Where will AI workloads go – when (and if) the big frontier language models, driving all the cap-ex now, finally get ‘trained’ (or sufficiently trained) and inference becomes the new AI workhorse? And what will this edge-wards shift mean for the new cloud boom – as described in the panel write-up from PTC in Hawaii last week? Because, as yet, training still drives most cap-ex – to the tune of £600-odd billion in 2026, just from the big five hyperscalers. Brute-force training is power-hungry and highly synchronized, and demands close-knit facilities in power-rich terrains. But the grid is practically bust, and vacancies are practically zero in ‘primary markets’. And so the build-out has migrated to secondary and tertiary markets. 

McKinsey says inference workloads will grow fastest and could account for more than half of AI compute demand by 2030. It doesn’t need the same power or synchronization – it’s more about throughput, latency, distribution. Workload placement will, on paper, be less about big campuses at arm’s length and more about small clusters at close hand. End-devices and local edge nodes will host, too. In the end, workloads will be split across the cloud/edge continuum, according to latency, privacy, and energy requirements. The big-build phase does not necessarily tally as sunk costs; long-term returns depend on future cloud-based use cases – unless the idea that inference is an edge-game is over-egged (per the PTC panel). Seems like no one is quite sure, yet.