Abstract:
As Anthropic's two largest enterprise customers, Meta and Microsoft are currently reducing their employees' use of Claude; this poses a challenge to Anthropic's efforts to maintain its explosive growth. Microsoft executives earlier this year expected spending on Anthropic technology within the company to reach at least $1 billion, according to people familiar with the matter.
One of the people familiar with the matter said that Microsoft has since lowered this expected number by more than a third because the company asked employees to use less Claude to save costs and switch to more use of Microsoft's self-developed AI tools. (This figure does not include the costs incurred by customers using the Anthropic model through Microsoft tools, which has been growing steadily.)
According to a person familiar with the matter, the number of employees using Anthropic programming assistant Claude Code within Meta has dropped significantly, recently to about 30,000. That's down from about 60,000 users earlier this year. Part of the decline in users was due to layoffs in the spring, when Meta cut about 10% of its workforce of 78,000.
But a more important factor is that Meta has developed its own AI programming tools and Muse Spark series of AI models, and actively promotes these products to other companies, while also encouraging internal employees to use them. Internal data shows that Muse Code, a tool built on the Meta model and designed to compete with Claude Code (Meta began testing the tool with external customers in August), has recently exceeded 6,000 employee users; while another similar tool, MetaCode, which is only for internal use, has more than 30,000 users.
Meta expects to spend billions of dollars this year on AI tools used internally, according to a June memo. While the memo didn't specify how the spending was divided among the different companies, Claude Code was undoubtedly the most popular external programming tool at the time. Meta spent more than $105 million on Claude Code in the last 28 days, internal data shows.
Meta also used Claude this year when developing its consumer-oriented Muse AI agent, but switched to its own model in the final product launched last month.
The two tech giants remain Claude's main users, but even a slight reduction in spending by core customers like them would create broader headwinds for Anthropic; after all, soaring costs for its AI tools have forced other enterprise customers like Cognition to look to cheaper alternatives.
A number of large companies with sensitive enterprise business, such as Palantir Technologies, Nvidia and Booz Allen Hamilton, have requested new safeguards or reduced or even stopped using Anthropic's most advanced (and expensive) models due to concerns about data privacy.
These challenges come as Anthropic prepares for what could be its largest initial public offering (IPO) ever, a process that requires the company to disclose more information about its business to potential investors.
These challenges come as Anthropic prepares for what could be its largest initial public offering (IPO) ever and requires it to disclose more information about its business to potential investors.
Nearly a quarter of Anthropic's revenue last year came from just two customers, according to data from Anthropic's draft IPO prospectus. Anthropic's revenue has grown significantly in the first half of this year: As of July, its annualized revenue was reportedly over $65 billion, up more than sevenfold from the start of the year, and the company has added a host of new customers during that time.
Anthropic told investors that about 6,000 customers spent $100,000 or more on its AI software in the 12 months ended in June, four times the number of such customers in all of 2025. The company also told investors that more than 1,000 businesses spent more than $1 million in the 12 months to June, and more than 100 businesses spent more than $10 million.
Meta and Microsoft are both competitors and customers of Anthropic. Meta CEO Mark Zuckerberg last week announced the establishment of a new division called Meta Enterprise Platform to focus on providing artificial intelligence products to enterprise customers. Meanwhile, Microsoft executives have been caught off guard by Anthropic's meteoric revenue growth and its raft of new software products for businesses over the past year.
This spring, the use of Claude within Meta surged. Employees have tried "tokenmaxxing" (maximizing the use of token quotas) to demonstrate their artificial intelligence technology capabilities and compete for ranking on a leaderboard called "Claudeonomics" created by employees voluntarily. This phenomenon prompted company executives to pledge in June to implement stricter controls on the use of artificial intelligence, including setting budgets, limiting employees' token consumption, and establishing a central dashboard to monitor excessive use of Claude Code and other external artificial intelligence tools.
Earlier this year, usage of Claude by Microsoft employees also surged. Annual spending by Microsoft employees on Claude-related models, including Claude Code, the Claude model integrated into Microsoft Copilot, and the Claude Mythos model for network security scanning, was expected to exceed $1 billion as of May, according to two people familiar with the matter.
Rising costs for engineers using Claude, coupled with pressure on Microsoft to bolster its own Claude Code competition, prompted some executives to direct employees to reduce their reliance on Anthropic's products in the coming months, according to these people.
Company executives cut subscriptions to Claude Code and instead required employees to use Claude models through Microsoft's GitHub Copilot tool. The tool allows users to choose which AI model to use, but also includes an "automatic" mode where queries are directed to OpenAI's model by default in most cases. This is a much lower operating cost for Microsoft because it doesn’t have to pay to use OpenAI models thanks to its commercial partnership with the AI developer.
Senior leaders, including Scott Guthrie, who oversees Microsoft's Azure cloud business, and Jay Parikh, who oversees its developer tools division, have asked employees to be more cautious when writing code and urged supervisors at all levels to monitor employee usage spending while prioritizing the use of models from non-Anthropic companies, according to people familiar with the matter. In addition, according to a current Microsoft employee, company executives can now view data on how much each employee spends on AI on a daily basis and the specific models they are using through GitHub Copilot.
"Internally, moving more workloads to OpenAI models will help us get greater value from our token investments," Parikh wrote in a memo to employees in August, which was reviewed by The Information. "This fiscal year, each department needs to manage its own token budget, so we all need to continue to learn and optimize."
Recently, Microsoft has lowered the overall quota limit for employees to use artificial intelligence (AI). According to people familiar with the matter, taking the cloud and artificial intelligence department with more than 60,000 employees as an example, each employee previously had a monthly budget of up to $100,000 for AI models; the department has recently lowered this budget to about $10,000 per month in most cases.
The cuts sparked dissatisfaction among some internal engineers, who complained that they could no longer use the models as freely as they had in the past, according to current employees.
While Microsoft has cut its internal spending on Anthropic, these figures do not include the much higher amount Microsoft pays Anthropic to provide its customers with features based on Claude's technology. Microsoft had planned to spend at least $2 billion a year using Anthropic's model to power the Copilot AI capabilities used by its productivity software customers. At the same time, Microsoft also replaced some of its own models with Claude in Copilot; according to people familiar with the matter, starting in July, Microsoft began to direct complex queries in Copilot to OpenAI's GPT-5.6 model by default.
However, the amount Microsoft pays Anthropic to provide these features to customers has been rising steadily this year as the number of Copilot subscribers grows, according to a person familiar with the matter. That increase has largely offset Claude-related cost reductions within Microsoft, leaving the total amount Microsoft pays Anthropic roughly flat compared with peak levels earlier this year, the person said.
A Microsoft spokesperson confirmed that the company has recently set OpenAI's model as the default model for internal AI programming, but said that engineers can still choose to use the Claude model. The spokesperson pointed out that Microsoft has been guiding employees to use its own GitHub Copilot AI programming tool instead of external tools such as Claude Code, because using its own tools helps the company improve its products.
A Microsoft spokesperson said in a statement: "Microsoft departments need to carefully and efficiently manage the use of AI based on their own business priorities and in compliance with standard engineering and operational specifications."
Comments