Part I - The Physical Substrate#
Contents
Everything downstream depends on four inputs: compute, energy, data, capital. Each has a different scaling law and a different ceiling.
The order matters. These are not four parallel concerns - they are a relay. Compute constrained the field through ~2024. Data exhaustion redirected it in 2024–25. Energy binds from ~2026. Capital tests the whole thing around 2027–29. And the character of each ceiling determines what the games in Part II can and cannot do.
Sections#
| Constraint | Binds when | Character | |
|---|---|---|---|
| Compute | Effective compute ~7–10×/yr; dollar term breaks at Apollo scale | 2029–30 | Financial, possibly evadable via efficiency |
| Energy | Interconnect queues, turbines, transformers | ~2026 onward | Physical and regulatory; skewed toward loosening - sector-side effects in energy sector |
| Data | High-quality human text exhausted | Already binding | Redirected rather than blocking - and it created the master asymmetry |
| Capital | Not a ceiling - a continuously re-priced verdict | 2027–29 | Financial and reversible; the only input that can run backwards |
| Inference economics | Not a ceiling - a floor collapsing | Continuous | Sets the two-year moat on any capability |
Two of these are not constraints#
Compute, energy, and data are ceilings - they cap what is possible. Capital and inference economics are not, and reading them as ceilings is a common error.
Capital is a verdict: there is no physical limit on the flow, only a continuously re-priced judgment about repayment. It is the fastest-moving input and the only one that can reverse. Inference cost is a floor collapsing: it does not constrain anything, it sets the depreciation schedule on every capability advantage anyone holds.
The practical difference: a ceiling tells you what cannot happen, a verdict tells you what someone currently believes, and a collapsing floor tells you how long your advantage lasts. Forecasts that treat all five rows alike get the timing wrong in both directions.
Relay + floor together: energy binds the build; inference collapse sets the moat life of whatever you built; capital re-prices whether the next build is allowed. → 2026–2028 capex test, Rates financial governor.
The inputs are coupled, not parallel#
Reading the five rows as independent misses most of the dynamics. The couplings do the work:
Energy × capital. The energy constraint is what converts capital from a verdict into a commitment. Money queued behind a 5-year interconnection wait cannot exit on an earnings cycle; it is already poured into shells and substations. So the tighter energy binds, the more the capital stack shifts toward long-lived physical assets - which is exactly the shift that bounds the downside of a correction (capital on the physical-asset residual).
Data × compute. Effective compute is only an engine where a training signal exists. Past data exhaustion, marginal FLOPs concentrate on verifiable domains, which narrows what the compute curve buys even as it steepens. This is why the 7–10×/yr figure and "progress feels uneven" coexist without contradiction.
Inference × capital. The collapsing floor is simultaneously the bull case and the bear case for the capex test: it destroys pricing power at any fixed capability (bear), while Jevons expansion of consumption is the only plausible route to the revenue bar (bull). Which effect dominates is close to the question of 2026–2028, and it cannot be settled from either page alone.
The relay metaphor's failure mode. A relay implies each constraint hands off cleanly to the next. The couplings above mean they can instead bind simultaneously - an energy-delayed build raises capex per delivered FLOP just as the capital verdict is being re-priced. The base case here assumes sequential binding; the compounding case is the pessimistic tail, and it is tracked jointly by A2 and A4 rather than by any single indicator.
Behind-the-meter routing. When A4 (queue) looks stuck, check whether new MW is arriving as captive generation that never enters the public queue (energy). Slack in the political series with tight physical capacity is a different story from both tight.
The one thing to take from Part I#
Data exhaustion forced the frontier onto verifiable-reward training. That produced the asymmetry that runs through this entire document: capability grows fastest where verification is cheap. Math, code, and formal reasoning race ahead. Taste, judgment, long-horizon physical interaction, and anything whose ground truth is expensive or contested lag badly.
Every domain prediction in Part III is downstream of that sentence. Uncertainty 5 is the claim that would reorder them.