Joey Brookhart of SemiAnalysis calls Anthropic's S-1 filing awesome, citing more than $70bn in annual recurring revenue, strong margins and free cash flow.
Brookhart says Anthropic's inference gross margin swung from negative 94% to positive 85% in two years, driven by compute efficiency gains on Trainium chips.
Nilay Patel doubts Anthropic's trillion-dollar case given open-weight models and rivals like Google DeepMind, Meta and SpaceX, while Brookhart says coding and vertical uses will drive growth.
Anthropic's IPO filing shows frontier AI can actually make money
Anthropic's S-1 shows a frontier AI lab can already run a real business, with $70bn in recurring revenue and inference margins that efficiency gains keep improving. The open question is whether adoption beyond coding and heavy early spenders grows fast enough to justify a $1.5 to $2tn valuation and trillions in compute commitments.
Takeaways
Anthropic's filing shows a genuinely profitable lab. Brookhart finds little to nitpick. The S-1 shows more than $70bn of annual recurring revenue, strong gross margins, free cash flow and an 11% non-GAAP EBIT margin, against negative 100% that The Wall Street Journal reported for OpenAI.
Inference margins swung from -94% to 85%. Compute efficiency drives the turnaround over two years, with token throughput on Trainium chips rising five to ten times a year. Only about 30% of compute serves inference, and Anthropic would move more into training if margins hold.
The $80bn loss is mostly non-cash. It reflects Anthropic's rising valuation after issuing warrants and convertible notes. Actual cash burn is about $10bn against more than $120bn of cash and securities, so Brookhart is not worried about dilution for equity investors.
Spending is concentrated and the Fortune 500 lags. The top 1% of customers spend about $100,000 per employee and top Claude Code users about $18,000 a year. Anthropic counts half the Fortune 500 as customers, yet its spending is 0.1% of customers' operating expenses.
Growth dipped when Meta cut back. Meta was probably over 10% of Q3 revenue before halving its Claude Code seats from 60,000 to 30,000. Monthly net new ARR fell from $10bn in Q2 to $4bn in Q3, with a recovery expected.
Hyperscalers sell half and supply the compute. About 48% of revenue flows through hyperscaler marketplaces such as Amazon Bedrock, and AWS's strong quarter reflects Anthropic's growth. Headline compute commitments reach $5.6tn over seven to ten years, which Brookhart expects revenue to cover.
Patel's doubts versus Brookhart's case
Patel argues open-weight models and rivals such as Google DeepMind and Meta threaten the trillion-dollar case. Brookhart answers that enterprises fine-tune open models and competition is intensifying, but coding and verticals like cyber and biology will drive growth.
Patel asks why Anthropic keeps chasing AGI instead of selling existing models into more industries. Brookhart says frontier models win in adversarial work such as cybersecurity and law, and Anthropic is verticalizing with Claude for Financial Services and Claude for Legal in the meantime.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Neoclouds and the OpenAI comparison
Elon Musk predicts an extinction event for neoclouds. Brookhart expects CoreWeave and Nebius to survive with backing from hyperscalers or chipmakers such as Nvidia, while former crypto miners face a tough financing market.
OpenAI has reportedly delayed its IPO, and Brookhart expects its consumer-heavy financials to look weaker when it files. He expects Anthropic to manage expectations so it beats estimates as a public company.