Anthropic is profitable — if you ignore the expensive parts

Anthropic told investors it booked roughly $559 million in operating profit on $10.9 billion of Q2 2026 revenue, up from $4.8 billion in Q1. The catch is the metric: that figure is EBITDA, and as the Wall Street Journal noted, Anthropic is not subject to public-company reporting, so it is unclear what accounting methods it used to book revenue and costs. Strip out the line items an AI lab actually lives and dies on — training runs, compute amortization — and "profitable" becomes a framing choice rather than a fact.

Skeptics point at the timing. The quarter coincided with a discounted compute arrangement reportedly worth about $1.25 billion a month at reduced rates, concentrated in May and June — a one-off tailwind, not an efficiency gain. The company's own revenue disclosures have also wobbled: $14B ARR claimed in February, $19B in March, $30B in April, while CFO Krishna Rao stated under oath in March that revenues were "exceeding $5 billion to date." ARR and recognized revenue are different animals, and the gap between them is where the argument lives.

SemiAnalysis models a Q3 operating profit above $1 billion at roughly a 6% margin — real, but thin. Anthropic confidentially filed for an IPO on June 1, which is the part that matters: an S-1 forces GAAP numbers and ends the metric-shopping.

Why it matters: the first AI lab to show audited profit sets the benchmark everyone else gets measured against — and right now nobody outside Anthropic has seen the books.