Anthropic ends the rationing era

FOR most of 2026, Anthropic's loyal users have been paying for less Claude than they wanted. Pro and Max subscribers ($20 to $200 a month) have burned through five-hour session windows in under an hour, then waited days for the weekly counter to reset. Thariq Shihipar, an engineer on Anthropic's technical staff, publicly confirmed in March that peak-hour throttling was deliberate. The signature product of the most-talked-about AI company of 2026 has been working at half-pace, by design.

On April 20th Anthropic announced $5bn in fresh capital from Amazon, with up to $20bn more contingent on commercial milestones, in exchange for $100bn of AWS spending over a decade and as much as 5 gigawatts of Trainium capacity. Four days later Google committed $10bn in cash and a possible $30bn on top, plus 5 more gigawatts of TPU compute starting in 2027. Combined with November's $5bn from Microsoft and its accompanying $30bn Azure pact, Anthropic now sits on roughly $65bn in pledged equity capital and 10 gigawatts of reserved AI training power. The financial press has read the news as a cap-table story.

The demand picture is what makes the gating involuntary. Anthropic's annualised revenue ran from roughly $19bn in March to roughly $30bn by April, propelled by a wave of consumer subscribers after a Pentagon-related ChatGPT backlash in February drove millions of users toward Claude. App-store downloads briefly surpassed ChatGPT's for the first time, and web traffic jumped more than 30% month-on-month. The product's success was not the problem; the product's success multiplied the problem, because every new Pro subscriber added another source of continuous, high-volume token consumption against a fixed silicon supply.