IBM CEO Arvind Krishna warns that rationing AI access is an ineffective strategy as inference costs begin to exceed training expenses. Global spending on inference is projected to reach $23.3 billion in 2026, while training costs are estimated at $19 billion.
Companies currently attempt to control AI budgets by capping tokens or limiting access to powerful models. Krishna advocates for a shift toward open-weight models, which he claims provide 90% of the performance of frontier models at a lower cost, allowing businesses to optimize expenses without restricting usage.
Krishna also highlights concerns regarding the scale of infrastructure investments by cloud providers, which he estimates at $8 trillion. He suggests that servicing the interest on such capital could require $800 billion in annual profits, raising questions about the return on investment for these data centers.
The transition from training to inference shifts financial risk from model developers to every business utilizing AI in production. Krishna remains skeptical of current large language model approaches for achieving artificial general intelligence, estimating the probability of success at between 0% and 1%.
