My prediction for the next six months: AI labs will face more pressure to explain which new workloads justify continued infrastructure expansion.

Better email drafts and meeting summaries are useful. But there is a limit to what a customer will pay to improve a task that was already cheap.

The stronger case would be work with a larger economic payoff: resolving difficult software problems, conducting valuable research, or reliably handling operations that currently require substantial human effort.

Those possibilities also push development in different directions.

A model optimized for research may need deeper reasoning and verification. An operational agent may need reliable tool use, recovery, and low latency. A mass-market assistant may need a much lower cost per interaction.

My guess is that commercial pressure will make those priorities diverge more visibly.…