Anthropic has a server problem, and it is a problem the company helped create.
The AI lab behind Claude posted $19 billion in annualized revenue in March 2026, up from $9 billion at the end of 2025 and just $1 billion 15 months earlier, The Information reported. That trajectory — roughly doubling every two months — is the kind of number most startups would celebrate. At Anthropic it is also the reason peak-hour usage limits tightened on March 26, affecting roughly 7 percent of users, particularly Pro tier subscribers burning through five-hour session allotments before noon Pacific, The Register noted.
"Overall weekly limits stay the same," Thariq Shihipar, a member of Anthropic's technical team, wrote in a post addressing the change. "Just how they are distributed across the week is changing. I know this was frustrating. We are continuing to invest in scaling efficiently."
The ironies here stack. Anthropic was built in part on the premise that frontier AI should be deployed carefully, with guardrails and usage policies designed to slow adoption when systems are not ready. Running into capacity constraints — the basic infrastructure problem of having more users than GPUs to serve them — is not the kind of friction the company's founders had in mind when they wrote about careful scaling. And it is happening precisely as Anthropic is winning the kind of usage that usually signals a company has crossed from interesting to inevitable.
Claude Code, the terminal-based coding assistant Anthropic launched in February 2025, reached $2.5 billion in annualized recurring revenue in twelve months, European Business Magazine reported, citing no primary source for the figure but noting the pace. In the same period, Anthropic's paid subscriptions more than doubled year-over-year, with the majority of new subscribers signing up at the $20 per month Pro tier, TechCrunch reported, citing consumer panel data from Indagari covering roughly 28 million U.S. users. Claude topped U.S. App Store download charts, web traffic rose more than 30 percent month-over-month, and the company says it now serves 18.9 million professional users, mlq.ai reported.
The growth is partly exogenous. OpenAI's announcement of a Pentagon partnership in February triggered a boycott that saw ChatGPT uninstalls surge 295 percent in a single day, with 2.5 million users joining a QuitGPT movement, mlq.ai reported. That is not a market dynamic Anthropic engineered. It is also not one the company can count on continuing, even as it absorbs the users OpenAI alienated.
The more durable pressure is Anthropic's own product traction. Developer signals consistently point to Claude gaining ground in coding tasks against OpenAI, though published market share data for code generation specifically is thin and imprecise. What is clearer is the funding picture: the lab confirmed $14 billion in annualized recurring revenue at the close of its Series G funding round in February 2026, AI Business Weekly reported, and announced a $30 billion raise at a $380 billion post-money valuation led by GIC and Coatue shortly after, European Business Magazine reported.
That capital is already being deployed toward the infrastructure Anthropic needs to stop rationing access. Google is financing more than $5 billion toward a Texas data center leased to Anthropic through Nexus Data Centers, with an initial capacity of 500 megawatts scalable to 7.7 gigawatts, Cointelegraph reported, citing sources familiar with the arrangement. That is not a company uncertain about its growth trajectory.
The usage limits Anthropic introduced on March 26 are a blunt instrument for a company that built its reputation on policy nuance. Peak hours, defined as 5 a.m. to 11 a.m. Pacific, are when Pro tier users are most likely to hit session limits they would not have hit a week earlier. The API, which serves developers rather than consumers, appears to have remained stable throughout. The consumer interfaces — particularly Opus 4.6, Anthropic's most capable model — bore the brunt of repeated outages, mlq.ai reported.
Anthropic ran a temporary countermeasure: a two-week promotion doubling off-peak usage limits from March 13 through March 28, mlq.ai reported. That promotion has now ended. The capacity problem has not.
The question worth asking is what a company that projects positive cash flow by 2027 — OpenAI does not — does with a server crunch that arrives before that milestone. The answer, for now, is rationing. The data center Google is backing will take years to build. The usage limits Anthropic put in place in March are the bridge instrument: not a permanent policy, but not a solved problem either.
There is something clarifying about watching a company built on careful, deliberate deployment discover that demand does not always cooperate with intention. Anthropic has been right about a great many things — about model capability scaling, about the commercial future of AI coding, about the importance of safety policies that resist political pressure. Being right has its costs. A server crunch at $19 billion in annualized revenue is one of them.
