Future of AI

The near future is unusually knowable, because it is under contract. Public roadmaps put NVIDIA's Rubin and AMD's MI400 in racks in 2026, with ~600 kW Rubin Ultra racks targeted for 2027. HBM4 is sold ahead of production. Announced campuses — Hyperion at up to 5 GW, the Stargate sites, Fairwater's replication — carry phased dates through 2028. Hyperscaler capital guidance funds all of it through at least 2026.

Scaling holds (the labs' case). Altman, Amodei and Hassabis have each publicly argued — in their own registers — that capability keeps converting compute into value: reasoning and agents deepen, data-centre electricity lands near the IEA base case (~945 TWh by 2030), gigawatt campuses become routine, and the first SMR fleets connect around 2030. In this world, today's buildout underestimates.

Returns flatten (the skeptics' case). LeCun argues current architectures cannot reach human-level intelligence; economic critics point to depreciation schedules and circular financing. In this scenario capex corrects the way telecom's did in 2001 — and, like fiber then, leaves behind cheap abundant compute for whatever comes next. Watchable markers: capex guidance cuts, GPU lease rates, release cadence.

Efficiency inversion (the DeepSeek lesson). Capability per dollar improves so fast that intelligence commoditizes; value migrates to distribution and integration; power demand undershoots forecasts. Marker: open models within months of the frontier, token prices falling faster than volumes rise.