Allocation Layer — Findings¶
Author: automated analysis pipeline Date: 2026-05-03 Status: Allocation layer complete. Effective-compute handoff parameters at the bottom.
1. Summary¶
Under sourced supply assumptions and base allocation parameters, the largest single frontier training run grows ~27.6% per year (2024→2040), reaching ~6.9e+28 FLOP by 2040. Across the 16 combined supply × allocation scenarios, the headline trajectory ranges from 14.1%/yr (chip-bottleneck × inference-heavy) to 48.1%/yr (capex-rich × training-race) — a roughly 50× spread in absolute 2040 FLOP.
The historical Rule A 2018+ trend (5.97×/yr per Phase 1) is unreproducible on supply fundamentals + reasonable allocation. The historical extrapolation crosses through the entire allocation envelope around 2027–2028 and reaches ~1e+37 FLOP by 2040, vs realistic projections in the 1e+27 to 1e+29 range.
The frontier-run share of total compute falls in every scenario from ~3.5% (2024) to <4% even in the most concentrated case by 2040. This is the central allocation insight: total usable compute scales faster than any single training run can absorb, regardless of allocation choice.
2. Why allocation was needed¶
The historical baseline measures single training-run FLOP; the supply capacity model measures total annual usable compute. These are different quantities, and comparing them directly was the "central reading mistake" flagged throughout the orientation docs.
The allocation layer bridges them by modeling what share of total usable compute goes to the single largest training run, year by year, scenario by scenario. With this in hand, the supply trajectory can be translated into a forward single-run trajectory and compared apples-to-apples with the historical trend.
The headline output, largest_frontier_run_flop, is the first
forward-looking quantity in this project that's directly comparable
with the historical Rule A 2018+ frontier-run trend.
3. Allocation assumptions¶
Source: data/assumptions/allocation_input_assumptions.yaml. Four
scenarios at three milestone years, linearly interpolated.
Bucket shares (sum to 1.0 at every milestone)¶
| Scenario | Year | Inference | Training | AI R&D | Post-train | Safety | Reserved |
|---|---|---|---|---|---|---|---|
| base | 2024 | 0.55 | 0.30 | 0.08 | 0.04 | 0.01 | 0.02 |
| base | 2030 | 0.60 | 0.24 | 0.10 | 0.04 | 0.01 | 0.01 |
| base | 2040 | 0.65 | 0.18 | 0.10 | 0.04 | 0.01 | 0.02 |
| inference_heavy | 2040 | 0.78 | 0.10 | 0.06 | 0.03 | 0.01 | 0.02 |
| training_race | 2040 | 0.45 | 0.35 | 0.12 | 0.04 | 0.02 | 0.02 |
| rnd_acceleration | 2040 | 0.55 | 0.18 | 0.20 | 0.04 | 0.01 | 0.02 |
Training-pool decomposition¶
| Scenario | Year | frontier_lab_training_share | largest_run_concentration | cluster_contiguity_factor |
|---|---|---|---|---|
| base | 2024 | 0.65 | 0.20 | 0.90 |
| base | 2030 | 0.60 | 0.10 | 0.90 |
| base | 2040 | 0.55 | 0.05 | 0.85 |
| training_race | 2040 | 0.65 | 0.15 | 0.95 |
| inference_heavy | 2040 | 0.50 | 0.03 | 0.80 |
| rnd_acceleration | 2040 | 0.55 | 0.07 | 0.80 |
All values are flagged confidence: medium and source scope_defaults.
Refinement targets in §8.
4. Compute by bucket¶
Under base supply × base allocation, the 6 buckets at 2024 / 2030 / 2040:
| Bucket | 2024 (FLOP/yr) | 2030 (FLOP/yr) | 2040 (FLOP/yr) |
|---|---|---|---|
| Inference | 2.18e+28 | 1.10e+30 | 1.07e+31 |
| Training | 1.19e+28 | 4.42e+29 | 2.96e+30 |
| AI R&D experiment | 3.18e+27 | 1.84e+29 | 1.65e+30 |
| Post-training | 1.59e+27 | 7.37e+28 | 6.58e+29 |
| Safety / eval | 3.97e+26 | 1.84e+28 | 1.65e+29 |
| Reserved / idle / fragmented | 7.94e+26 | 1.84e+28 | 3.29e+29 |
| Total usable | 3.97e+28 | 1.84e+30 | 1.65e+31 |
Inference dominates throughout. Training holds at ~30% in 2024 but shrinks in absolute share toward 2040, even though the absolute training-compute number grows ~250× over the horizon.
Chart: outputs/charts/allocation_compute_by_bucket.png.
5. Largest frontier training-run projections¶
Under each combined scenario, the largest single frontier training run by year. Sorted by 2040 value (highest first):
| 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 |
| power_dc_bot × training_race | 1.39e+27 | 4.03e+28 | 2.15e+29 | 37.0%/yr |
| chip_bot × training_race | 1.53e+27 | 4.99e+28 | 2.12e+29 | 36.1%/yr |
| capex_rich × rnd_acceleration | 1.53e+27 | 4.40e+28 | 1.60e+29 | 33.7%/yr |
| capex_rich × base | 1.53e+27 | 4.66e+28 | 1.22e+29 | 31.4%/yr |
| base × rnd_acceleration | 1.39e+27 | 2.25e+28 | 9.13e+28 | 29.9%/yr |
| base × base | 1.39e+27 | 2.39e+28 | 6.93e+28 | 27.6%/yr |
| power_dc_bot × rnd_acceleration | 1.23e+27 | 1.16e+28 | 3.68e+28 | 23.7%/yr |
| chip_bot × rnd_acceleration | 1.34e+27 | 1.44e+28 | 3.62e+28 | 22.9%/yr |
| capex_rich × inference_heavy | 1.09e+27 | 2.12e+28 | 3.47e+28 | 24.2%/yr |
| power_dc_bot × base | 1.23e+27 | 1.23e+28 | 2.79e+28 | 21.6%/yr |
| chip_bot × base | 1.34e+27 | 1.52e+28 | 2.75e+28 | 20.8%/yr |
| base × inference_heavy | 9.88e+26 | 1.09e+28 | 1.98e+28 | 20.6%/yr |
| power_dc_bot × inference_heavy | 8.69e+26 | 5.58e+27 | 7.97e+27 | 14.9%/yr |
| chip_bot × inference_heavy | 9.52e+26 | 6.91e+27 | 7.84e+27 | 14.1%/yr |
Headline: range is 14.1% → 48.1% CAGR; absolute 2040 spread is ~50×. Allocation choice (color in the chart) dominates supply choice (marker) — the four allocation scenarios stack ~0.7 OOM apart by 2040, while supply variation within an allocation accounts for ~0.3 OOM.
Chart: outputs/charts/allocation_largest_frontier_run.png.
6. Historical comparison¶
The vs-historical chart
(outputs/charts/allocation_vs_historical_training_compute.png)
overlays the historical Rule A 2018+ extrapolation (5.97×/yr,
rebased to 1e+25 FLOP at 2024) on all 16 allocation projections.
Key observations:
- Crossover ~2027–2028: the historical extrapolation reaches the upper edge of the allocation envelope (training_race scenarios) around 2027–28 and rapidly leaves it behind.
- 2030 gap: historical extrapolation ~4.5e+29 FLOP vs best-case allocation ~1.5e+29 FLOP (capex_rich × training_race). Roughly 3× gap.
- 2040 gap: historical extrapolation ~1e+37 FLOP vs best-case allocation ~1e+30 FLOP. Roughly 7 OOM gap.
The gap_ratio column in
outputs/tables/allocation_vs_historical_trend.csv quantifies this
year-by-year for every combined scenario.
Reading note: the historical trend is descriptive — it fits the 2018–2024 frontier-model corpus. It is not a forecast. The allocation model is a forward projection on supply fundamentals and assumed allocation choices. The "gap" represents three things blended together: (a) the historical trend was already slowing; (b) our allocation parameters may be conservative; (c) supply fundamentals genuinely cap single-run growth. Phase 4 (effective compute) and Phase 5 (capability mapping) will reframe this from raw FLOP to effective FLOP and may close some of the apparent gap.
7. Scenario sensitivity¶
Sensitivity by parameter, holding all else at base × base:
- Allocation scenario (largest effect): swapping base allocation for training_race raises 2040 largest_run by ~7.7×; swapping for inference_heavy reduces it by ~3.5×.
- Supply scenario (smaller effect within an allocation): swapping base supply for capex_rich raises 2040 largest_run by ~1.8×; swapping for chip_bottleneck reduces by ~2.5×.
- Implied joint range across all 16 scenarios: 50× in absolute 2040 FLOP terms.
Conclusion: allocation choice dominates supply choice for single-frontier-run projections. This is the inverse of what intuition might suggest if you only think of the model as "supply-constrained" — once supply is generous (any of the four supply scenarios except chip_bottleneck have ample compute by 2030), the binding question becomes how that compute is used, not how much exists.
8. Key uncertainties¶
- largest_run_concentration is the highest-leverage unsourced
parameter. A swing from 0.10 to 0.20 (within plausible range)
changes 2040 largest_run by 2× linearly. Currently flagged
mediumconfidence; the actual historical concentration parameter is debated — top-10% of frontier-lab compute? top single 5%? Sensitivity analysis would tighten this. - frontier_lab_training_share is the second-highest leverage parameter. Currently 0.55–0.70 across scenarios. The actual share depends on the definition of "frontier lab" (3 labs? 10? 30?) which the model doesn't separately address.
- Allocation parameters are not backcast-calibrated. The 2024 starting values were chosen to match plausible 2024 reality but were not formally fit against the historical 2018–2024 frontier-model record. A backcast calibration could tighten the 2024 anchor.
- No correlation across parameters. If allocation is correlated with supply (e.g. chip_bottleneck → forced higher concentration to make any training run viable), our scenario joint is too uniform. Sensitivity to the joint structure is unmodeled.
- No fragmentation in the cross-product. Each combined scenario assumes one supply scenario across all 17 years; real-world supply may shift binding constraints mid-horizon.
- Historical Rule A 2018+ trend may overweight 2018-2022. The Rule A trend gives equal weight to the steeper early-deep- learning era; recent (2023-2026) frontier models have been clustering between 1e25 and 1e26 FLOP rather than continuing the historical 5.97× pace. Re-fitting Rule A on 2022+ data would give a slower trend that the allocation projections could plausibly catch up to.
9. Effective-compute layer handoff parameters¶
These are the explicit handoff parameters for the effective-compute layer (Phase 4 in the upstream spec; renamed to "the effective- compute layer" in this project's docs).
=== Largest frontier run by year (FLOP) ===
Year Slow envelope Base case Fast envelope
(chip_bot × (base × base) (capex_rich ×
inference_heavy) training_race)
2024 9.52e+26 1.39e+27 1.74e+27
2025 2.55e+27 4.74e+27 5.71e+27
2026 4.14e+27 9.27e+27 1.18e+28
2027 5.19e+27 1.31e+28 1.83e+28
2030 6.91e+27 2.39e+28 1.53e+29
2035 7.31e+27 4.16e+28 4.18e+29
2040 7.84e+27 6.93e+28 9.38e+29
Read full CSV: outputs/tables/allocation_largest_frontier_run.csv
=== Recommended effective-compute layer envelope ===
Use base supply × base allocation as the central case.
Use chip_bottleneck × inference_heavy as the slow / pessimistic floor.
Use capex_rich × training_race as the fast / optimistic ceiling.
Use base × inference_heavy and capex_rich × base as alternative-stress
cases (capacity exists but allocation favors / disfavors training).
=== Bucket-level annual compute by year (base × base, FLOP/yr) ===
2024: inference 2.18e+28, training 1.19e+28, ai_rnd 3.18e+27,
post_training 1.59e+27, safety 3.97e+26, reserved 7.94e+26
2030: inference 1.10e+30, training 4.42e+29, ai_rnd 1.84e+29,
post_training 7.37e+28, safety 1.84e+28, reserved 1.84e+28
2040: inference 1.07e+31, training 2.96e+30, ai_rnd 1.65e+30,
post_training 6.58e+29, safety 1.65e+29, reserved 3.29e+29
=== Frontier run share of total compute (base × base) ===
2024: 3.51%
2030: 1.30%
2040: 0.42%
The share trajectory is nearly identical across supply scenarios
(within an allocation scenario) because supply changes both
numerator and denominator proportionally. The only allocation
with materially different share is training_race, which keeps
~3-4% through 2040.
Known weaknesses (carry forward)¶
- 7-OOM 2040 gap to historical extrapolation is real — the effective-compute layer needs to address whether this is capability-relevant or whether algorithmic / architectural improvements close most of it.
- All allocation parameters are
confidence: mediumfrom upstream scope defaults; sourcing pass needed. - No backcast calibration; the 2024 anchor is plausible but unverified.
- 16 combined scenarios is a lot; downstream layers should pick a small subset (slow / base / fast) rather than carry all 16.
10. Open questions¶
- What's the right way to compare against the historical trend? Options: (a) compare raw FLOP (current approach, shows large gap); (b) compare effective FLOP after Phase 4 adjustments; (c) re-fit the historical trend on 2022+ data only and compare against that. All three are defensible; the project should pick one as the headline.
- Should the cluster_contiguity_factor be supply-scenario- dependent? Currently allocation-only. Under the power_datacenter_bottleneck supply scenario, contiguity should plausibly fall (more fragmentation when DC slots are scarce). Joint structure would tighten the projection.
- Is the largest-run concentration parameter the right handle? An alternative: model the number of frontier-run- class training runs per year (currently implicit in 1 / largest_run_concentration). Could be more interpretable.
- How should the effective-compute layer use these envelopes? The handoff is 16 scenarios; the next layer probably wants 3 (slow / base / fast). Recommended subset above; revisit when building the effective-compute layer.
Appendix: deliverable checklist¶
| Spec deliverable | File | Status |
|---|---|---|
| Allocation assumptions YAML | data/assumptions/allocation_input_assumptions.yaml |
✓ |
| 4 scenario YAMLs | scenarios/allocation_*.yaml |
✓ |
| Engine module | model/allocation_engine.py |
✓ |
| Pipeline | pipelines/allocation.py |
✓ |
| Chart helpers | pipelines/allocation_charts.py |
✓ |
| 6 charts | outputs/charts/allocation_*.png |
✓ |
| 5 tables | outputs/tables/allocation_*.csv |
✓ |
| Tests | tests/test_allocation_engine.py |
✓ (9/9 passing) |
| Initial notes | docs/allocation_initial_notes.md |
✓ |
| Findings memo | docs/allocation_findings.md |
✓ (this file) |
| Scope section | docs/scope.md §3 |
✓ |
| README updated | README.md |
✓ |
| Orientation docs updated | docs/{model_state,model_map,output_guide,component_contracts}.md |
✓ |