ai-compute-users

Estimates of the compute quantities used by major AI players: Monte Carlo models of frontier labs' compute (Google DeepMind, Meta Superintelligence Labs, OpenAI, Anthropic, SpaceXAI) in H100-equivalents (H100e), plus Alphabet-level activity estimates.

Split out of epoch-research/ai-chip-counts (its ai-lab-compute/ directory, history preserved).

Research notebooks vs. the estimate script

The repo has two layers that share one set of priors:

• Research notebooks (notebooks/) — the canonical per-lab walkthroughs. Each notebook develops one model in full: the reasoning behind every input, sources, intermediate charts, and sensitivity sweeps. They are jupytext-paired (.ipynb ↔ .py): edit the .py, then run notebooks/sync_notebooks.sh --execute (or python3.11 -m jupytext --sync --execute notebooks/<name>.py) so outputs land in the .ipynb. Commit both files.
• The estimate script (frontier_lab_compute_model.py, repo root) — the consolidated, importable versions of the same models: end-2025 estimates for the five labs, the end-2024 backcasts for Google DeepMind, Meta, and Anthropic, and the SpaceXAI model (end-2024, end-2025, and mid-2026, anchored on Epoch's Colossus data-center capacity estimates). It restates no judgment priors — like the notebooks, it loads them from lab_model_params.csv (via lab_compute_utils.load_lab_params()), the single source of truth for the model priors. Change a prior in the sheet and both layers pick it up.

The notebooks:

Notebook	Estimates
deepmind_compute_model	Google DeepMind compute, end-2025
msl_compute_model	Meta Superintelligence Labs compute, end-2025
openai_power_model	OpenAI compute at year-ends 2023–2025 (power-based, mainline)
anthropic_power_2025	Anthropic compute, end-2025 (power-based, mainline)
anthropic_cloud_spend_2024	Anthropic compute, end-2024 (canonical) + end-2025 cloud-spend cross-check
openai_cloud_spend	OpenAI compute 2024–2025 via cloud spend (cross-check)
alphabet_level_activities_model	Alphabet-level activity estimates
lab_2024_backcasts	End-2024 backcasts across the four labs
spacexai_compute_model	SpaceXAI compute, end-2024 / end-2025 / mid-2026 (Colossus-anchored)

Supporting modules (repo root): lab_compute_utils.py (prior loader, fleet buildout helper), epoch_data.py (chip fleet data — Nvidia per-owner fleets, TPU and AMD cumulative sales — plus data-center capacity timelines, fetched at runtime from the Epoch AI data hub, cached one download per day under .cache/), and data/ (hand-maintained inputs; see data/README.md).

Table exports

generate_lab_compute_tables.py (repo root) turns the estimate script's Monte Carlos into export tables: point-in-time total compute per lab in H100e, with P5 / median / P95 uncertainty. It holds no model structure of its own — it runs the models from frontier_lab_compute_model.py and shapes their sample arrays.

Usage

from generate_lab_compute_tables import get_all_tables

tables = get_all_tables()          # dict of two DataFrames
tables["year_end_by_lab"]          # (end-of-year × lab) headline estimates
tables["intermediates_by_lab"]     # the quantities behind each estimate

Or the individual getters:

from generate_lab_compute_tables import get_year_end_by_lab, get_intermediates_by_lab

df = get_year_end_by_lab()

From the command line (writes both data/lab_compute_*.csv files):

python3.11 generate_lab_compute_tables.py

Results are deterministic: each lab model reseeds 42 internally so it reproduces its canonical notebook run — hence no seed parameter.

Schema — year_end_by_lab

Column	Type	Notes
Name	str	"{Lab} end-{Year}", e.g. "OpenAI end-2025"
Lab	str	Google DeepMind / Meta Superintelligence Labs / OpenAI / Anthropic
Year	int	Calendar year of the snapshot
Date	str	YYYY-12-31 — the point in time the estimate refers to
h100e_p5 / h100e_med / h100e_p95	float	Percentiles of total H100e
Notes	str	Generation timestamp

H100e converts each chip at its dense 8-bit peak FLOP/s divided by the H100's 1979 TFLOP/s. Estimates cover compute rented or used by each lab (not owned), at the stated moment in time — they are operational-stock snapshots, not flows, so consecutive years must not be summed.

The 2024 rows for Google DeepMind, Meta, and Anthropic are backcasts. They keep the same Lab labels for continuity, but "the lab" in 2024 means frontier-AI compute at the company: Meta Superintelligence Labs did not exist in 2024 (its predecessor was Meta AI / GenAI plus FAIR), and the backcast share priors are for those predecessor scopes — see notebooks/lab_2024_backcasts.ipynb. Anthropic's 2024 row converts reported cloud spend at 2024 prices (notebooks/anthropic_cloud_spend_2024.ipynb).

Schema — intermediates_by_lab

How each lab's final distribution is computed: one row per intermediate quantity (owned fleets, deployment ratios, shares, power, chip counts, ...), in model order. The traces come from MODEL_STEPS in frontier_lab_compute_model.py — each model records its steps as pure bookkeeping, so adding a step(...) entry there is all it takes to extend this table (and the walkthrough page below).

Column	Type	Notes
Name	str	"{Lab} end-{Year} · {Label}"
Lab / Year	str / int	Same conventions as year_end_by_lab; one trace per modelled (lab, year) snapshot
Step	int	1-based position in the model's computation
Variable	str	Machine name; sheet-prior steps match their row name in lab_model_params.csv
Label	str	Human-readable name
Kind	str	input (sampled prior) / constant (fixed scalar) / derived / final
Units	str	H100e, MW, share, ratio, quarters, chips, USD B/yr, ...
Expression	str	For derived/final rows: how the step combines earlier ones, by Variable name
value_p5 / value_med / value_p95	float	Percentiles in the row's own units; constants repeat the same value
Notes	str	Generation timestamp

Each snapshot's final row equals its year_end_by_lab row (tested).

Coverage

Lab	Snapshots
OpenAI	2023, 2024, 2025
Anthropic	2024, 2025
Google DeepMind	2024, 2025
Meta Superintelligence Labs	2024, 2025
SpaceXAI	2024, 2025, mid-2026

OpenAI's power model yields a snapshot per disclosed year-end; DeepMind, Meta, and Anthropic add end-2024 backcast models (Anthropic's converts reported cloud spend at 2024 prices; an experimental power-model backcast stays notebook-only in notebooks/anthropic_cloud_spend_2024.ipynb). Rows are omitted (not zero-padded) where no model exists — treat a missing (lab, year) as "no estimate", not zero. SpaceXAI's 2026 row is dated June 30 (not Dec 31) and is net of the Colossus capacity SpaceX sells to Anthropic, Google, and Reflection AI — its 2024/2025 rows have no such subtraction because the sale agreements all start May–July 2026.

Caveats for downstream use

• Don't derive cross-row totals or ratios from the percentile columns. The lab models share one RNG stream (each reseeds 42), so per-sample draws are artificially aligned across labs and years: per-row percentiles are valid, but a stacked "all labs" bar with a credible interval — or a year-over-year growth CI — is not supported by this table. Stacking the medians for display is fine.
• Missing (lab, year) rows mean "no estimate", so a grouped (not stacked) layout reads best for the earlier years with partial lab coverage.

Consolidated page (draft)

python3.11 build_compute_page.py

regenerates lab_compute_page_draft.html, one self-contained page in the Epoch website style holding both views of the data: a grouped bar chart of the year-end estimates (median bars, 90%-CI whiskers, collapsible data table), and one walkthrough section per (lab × year-end) snapshot showing every intermediate quantity in model order. The page is a pure view over the two tables from get_all_tables() — new steps or snapshots added to MODEL_STEPS appear on the page (and in the intermediates CSV) with no changes to the viz code (a new snapshot also needs its (lab, year) → trace-key entry in LAB_YEAR_KEYS in generate_lab_compute_tables.py). Re-run it whenever the models or priors change.

Setup

Python 3.11 with pip install -r requirements.txt. Runs need network access for the Epoch data hub fetch.

Tests:

python3.11 -m pytest tests/