Anthropic has spent the past two years signing up the companies that once looked like its most natural rivals. Microsoft invested billions. Meta’s engineers adopted Claude Code by the tens of thousands. Both are now walking that back, quietly and on the eve of Anthropic’s public debut.
The details come from The Information, which reported this week that the two companies, among Anthropic’s largest corporate customers, are cutting their internal use of Claude. Microsoft had projected it was on pace to spend at least $1 billion this year on its own internal use of Anthropic’s technology, according to a person familiar with the matter. That estimate has since been slashed by more than a third, the report said, as leadership steered employees toward the company’s own tools.
The mechanism at Microsoft is a budget, and the cuts are sharp. Executives Scott Guthrie and Jay Parikh told employees to use in-house offerings such as GitHub Copilot and OpenAI’s models, according to The Information. In the cloud division, the monthly per-employee budget for outside AI tools reportedly fell from $100,000 to about $10,000. Engineers in the Experiences and Devices group have been migrated onto the GitHub Copilot command-line tool.
Meta tells a similar story with different numbers. Internal use of Claude Code has fallen to roughly 30,000 employees from about 60,000 earlier this year, The Information reported. Part of that drop is the spring layoffs that cut about 10% of Meta’s 78,000-person workforce. The larger part is substitution. MetaCode, the company’s internal coding tool, has passed 30,000 users, and Muse Code, an external competitor to Claude Code the company began testing in August, counts more than 6,000 internal users. Meta’s own summary, quoted in the report, is blunt: the rapid advance of its proprietary coding tools has largely displaced the need for Claude.
Meta’s retreat fits a longer pattern inside the company. It introduced AI cost-cutting measures in June as internal spending on AI tools reportedly climbed into the billions, and in one 28-day stretch the company spent more than $105 million on Claude Code alone, according to The Information. Even so, Meta had separately projected it could spend up to $10 billion a year on Anthropic’s models, a number that now reads less like a plan than an upper bound it intends to avoid.
The pullback is narrower than the headline numbers suggest, and the boundaries matter. Microsoft’s investment in Anthropic, reported at up to $5 billion, is untouched, as is Anthropic’s roughly $30 billion commitment to buy compute capacity from Microsoft’s Azure. Customer spending on Claude through Azure and Amazon’s Bedrock continues to grow. What is shrinking is the part Anthropic cannot build a durable business on anyway: two big tech companies paying to use a rival’s model for their own employees.
The timing is what gives the move its edge. Anthropic is preparing to list its shares, with a target it has put before Thanksgiving, and its prospectus has begun to surface numbers that investors are still digesting. Two of its marquee customers choosing this moment to reduce their internal reliance sends a signal the company would rather not have attached to its roadshow. Analysts said the retrenchment adds a headwind to Anthropic’s revenue expectations, even if the direct financial impact is modest.
Meta’s chief executive, Mark Zuckerberg, has attacked Anthropic in public while his company kept paying for its tools, a contradiction the internal cuts begin to resolve. Microsoft sits in a gentler version of the same bind: it owns a stake in OpenAI and sells Copilot, yet its own engineers were choosing Claude often enough that executives felt the need to set budgets. The companies are not severing ties with Anthropic; they are ending the habit of paying it to be better than them.
There is a logic to the timing from the customers’ side as well. Microsoft and Meta are both racing to make their own models good enough that they no longer need to pay for someone else’s. Every employee who switches from Claude to an in-house tool is a dollar kept and a vote of confidence in a product they are trying to sell. Meta has gone further, signaling it wants to turn Muse into a moneymaker by selling AI services to businesses, which puts it in direct competition with the company whose tools it is still buying.
Anthropic has grown fast enough to become a problem for the incumbents, and that is part of the point. Its products have begun to encroach on the office-software territory Microsoft and Meta care about, and its coding tools have become the default in places that used to belong to them. The two companies now find themselves in the awkward position of being Anthropic’s investors, its competitors, and, for a while longer, its customers. This week’s numbers are the first clear sign of which of those roles they intend to keep.
