The Biggest Anthropic Customers Just Started Shopping Elsewhere — Inside Their Own Buildings
October 6, 2026 · 5 min read
Some of Anthropic's biggest buyers just told their own employees to use someone else's tools — or rather, their own. According to The Information, Meta's internal Claude Code user base has fallen from about 60,000 at the start of the year to roughly 30,000, while the company's in-house MetaCode has crossed the 30,000 mark and is still climbing. Meanwhile, Microsoft has cut its internal budget for Anthropic technology — at least $1 billion — by more than a third. (via The Information, via AIStockWire)
Two of the industry's largest wallets, shrinking at the same time, for the same reason: the people who know AI infrastructure best are the first ones replacing it with their own. This is not a quality complaint. Nobody is saying Claude Code got worse. This is about big tech deciding that coding assistants are too strategic to rent — and that their own engineers are the cheapest beta testers they will ever get.
The in-house swap, explained
Meta's numbers tell the cleanest story. Claude Code went from 60,000 internal users to 30,000; MetaCode crossed 30,000, overtaking the tool it replaced. That is not drift — it is a managed migration, and the driver is obvious once you think like a CFO. Every internal Claude Code seat is money leaving the building, per user, per month, forever. Every MetaCode seat is sunk cost turning into scale advantage. Your own engineers dogfood the product, file bug reports for free, and the tool improves on data that never leaves your walls. For a company with tens of thousands of engineers, building your own coding assistant is no longer a vanity project; it is the cheapest procurement strategy on the menu.
Two sets of books
Here is the twist both companies push hard: internal cuts are not customer cuts. Meta and Microsoft stress that what enterprise clients spend on Anthropic models through their platforms is still going up. So we get the oddest of all charts — Anthropic's revenue rising while its two biggest landlords quietly move their own staff off its tools. Read it as two separate markets. The internal market runs on cost and control: your own tool wins by default. The external market runs on trust and inertia: enterprise customers are still buying, still renewing, still building on Anthropic's models. One ledger shrinking and the other growing can both be true. The question is which ledger sets the trend.
The moat question for Claude Code
This is the uncomfortable part for Anthropic. Claude Code's reputation was built inside the world's best engineering orgs — the teams that tried everything and kept coming back. If those same teams leave anyway, what was the moat? A product lead, or just a head start? Coding assistants may be the layer that commoditizes fastest: the model can be swapped, the IDE integration rebuilt, and the workflow data that makes the tool smarter stays inside the company that generates it. Anthropic's best hope is the second ledger: enterprises that can't build their own will keep renting. But "our customers stay because they can't build it themselves" is a fragile loyalty — it lasts until building it gets cheap enough.
What it means for everyone else
Watch your own stack the way a landlord watches rent. If you are a team of twenty and Claude Code saves you hours a week, these headlines change nothing — you are not in the market where this fight happens. But the direction is clear: the coding-assistant layer is becoming infrastructure that big companies own rather than rent. The safe bet is staying portable — standard configs, no deep lock-in to any one assistant — because the next migration wave will not ask permission. And for the model labs watching: your loudest champions inside big tech are also your future competitors. Price accordingly.