How Did Anthropic's Inference Margins Turn Positive?

Joey Brookhart says Anthropic's gross margin on inference went from negative 94% two years ago to positive 85% now. The unit he uses is revenue per megawatt of inference. SemiAnalysis puts Anthropic at about $65 million of revenue per inference megawatt, up from the mid-teens a year ago, and an Opus 5.5-class model on Blackwell could reach $100 million to $120 million. Compute cost is a little over $10 million per megawatt. Token throughput on Amazon Trainium chips is up 5 to 10 times year over year, from the hardware and from how the model runs on it.

Only about 30% of the compute is serving paying inference. The rest is experimentation, research, and training the next model, which is where the returns and the defensibility sit. Nilay Patel expected the S-1 to show current models throwing off margin that would be poured into training, because Anthropic is chasing AGI and the filing says Claude might be conscious. Training cost is instead managed against the inference margin. With gross margins of 85% or better, the lab can raise the share of compute that goes to training. Growth came in faster than planned. An 11% non-GAAP EBIT margin was not the goal. They would rather sit closer to zero, or even below, if that compute went to training. The roadshow has pointed investors at the profitability curve and at how little cash was burned to get here.