Its IPO filing shows what a missed forecast costs, and what smaller AI teams can learn.
Anthropic’s amended IPO filing of 2 November shows what a missed forecast can cost. When demand outpaced the compute it had booked, the company bought $5.0 billion of short-term capacity that now costs significantly more than its multi-year compute, and it reported a $567 million operating loss for the quarter.
The scale is far beyond most teams, but the trade-off is not. Any team that buys GPUs chooses between committed and short-term capacity, and the filing is a rare primary-source look at what that choice costs.
What does the filing say?
The amended filing adds third-quarter results and a full count of compute obligations. These are the main points, as reported from the filing:
- Revenue: about $17 billion in the quarter, up about 47% from the second quarter, against about 144% growth the quarter before
- Compute and infrastructure cost: $9.86 billion, up from $5.56 billion, or 58 cents of each revenue dollar, compared with 48 cents in the second quarter
- Short-term compute: $5.0 billion in the quarter, against about $1.5 billion in the whole first half
- Operating result: a $567 million operating loss, which the filing links to short-term compute costs and a $2.0 billion non-cash charge tied to a donation-matching program
- Commitments: about $895 billion, almost all for computing and data centers, with about $70 billion due within a year
Why does short-term compute cost more?
Anthropic buys most of its compute well ahead of need. It books most of its computing 12 to 18 months ahead. When demand runs past that plan, it fills the gap with short-term capacity, which the filing says costs significantly more than the multi-year equivalent.
The filing gives no unit price or percentage premium, so the size of the gap is not public. It does say Anthropic expects to use more short-term compute in the near term, and that compute and infrastructure will remain its largest operating expense for the foreseeable future.
The pattern shows up in public price lists too, at a smaller scale. Across 96 reserved H200 configurations, the median reserved rate is $3.62 per GPU-hour, against a median of $4.77 on demand. That is about 24% lower.
What is the catch with committed capacity?
Commitments lower the unit cost, and they fix the spending. The filing’s commitments total almost seven times the roughly $132 billion in cash and short-term investments Anthropic held at the end of September.
Terms differ between contracts. According to a summary of the filing, one Microsoft commitment of at least $44.9 billion is due whatever Anthropic uses, while Lambda’s fees accrue only on delivered and accepted capacity. The terms decide how much risk the buyer carries.
The filing describes the short-term lines as cancelable without penalty within a year, and it refers to moving to more cost-effective, long-term compute as dedicated infrastructure comes online. Committed capacity lowers the cost of a steady base, and short-term buying covers a forecast that proved too low.
What can smaller AI teams learn?
The filing covers compute across several chip types, including TPU-based and NVIDIA-based capacity, and Anthropic buys at a scale few can match. Even so, five lessons carry over to any team that rents or commits to GPUs:
- Commit to the base, not the peak: A commitment pays off on certain hours, and idle committed capacity costs the same as busy capacity
- Treat short-term capacity as insurance: It covers demand the forecast missed, at a premium
- Forecast with a margin: Anthropic books 12 to 18 months ahead and still ran short, so a forecast that is too low costs a premium and one that is too high leaves capacity idle
- Read the terms: One commitment in the filing is due whatever is used, while another accrues only on delivered capacity, and cancellation rules differ
- Measure utilisation: The break-even between hourly and committed rates depends on how many hours a GPU is actually used
For a team with a 12% annual discount, for example, the committed option only wins above about 88% utilisation. The break-even moves with every contract.
What to watch next
The next disclosures will show how the pattern develops:
- The share of short-term compute: Whether it falls as dedicated capacity comes online
- Unit pricing: Whether any buyer or provider publishes the real premium for short-notice capacity
- Other filers: Whether more AI companies disclose how much of their compute is committed against bought as needed
For now, the filing offers one clear fact: even a buyer with about $895 billion in commitments paid a premium when its forecast fell short.
