Executive Summary¶
The model in one sentence¶
A scenario-based model of frontier AI compute that combines an empirical historical baseline (what single frontier training runs did historically) with a forward supply-capacity projection (how much total AI compute could exist per year through 2040), with downstream allocation, capability, and economic-feedback layers planned next.
What has been built¶
| Component | Purpose | Status |
|---|---|---|
| Historical baseline | Reconstruct historical frontier-model training compute and estimated training cost from Epoch's "Notable AI Models" dataset; fit log-linear curves with rule and cost-variant sensitivity | ✓ complete |
| Supply capacity model | Project annual usable global AI compute capacity 2024–2040 from chips, power, data centers, capex, and utilization, under four scenarios | ✓ complete |
| Allocation layer | Split usable compute into 6 buckets; decompose training pool into largest single frontier run; cross supply × allocation = 16 combined scenarios | ✓ complete |
| Effective compute | Convert raw frontier-run FLOP into algorithmically-adjusted effective compute | ✗ next |
| Capability mapping | Map effective compute to task horizons / benchmark performance / automation levels | ✗ future |
| Projection engine | Probabilistic projections combining all above | ✗ future |
| Economy feedback | Revenue / reinvestment loops back into supply-side capex | ✗ future |
What the model currently projects¶
- Historical training-compute and training-cost growth rates under three frontier-model definitions and full / 2018+ time windows (
outputs/tables/historical_trend_estimates.csv). - Annual usable AI compute capacity by scenario, 2024–2040, with per-year breakdowns of installed H100-equivalent stock, power-limited stock, DC-limited stock, capex-limited stock, and the binding constraint (
outputs/tables/supply_fundamental_inputs_by_year.csv). - Capex required vs available per scenario per year.
- Sensitivity bands on the three highest-leverage supply inputs (shipments, AI-DC capacity, capex).
What the model does not yet project¶
- The size of the largest single frontier training run in any future year. The supply-capacity model gives total annual usable compute; converting that into "the largest training run" requires an allocation layer that is not yet built. This is the single most important gap.
- Algorithmically-adjusted effective compute (after architectural / data-quality / post-training improvements).
- AI capabilities (task horizons, benchmark scores, automation levels).
- AI-economy feedback loops (revenue → reinvestment → more supply).
- Geography splits — currently all global aggregate.
Main findings so far¶
- Historical (Rule A 2018+): frontier training compute grew ~5.97× per year (R²=0.84, n=113), doubling every ~4.7 months. Frontier training cost grew ~4.89× per year (R²=0.72, n=74). Cost per FLOP fell ~24% per year.
- Supply (sourced base case): total usable AI compute grows ~45.7% per year 2024→2040 (CAGR), reaching ~1.65e+31 FLOP/year by 2040. Capex is the binding constraint 2024–2036; chips become binding 2037–2040.
- Allocation (base × base): the largest single frontier training run grows ~27.6% per year 2024→2040, reaching ~6.93e+28 FLOP by 2040. Across all 16 combined supply × allocation scenarios, CAGR ranges from 14.1%/yr (chip-bottleneck × inference-heavy) to 48.1%/yr (capex-rich × training-race) — ~50× spread in absolute 2040 FLOP. Frontier-run share of total compute falls in every scenario from ~3.5% in 2024 to <1% by 2040 in most cases.
- Historical-vs-projection gap (now quantified): the historical 5.97×/yr extrapolation crosses through the entire allocation envelope around 2027–2028 and reaches ~1e+37 FLOP by 2040 vs realistic projections ~1e+28-29. ~7 OOM gap by 2040 in raw FLOP — a real signal that the historical trend was already slowing, allocation parameters may be conservative, and supply fundamentals genuinely cap single-run growth. The effective-compute layer (next) may close some of this by adjusting for algorithmic-efficiency gains.
Main conceptual caution¶
The historical baseline measures one training run. The supply-capacity model measures all global AI compute. They are not the same quantity, and treating them as comparable trends is the most common reading mistake.
The historical 5.97×/yr is the growth rate of the largest single training run released in each window; the supply 45.7%/yr is the growth rate of total annual usable compute across every chip on earth. Frontier runs use a share of total compute, and that share has likely been growing — which is why one trend can outpace the other for a while. Fixing this dependency is exactly what the allocation layer exists to do.
What comes next¶
The effective-compute layer. Convert raw largest_frontier_run_flop_by_year (now produced by the allocation layer) into algorithmically-adjusted effective compute. Epoch's published estimate is ~3×/yr efficiency gain for language-model training; this layer's job is to make capability-relevant FLOP numbers comparable across years. Once effective compute is in hand, the capability-mapping layer can translate it into task horizons and benchmark performance.
For details on each of these points: docs/model_map.md (architecture), docs/model_state.md (build status), docs/output_guide.md (output interpretation), docs/historical_findings.md, docs/supply_findings.md, and docs/allocation_findings.md (component memos).