Skip to content

Allocation Layer — Sprint 1 Initial Notes

Date: 2026-05-03 Sprint goal: working allocation layer producing largest_frontier_run_flop from supply outputs across the 4×4 = 16 combined-scenario cross-product.

What's in this sprint

  • data/assumptions/allocation_input_assumptions.yaml — 4 scenarios × milestones (2024 / 2030 / 2040) × 9 parameters per milestone (6 bucket shares + 3 training-decomposition multipliers).
  • scenarios/allocation_*.yaml — 4 registration files (display name, description, assumption_scenario reference). Same pattern as supply.
  • model/allocation_engine.py — load / interpolate / validate, plus the projection pipeline (cross-product, allocate buckets, frontier-lab decomposition, largest-run estimate, historical comparison).
  • pipelines/allocation.py — uv run allocation. Reads supply CSV + historical trend, runs engine, writes 5 tables + 6 charts.
  • pipelines/allocation_charts.py — 6 chart helpers.
  • tests/test_allocation_engine.py — 9 tests covering all 8 invariants from the scope plus a negative test on share validation.

Headline numbers (base supply × base allocation)

Year Largest run (FLOP) Share of total
2024 1.39e+27 3.51%
2030 2.39e+28 1.30%
2040 6.93e+28 0.42%

CAGR 2024→2040: 27.6% per year for the largest single frontier run. For comparison, the historical Rule A 2018+ extrapolation is 5.97× per year (497% CAGR). The historical extrapolation crosses through the realistic allocation envelope around 2027–2028 and continues skyward. By 2040 the historical extrapolation reaches ~1e+37 FLOP, versus realistic allocation projections in the 1e+27 to 1e+29 range.

Range across all 16 combined scenarios

The largest_run trajectory is dominated by allocation choice, not supply choice. From outputs/tables/allocation_scenario_summary.csv:

Combined scenario 2024 2030 2040 CAGR
capex_rich × training_race 1.74e+27 1.53e+29 9.38e+29 48.1%/yr
base × training_race 1.58e+27 7.84e+28 5.34e+29 43.9%/yr
capex_rich × rnd_acceleration 1.53e+27 4.40e+28 1.60e+29 33.7%/yr
base × rnd_acceleration 1.39e+27 2.25e+28 9.13e+28 29.9%/yr
capex_rich × base 1.53e+27 4.66e+28 1.22e+29 31.4%/yr
base × base (headline) 1.39e+27 2.39e+28 6.93e+28 27.6%/yr
chip_bottleneck × base 1.34e+27 1.52e+28 2.75e+28 20.8%/yr
chip_bottleneck × inference_heavy 9.52e+26 6.91e+27 7.84e+27 14.1%/yr
(... 8 more)

The ratio of fastest (capex_rich × training_race, 48.1%/yr CAGR) to slowest (chip_bottleneck × inference_heavy, 14.1%/yr CAGR) is roughly 4× over 16 years, or ~50× by 2040 in absolute FLOP.

Three observations worth flagging now

1. Allocation choice dominates supply choice for the largest-run trajectory.

In the cross-product chart (outputs/charts/allocation_largest_frontier_run.png), the four allocation-scenario "color bands" stack vertically by ~0.7 OOM by 2040, while the four supply-scenario "marker variants" within each color band only spread by ~0.3 OOM. This means: the political / strategic question of how compute is allocated matters more for frontier-run trajectories than the physical question of how much compute exists in total.

2. The frontier-run share of total compute falls in every scenario.

This is the central allocation insight: even under the training-race scenario where labs prioritize frontier runs, total usable compute scales faster than any single training run can absorb. By 2040 the largest run is < 4% of total compute even in the most concentrated scenario, vs ~3.5% in 2024. The "compute is consumed by inference serving" narrative holds across scenarios.

3. The historical Rule A 2018+ trend is unreproducible on supply fundamentals + reasonable allocation.

The vs-historical chart (outputs/charts/allocation_vs_historical_training_compute.png) shows the historical extrapolation crosses above every allocation projection by ~2027 and reaches ~1e+37 FLOP by 2040. No combination of supply + allocation under our assumptions reproduces 5.97×/yr single-run growth. This is what the orientation docs warned about, now made concrete by an actual single-run forward projection. Three honest readings:

  • The historical trend was already slowing. Recent frontier models are clustered between 1e25 and 1e26 FLOP rather than continuing the 5.97× pace; the historical fit gives equal weight to the steeper 2018-2022 era.
  • Allocation parameters are conservative. The training-race scenario at 35% training share + 65% frontier-lab share + 15% largest-run concentration is at the upper bound of plausible — but not implausible. If realistic concentration is closer to 20-25%, the gap closes by ~1 OOM.
  • Supply fundamentals genuinely cap single-run growth. Even capex_rich + training_race at 48.1%/yr CAGR is one full OOM/yr slower than the historical 5.97×/yr.

All three are probably partly true. Phase 4 (effective compute) and Phase 5 (capability mapping) will reframe this from "single-run FLOP" to "effective single-run training-equivalent FLOP" and may close some of the apparent gap.

What the next pass should do

Not in this sprint, but candidates for refinement:

  1. Source the allocation parameters. Currently all values are confidence: medium from the upstream scope defaults. Hyperscaler 10-Ks, lab disclosures, and academic estimates of the largest-run concentration parameter could move the central estimates and tighten the bounds.
  2. Add a fragmented-market scenario. The scope offers allocation_fragmented_market as optional; could add as a 5th allocation scenario if cluster-contiguity becomes a bigger story.
  3. Time-vary the supply scenario weights. Currently each combined scenario assumes the same supply scenario across all 17 years; a scenario where the binding constraint shifts mid-horizon would be more realistic.

Open questions

  1. Should the largest-run concentration parameter be calibrated to 2018-2024 backcast? The historical Rule A 2018+ trend is observable; the allocation model should be able to reproduce the historical single-run trajectory in backcast. The current assumptions are forward-only; a backcast calibration would tighten the 2024 starting values.
  2. How should the cluster_contiguity_factor evolve? Currently modeled as monotonically declining (geographic / cluster-size fragmentation grows). Could go the other way if frontier labs keep building larger dedicated campuses (Stargate-style).
  3. What's the inference / training split within a frontier lab? We model frontier_lab_training_share as one number per scenario; in practice some labs (e.g. inference-heavy commercial) will have a very different split from others (e.g. capability-focused safety labs). A weighted-by-lab decomposition could be a future refinement.

These will live in docs/allocation_findings.md once the memo is finalized. Provisional from sprint 1:

  • largest_frontier_run_flop_by_year — already emitted; the headline Phase 4 input.
  • training_compute_flop_year, ai_rnd_experiment_compute_flop_year, post_training_compute_flop_year, inference_compute_flop_year — bucket-level totals.
  • frontier_run_share_of_total_compute — for sensitivity testing.
  • Recommended scenario envelope:
  • Slow: chip_bottleneck × inference_heavy (CAGR 14.1%)
  • Base: base × base (CAGR 27.6%)
  • Fast: capex_rich × training_race (CAGR 48.1%)