IBM CEO Arvind Krishna told investors that open-weight models now provide about 90 percent of frontier-model effectiveness at 10 to 25 percent of the token cost. He said enterprises can use them for tasks such as email summaries, contract analysis and customer responses. IBM is seeing increased interest in on-premises AI infrastructure and plans software to simplify deployments.

Why it matters

Lower model costs could make private, on-premises AI more practical for enterprises. Teams deploying these systems will still need expertise in GPU clustering, networking, caching and kernel engineering.

Key points

  • Open-weight models deliver about 90% of frontier-model effectiveness, according to IBM’s CEO.
  • IBM estimates open-weight models cost 10% to 25% as much per token as frontier models.
  • Useful enterprise AI infrastructure may fit in two or three racks, IBM says.