Anthropic reports ~80× surge in usage and revenue

Anthropic’s latest growth number is the kind that stops being a brag and starts becoming an operations problem. At its developer conference in San Francisco this week, CEO Dario Amodei said revenue and usage grew 80-fold in the first quarter of 2026 on an annualized basis. The company had planned for something closer to 10-fold. Instead, Claude demand blew past its own infrastructure plans and forced a scramble for more compute. (cnbc.com) Why does “80-fold” matter so much? Because this is not normal software growth. Anthropic is saying both customer usage and revenue accelerated at roughly the same extreme pace, which means the problem is not just hype or free-tier traffic. People are actually paying for a lot more model output. One outlet pegged that pace at a roughly $30 billion annualized revenue run rate by April 2026, which is a snapshot rather than booked yearly revenue — but it shows the scale of the jump. (cnbc.com) What broke first? Compute. Amodei told attendees the growth explains Anthropic’s recent “difficulties with compute,” basically admitting the company underbought infrastructure for the demand it ended up getting. That matters because frontier AI labs are constrained by chips, power, networking, and data-center space all at once. If any one of those lags, product demand turns into outages, queues, or usage caps. (cnbc.com) So what did Anthropic do? It appears to have gone shopping outside the usual cloud giants. Multiple reports say Anthropic signed a $1.8 billion, seven-year computing deal with Akamai to expand capacity for Claude. Akamai had disclosed a major long-term cloud agreement with an unnamed frontier-model customer, and later reporting identified Anthropic as the buyer. The size matters, but the weirder part is the vendor choice — a company best known for content delivery is now landing one of the biggest AI infrastructure contracts around. (forbes.com) Why Akamai? Because AI demand is spilling beyond the classic hyperscalers. Akamai has been trying to turn its distributed cloud business into a serious compute platform, and Anthropic needs capacity wherever it can find reliable GPUs and networking. This is the same pattern showing up across the market — frontier labs are no longer treating Amazon, Microsoft, and Google as the only places to run serious workloads. When growth is this fast, “good enough and available now” can beat “perfect but booked out.” (forbes.com) What was Anthropic launching at the conference? More agent infrastructure. Anthropic has been pushing Managed Agents — a hosted service for long-running AI work — and its broader pitch is that Claude should not just answer prompts but actually carry out multi-step tasks over time. Conference coverage highlighted new features for agents to coordinate with each other and review prior work, which fits Anthropic’s longer arc toward more autonomous coding and workflow systems. (anthropic.com) Why were developers uneasy? Because the products on stage were aimed straight at software work. Semafor’s dispatch from the event captured the mood: plenty of developers came to learn the tools, but also to gauge how quickly those tools could eat into their jobs. That tension is not separate from the growth story — it is the growth story. Claude Code and related agent products seem to be driving a lot of the surge. (semafor.com) What’s the real catch here? Hypergrowth in AI is not just a sales challenge. It is a reliability challenge. When usage jumps faster than expected, companies have to bolt on compute, permissions, evals, telemetry, and safety controls at the same time. Anthropic’s own engineering blog recently described quality issues in Claude Code that traced back to multiple changes landing together — a useful reminder that fast product expansion can outrun mature production discipline. (anthropic.com) The bottom line: Anthropic is no longer just proving people want Claude. It is proving that demand for frontier AI can outrun even the labs building it — and that the next bottleneck may be infrastructure and operational control, not model quality. (cnbc.com)