Don’t expect supersize profits from Anthropic any time soon

BREAKNECK growth is commonplace in Silicon Valley. The rise of Anthropic is something else. In the first quarter of last year – when Claude Code, its hit programming tool, was first released – the artificial-intelligence lab pulled in US$376 million of revenue.

By the second quarter of this year that figure had risen to 30 times its original level, to US$11.6 billion, exceeding corporate mainstays such as McDonald’s.

In a few weeks this darling of the AI revolution will undergo an initial public offering that will be, if not the largest of all time, then certainly close to it. On Monday (Oct 26), it publicly filed supporting paperwork that gave the outside world a closer look at its finances.

Anthropic’s growth is all the more remarkable because, unlike many zippy startups, it has already reached profitability. In the second quarter, it eked out an operating profit of just under US$1 billion, having lost US$1.7 billion in the preceding three months.

The company is particularly eager to draw attention to its spending on computing power, which fell from 251 per cent of sales in the first quarter of 2025 to just 48 per cent in the second quarter of 2026.

Anthropic claims that improvements in its AI models have made it faster and more efficient to develop and serve its technology, allowing the company to make the most of its limited resources.

But can Anthropic’s stunning performance continue?

A closer look at its filing suggests caution is warranted: growth is slowing; the company is paying higher prices to access scarce computing power; and a resulting shift towards directly purchasing AI chips will guzzle up much of the cashflow Anthropic would otherwise be on course to generate.

Earlier investors who had hoped for a public valuation as high as US$2 trillion may thus be disappointed.

Anthropic’s scorching revenue growth has put a target on its back. Claude Code made it an early leader in the race to automate programming. Tech giants such as Meta and Microsoft became huge customers, spending billions of dollars on Anthropic’s technology.

More recently, those customers have been pushing staff to use internally developed alternatives. At the same time, OpenAI, Anthropic’s arch-rival, has been gaining ground with its own offering, Codex, the daily usage of which has risen to 10 times its early-August level.

And after the short-lived craze for “tokenmaxxing” earlier this year, companies of all stripes have been more closely monitoring their spending on AI and imposing limits on their engineers.

This month, Anthropic told shareholders during a private gathering that revenue in the third quarter was up by about 45 per cent from the preceding three months, a third of the rate it achieved in the second quarter.

In its IPO prospectus, Anthropic warned that, as the market matures, growth could slow “more rapidly than we anticipate”. Moreover, it has become increasingly reliant on the tech giants to push its products.

Almost half of Anthropic’s revenue in the first six months of this year came from sales made through the cloud businesses of Alphabet, Amazon and Microsoft. Anthropic’s spending on sales, marketing and distribution partnerships such as these chewed up 19 per cent of revenue in the first half of 2026, up from 10 per cent in the same period a year before.

Compute limitations

Another problem Anthropic faces is that, like other AI developers, it is struggling to secure the computing power it needs. In its prospectus, it warned investors that its use of “short-term compute” could go up in the near term, weighing on margins.

For example, Anthropic has agreed to pay SpaceX about US$3.1 billion a month to use its Nvidia chips into 2029. (Anthropic is able to back out from early next year with 90 days’ notice.)

That explains why Anthropic is starting to develop more of its own computing infrastructure, rather than relying solely on cloud providers.

This strategy should help keep operating costs under control, but also provides forewarning that the lab will become a more capital-intensive business in the long run.

The initiative, called “Anthropic Direct”, involves procuring chips from providers such as Broadcom, which develops custom silicon, and installing them in rented data centres. That could make it harder to start throwing off the kind of free cashflow investors might hope for, and require the business to raise more capital.

In Anthropic’s prospectus, its founders argued that its spending plans “may seem irresponsibly large at first glance”, but are justified by the exponential trajectory that AI is on.

“We will aim to be bold but sane,” they write. With a technology so mind-bending, striking that balance may prove surprisingly difficult. ©2026 The Economist Newspaper Limited. All rights reserved