According to Arvind Krishna, inference spending, or the cost of using AI systems once deployed, is expected to reach $23.3 billion in 2026, compared with $19 billion for training. This shift makes AI a daily operating expense for businesses rather than a capital investment concentrated among a few frontier-model providers. Krishna argues that companies should lower the cost per token through open-weight models, which he says deliver about 90% of frontier-model performance at 10% to 25% of the cost, instead of rationing employee access. He also estimates that major cloud providers’ announced data-center commitments total about $8 trillion, with annual interest servicing potentially requiring $800 billion in profits. For companies and investors, the shift moves financial risk toward AI usage budgets and raises questions about how quickly data-center investments can generate returns.
Source: Read the original article
