Frontier AI labs look overvalued right now. Andrew Ho’s post nails the core problem: these companies are locked in a punishing treadmill. They must keep pouring ever-larger sums into training the next model just to stay ahead of open-source and second-tier competitors. Revenue grows, but the required training spend grows faster. High gross margins on inference alone don’t save you when the bulk of the capital is still going into the next training run. The market is pricing in rapid, near-term arrival of highly autonomous agents that autonomously found and run companies. That outcome is possible eventually, but the evidence so far points to slower, spikier progress constrained by data, reliability, and the hard problem of economic calculation. Diffusion of even today’s capabilities into real workflows is taking years, not months. Freezing model quality at current levels would still require a decade-plus of product and process invention before the full economic value is realized. Private-market valuations at the $100B–$1T scale assume the opposite: that the labs can stay unprofitable for a long time while the world rearranges itself around them at high speed. Public markets have historically been less forgiving once lockups expire and the cash-burn numbers become continuous reporting obligations. AI will be a massive industry. The labs themselves, at current prices, look like they are pricing in the most optimistic path rather than the median one.