The cost curve is splitting AI into two markets

The AI capital cycle has split into two tracks, and the cost ceiling is the visible signal of an underlying split. Productivity-AI is running into a ceiling that customers are now publicly routing around. The most credentialed research names in the field are reorganizing around a different bet: the next decade's returns will come from machines that generate knowledge, not from machines that speed up work.

Khosla's tweet, posted the same week Databricks published its warning that AI coding costs are on track to "overtake revenue" if unchecked and Jeff Dean announced his exit from Google to cofound Discovery Loop, names the productive frame in one sentence. These three moves are consistent with a directional bifurcation thesis that the next six to twelve months will test.

Most observers will read these as two unrelated stories: enterprise plumbing and a personnel move. Read together, they describe a single market. As inference costs scale with use, customer routing pressure pushes the productivity track toward commoditization. Every dollar saved on a code completion is a dollar not paid to a frontier lab. The discovery track trades on a different scarcity: the small population of researchers who can build systems that automate the scientific method. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le leaving Google together is the labor market pricing that scarcity in real time.

If productivity-AI cost curves bend the way Stripe, Coinbase, Uber, and Ramp are now bending them, more frontier capital reallocates to discovery bets. The cost ceiling is the visible signal of the split. The discovery reallocation is what it produces.