Former OpenAI researcher Andrew Ho is poking holes in the trillion-dollar valuations being tossed around for frontier AI labs. His back-of-the-envelope math: even if an inference business ran at 80% gross margins, and you gave it a 20x price-to-earnings multiple, a company would still need $100 billion to $200 billion in annual revenue to justify a $1 trillion valuation. And that's before spending a cent on training the next model.
In reality, OpenAI, Anthropic, and their peers have no choice but to keep training. The moment they stop, users might jump ship to cheaper alternatives like Qwen or Kimi. Revenue is growing, but so is the cost of training the next generation of models...
