When Anthropic filed for its IPO in October, my quick reaction was that the business is real, but the hard part comes next. The amended prospectus adds September-quarter results that sharpen my concern. Anthropic commits to compute long before it knows how much customers will use or pay for. Getting the timing wrong is expensive in either direction.
What Changed in Q3
June’s operating profit looked promising. September shows how quickly those numbers can change.
| Metric | Q2 2026 | Q3 2026 |
|---|---|---|
| Revenue | $11.56B | $16.96B |
| Compute and infrastructure | $5.56B | $9.86B |
| Revenue per $1 of compute | $2.08 | $1.72 |
| GAAP operating income | $974M | ($567M) |
| GAAP operating margin | 8% | (3%) |
| Noncash donation-matching charge | $0 | $1.96B |
| Adjusted operating income | $2.24B | $2.46B |
| Adjusted operating margin | 19% | 15% |
Revenue grew 47% from June, but compute spending rose 77%. The swing to a reported operating loss included a $1.96 billion noncash donation-matching charge. Adjusted operating income, which excludes that charge and stock-based compensation but still includes model-training costs, actually increased. Yet adjusted margin fell from 19% to 15%. One quarter does not establish a trend, but it reveals the sensitivity to compute costs.
Paying More to Catch Up
Anthropic spent $5 billion on compute under short-term commitments in Q3. That is the cost of the capacity, not the premium it paid. The company says this compute is significantly more expensive than comparable capacity secured in multi-year deals. When customer demand or research needs outrun contracted capacity, Anthropic has to fill the gap at higher prices. It expects short-term purchases to increase in the near term.
Higher usage in Q3 was also partly offset by about $600 million in price cuts and contractual discounts. Anthropic tightened enforcement against customers violating its usage policies. Such customers accounted for an estimated 5% to 7% of revenue in the first nine months of 2026. Enforcement could slow growth, but the effect is unquantified.
The Risk of Buying Too Much
The opposite mistake is committing to capacity customers never fully use. Anthropic usually buys compute 12 to 18 months before delivery. It now reports $900.9 billion in aggregate commitments, mostly for compute and infrastructure. About $70.7 billion is due in the 12 months after September 2026, with $451.6 billion expected before December 31, 2030. Arrangements without fixed dollar commitments could add another $150 billion to $175 billion in payments.
These are not debts due today. About $186.2 billion of the $900.9 billion involves termination rights or conditions that must be met before certain obligations start. But many agreements require payment regardless of usage. Meanwhile, 79% of revenue in the first nine months of 2026 was consumption-based, although some enterprise contracts include minimum spending commitments. Customer spending can change faster than infrastructure costs.
Anthropic’s Attempt to Improve the Timing
Anthropic Direct is designed to give the company more flexibility. It leases chips directly and hosts them in dedicated facilities instead of relying solely on cloud providers. Data centers take longer to build but account for a smaller share of total compute costs. Anthropic can reserve facilities early and delay some of the bigger chip purchases until demand is clearer. As the facilities come online, the company expects to replace expensive short-term capacity with cheaper long-term compute. That adds more direct commitments and potentially earlier cash outlays.
Who Carries the Risk?
Anthropic ended September with $146.6 billion in liquidity, including undrawn credit, and expects roughly $84.6 billion in net proceeds from the IPO and concurrent private placement combined. That is a substantial buffer. Still, a default on certain commercial agreements can trigger defaults under its financing arrangements, and its credit facility requires $5 billion in minimum liquidity starting with the December 2026 quarter.
Public investors will help finance the expansion but have limited control. The founders retain 50.1% of the general shareholder vote, a separate trust chooses most directors, and the public benefit structure allows tradeoffs that may not maximize short-term returns.
I’ll watch revenue per dollar of compute, adjusted margin, and whether short-term compute purchases decline as Anthropic Direct comes online. For AI buyers, changing prices and capabilities are another reason to favor a hybrid model portfolio rather than commit every workload to one provider. Anthropic has demonstrated remarkable demand. The harder question is whether it can keep compute costs aligned with that demand.