Anthropic’s revised IPO paperwork puts a staggering number on the infrastructure needed to build and run advanced AI systems: approximately $895.4 billion in purchase commitments and lease agreements. The company filed the amended Form S-1 on November 2, 2026, as it prepares to offer Class A common stock.
Substantially all of that spending relates to compute and infrastructure. The total includes agreements entered into by September 30, arrangements signed afterward, short-term commitments, and leases that have not yet started. The filing also adds third-quarter 2026 financial results that were missing from the original paperwork filed last week.
A long-term infrastructure bill
Anthropic expects about $447.5 billion of the commitments to be incurred before December 31, 2030. Substantially all of the remaining $447.9 billion is expected between 2031 and 2045, which shows how far into the future the company has planned its compute capacity.
About $69.8 billion is due during the 12 months after September 30, while approximately $213.8 billion is tied to arrangements with termination rights or conditions that must be met before work begins. Those conditions include financing, facilities, and regulatory approvals.
The September 30 contractual-obligations table accounts for $319.993 billion. That figure includes $4.224 billion in operating leases, $1.950 billion in finance leases, and $313.819 billion in purchase commitments. The purchase commitments are mainly future minimum payments under cloud hosting and other infrastructure agreements.
Several large amounts sit outside that table. They include approximately $17.4 billion in short-term commitments, $451.2 billion in estimated undiscounted obligations under executed leases that have not started, $100.2 billion in commitments entered after September 30, and $6.6 billion in later short-term commitments.
Separate arrangements without fixed dollar commitments carry estimated payments of $150 billion to $175 billion. That range includes $125 billion to $150 billion for additional Google TPU compute expected for delivery between 2028 and 2031. Those payments depend on capacity, delivery method, pricing, and financing mechanics.
Cloud contracts and TPU capacity
The filing lists minimum commitments of at least $44.9 billion to Microsoft between November 2026 and June 2033, at least approximately $121.3 billion to Google between April 2026 and July 2033, and $110.0 billion to Amazon between May 2026 and April 2036.
Each commitment applies whether Anthropic uses the capacity or not, subject to the stated terms. Google and Amazon’s agreements include true-up provisions if spending falls short and cannot be canceled unless the counterparty commits an uncured material breach. Microsoft’s commitment has similar limits, with an exception for undelivered tranches on a per-tranche basis.
Five-year master equipment leases with Broadcom and financing entities provide approximately $161.2 billion in rent, excluding interest. Anthropic will generally pay that rent quarterly as AI infrastructure equipment becomes ready for service. The leases provide access to TPU-based compute, with part of that capacity coming online in 2026 and more expected from 2027.
Other disclosed spending figures include $45 billion for data center capacity from Microsoft, $110 billion with Amazon, $121 billion with SpaceX, and $115 billion with Google. The list also includes $45 billion with neocloud Nscale, $41 billion with Lambda, at least $20 billion with AMD for chips, and a $161 billion leasing deal with Broadcom.
Revenue is surging, but costs will arrive first
Anthropic generated $16.956 billion in third-quarter revenue, up from $1.382 billion a year earlier. The filing describes that as 1,127% year-over-year growth. Revenue also rose by more than $5 billion from the prior quarter, reaching nearly $17 billion and rising nearly 47% from $11.6 billion.
For the first nine months of 2026, revenue reached $33.2 billion, compared with $2.5 billion during the same period in 2025. Anthropic held $131.8 billion in cash and short-term investments at the end of September.
The growth rate has cooled from the 147% increase between the first and second quarters, but the business remains far larger than it was a year ago. Revenue for the three months ended September 30 was 86 times the figure for the same period in 2024, producing about $16.8 billion, the same figure that leaked from the investor day.
Anthropic spent $5.6 billion on compute and infrastructure during the six months through June 2026. As more capacity comes online, the company warns that expenses will arrive before the full revenue that the compute ultimately supports. Anthropic says operating margin and Adjusted Operating Margin may compress during those periods, while price competition could hurt margins or affect its ability to achieve or sustain profitability.
The company generated $3.9 billion in free cash flow during the first half of 2026. As of June, it was generating more than two dollars in revenue for every dollar spent on compute, while its adjusted operating margin stood at a 19% profit margin in the most recent quarter reported. That measure excludes stock-based compensation, non-cash charitable contributions, and payments to settle copyright claims from book authors.
Privacy, safety, and control
The amended filing also says Anthropic has $2.1 billion in donation-matching expenses this year, equal to 6% of operating expenses for the period, and expects that amount to rise. The company says it cannot retain prompts or outputs from a substantial share of traffic and warns that zero data retention, or ZDR, could become “incompatible” with certain advanced models, forcing users to accept less than frontier intelligence or move to competitors with fewer monitoring requirements.
Anthropic says AI could help authoritarian governments create high-tech surveillance states, deepen repression, and make themselves much more difficult to dislodge. The company says it is “committed to working with free societies to provide a counterweight to authoritarian abuse and promote freedom and democracy.” Its filing also says concerns have been raised that advanced AI could pose catastrophic or existential risks if used without proper safeguards, oversight, and security.
Control of the company rests with its Long-Term Benefit Trust, whose five members can appoint a majority of the corporate board. Anthropic’s seven cofounders, through Founders LLC, hold 50.1% voting power to elect the remaining directors. In limited deadlock situations, employees holding Class E voting stock can help decide, and employees can vote on certain changes to the trust. Regular investors receive Class A shares, while Amazon and Google receive Class N shares with no voting power.
Based on
- Anthropic Details Long-Term Compute Commitments in Amended IPO Filing — unite.ai
- Anthropic’s Amended S-1 Reveals Q3 Revenue, Massive Compute Bill – Business Insider — businessinsider.com
- Anthropic’s spending commitments surge to nearly $900 billion — axios.com
- The big takeaways and big questions from Anthropic’s S-1 | Fortune — fortune.com
- Anthropic releases amended IPO prospectus with third-quarter financials — cnbc.com
