Mark this page's rule: every claim below is attributed and conditional. The future section of a database earns trust by labeling speculation, not avoiding it. See also Future and AGI timelines.
What is roughly knowable to 2027
Hardware roadmaps are public commitments: Rubin and MI400 in 2026, Rubin Ultra's ~600 kW racks targeted 2027, HBM4 ramping, N2/18A silicon underneath. Announced campuses (Hyperion, Stargate sites, Fairwater replication) carry 2026–28 phase dates. The capex disclosures fund through at least 2026. Short of shock, the 2027 stack is largely ordered already.
The 2030 scenario fork
Scenario A — scaling holds (the labs' own base case: Altman, Amodei, Hassabis in their published remarks): reasoning + agents keep converting compute into capability; data-center electricity lands near the IEA base case (~945 TWh); gigawatt campuses become routine; first SMRs connect (~2030 targets); AI-designed chips and automated research begin compounding. In this world today's buildout looks conservative.
Scenario B — returns flatten (the skeptics' case: LeCun on architecture limits, Marcus on reliability, the bubble analysts on economics): pre-training gains stall, agents plateau at partial reliability, revenue " "disappoints the depreciation schedule; capex corrects the way telecom's did in 2001 — leaving, as fiber did, cheap abundant compute for the next cycle. Falsifiable markers: hyperscaler capex guidance cuts, GPU lease rates collapsing, model-release cadence slowing.
Scenario C — efficiency inversion (the DeepSeek lesson generalized): capability per dollar improves so fast that intelligence commoditizes; value migrates from training compute to distribution and integration; power demand undershoots every forecast. Markers: open models within months of frontier, token prices falling faster than volumes rise.
Wildcards with real probability mass
Taiwan (the single point of failure every scenario shares), policy ruptures (export regimes, energy politics), a safety incident that changes deployment law, and architecture surprise (world models, post-transformer designs) — each rewrites the map faster than any trend line.
How this site will keep score
We date every record, label every basis, and keep the timeline current. When 2030 arrives, the right way to have been wrong is legibly — that is what a database owes its readers that a prediction never can.
