Anthropic raised $81.3 billion in the largest IPO on record while continuing to change how much Claude subscribers could use within a five-hour window. The stock sale prices years of expected growth; the window allocates a finite supply of machine-hours today.

Key takeaways

  • Anthropic’s amended S-1 disclosed $895.4 billion in purchased commitments and lease agreements, substantially all tied to compute and infrastructure.
  • Anthropic’s reported compute and infrastructure spending rose from $2.3 billion in the first half of 2025 to nearly $9 billion in the first half of 2026.
  • The reported Anthropic Direct plan included leases for roughly 875,000 TPUs at about $7.2 billion a year.
  • Amazon and Google are set to hold nearly 15% and 14%, respectively, of Anthropic’s outstanding capital stock after the IPO and private placement.
  • Anthropic made a 25% increase in standard weekly Claude Code limits permanent after initially offering a larger temporary increase.

The IPO buys time for capacity to earn its cost

In the coverage record, Anthropic’s 2021 Series A raised $124 million. Five years later, the company priced its IPO at $83 a share. The Series A funded a young model developer. The IPO arrived after the developer had made commitments to an AI infrastructure system measured in leases, power, buildings, and chips.

Anthropic’s amended S-1 disclosed $895.4 billion in purchased commitments and lease agreements, substantially all related to compute and infrastructure. That is not a bill payable at once, and it is not a measure of capacity already running. Anthropic has committed to obtaining a great deal of capacity before the full economics of selling work on it are visible.

The IPO changes who can fund the wait, not whether the reserved capacity must earn a return. Public shareholders can provide equity runway while Anthropic builds demand; cloud partners and lenders can finance parts of the physical system. None of those transactions establishes that a leased accelerator will be busy at a price that covers its share of the cost.

Scarce chips once made securing supply the urgent question. Anthropic still has to secure it, but a large reservation book adds another task: selling enough useful inference on the machines it has agreed to take. The financial capacity to keep expanding before cash flows validate each commitment is an advantage. It also allows the interval between commitment and proof to grow.

A leased TPU makes forecasting an operating job

An analysis of Anthropic’s agreements counted at least 14.8 gigawatts of compute capacity and estimated spending of as much as $517 billion over the following decade. Anthropic’s reported compute and infrastructure spending had already risen from $2.3 billion in the first half of 2025 to nearly $9 billion in the first half of 2026. That spending is a strong reason not to dismiss the buildout as a paper exercise. An agreement for power, though, does not say how much paid work will run on the resulting hardware.

The reported Anthropic Direct plan makes the mismatch of clocks more concrete. Documents described more than $900 billion in planned commitments through 2045, including leases for roughly 875,000 TPUs at about $7.2 billion a year. The plan’s figure and the S-1’s $895.4 billion have different reported scopes and may overlap; adding them would manufacture a larger obligation, not clarify the one Anthropic faces.

Anthropic has to forecast demand by model, workload, time of day, and price against capacity obtained on a much longer schedule. It faces a capacity lag when a customer’s preferred model or response speed changes after a lease has been arranged. A human waiting for an answer values latency differently from an agent working through tasks without someone watching each response. Anthropic could schedule some of that agent work for less crowded hours, but scheduling can fill an idle period only if customers have work they will pay to run.

Physical scarcity has not vanished. US data-center builders have faced shortages of power, property, and parts, with reported cooling-system lead times five times longer than before. A lab can therefore encounter a delay getting one site ready while trying to keep another site’s aging machines occupied. An airline cannot carry yesterday’s empty seat into tomorrow; Anthropic cannot carry an idle TPU-hour forward either.

Claude’s subscription is a claim on metered work

Claude Code customers on the $200-per-month Max plan encountered unexpectedly restrictive limits in July 2025. In March 2026, Anthropic adjusted session limits during peak hours as Claude’s popularity strained compute. Those changes exposed what a flat software price can obscure: each additional session asks the provider to deliver variable, time-sensitive machine work.

Anthropic did not only tighten access. It later made a 25% increase in standard weekly Claude Code limits permanent after a larger temporary increase, and raised five-hour limits by 20% for Pro, Max, and Team plans while offering a rate-limit reset. More generous access matters. It shows that supply, scheduling, and product policy can change rather than following a one-way path toward rationing.

The price still has to fit the workload. Anthropic says a typical Claude Code review consumes $15 to $25 in tokens. A separate analysis estimated that a $200 Max subscription could provide up to roughly $8,000 a month of tokens at API prices. That upper-bound comparison is neither a typical customer’s usage nor Anthropic’s cost to serve it. It does show why subscription revenue alone cannot describe the exposure created when customers delegate more work to agents.

Recurring enterprise tasks offer a different kind of evidence from sign-ups or generous limits: customers must find the completed work valuable enough to keep buying it. Anthropic can use prices, credits, and usage windows to distinguish those tasks from occasional experimentation. Each control determines which customer work gets a scarce hour on the fleet.

Cloud partners own part of the wait

Amazon is building an Indiana data-center complex for Anthropic’s computing needs. An SEC filing indicates that Amazon and Google are set to hold nearly 15% and 14%, respectively, of Anthropic’s outstanding capital stock after the IPO and private placement. They are not merely vendors quoting a price for capacity; they also have ownership exposure to the customer using it.

Other companies have extended the frontier-lab financing chain. Anthropic committed to buy $30 billion of Azure capacity in a deal accompanied by potential investments of up to $5 billion from Microsoft and $10 billion from Nvidia. Separately, Oracle moved $66 billion of debt for AI data-center construction off its balance sheet through special-purpose vehicles. That Oracle debt is not Anthropic’s obligation. It illustrates how builders can place financing for this asset class beyond a cloud company’s own balance sheet, while lenders and insurers still have to assess the buildings and their prospective tenants.

Amazon’s ownership stake, a cloud contract, and a lender’s claim do not carry identical risks. They do tie several institutions to the same practical question: whether customers use enough capacity, for long enough, to support what each institution supplied. The IPO adds public equity to that chain without collapsing its different claims into one balance sheet.

Older chips can make a full reservation less valuable

Supply alternatives sharpen the test. Inferact, founded by the creators of the open-source vLLM inference engine, raised $150 million to commercialize serving infrastructure. DeepInfra raised $107 million and supported more than 190 open models. Mistral trained Mistral Large 4 on 3,800 Nvidia Grace Blackwell GPUs in European data centers. None of those facts proves that customers can replace every Anthropic workload. They do show that more companies are building ways to train, serve, and buy model capacity outside a single frontier lab’s reservations.

Hardware age can matter before a long contract ends. Microsoft has reportedly been turning off rented Oracle capacity after declining to renew a 2023 deal because it used older Nvidia chips. Microsoft’s decision concerns Oracle’s equipment, not Anthropic’s fleet. It shows the operating condition that a large reservation must survive: a machine can remain available after a buyer decides newer capacity is worth more.

One analysis estimates that Anthropic would need roughly $602 billion in annual revenue to reach a 20% return on invested capital comparable to Big Tech peers. That is an analyst’s benchmark, not a company target or a demonstrated shortfall. Investors lack fleet utilization, inference gross margin, and customer-cohort profitability figures in this record, so they cannot establish whether Anthropic’s commitments will earn their cost of capital.

Frequently asked questions

How much Anthropic IPO stock was sold to retail investors?

Anthropic sold $1.1 billion of stock to retail investors in its IPO. The disclosed retail group included investors in the UK, EU, Norway, and Switzerland.

What corporate form is Anthropic expected to have as a public company?

Anthropic is set to become the largest public benefit corporation in the US stock market. A public benefit corporation has a stated public-purpose obligation alongside its financial objectives.

Who said they invested in Anthropic to promote AI safety?

Dustin Moskovitz said he invested in Anthropic to promote AI safety.

Which pro-AI safety organization did Anthropic financially back?

Anthropic backed Public First Action, described in the record as a pro-AI safety nonprofit.

Reported scale of Anthropic’s capacity commitments

MeasureReported figureTimeframe or scope
Purchased commitments and lease agreements$895.4 billionSubstantially all related to compute and infrastructure
Compute capacity counted in an agreement analysisAt least 14.8 gigawattsReported agreement capacity
Estimated spending in that analysisAs much as $517 billionOver the following decade
Planned commitments in the reported Anthropic Direct planMore than $900 billionThrough 2045; reported scope may overlap with the S-1 figure
TPU leases in the reported Anthropic Direct planRoughly 875,000 TPUs at about $7.2 billion a yearAnnual lease cost
Reported compute and infrastructure spending$2.3 billion to nearly $9 billionFirst half of 2025 to first half of 2026

Claude’s five-hour counter resets. The annual lease clock on roughly 875,000 planned TPUs does not.