Anthropic's IPO Makes the Compute Bill Public
The Short Version
Anthropic filed to go public.
The interesting part is not the ticker.
It is the electricity bill with footnotes.
On October 26, Anthropic submitted a preliminary S-1 registration statement for a Nasdaq listing under the symbol ANT. The number of shares, price range, and expected proceeds are still blank. This is a filing, not a completed IPO.
It is also the first detailed public map of how a frontier AI lab turns capital into compute, compute into intelligence, and intelligence into consumption revenue.
The growth is extraordinary. Anthropic reports $16.3 billion in revenue for the first six months of 2026, up from $1.2 billion a year earlier. Revenue for all of 2025 was $4.6 billion.
The commitments required to support that curve are extraordinary too.
Anthropic has signed enormous multi-year arrangements with Amazon, Google, Microsoft, SpaceXAI, Nscale, Lambda, Broadcom, AMD, and other infrastructure providers. Some are minimum-spend or take-or-pay commitments. Some depend on capacity being delivered. Some can be cancelled after an initial period. Some estimates are not included in the filing's contractual-commitments table.
So do not compress them into one clean debt number.
The useful conclusion is simpler.
An AI lab can sell software by the token while buying its factory years in advance.
The IPO helps finance the gap.
The $80.7 billion loss is real accounting, not the whole operating story
The number that will travel fastest is Anthropic's $80.7 billion net loss for the first half of 2026.
It needs a second sentence.
Anthropic says $72.4 billion of that loss came from changes in the fair value of convertible notes, derivative liabilities, and SAFEs, primarily because the company's enterprise valuation increased. When the value of securities that will convert into equity rises, accounting can record a larger liability and a large non-cash loss.
That does not make the loss imaginary.
It makes it a poor proxy for the cash consumed by the business.
For the same six-month period, Anthropic reported a $693 million operating loss and $3.9 billion in cash provided by operating activities. It also reported $1.6 billion in adjusted operating income after excluding stock compensation and other items. That final number is management's non-GAAP view, not a substitute for the audited statements.
All three figures describe different things.
- Net loss includes the valuation movement in financing instruments.
- Operating loss is closer to the economics of running the company before those financing effects.
- Operating cash flow includes working-capital timing and does not pay for every infrastructure investment.
The filing says exactly that last part. Anthropic expects capital expenditures to increase significantly and warns that operating and investing cash use may rise even if financial performance improves.
This is why “profitable” and “self-financing” are dangerous shortcuts for an AI lab.
The income statement can improve while the infrastructure machine asks for more capital.
Compute is being contracted before the customers arrive
Anthropic's named agreements show the scale of the advance planning.
The company says it committed to spend at least $44.9 billion with Microsoft through 2033, at least $114.5 billion with Google through 2033, and at least $110 billion with Amazon through 2036. Those agreements generally require the spend regardless of actual usage, subject to their specific terms and exceptions.
The filing also describes up to $120.5 billion of SpaceXAI capacity through 2029, with about $3.1 billion per month once all capacity is delivered, which Anthropic expects in December 2026. Those agreements can be cancelled after initial non-cancelable periods and 90 days' notice.
Other arrangements include up to $44.6 billion with Nscale, about $41.3 billion with Lambda, more than $20 billion of expected spending if an initial AMD capacity tranche is fully procured, and roughly $161.2 billion of rent payments under five-year Broadcom equipment leases. Anthropic separately estimates another $125 billion to $150 billion for additional TPU compute whose final payment depends on capacity and deployment conditions.
These numbers are not interchangeable.
They span different time periods, conditions, cancellation rights, accounting treatments, and delivery assumptions. Hardware leases may support capacity delivered through another arrangement. “Up to” is not the same as “minimum.” An estimate outside the commitments table is not the same as a fixed obligation.
But the pattern is impossible to miss.
Anthropic is reserving a physical production system before it knows exactly which models, customers, and workloads will consume it.
That creates the central risk of the business.
Underinvest and the lab cannot train, launch, or serve enough intelligence.
Overinvest and it owns a magnificent supply of expensive capacity that customers may not use on the expected schedule.
The company says its consumption-based “pay-for-intelligence” model produces the substantial majority of revenue. Much of that revenue is not protected by contracts as long as the infrastructure commitments supporting it.
Long-duration supply meets short-duration demand.
The S-1 is the financing plan for that mismatch.
The cloud providers sit on every side of the table
Amazon, Google, and Microsoft are not ordinary vendors here.
They provide compute. They distribute Claude. They buy Anthropic's intelligence service. They hold equity. They finance capacity. And they build competing AI models and applications.
Anthropic says these relationships give the cloud providers visibility into its pricing and commercial terms, while potentially influencing compute allocation, service quality, product placement, sales priorities, and incentives.
That paragraph may be the most useful risk disclosure in the filing.
The standard platform story says a startup uses cloud partners to scale distribution and infrastructure.
The frontier-lab version is stranger.
Your investor can be your landlord, reseller, customer, lender, and competitor at the same time.
That does not mean the partnerships are bad. They helped Anthropic reach a scale that would have been almost impossible to finance alone.
It means “independence” is not a binary state.
Anthropic can own its models and product roadmap while remaining deeply dependent on companies with their own model roadmaps.
Public investors are buying economics, not ordinary control
The governance structure is just as engineered as the compute stack.
Anthropic is a public benefit corporation. Its Long-Term Benefit Trust holds a special class of stock that is designed to elect and remove a majority of the board. A Founder LLC controlled initially by seven co-founders will hold a single Class F share with 50.1% of the voting power on matters submitted to stockholders generally.
The Class F share has no economic rights.
Public Class A shares get one vote each, but Anthropic will qualify as a “controlled company” under Nasdaq rules.
That separation is deliberate.
Public investors can participate in the economics without receiving ordinary public-company control over the mission or the founders.
The filing also answers one question from last week's article about Anthropic's Public Benefit Fund. The 5% contribution is no longer only a proposal subject to stockholder approval. Anthropic says it has contributed 846,840,960 non-voting Class N shares, representing 5% of fully diluted shares as of June 30 under the filing's assumptions.
The future 2% additions remain discretionary. The filing says they can occur when the share price clears 50% appreciation milestones, beginning with the IPO price, for at least 90 of the previous 120 trading days and also clears a ten-day average test.
This is more concrete than the original announcement.
It is still controlled by Anthropic's board and, for now, managed by Dario Amodei, Daniela Amodei, Jack Clark, and Neil “Buddy” Shah.
The public-benefit machinery is real.
So is founder control.
Claude may be improving the factory that builds Claude
Anthropic says more than 90% of contributions to its codebase in September were written by Claude. It also says average daily code volume per engineer was more than 45 times its 2024 level.
That is a remarkable internal claim.
It is not evidence that every engineer became 45 times more productive.
Lines of code are not value. AI-generated code can create review work, security risk, duplication, and maintenance costs. The filing itself warns that internal use of Claude Code can introduce errors or vulnerabilities.
But the direction matters.
Anthropic is not only selling AI labor. It is using AI labor to increase the speed at which it builds the product that it sells to finance more compute.
That is the compounding loop public investors are being asked to fund:
capital buys compute, compute improves models, models generate revenue, models accelerate the company, and the company signs for more compute.
The loop is powerful.
It is also unforgiving when any link slows down.
The IPO is a capacity decision
The preliminary filing does not yet tell investors the offering price, share count, or valuation. Early coverage has therefore focused on the public debut, but the more durable story is already visible in the S-1.
Anthropic is not going public simply because private investors want liquidity.
It is entering public markets while converting an asset-light software story into a capital-intensive infrastructure story.
Anyone evaluating the company should track more than revenue growth and model benchmarks.
Watch revenue per dollar of compute.
Watch how much contracted capacity actually arrives and becomes usable.
Watch the gap between customer contract duration and infrastructure duration.
Watch whether operating cash generation survives the next wave of upfront investment.
Watch how much negotiating power remains when the same partners provide capital, chips, distribution, and competing models.
And watch the distinction between accounting losses, operating economics, and cash committed years into the future.
Anthropic's S-1 makes one thing very clear.
The frontier model is only part of the product.
The other product is a financing system capable of keeping the machines turned on.