Their costs do not scale linearly with revenue. Inference is expensive, but it's a variable cost. Anthropic's overall costs include massive fixed costs in training, which are the same regardless of usage.

It's easy to falsify the claim with a simple experiment: imagine they had no customer at all, $0 in revenue. Their costs would still be massive. If the claim were true, $0 revenue should mean $0 costs, right?

If people are sure they can always short NVIDIA