Opinion · AI infrastructure · September 10, 2026
Backlog is not revenue. It is a set of questions.
A $104 billion backlog number is impressive. It is not $104 billion of revenue sitting in an account, and it is definitely not $104 billion of profit. I used to treat backlog as proof that demand had been settled. Now I treat it as the beginning of the work: what is actually committed, when does it convert, who pays to build it, and what remains after depreciation and interest?
Figure 1 · A backlog is a timeline, not a bank balance
CoreWeave disclosed $103.7B of unsatisfied RPO at June 30, with only 41% scheduled for the next 24 months.
Source: CoreWeave Q2 2026 10-Q. Dollar amounts above are AEA calculations using the company’s disclosed percentages and $103.7B RPO.
Five labels that should not be interchangeable
Revenue is what has been recognized. Deferred revenue is generally cash billed or received ahead of recognition. RPO is the transaction price allocated to unfinished performance obligations. Backlog may be a broader management label. Total contract value can include years of potential services. These terms overlap, but they are not synonyms; every headline gets easier to read once that is clear.
The CoreWeave case: demand plus a financing test
CoreWeave reported $103.7 billion of unsatisfied RPO at June 30, 2026. The company says its committed contracts are generally take-or-pay, and that 98% of Q2 revenue came from them. That deserves credit. But the same filing shows $14.1 billion of first-half capex, approximately $35.1 billion of total debt including recourse and non-recourse debt, and top-three customer concentration of 72% of Q2 revenue. The RPO is real evidence of demand. It is also evidence of a huge amount of delivery, financing, and customer-concentration work still ahead.
| Question | What to look for | Why it changes the headline |
|---|---|---|
| Is it committed? | Take-or-pay, cancellation clauses, service-level credits | Not every contract dollar has the same certainty. |
| When does it convert? | Disclosed recognition schedule | Duration is not an annual revenue number. |
| Who funds capacity? | Prepayments, customer-owned GPUs, provider capex | Two similar backlogs can require wildly different capital. |
| How concentrated? | Revenue and receivable concentration | One customer changing course can matter more than the aggregate. |
| What remains? | Depreciation, interest, and service credits | A real contract can still become expensive revenue. |
Same headline, different contract
Nebius’s Meta agreement is a useful contrast. Its disclosure described roughly $12 billion of dedicated GPU-cluster orders plus up to $15 billion of capacity backstop: Meta pays for the latter only if specified capacity remains unsold. The headline can be presented as “up to $27 billion,” but the economics are not identical across those dollars. Oracle offers another contrast: it reported that customers prepaid GPU purchases or supplied GPUs for a large portion of its AI-contract increase. Demand quality and funding quality are separate questions.
The fair bull case
The skeptical reading should not become “backlog is fake.” Long-duration, take-or-pay commitments with disclosed schedules and prepayments can make future revenue more visible and support asset-backed financing. The discipline is not to dismiss the number. It is to discount it for time, delivery, concentration, capital needs, and margins instead of stopping at the biggest number on the slide.
MonitorMy next-quarter dashboard
RPO growth; the percent due within 24 months; deferred revenue and prepayment; capex per added active megawatt; active versus contracted power; top-customer concentration; interest relative to operating cash flow; and actual margins after depreciation. If those move in the wrong direction, a backlog headline will not rescue the economics.