Anthropic now generates $2.08 in revenue for every dollar it spends on compute, according to its S-1 filing with the U.S. Securities and Exchange Commission (SEC). The filing, submitted on October 26, 2026, comes as the AI lab prepares for a potential initial public offering (IPO). An S-1 is the registration document U.S. companies file with the SEC when preparing to sell shares to the public, and it requires standardized disclosure of financial results for regulators and prospective investors to examine.

The figure marks a sharp reversal from the first quarter of 2025, when the same dollar of compute brought in just $0.40. For a business whose primary raw material is GPU time, the filing frames the change as the difference between a science project and a company.

The numbers inside the filing

Anthropic posted revenue of $11.555 billion for Q2 2026, while compute and infrastructure expenses for the quarter came to $5.558 billion. Dividing one by the other yields the $2.08 ratio.

Growth rates tell the same story from a different angle. Revenue jumped 1,368% year-over-year in Q2 2026, while compute costs rose 300% over the same stretch.

As a share of revenue, compute fell to 48% in Q2 2026. In Q1 2025, it stood at 251% of revenue, meaning Anthropic was spending roughly two and a half dollars on infrastructure for every dollar customers paid.

The improvement also showed up on the operating line. Anthropic recorded an operating profit of $974 million for Q2 2026, an 8% margin.

The company nonetheless reported a net loss of $66.3 billion for the quarter, driven primarily by fair-value charges. Fair-value accounting requires companies to periodically reprice certain assets and liabilities to current market values, so headline net income can swing sharply even when day-to-day operations are unchanged. Because those charges sit below the operating line, they reflect accounting revaluations rather than the cost of serving customers.

Why compute efficiency matters so much

Anthropic's filing credits much of its improvement to more efficient use of a shared compute pool. Rather than tying specific workloads to specific hardware, the company draws from a single pool that spans NVIDIA and AMD GPUs, Google TPUs, and AWS Trainium chips.

NVIDIA CEO Jensen Huang has been making a related argument in public. Huang says software optimization can double AI model speed on existing GPUs and double revenue.

Anthropic's numbers line up with that thesis. Compute spending kept climbing, yet revenue grew far faster, which suggests each unit of hardware was doing considerably more productive work by Q2 2026 than it was in early 2025.

What this means for Anthropic's IPO

For prospective investors, the ratio of revenue to compute cost may be the single most useful number in the filing. It speaks directly to whether AI model businesses can earn their way out of their infrastructure bills. A move from $0.40 to $2.08 in roughly five quarters is a strong answer.

The 8% operating margin is modest by software standards. But it marks a turn from a cost structure in which compute alone exceeded revenue several times over.

The $66.3 billion net loss is the obvious counterweight. Investors will want to understand exactly what drove the fair-value charges, whether they recur, and how they might affect shareholders after a listing. Filing an S-1 also starts, rather than concludes, the listing process, with SEC staff review and investor roadshows typically preceding any pricing of shares.

For chip suppliers, the picture is more nuanced. Anthropic's compute costs still rose 300% year-over-year, so efficiency gains have not meant smaller hardware budgets.

The multi-vendor approach adds another wrinkle. With NVIDIA, AMD, Google, and AWS hardware all in the pool, no single supplier holds all the leverage.

Reporting: CryptoBriefing. Edited by Diego Almada Lopez.