Anthropic has disclosed $895.4 billion of spending commitments and holds $131.8 billion in cash and short-term investments. That is an extraordinary commitment to future capacity, backed by an extraordinary reserve. The question is whether it can turn that capacity into durable profits: Q3 revenue grew 46.7%, but compute expense grew 77.5% and adjusted operating margin declined.
The capacity-cost test has arrived
Our Q2 review identified a clear question: could revenue continue to absorb the cost of delivered compute? Q3 provides a less comfortable answer. Revenue increased by $5.401 billion, but compute and infrastructure expense increased by $4.305 billion. That single expense category absorbed 79.7% of the incremental revenue before changes in personnel, distribution and other expenses.
| Quarterly measure | Q3 2025 | Q2 2026 | Q3 2026 |
|---|---|---|---|
| Revenue | 1,382 | 11,555 | 16,956 |
| Compute and infrastructure | 2,047 | 5,558 | 9,863 |
| Personnel | 489 | 1,857 | 1,711 |
| Sales, marketing and partnerships | 202 | 2,135 | 2,361 |
| Other operating and administrative | 233 | 1,031 | 3,588 |
| Bartz settlement | 1,500 | 0 | 0 |
| Total operating expenses | 4,471 | 10,581 | 17,523 |
| GAAP operating income (loss) | (3,089) | 974 | (567) |
| Adjusted operating income (loss) | (1,284) | 2,235 | 2,463 |
| GAAP operating margin, calculated | (223.5%) | 8.4% | (3.3%) |
| Adjusted operating margin, calculated | (92.9%) | 19.3% | 14.5% |
The annual comparison remains extraordinary: revenue rose 1,126.9%, versus 381.8% growth in compute expense. But that comparison obscures the sequential squeeze. Compute rose from 48.1% of revenue in Q2 to 58.2% in Q3. Both statements are true: Anthropic has become much more efficient than a year ago, and its latest quarter became more expensive to serve and develop.
Management supplies a specific explanation. It says Q3 included $5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable capacity bought under multi-year arrangements. That is roughly half of total quarterly compute expense. It is the amount spent under those commitments, not a disclosed $5 billion premium or a removable exceptional charge. The filing does not quantify the cost of an equivalent long-term replacement.
Anthropic generally contracts long-term compute 12 to 18 months before availability. When customer demand or research requirements exceed that capacity, it buys the more expensive short-term supply. Management continues to expect increasing short-term use to compress margins in the near term. The important economic question is how quickly cheaper capacity becomes available, and whether demand remains sufficient to utilize it.
Pricing adds another hurdle. Management says new-model price reductions and contractual discounts reduced revenue by approximately $600 million in the Q2-to-Q3 comparison, partially offsetting higher usage. This is a disclosed revenue headwind, not a separately measured reduction in profit. More capable or cheaper models can encourage adoption, but usage must expand enough to offset lower prices as well as the capacity bill.
There is also progress on distribution economics. Sales, marketing and partnerships fell from 22% of revenue in Q1 to 14% in Q3. Management attributes that improvement primarily to a lower share of revenue through channel partners and changes in distribution terms. The expense still increased in dollars in Q3. Better distribution economics help, but did not offset the latest compute-cost pressure.
The GAAP reversal needs two separate explanations
Q3 moved from a $974 million operating profit to a $567 million operating loss. Compute pressure is part of that story, but so is a $1.960 billion non-cash charitable contribution adjustment associated with the employee donation matching program. The company matches qualifying employee donations with Class N shares. Non-cash does not mean economically costless, and management expects public-benefit initiatives to remain a factor in future expenses and potential dilution.
The company's bridge is explicit: the $567 million operating loss, plus $1.070 billion of stock-based compensation and $1.960 billion of non-cash charitable contributions, produces $2.463 billion of adjusted operating income. Adjusted profit rose 10.2% from Q2, while revenue rose 46.7%. The resulting adjusted margin fell approximately 4.8 percentage points. Removing the charitable contribution alone would produce $1.393 billion of operating income, or 8.2% of revenue, but that is an analytical illustration, not the company's adjusted measure.
The $23.291 billion Q3 net loss answers a different question again. It includes a $21.936 billion fair-value charge on convertible notes, derivative liabilities and SAFEs. That valuation charge is not an equivalent cash payment. Neither the large net loss nor the adjusted operating profit should be presented as cash burn or cash generation.
Cash generation improved, with financing and working capital kept visible
The nine-month cash-flow statement reports $6.569 billion of operating cash inflow. Subtracting the previously reported first-half $3.929 billion gives $2.640 billion of derived standalone Q3 operating cash inflow. The corresponding calculation for property and equipment purchases is $925 million minus $482 million, or $443 million in Q3. These are differences between cumulative filings, not a separately presented quarterly cash-flow statement.
For the nine months, operating cash flow less property and equipment purchases equals $5.644 billion. That narrow calculation misses $1.955 billion of prepayments for leased assets and equipment and $88 million of finance-lease principal payments. Subtracting those amounts leaves $3.601 billion, an ECI cash-investment illustration rather than a company-defined free-cash-flow measure. It still does not represent all future compute requirements.
Working capital also matters. Nine-month receivables growth absorbed $7.076 billion of cash, while changes in payables, accrued expenses and other liabilities contributed $4.377 billion and deferred revenue contributed $1.702 billion. Those timing effects should remain visible when evaluating the sustainability of operating cash generation.
September cash and equivalents of $95.562 billion plus short-term investments of $36.283 billion total $131.845 billion, compared with $131.651 billion in June. Cash alone fell while short-term investments rose. The nine-month statement still reports $110.432 billion of net preferred-stock issuance proceeds, the same amount reported for the first half. The large liquidity reserve reflects earlier capital raising as well as operations.
The blank commitment fields are now a concrete schedule
The amendment reports approximately $895.4 billion of aggregate committed spending as of and subsequent to September 30, substantially all related to compute and infrastructure. About $447.5 billion is expected before December 31, 2030, with substantially all of the remainder between 2031 and 2045. About $69.8 billion is due in the 12 months following September 30.
| Bridge to disclosed aggregate | $ billions |
|---|---|
| September contractual-obligation table | 319.993 |
| September short-term commitments excluded from table | 17.4 |
| September leases not yet commenced | 451.2 |
| Subsequent longer-term commitments | 100.2 |
| Subsequent short-term commitments | 6.6 |
| Calculated sum, subject to component rounding | 895.393 |
The scope qualifications are material. Approximately $213.8 billion of the aggregate relates to arrangements with exercisable termination rights or contingent conditions before commencement. Of the $69.8 billion due over the next 12 months, about $4.3 billion is associated with such conditions or rights. These provisions do not make every commitment freely cancelable.
There is also spending outside the aggregate. The filing estimates another $150 billion to $175 billion under certain compute arrangements without fixed dollar commitments. Amounts depend on capacity, delivery, pricing and other deployment conditions. That estimate should be displayed separately, not silently added to a fixed contractual-obligation total. The table also excludes $6.2 billion of Amazon Notes principal expected to settle in shares.
Against the schedule, management identifies $146.6 billion of liquidity, including $14.8 billion of available revolving credit alongside cash and short-term investments. It separately identifies up to $77 billion of financing capacity under note purchase agreements and up to $5 billion of equity commitments. These are different forms of funding, subject to their respective conditions, not one unrestricted cash balance. Nor is the lifetime spending total a present funding deficit: the timing, future receipts, financing costs and delivery conditions must be modeled together.
The scale is clearer beside the cash reserve. The $69.8 billion due over the next 12 months equals approximately 52.9% of September cash and short-term investments. It also slightly exceeds $67.824 billion, or four times Q3 revenue. That annualized revenue figure is an illustration, not a forecast, and these contractual payments are not all current operating expenses. Future revenue and cash receipts matter. Still, a large opening cash balance does not eliminate the need for strong cash generation as capacity arrives.
The $895.4 billion aggregate plus the separately estimated $150 billion to $175 billion implies a combined potential spending envelope of approximately $1.045 trillion to $1.070 trillion. That arithmetic combines commitments with variable estimates across different periods; it is not a newly disclosed fixed contractual total. The $100.2 billion of subsequent longer-term commitments also shows how quickly the obligation base is expanding beyond the September balance-sheet date.
Delivery conditions and financing costs belong in the margin discussion
Several contracts illustrate why the aggregate needs its footnotes. Google's hosting commitment is at least approximately $121.3 billion from April 2026 through July 2033, regardless of actual usage, with a true-up if spending falls short, subject to specified adjustments and exceptions. Microsoft's minimum $44.9 billion commitment runs from November 2026 through June 2033. These named contracts are components of the disclosed arrangements, not amounts to add again to the aggregate.
SpaceXAI is providing another substantial source of GPU capacity. The filing describes approximately $3.1 billion per month once all contracted capacity is delivered, expected in December 2026, with different fees during the ramp. That would equal approximately $9.3 billion over three fully delivered months, not a Q4 expense forecast. The agreements have initial non-cancelable periods ending in January and March 2027, after which either party can cancel on 90 days' notice. Delivery can support Anthropic's growth, while the amount actually incurred depends on timing and contract duration.
Some other conditions protect against paying for capacity that never becomes available. Lambda's $34.9 billion August arrangement permits either party to terminate if Lambda does not timely secure the facility lease. Nscale must use best efforts to obtain qualifying financing for its GPUs and infrastructure. Reduced obligations under these provisions could also mean less available compute. They are not a blanket option for Anthropic to walk away simply because demand slows.
Supplier financing has an economic cost. Broadcom's $42 billion note facility is available, subject to conditions, solely to finance quarterly rent under the relevant lease arrangement. The notes carry 9% annual interest paid in kind and may convert at 90% of the applicable 30-day trailing share-price average. Notes issued when specified revenue conditions are not met carry 13% interest and an 85% conversion price. Under a change-of-control provision, the holder can elect cash equal to 150% of outstanding principal and accrued interest or conversion. These are financing terms, not cash already borrowed. They connect growth performance to future financing cost and potential dilution.
Anthropic Direct changes the risk as well as the cost structure
Anthropic is expanding beyond cloud contracts into directly leased chips hosted in dedicated facilities operated on its behalf. Management calls this Anthropic Direct. Separating facilities from accelerator commitments can secure long-lead-time space while delaying larger chip decisions until demand is clearer. It could replace expensive short-term supply with more cost-effective long-term capacity.
The tradeoff is greater capital intensity and delivery exposure. Management expects the buildout to become a significant use of operating and investing cash. The operational test is whether delivered capacity lowers compute expense relative to revenue without creating underutilized obligations. Customer demand, research requirements and infrastructure timing all matter; the filing does not isolate their individual contributions to quarterly compute costs.
Contract duration also differs sharply from product cadence. Model versions released in the six months ended September 30 produced more than 80% of Q3 revenue, while substantial infrastructure obligations extend through 2045. This does not establish that the infrastructure will become obsolete on the same schedule as a model. It does mean Anthropic must repeatedly deliver useful new models and sustain customer demand over a much longer investment horizon.
Adoption remains strong, but the measures answer different questions
At September 30, more than 11,900 customers generated at least $100,000 of trailing-12-month revenue, versus more than 6,000 at June 30. There were more than 1,900 at the $1 million threshold and approximately 250 at $10 million. The September counts had already appeared as later-dated adoption evidence in the original S-1, so they are now aligned with the reporting quarter rather than newly discovered customer growth.
Reported net revenue retention was approximately 900%. Its definition compares a qualifying existing customer's revenue in the latest three months with the corresponding three months a year earlier, excluding subscribers to Claude Pro, Max and Ultra. It is an aggregate revenue expansion measure, not a customer-retention percentage, a median customer outcome or a profitability measure.
Remaining performance obligations increased from $10.8 billion in June to $13.8 billion in September. Management expects about 75% of the September balance, approximately $10.35 billion, to become revenue in the next 12 months. That is a new backlog cohort and recognition window. It does not prove fulfillment of the earlier June-cohort expectation or forecast total next-year revenue.
Two accounting items complicate the next profitability comparison
Using the supplemental functional expense note, nine-month cost of revenue was $4.812 billion. Subtracting first-half cost of revenue of $2.380 billion gives $2.432 billion for Q3, implying a calculated 85.7% gross margin. This is not inconsistent with compute consuming 58.2% of revenue: compute spans research, development and other workloads, while channel partner costs are classified separately. Gross margin does not capture the full cost of developing and distributing Claude or establish per-token economics.
The IPO compensation overhang has also become more concrete. Had the offering completed on September 30, Anthropic says it would have recorded $8.2 billion of cumulative RSU compensation expense. It reports another $24.1 billion of unrecognized RSU compensation expected over approximately two years, plus $2.1 billion for other equity awards over approximately 1.4 years. The $8.2 billion is a hypothetical September-date charge, not a Q3 booked expense or a final offering-date charge.
The amendment still leaves the offering price range and share count blank. Anthropic also intends to use part of the IPO proceeds for tax withholding and remittance associated with settling RSUs; the amount remains unspecified. Gross fundraising should therefore not be treated as cash entirely available for infrastructure. A reported valuation or an annualized-revenue multiple would be a separate scenario, not established IPO terms.
Update to our accountability record
Capacity costs: the previously identified margin-pressure mechanism now has a reported quarterly outcome and management attribution. However, the original S-1 was filed October 26, after Q3 ended. This is a follow-up to our Q2 review, not evidence that management made and met a pre-Q3 forecast. No binary promise score is assigned.
Commitment disclosure: the aggregate and payment-timing blanks flagged in the prior review are filled. This closes that specific disclosure question while opening a more concrete delivery-and-funding test. June and September obligation balances remain dated snapshots, and the new aggregate also includes subsequent commitments.
Backlog recognition: keep the June 2027 endpoint for the original June cohort. Add a separate September 2027 endpoint for the September cohort. Both remain pending unless future evidence identifies recognition from the relevant opening balance.