Enterprise AI budgets shift from pilots to reserved cloud capacity

NEW YORK — The first year of generative AI in the enterprise was a thousand demos. The second is a line item. CIOs said they are consolidating vendor sprawl into two or three cloud contracts with reserved capacity, because on-demand inference has become a budget surprise.

That shift favors the hyperscalers that can promise both a model catalog and a power-backed cluster. It punishes startups that sold a clever wrapper without a capacity story.

Consultants who made a living standing up pilots are being asked a harder question: which workflows survive a 2026 budget review. The ones that do will look less like magic and more like a unit-cost spreadsheet.