Skip to content

Input Provenance

Every input the model reads, where it came from, and how confident we are in it.

Confidence rubric

  • high — directly sourced from a named public source (e.g. NVIDIA H100 datasheet, IEA published TWh figure, the Epoch CSV itself).
  • medium — derived or synthesized from named public sources (e.g. global H100-eq shipments inferred from NVIDIA's reported datacenter revenue + AMD MI300 production guidance + TPU pod estimates).
  • low — round-number placeholder or long-horizon extrapolation. Most remaining low rows are 2040 values that extrapolate from the 2030 anchor.

The full source/confidence metadata lives in the assumptions YAML itself; see data/assumptions/supply_input_assumptions.yaml for per-row provenance.

Inputs by component

Input Component File Source Source type Confidence Used for Notes
Epoch "Notable AI Models" dataset Historical baseline data/raw/epoch_notable_ai_models_raw.csv Epoch AI public dataset high historical training compute, cost, parameters, hardware, frontier flag raw immutable source; retrieved 2026-05-02
Epoch "Frontier" subset Historical baseline (cross-check) data/raw/epoch_frontier_ai_models_raw.csv Epoch AI public dataset high sanity-check our own frontier rules against Epoch's curated subset not used directly in fits
Epoch "Large-Scale" subset Historical baseline (cross-check) data/raw/epoch_large_scale_ai_models_raw.csv Epoch AI public dataset high reference only not used directly in fits
H100/H200/B200 shipments + AMD/TPU/Trainium Supply capacity supply_input_assumptions.yaml h100_equivalent_shipments Epoch "Can AI scaling continue through 2030?" (Aug 2024) + author synthesis public estimate + synthesis medium accelerator stock projection, chip-limited stock NVIDIA reported H100/H200 ~1.5–2M for 2024 (Epoch); AMD MI300, TPU, Trainium, China-domestic added by synthesis
Accelerator lifetime (years) Supply capacity supply_input_assumptions.yaml accelerator_lifetime_years hyperscaler 10-K depreciation schedules + author synthesis public financial disclosures + synthesis medium retirement curve in stock projection base 5y; capex-rich 4y; chip/power-DC bottleneck 6y (forced longer holds)
H100 peak FLOP/s Supply capacity supply_input_assumptions.yaml peak_flops_per_h100e NVIDIA H100 datasheet public spec high theoretical compute calculation 989 TFLOP/s FP16 dense; constant by H100-eq definition
H100 power draw (kW) Supply capacity supply_input_assumptions.yaml power_kw_per_h100e NVIDIA H100 datasheet (2024); Epoch 24× efficiency improvement (2030) public spec + public estimate high (2024) / medium (2030) per-chip effective power for the power constraint H100 SXM 700W chip TDP; 0.25 kW by 2030 implied by Epoch's 24× perf-per-watt improvement
Server / cluster power overhead Supply capacity supply_input_assumptions.yaml server_power_overhead industry-typical conservative industry figure medium converts chip TDP → server-level power 1.5× multiplier (CPUs, NICs, storage, power-delivery losses)
PUE Supply capacity supply_input_assumptions.yaml pue industry-typical published hyperscale-DC ranges high (2024) / medium (2030+) converts server power → datacenter total Modern hyperscale AI DC ~1.15–1.25 (2024), trends toward 1.10 with liquid cooling
AI data-center capacity (MW, global) Supply capacity supply_input_assumptions.yaml ai_datacenter_capacity_mw IEA "Energy and AI" report (April 2025) + Epoch + author synthesis public report + public estimate + synthesis medium power and DC constraints IEA: 415 TWh global DC 2024 → 945 TWh 2030; AI share derived; 12 GW (2024) → 80 GW (2030) base case
AI share of DC power Supply capacity supply_input_assumptions.yaml ai_share_of_dc_power industry-typical conservative figure medium fraction of AI-dedicated DC power that's actually AI workloads 0.85 base; 0.80 power-DC bottleneck; 0.90 capex-rich
DC packing efficiency Supply capacity supply_input_assumptions.yaml dc_packing_efficiency author synthesis synthesis medium (2024) / low (2040) DC slot/cooling/transformer slack — separate from raw grid power 1.0 most scenarios; falls to 0.65 by 2040 in power_datacenter_bottleneck
Cluster utilization (MFU) Supply capacity supply_input_assumptions.yaml cluster_utilization industry-typical published MFU benchmarks (Llama-3, GPT-4 estimates) medium converts theoretical → usable compute 0.40 (2024) → 0.55 (2040) base; transformer training MFU ~35–50%
Accelerator unit cost (USD per H100-eq) Supply capacity supply_input_assumptions.yaml accelerator_unit_cost_usd NVIDIA Investor Relations (H100 ASP) + author synthesis public IR + synthesis high (2024) / medium (2030+) capex constraint $30K (2024) → $15K (2030); H100 list ~$25–40K
Cluster capex multiplier Supply capacity supply_input_assumptions.yaml cluster_capex_multiplier author synthesis from Epoch's "power infra ~40% of GPU cost by 2030" public estimate + synthesis medium converts chip cost → installed-cluster cost (servers + networking + DC + power infra) 2.2× (2024) → 2.5× (2030); scenario-keyed
Hyperscaler AI infrastructure capex Supply capacity supply_input_assumptions.yaml ai_infrastructure_capex_usd Microsoft/Alphabet/Meta/Amazon 10-K filings + Stargate announcement + author synthesis public financial disclosures + synthesis medium capex constraint $210B (2024) → $1.5T (2030) base; AI share of total hyperscaler capex assumed ~75%
Cloud rental rate per H100-eq per year Supply capacity supply_input_assumptions.yaml cloud_rental_usd_per_h100e_year industry-typical (~$2/hr × 8760 × 80% util) + author synthesis published cloud-rental rates + synthesis medium cost-per-H100e-year (cloud variant), preserves Phase 1 cost-variant insight $15K/yr (2024) → $5K/yr (2040) base

Source-quality breakdown

Per the YAML's per-row confidence flags, sourced supply assumptions are distributed roughly:

Confidence Approximate row count Notes
high ~25% NVIDIA H100 spec, IEA TWh figures, NVIDIA IR ASP
medium ~55% Synthesis from cited public sources
low ~20% Mostly 2040 long-horizon extrapolations

The single highest-leverage medium-flagged input is 2030 H100-equivalent shipments (and therefore stock). Epoch's published range is 20M–400M H100-eq for "training stock by 2030" — a 20× spread. The base case anchors to Epoch's median of 100M; the chip_bottleneck scenario anchors to the 20M floor. This single input dominates the supply-capacity output ranges.

Per-component external dependencies

  • Historical baseline: depends only on Epoch's "Notable AI Models" CSV. Refresh by re-downloading the CSV (it's updated periodically) and re-running uv run historical.
  • Supply capacity model: depends on the assumptions YAML; upstream sources cited in the YAML's per-row source field are not directly read. To incorporate updated Epoch AI scaling estimates, IEA reports, or hyperscaler 10-K disclosures, edit the YAML and re-run uv run supply.

Future component inputs (not yet sourced)

When the allocation layer is built, it will need:

  • Training-vs-inference split. Currently estimated at ~35/55/10 (training/inference/reserves) for 2024 — to be sourced from public hyperscaler / lab disclosures and academic estimates.
  • Largest-run concentration. Fraction of total training compute that goes to the single largest run. Currently unmodeled. Probably 5–15% historically, with wide bands.
  • AI R&D experiment share. ~10–30% of frontier-lab compute budgets historically — to be refined when allocation lands.