In 2026, for the first time, more is spent on using models than on training them: according to Gartner inference is worth 55% of AI infrastructure spending, and will rise to 59% in 2027. In the same year, in a Broadcom survey of 1,800 IT leaders, 56% of companies say they run or want to run inference on private cloud.
Training needs huge, concentrated compute centers. Inference does not: it can be distributed, close to the data and to those who use it. It is the ground for a small, private, solar-powered data center.
The cost of achieving the same quality falls every year. What in 2024 required a large cluster, in 2026 fits in a single machine.