Anthropic’s Compute Bets Put Safety Under Pressure

Anthropic warns that using AI to develop AI could push its models’ capabilities ahead of its ability to build safety, alignment and security measures. The warning appears in a business filing centered on rapid growth and enormous commitments to compute, the infrastructure needed to train and run its models. The company’s contracts make growth less a choice than a condition of its plan.

Safety on the company’s clock

Anthropic’s risk factors say AI-assisted development “may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures.” The filing also refers to catastrophic harm and the development of forms of sentience. These are the company’s own warnings about what could happen as it deploys increasingly capable systems.

Claude is already helping build Claude, under human direction, according to the article. Anthropic says, “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.” That’s the loop the company is betting on: more capability, faster development, then more deployment. The filing’s safety warning describes the risk of that loop outrunning safeguards.

Anthropic’s revenue grew fourteen times to $16.3 billion between the first half of 2025 and the first half of 2026. The company’s net revenue retention was 900 percent, and about three-quarters of its customers in the first half of 2026 spent at least $100,000 a year. Anthropic sells mainly to enterprise customers and focuses on coding and cybersecurity, naming only OpenAI and Google as rivals in the filing.

Contracts that demand growth

Anthropic disclosed $265.6 billion in fixed minimum non-cancelable purchase commitments, with terms between one and 20 years. Lease agreements, primarily related to compute and infrastructure capacity, that hadn’t yet commenced totaled $321.4 billion in unaudited figures as of June 30. After that date, the company entered compute and infrastructure agreements totaling about $245 billion, also with terms between one and 20 years.

Other transactions not yet finalized carried estimated costs of $125 billion to $150 billion, in addition to variable compute commitments. The Verge’s Elizabeth Lopatto estimated minimum commitments of about $850 billion over the next 20 years, rising to about $1 trillion if unfinished transactions were included at the current estimated cost. She noted she might have missed items. The commitments run through 2045.

PitchBook analyst Harrison Rolfes modeled spending at about $90 billion a year from 2027 to 2030, implying $188 billion in revenue to maintain the current cost-to-compute ratio. “Growth is not optional for this company,” he wrote. “The contracts assume it.” A corporate strategy backed by obligations on this scale has little room to slow down without confronting costs the company has already locked in.

Efficiency, competition and the bill

Compute and infrastructure costs associated with customers, along with other items, totaled $2.3 billion in the first half of 2026. Model development and related items cost $9.7 billion. Rolfes said Anthropic was spending less on compute for each dollar of revenue, while the company still needs to train models efficiently and generate enough income to meet its commitments.

Anthropic faces competition from other frontier AI labs and open-source models. Claude Opus 5.5 topped the Epoch Capabilities Index at the time of the article, though the leaderboard changes frequently. Google has multiple paths to revenue and can bundle AI into products businesses already use, giving it a distribution advantage. Anthropic’s revenue tends to arrive in spurts after it releases new models.

Cybersecurity is one area Anthropic is targeting for growth. Hackers can use AI to conduct attacks, and security problems have also arisen from a partner’s handling of protocols around AI agents. Opus 5.5 reduced customer costs by 40 percent compared with Opus 5; Anthropic said lower costs could increase usage and preserve margins.

The filing showed an $80 billion loss, described in the article as largely an accounting matter that didn’t reflect the underlying business. Excluding that and other charges, including employee compensation, Anthropic had $1.62 billion in adjusted income in the first half of 2026. Futurum Group analyst Brendan Burke said the filing captured “a happy period” when early models surged in usage and limited compute let Anthropic charge more. DA Davidson analyst Gil Luria said greater efficiency doesn’t necessarily reduce margins, pointing to cloud providers including Amazon Web Services. Burke said frontier labs need high-margin products to avoid a race to the bottom in token prices. The company’s own warning remains: capability may advance faster than the measures meant to contain it.