Base Rates - Technology and Capital Cycles (classes 1–2)#
Contents
The two classes that anchor the corpus's timing claims: how long general-purpose technologies take to show up in the productivity statistics, and what happens when the capital markets that finance them run ahead of demand.
1. General-purpose technology diffusion#
| Technology | Invention → measurable TFP effect | Rate limit |
|---|---|---|
| Steam | ~80 years | Capital stock replacement |
| Electricity | ~40 years | Factory redesign around unit drive |
| Computing | ~25 years | Software, and complementary skills |
| ERP / enterprise software | ~10 years | Process re-engineering |
| Internet | ~10–15 years | Business model discovery |
The trend is real: each cycle is faster. Extrapolating gives AI something like 5–10 years to measurable aggregate effect, which puts the signal in 2029–33 - the basis for the 2028–2032 claim and for Part V row 2's low 2030 number.
Adjustment, faster: AI ships over existing rails, self-serves, needs no physical retooling, and has by far the fastest adoption curve of any technology in the table.
Adjustment, slower: the rate limit in every row above was never the technology - it was organizational redesign, which runs at the speed of management turnover and has not obviously accelerated. The METR result is direct evidence that adoption and benefit realization are separate variables.
These two adjustments point in opposite directions and roughly cancel, which is why the base case sits close to the naive extrapolation. That is a weak reason to believe it, and it is stated as such.
Why the lag is structural, not measurement noise#
The Solow paradox ("you can see the computer age everywhere but in the productivity statistics") was not a failure of statistics so much as a failure of complementary investment. Electricity's TFP contribution only became visible after factories were redesigned around the unit drive rather than the line shaft - a rebuild that took a generation even after motors were cheap. Computing's contribution waited on software, networks, and a workforce that could use them. In every row of the table, the invention is early and the reorganization is late.
AI inherits the same structure with one twist: the reorganization cost is mostly managerial and process, not physical, so it should be cheaper than electrification and still slower than the software shipping cycle. That is the J-curve the B2 indicator is built to survive. Reading early flat TFP as falsification of capability is the most common error this reference class exists to prevent.
What would break the cancellation. If reorganization itself accelerates - if agents redesign workflows without waiting for management turnover - the faster adjustment wins and the 2029–33 window is too late. That is one path by which Uncertainty 1 would force a re-score of row 2, and it is distinct from "models got smarter." Capability without reorganization still sits in the Solow paradox; reorganization without capability is empty. The product is what B2 watches.
2. Infrastructure capex booms#
Railways (1840s, 1870s), electrification, telecoms (1990s), shale (2010s). The pattern is consistent enough to be useful:
- Over-build relative to near-term demand is the norm, not the exception
- A financial correction occurs, usually triggered by credit rather than by demand
- The physical asset survives the correction and gets used
- Equity holders are largely wiped out; the technology is not
This is where the "~40% chance of a serious correction that resets the timeline by 2–4 years without killing the technology" figure comes from. Note it is a conjunction of a high-probability event (correction) and a specific mechanism (credit) - see Capital.
Adjustment: AI capex has an unusually short-lived core asset. Rail track lasts a century; accelerators last 3–6 years. That makes the correction sharper and the surviving-asset argument weaker than the reference class implies - except for the rising share in power and shells, which behaves like the classic case.
How to read the boom against this class#
Three features of the present boom map cleanly onto the table, and one does not. Over-build is already visible in announced campus capacity against near-term revenue (the A2 / revenue-bar question). Credit migration from hyperscaler cash flow into private credit and SPVs is the classic pre-correction financing structure. Long-lived residual assets (powered land, interconnection rights) are a rising share of the bill, which is the feature that historically protected technology continuity through equity wipeouts. The feature that does not map is asset half-life of the silicon itself: every prior boom left infrastructure that still worked a decade later; GPUs leave a depreciation schedule. That is why row 6 prices a credit event as more likely than a pure demand miss, and why the correction - if it comes - resets financing capacity and secondary labs more cleanly than it resets the research line.
Failure mode of this class applied to AI: treating every drawdown as confirmation. A pure equity re-rating with no credit event and no capacity cull is not the reference-class correction; it is a price move. The class predicts stranded over-build and surviving useful capacity and a multi-year financing chill. Score the conjunction, not the headline.
Pair with winters: class 2 is the capital-markets reference; class 4 is the capability-funding archaeology. Full winters require capability stuck below a paying use case. This cycle has revenue; class 2 is therefore the central case and class 4 is the tail.
Investor scrutiny is not the correction. 2026 earnings seasons already price skepticism of the spend; that is the political economy of the capex test arriving early, not class 2 firing. Class 2 needs credit stress, capacity cull, and multi-year financing chill - score the conjunction.