Anthropic’s Compute Pledges Near $850 Billion Face Profit Test

Grace Kim

Anthropic has $265.6 billion in minimum, non-cancelable purchase commitments and $321.4 billion in leases not yet commenced.

The leases primarily cover compute and infrastructure capacity, based on unaudited June 30 figures.

Together, newer compute agreements and the disclosed obligations imply a floor of about $850 billion through 2045, before adding an estimated $125 billion to $150 billion in pending transactions.

That scale makes sustained revenue growth, not model rankings, the central commercial issue in the S-1.

Can Anthropic sustain this growth?

Revenue grew fourteenfold to $16.3 billion in the first half of 2026 from the first half of 2025.

Anthropic reported an $80 billion loss. The filing review described that charge as largely accounting-related.

After removing that loss and other charges, including employee compensation, adjusted income reached $1.62 billion, suggesting AI revenue can currently cover Anthropic’s operating base.

About three-quarters of customers spent at least $100,000 annually in the first half of 2026.

Net revenue retention reached 900%.

Anthropic names only OpenAI and Google as rivals. It has not released a consumer image generator, a TikTok clone or an anime-style character product.

Instead, it concentrated on enterprise coding and cybersecurity, targeting corporate IT budgets.

How does competitive pressure shape the strategy?

Other frontier labs continue to target the leading edge. Claude Opus 5.5 led the Epoch Capabilities Index when the filing was published, although rankings change frequently.

Open-source alternatives lagged by roughly four months.

Google has several routes to AI revenue and can wrap AI into products that businesses already use, while Anthropic must rely on fewer distribution channels and stronger model performance.

Microsoft Teams displaced Slack through bundled distribution, rather than decisive product superiority.

That distribution gap matters less in cybersecurity, where model capability remains critical. Customers must update constantly because attackers can use AI for brute-force attacks.

A partner’s flawed security protocols around AI agents have also caused product problems.

Revenue arrives in bursts after major model releases. Anthropic cannot stop training, and must improve efficiency against larger competitors.

What do the commitments require?

The S-1 leaves aggregate purchase and lease figures blank on page 64. It also omits expected spending through Dec. 31, 2030, on page 137.

Despite those gaps, the filing identifies the following obligations:

• $265.6 billion in fixed minimum, non-cancelable purchase commitments, with terms of one to 20 years.
• $321.4 billion in leases not yet commenced, primarily for compute and infrastructure capacity, based on unaudited June 30 figures.
• About $245 billion in newer compute and infrastructure agreements, also with one- to 20-year terms.
• $125 billion to $150 billion in estimated costs for transactions not yet finalized.
• Additional variable compute commitments without a stated total.

These items imply roughly $850 billion in minimum obligations over 20 years. Including pending transactions at current estimates raises the total toward $1 trillion.

The calculation may omit other obligations.

Anthropic did not disclose a payment schedule. PitchBook analyst Harrison Rolfes models about $90 billion in annual payments from 2027 through 2030, a pace that would require $188 billion in revenue to preserve the current cost-to-compute ratio.

“Growth is not optional for this company. The contracts assume it.”

Can efficiency preserve margins?

Customer-associated compute, infrastructure and related items cost $2.3 billion in the first half of 2026. Model development and related items cost $9.7 billion.

Rolfes said Anthropic simultaneously spent less on compute for each dollar of revenue.

Futurum Group analyst Brendan Burke cautioned that the S-1 captures an early period of surging model usage, constrained compute capacity and higher prices, leaving the durability of the efficiency trend unconfirmed without third-quarter figures.

Opus 5.5 reduced customer costs by 40 percent from Opus 5. Anthropic says lower prices drive more usage, which preserves margins.

If each new model repeats that reduction, customers could perform tasks at a fraction of current prices. Burke said frontier labs need high-margin products to prevent token-price competition from eroding margins.

DA Davidson analyst Gil Luria offered a counterpoint. AWS and other cloud providers have become more efficient and cut prices annually while sustaining growth and rising margins.

Is recursive self-improvement essential?

Anthropic already uses Claude to develop Claude under human direction. The risk factors frame that dependency as a commercial necessity, with clear safety tradeoffs.

The company describes the model-development loop this way:

“More capable models help our researchers discover these improvements faster; in turn, more powerful compute multipliers enable us to train even more capable models and deploy them more efficiently.”

The filing’s broader language points to an expectation of dramatic recursive self-improvement with minimal or no supervision, perhaps soon.

Its risk factors warn that AI-assisted development may cause model capabilities to advance faster than safety, alignment and security measures can be developed, validated and implemented, leaving Anthropic exposed if its controls lag.

The filing also discusses catastrophic harm and possible forms of sentience.

Through 2030, the commercial test is whether enterprise demand and faster training can support the $188 billion revenue level Rolfes says this spending pace requires, while lower model prices preserve margins against Google, OpenAI and open-source rivals.