Epoch AI

“Aggregate cash capital expenditure across the largest hyperscalers (Microsoft, Amazon, Alphabet, Meta and Oracle) is on track to overtake their operating cash flow around Q3 2026 — capex is growing roughly 70% per year against about 23% annual growth in operating cash flow, pushing most hyperscalers toward external financing.”

Q3 2026 results are not reported until late October 2026, so the crossover cannot yet be confirmed; corroborating evidence is mounting (FactSet reported in July 2026 that hyperscaler free cash flow was under mounting pressure, with incremental annual debt rising from 9% of capex in FY24 to 32% by mid-2026).

“Overall, we conclude that power demand for frontier training will likely grow by 2.2x to 2.9x per year in the coming years, implying that the largest training runs will reach 4-16 GW by 2030.”

Largest known training runs were in the low hundreds of MW as of 2025; the 4-16 GW range resolves at the end of 2030.

“GPU energy efficiency has been improving by about 40% annually and is likely to continue at a similar pace — partially offsetting continued growth in training compute, which we expect to persist at 4-5x per year [through 2030].”

Resolves against measured FLOP/watt trends for frontier AI accelerators through 2030.

“In our median projection, we expect enough manufacturing capacity to produce 100 million H100-equivalent GPUs for AI training by 2030, sufficient to power a 9e29 FLOP training run — even after accounting for GPUs being split between multiple AI labs and in part dedicated to serving models.”

Epoch's uncertainty range spans 20M-400M H100-equivalents (1e29-5e30 FLOP); resolves against actual AI accelerator production and single-run scale in 2030.

“The model estimates a 10% chance of transformative AI — defined as 'AI that if deployed widely, would precipitate a change comparable to the industrial revolution' — being developed by 2025, and a 50% chance of it being developed by 2033.”

The model's 10%-by-2025 leg did not come in — no industrial-revolution-scale transformation occurred by 2025 — but the headline 50%-by-2033 estimate resolves at the end of 2033. TIME notes Epoch researchers themselves expect slower progress than the model implies.