IBM CEO Warns Against AI Rationing

This year, companies will spend more on running AI than they ever did training it. Arvind Krishnamurthy says that rationing those costly models is a slow way to lose. He notes most firms now cap tokens, add spending meters, and reserve the best models for a tiny approved group. The real cost driver, he explains, is engineering - AI agents read about six hundred words for every word they write. IBM's open-source caching layer kept ninety-eight percent of data in cache, letting chips do seven times more work. In a test, an open-weight model ran three thousand agents and delivered a fifty-two-fold speedup for just twenty-six dollars. A comparable proprietary model gave a two-hundred-four-fold boost but cost two hundred ninety-nine dollars, over ten times more. He urges firms to own the AI operating layer and control where workloads run. The goal, he says, is maximizing token yield rather than chasing the cheapest token.