What is in Anthropic's IPO filing?
Anthropic filed its Form S-1 with the US Securities and Exchange Commission on 26 October 2026. The company plans to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The document runs to 335 pages, according to the BBC, and leaves both the price range and the share count blank.
The headline numbers are unusual even by the standards of the AI industry. Revenue reached $16.3 billion in the six months to 30 June 2026, up from $1.2 billion a year earlier. That is year-over-year growth of 1,300%, a rate Anthropic says it does not expect to sustain.
The company reported a net loss of $80.7 billion for the same period. It lost $42.0 billion across the whole of 2025, $8.3 billion in 2024 and $1.5 billion in 2023. Cash and short-term investments still came to $131.7 billion at the end of June.
The offering itself is still incomplete. The cover page leaves the price per share blank, and selling stockholders will sell an unspecified number of shares. The filing does include unaudited figures through 30 September 2026, which is why several metrics run three months past the audited half year.
Anthropic is incorporated as a public benefit corporation, which obliges it to weigh its stated mission alongside shareholder returns. A Long-Term Benefit Trust holds special rights over company leadership. Dario Amodei is chief executive and Daniela Amodei is president.
The filing also confirms the company's scale in enterprise software. More than 11,900 customers were spending at least $100,000 a year by 30 September 2026, up from about 1,500 at the end of 2025. Roughly 250 of them spend $10 million or more.
Why is the loss $80.7 billion on $16.3 billion of revenue?
Most of the loss is an accounting entry rather than money leaving the business. Anthropic recorded a $72.4 billion charge for the rising value of convertible notes held by Amazon. Those notes turn into shares, so their fair value climbs as Anthropic's valuation rises.
Remove that charge and the picture changes. The operating loss for the half year was $693 million, down from $2.1 billion a year earlier. Adjusted operating income, a non-GAAP measure the company defines, was $1.6 billion over the same period.
The cost lines show where the money actually goes. Cost of revenue, mostly compute for customer workloads, was $2.38 billion. Research and development, largely compute for model development, reached $9.74 billion. Sales and marketing rose to $3.81 billion from $239 million a year earlier.
Research spending is the line that moved most. It was $1.85 billion in the first half of 2025, so it roughly quintupled in a year. Sales and marketing grew even faster, from a small base, as Anthropic pushed into large enterprises and free consumer usage.
For comparison, cost of revenue was $2.77 billion for all of 2025, when revenue was $4.6 billion. Customer-serving compute is now the largest variable cost, and spending on model development is roughly four times larger again.
The Verge's Elizabeth Lopatto treats the $80 billion loss as largely an accounting matter that is not relevant to the underlying business. She also notes that some totals in the filing are left blank, which makes outside reconciliation harder.
Analysts quoted by the BBC are more cautious. Troy Hooper of Mergermarket said investors lack a clear view of underlying profitability because of the accounting charges. In his reading, buyers would be paying for future dominance rather than present-day economics.
Who is Anthropic paying for compute?
The S-1 lists long-term infrastructure agreements with seven named suppliers. Several carry minimum spending guarantees, which Anthropic must honour whether or not it uses the capacity. These are the figures the filing gives for each one.
- •Google: at least $114.5 billion between April 2026 and July 2033, with a true-up payment if spending falls short.
- •Amazon: at least $110.0 billion between May 2026 and April 2036.
- •Microsoft: at least $44.9 billion between November 2026 and May 2033.
- •SpaceXAI: up to $120.5 billion through 2029, at roughly $3.1 billion a month once capacity is fully delivered.
- •Nscale: up to $44.6 billion on terms extending into 2033.
- •Lambda: about $41.3 billion on terms extending into 2034.
- •AMD: more than $20 billion expected for Helios rack systems in 2027, with no fixed minimum.
How large are Anthropic's commitments, and who is on the other side?
As of 30 June 2026, Anthropic had $265.6 billion of fixed or minimum purchase commitments, with terms of one to 20 years. It also had lease agreements for compute and data centre capacity that had not yet started, worth $321.4 billion. That lease figure was $50.2 billion at the end of 2025.
The filing then sets out agreements signed after that date. New and amended leases and commercial contracts carry minimum payments of about $417 billion. A further $125 billion to $150 billion is estimated for deals whose terms are not final.
The Verge added the disclosed categories and arrived at roughly $850 billion of commitments over 20 years. That total would approach $1 trillion if the unfinalised transactions are counted at current estimates. This is the publication's arithmetic, not a figure the company reports.
PitchBook analyst Harrison Rolfes models payments of about $90 billion a year from 2027 to 2030. On his arithmetic, Anthropic needs roughly $188 billion of revenue to keep its current ratio of compute cost to revenue. He told The Verge that growth is not optional, because the contracts assume it.
Two features of the supplier list stand out. First, Anthropic is buying capacity from companies that also compete with it. Google, Amazon and Microsoft are suppliers, investors and rivals at once, and the filing names OpenAI and Google as direct competitors in frontier models.
Second, the list now includes Elon Musk's AI business. The filing refers to SpaceX.AI LLC, formerly X.AI LLC, as a wholly owned subsidiary of SpaceX. Anthropic expects that capacity to be fully delivered in December 2026.
Google's arrangement covers tensor processing units as well as cloud capacity. Anthropic also agreed to buy additional TPU compute for delivery between 2028 and 2031, estimated at $125 billion to $150 billion. Those payments depend on capacity delivered, so they sit outside the fixed commitments table.
The BBC quotes Michael Monaghan of Founder ETF, who points out that many of Anthropic's largest customers have no long-term contract. Concentration has fallen, but the company still relies on a small group of buyers to support its growth.
For context, the company ended June with $131.7 billion of cash and short-term investments, having raised $141.1 billion in total. The commitments are spread across two decades, so they do not all fall due at once. Even so, they convert future growth from an ambition into an obligation.
Why is growth not optional for Anthropic?
Anthropic's enterprise sales model is the engine behind those contracts. Net revenue retention was approximately 900% as of 30 September 2026, meaning existing customers spent about nine times what they spent a year earlier. Retention above 100% is strong for enterprise software; 900% reflects a very young customer base.
The customer mix is broadening. More than half of the Fortune 500 and over 30% of the Global 2000 were customers spending at least $100,000 a year, the filing says. Revenue from those Fortune 500 accounts still averages less than 0.1% of their total expenses.
Revenue is also tied to release timing. Model versions launched in the six months to September 2026 generated more than 80% of third-quarter revenue. Anthropic warns that this cadence creates variability between periods, because usage follows each new model.
Concentration is falling but still matters. About 79% of third-quarter revenue came from outside the top five customers, up from 64% in the first quarter. No single customer accounted for more than 10% of revenue in the nine months to September.
Efficiency is the counterweight. Anthropic reports trailing-twelve-month revenue per employee above $6.6 million as of September 2026, against $1.4 million a year earlier. It also says revenue grew 1,368% in the June quarter while compute and infrastructure expenses grew 300%.
Distribution is another lever. Claude is sold through all three major clouds and through systems integrators including Accenture, Deloitte, KPMG and PwC. Data platforms such as Salesforce and Snowflake also carry it, and Anthropic has launched a services venture called Ode with Anthropic for mid-sized companies.
The company argues that cheaper tasks expand usage rather than shrink revenue. It says Claude Opus 5.5 cut customer cost per task by about 40% against Opus 5, while preserving margins through better compute efficiency. Whether that holds as models get larger is the open question.
What does 'closing the loop' mean in the filing?
Anthropic says Claude now helps build Claude. The filing states that research, training, testing and evaluation for new models involve Claude itself, including writing and reviewing code. The company calls this recursive dynamic 'closing the loop'.
In its own words, more capable models help researchers find improvements faster, and those improvements allow the training of more capable models. That describes a feedback cycle rather than a process that runs without people. The filing says the work happens under human direction.
The same section carries a warning. Anthropic says its use of AI to accelerate research may cause capabilities to advance faster than its ability to develop, validate and implement safety measures. It adds that much of its evidence on acceleration comes from lagging indicators.
The risk factors go further. They discuss catastrophic or existential risks to humanity and the possibility that systems could develop forms of sentience or moral status. They also note that recursive self-improvement is monitored under the company's Responsible Scaling Policy.
The filing also defines what it calls Powerful AI. It describes systems that match top human experts across most disciplines, control computers directly, work autonomously for days and coordinate with other agents. Anthropic says such systems will be widely available within a few years.
The Verge reads the business plan as depending on rapid recursive improvement, because growth is what pays for the compute contracts. That is an interpretation of the filing rather than a statement by Anthropic.
Capability claims in the document are concrete in places. Anthropic says Claude's score on ExploitBench, a Carnegie Mellon benchmark for exploiting software vulnerabilities, rose from 40% with Opus 4.8 to 91% with Opus 5.5 in about four months. The benchmark measures offensive security skill, which the company frames as a defensive tool.
Who gains most if Anthropic lists, and what happens next?
Amazon holds $6.2 billion of convertible notes, which is Anthropic's entire debt load. Amazon valued those notes at $97.9 billion as of 30 June, according to TechCrunch. They can convert into as many as 1.84 billion shares, capped at 12% of the partially diluted share count.
TechCrunch reports that Amazon has invested about $18 billion in total, including $10 billion of preferred stock in the second quarter. Anthropic has also committed to spend at least $110 billion on AWS over ten years. That cloud contract is worth as much to Amazon as the equity stake.
Google's position is smaller but still large. Court filings reported by the New York Times put its stake at about 14% in early 2025, and PitchBook's Rolfes estimates it nearer 8% now. Google also holds a $114.5 billion compute commitment from Anthropic.
Venture firms will make money, but each holds less than 5% of the company, which TechCrunch calls highly unusual for an IPO of this size. At a $965 billion private valuation, a 1% stake is worth almost $10 billion.
The note cap matters for the listing. Amazon's notes that cannot convert stay outstanding and convert in later quarters under the same limit, so its stake builds over time rather than all at once.
The BBC reports that Anthropic is widely expected to list at a valuation of $2 trillion or more. That would make it the largest stock listing in history, but the filing sets no price and the expectation remains unconfirmed.
The first thing to watch is pricing. No roadshow schedule has been published, and the share price and share count are blank. Until an amended prospectus sets a range, every valuation figure is an estimate.
The second is competition. Anthropic names OpenAI and Google as direct rivals and acknowledges open-weight models that trade intelligence for cost. Cheaper rivals put pressure on the pricing power that its compute contracts depend on.
The filing describes a company growing faster than any enterprise software business of its size, and one that has already promised much of its future cash to compute suppliers. Whether that is prudence or a bet on rapid self-improvement is the question investors will price.