Meta and Microsoft Cut Internal Claude Usage, Anthropic Faces Pressure from Major Client Losses
nashnova research
Meta and Microsoft are actively slashing internal use of Anthropic's Claude — one halving its user count, the other trimming over a third of its budget. This means Anthropic's biggest revenue pillars are being replaced by in-house tools just as it prepares what could be the largest IPO in history.
How deep are Microsoft's cuts?
Microsoft executives had projected at least $1 billion in internal Claude spending this year. That figure has since been cut by more than a third.
The reason is straightforward: Microsoft told employees to reduce Claude usage to control costs and shift to its own AI tools.
This means → Microsoft, one of Anthropic's largest customers, is redirecting its internal AI budget from "buy" to "build."
An important distinction: the cuts apply only to internal employee usage. Spending by Microsoft's customers who access Anthropic models through Microsoft products continues to grow.
Why did Meta's user count halve?
The number of Meta employees using Claude Code — Anthropic's coding assistant — dropped from roughly 60,000 earlier this year to about 30,000, a 50% decline.
Meta attributed part of the drop to spring layoffs (roughly 10% of staff), but the bigger driver is accelerating in-house replacement.
Meta launched two alternatives: Muse Code (built on its own models, now in external testing, with over 6,000 internal users) and MetaCode (internal only, over 30,000 users).
In plain terms = Meta's two homegrown tools combined now match Claude Code's former user base.
Users halved — so why is the bill still huge?
Internal data show Meta spent over $105 million on Claude Code in the most recent 28-day period.
This reflects a key dynamic: even with half the users gone, the remaining ones are extremely heavy consumers — power coders generate far more API calls than casual users.
A related signal: Meta used Claude while building its consumer-facing Muse AI assistant, but the shipped product has already switched to Meta's own models.
Who else is pulling back?
Palantir, Nvidia, and Booz Allen Hamilton — large clients handling sensitive enterprise data — have demanded stronger privacy guarantees, and some have reduced or stopped using Anthropic's most expensive models.
Meanwhile, OpenAI and Google keep competing on price, and some customers are shifting to open-source or rival products.
OpenAI's annualized revenue surged roughly 70% over recent months to nearly $70 billion; part of that gain is believed to come from Anthropic defectors.
What do Anthropic's own numbers show?
In the 12 months through June, about 6,000 customers spent over $100,000 annually — four times the full-year 2025 figure.
More than 1,000 enterprises spent over $1 million a year; over 100 spent more than $10 million.
Annualized revenue hit $65 billion by July, up more than sevenfold from the start of the year.
This means → Anthropic's overall pie is expanding fast, but its biggest clients are actively reducing dependence — the "quality" of that growth is what IPO investors will probe hardest.
What does this mean for Anthropic's IPO?
Anthropic is preparing what could be the largest IPO in history, which will require disclosing its revenue composition in greater detail.
In plain terms = total revenue is climbing, but the biggest customers are walking toward the exit — investors will ask: how much of this growth is repeatable?
This reflects a structural trend across AI: big tech companies adopt external models for a fast start, then gradually replace them with in-house tools, ultimately internalizing AI as proprietary infrastructure.
市场有风险，内容仅供研究参考，不构成投资建议。