. We asked the same question from the other end: what does a unit of compute earn, read from CoreWeave's own statements (the workings are here: capefearadvisors.substack.com/p/coreweave-twenty-seven-years). Your $8.05M per MW at 55% realisation brackets their filed $8.31M. The revenue sides of the two models agree, which means the whole distance between your 13.4% and our result has to live on the cost side. It does, and it's worth walking.
Your campus runs on roughly $237M of cash opex against $2.0B of revenue, an implied operating margin near 88%. The company you calibrate against files 48 to 56. At your own replacement cadence, the 250MW campus consumes $1.36B of capital a year; at the filed margin it generates $0.97B to $1.15B, and the surplus inverts. Granting ownership fully, so rent leaves the filed number, your campus breaks even at a cash margin between 57 and 68 percent, inside the corridor between the filed figure and your model. At your most generous corner the surplus is approximately zero. Our grid found the same shape from the ledger: ninety-nine cents on the dollar over the schedule, never a dollar.
One assumption worth updating, and it runs your way: the AI-native you calibrate against books six years on GPUs, extended from five in January 2023, not three. And since book schedules touch no cash line, your construction survives any schedule anyone chooses. The debate over lives is about what readers are shown. The cash math never consulted it.
One number for your hurdle: their filed average cash cost of debt ran 9.1% in Q1, their newest unsecured coupon priced at 9.625% at par in June, and those notes last printed at a 10.47% yield on July 10. An 8% unlevered bar sits below the coupon.
Your decay data is the first empirical read on the residual their schedule requires at year six, and "the same exposure, observed twice" deserves to travel. So does your split: fifty-seven cents of every capital dollar never reaches a semiconductor company. Set that beside our most generous corner, which only approaches a dollar by assuming a margin over cash operating costs near 97 percent, power included. Your model prices the plant. Ours priced the concession. The next question belongs to both: what does it cost to run the fifty-seven cents?
Really thoughtful analysis — and I like the framing of pricing the plant versus pricing the concession. I agree the cost side is where the remaining debate sits, particularly around the ongoing cost of operating and replacing the non-GPU infrastructure. The six-year GPU depreciation update is also a helpful correction. Thanks for digging into the numbers and sharing this.
The AI buildout is really a duration mismatch disguised as a capex boom.
Power arrives on a 4–7 year clock. Buildings take years. GPUs can lose most of their economic value in 3–4.
That means utilisation matters more than almost any headline capex number. If fleet utilisation slips from 90% to 70%, the model here takes project IRR from 13.4% to roughly 1%. The infrastructure is still there. The debt is still there. The silicon replacement cycle is still coming.
That is why the cleanest trade may sit one layer upstream: own the suppliers selling into a 15-year physical asset, not the operators carrying the residual risk of rapidly depreciating compute.