NVIDIA in the piece «How XPUs Meet a World-Class AI Factory» described an approach to infrastructure for AI factories operating continuously. The focus is on the output: token-generation rate, tokens per watt, token cost, utilization, and uptime.
According to NVIDIA, such an economy of infrastructure should be designed and built as a full factory, not as a set of individual accelerators.
This matters because comparing individual chips does not reveal the full system efficiency. However, the available material is presented as metadata and a short synopsis of the publication, so details of the architecture, list of participants, and practical results remain unknown.
editorial commentary
Why it matters
Probable consequence: AI-infrastructure providers will more often compare systems by output and cost of result, not only by the characteristics of individual accelerators. The next observable signals will be published measurements of utilization, token cost, and uptime. Substantial uncertainty arises from the fact that the available source lacks technical details and independent verification.