The AI industry has spent three years rewarding companies that moved fastest to adopt outside models.
That phase is not over. But a second phase is beginning, and it favors companies that are replacing what they once bought with something they built.
Paying another company to power your AI tools works until it doesn’t. At the scale Microsoft (MSFT) and Meta (META) operate, those bills get large fast.
Building in-house is a bigger investment upfront, but the data stays yours and the tools can be shaped around what your employees actually do. That is a different proposition than a vendor relationship.
Both Meta Platforms and Microsoft are now steering employees away from Anthropic’s Claude, according to The Information. The specifics differ between the two companies, but the underlying logic is the same.
Also read: Mark Zuckerberg sends bold message to Meta investors after Muse
Microsoft slashed its internal Claude budget by a third
Microsoft had projected at least $1 billion in internal spending on Anthropic’s tools earlier this year. That estimate has since been reduced by more than one-third. Company leaders have directed employees to limit their Claude usage and prioritize Microsoft’s own AI products.
That cut applies only to Microsoft’s internal use. Enterprise customers can still access Anthropic models through Microsoft’s platforms, and demand from those customers has reportedly held steady.
Microsoft can reduce what it spends internally while still selling Claude access to businesses that want it. The two decisions are not the same one.
Microsoft has spent the past year pushing its own AI products harder across its enterprise portfolio. GitHub Copilot has reached 50 million users and paid Copilot seats have been growing.
Internal tools that run on the company’s own infrastructure become easier to justify when they are improving fast enough to compete with outside alternatives.
Meta’s Claude Code user count dropped from 60,000 to 30,000
Meta’s pullback is concentrated in Claude Code, Anthropic’s AI coding assistant. Usage inside the company has fallen from roughly 60,000 employees to around 30,000.
The company’s spring layoffs, originally announced by Bloomberg, cut approximately 10% of Meta’s workforce. This round of layoffs account for part of that decline. The primary driver, per the report from The Information, is Meta’s push to develop its own AI tools.
The replacements are already running. MetaCode, an internal coding assistant, has surpassed 30,000 users inside the company.
Meta also began external testing of Muse Code in August. Powered by Meta’s own models, Muse Code had already reached more than 6,000 internal users by the time of the report.
Meta is not simply reducing Claude usage. It is replacing it with tools it owns and controls.
The competitive dynamic between Meta and Anthropic is layered. Meta has also been in discussions about a compute deal with Anthropic, in which Anthropic would pay for access to Meta’s data center infrastructure. Commercial relationships and competitive ones can exist simultaneously in this industry.
Why the economics favor building over licensing at this scale
When a company is staring at a nine-figure internal AI bill, building its own tools starts to look a lot more compelling. You pay more upfront, but you own what you build, and it gets cheaper to run over time.
The strategic reasons are at least as compelling as the financial ones. Microsoft can integrate its own models more directly into GitHub Copilot, Microsoft 365, and Azure.
More AI:
- Nvidia just made a move Wall Street wasn’t ready for
- Microsoft just took sides in AI policy fight
- OpenAI just disclosed something genuinely alarming
Meta can train internal tools on workflows specific to how its engineers build and ship products. A company using its own AI stack generates usage data that helps those tools improve. A company licensing someone else’s model does not get the same feedback loop.
There is also a bigger strategic picture here. Having the best model used to be enough. That is less true now.
The companies consolidating real advantage are the ones owning the cloud compute, the developer tools, the enterprise software layer. The more of that stack you control, the more of the revenue AI creates flows back to you. Both companies are playing that game.
Related: Microsoft CEO names one key concern with its AI
What this pullback does not tell you about Anthropic
The internal reductions at Meta and Microsoft are not a verdict on Anthropic’s commercial position. Enterprise demand for Claude through Microsoft’s platform has held steady, according to The Information.
Businesses often prefer access to multiple AI providers over a single-vendor relationship. A platform that offers both Microsoft’s own models and Anthropic’s is more useful to those customers than one that offers fewer options.
The signal that matters for investors is which customers are sticking. Hyperscalers with the engineering capacity to build their own AI tools will eventually do so.
The enterprises that cannot build for themselves represent more durable demand for a provider like Anthropic. That customer segment may ultimately offer more predictable revenue than relationships with large tech companies actively building internal alternatives.
What the report establishes is that AI partnerships have natural limits. Large technology companies will use outside models where they represent the most efficient option and build their own where it makes more sense.
As internal alternatives at Meta and Microsoft continue to improve, the balance will keep shifting in the same direction. For Anthropic, the commercial opportunity going forward may lie less with companies that can build for themselves and more with the enterprises that cannot.
More on Meta and Microsoft
- Does Meta pay dividends? Its yield and payouts explained
- Microsoft’s logo over the years: A timeline of innovation
- History of Meta: Company timeline, facts & milestones
- Is Microsoft a good long-term investment in 2026?
- Meta’s stock buybacks: How the company’s AI spending could affect shareholder returns
