Anthropic ARR Jumps to $44B as Model Labs Grab the Margin

The interesting number in SemiAnalysis's latest teardown of the AI stack is not the top-line ARR figure, though that one is dramatic enough. Anthropic's annualized revenue reportedly ran from $9B to over $44B in about a year, and the gross margins the company is booking on inference infrastructure have climbed from 38% to over 70% in the same window. Take the specifics as reported by Dylan Patel, Daniel Nishball, and their co-authors, not as audited numbers, but the direction is the point: the piece frames this as the AI labs capturing all the value now, from almost none last year.

The reason margins moved that hard is the plumbing under agentic workloads, not the sticker price. The SemiAnalysis piece walks through Opus 4.7 usage where the headline is $5 per million input tokens and $25 per million output, but the true blended cost the authors compute is $0.99. The trick is that cached input is only $0.50 per million, agentic tasks are heavily input-skewed (SemiAnalysis puts their own Claude Code usage at about a 300:1 input-to-output ratio), and cache hit rates run above 90%. Layer on a Blackwell generation that reportedly produces roughly 30x more tokens per second on frontier workloads than Hopper did a year ago, and the unit economics stop looking like a services business.

The sharper argument in the piece is that Nvidia and TSMC have not repriced to match. The authors call TSMC's restraint a strategic error and expect Nvidia's correction to come through SOCAMM memory modules rather than headline GPU pricing. Meanwhile the intermediaries that did move have already moved: memory prices up roughly 6x in the past year and one-year H100 rental contracts up around 40%.

The honest caveat is that this is one analyst shop's read, part of it seeded from their own internal spend (SemiAnalysis says its Claude bill is running at roughly $10.95 million annualized, about 30% of employee compensation), and Anthropic's ARR and margin figures are not filings you can pull. What the reporting does not settle is whether the margin expansion is a plateau or a moment: if Nvidia does raise SOCAMM prices as the authors expect, or if a rival lab drops pricing to buy share, the picture flips fast. The forward read for anyone selling picks and shovels is that pricing power has, at least right now, moved up the stack to whoever runs the model.