# 2032–2040 - the physical world

← [Part IV](README.md) · [Index](../README.md)

---

**If robotics cost curves hit, this is where growth rates actually change.**

If they don't, we get a very rich information economy sitting on a normal physical one, with the gap producing sustained political conflict about distribution.

That is the whole fork. Everything before this period is redistribution and reorganization within an existing economic structure. This is the period where cheap cognition either crosses into atoms or it doesn't - and that determines whether this decade is remembered as an important one or a decisive one.

The rate-limiting factor is **physics and atoms**: data that has to be collected in the world, actuators that have to be manufactured, power that has to be permitted and delivered, and care work that still requires a body in a room.

The reason the *growth rate* hinges on atoms is Baumol's arithmetic, and it is worth stating because it is the fork's engine rather than a background fact. When one sector's productivity explodes, its prices collapse and its share of nominal GDP shrinks; aggregate growth converges toward the growth rate of the sectors that did *not* improve, which come to dominate spending precisely because they stayed expensive. Cheap cognition without cheap manipulation therefore produces a one-time level gain and a large relative-price rearrangement - the [two-economy split](../09-macro/prices.md) - but the aggregate growth rate ends up bounded by construction, care, food, and energy. Robotics is the only lever that attacks that stagnant remainder directly, which is why a technology fork in one industry decides a macroeconomic question for all of them. The failure mode of this arithmetic: it assumes the stagnant sectors' demand is inelastic and their labor cannot be augmented short of full automation. If cognition raises physical-sector productivity *indirectly* at scale - permitting, logistics routing, preventive medicine reducing care demand, design cutting material inputs - the growth dividend arrives without the humanoid fork resolving, and the 40/45/15 split above under-weights the middle-into-high path.

## The split

> **~40%** - materially higher growth (global real GDP growth sustained >4%)
> **~45%** - meaningful but unspectacular (+0.5–1.5pp)
> **~15%** - disappointment or disruption-dominant

The middle case is the modal one and the least discussed. It looks like: real productivity gains, real welfare gains, badly distributed, with the political salience of the disruption exceeding the political salience of the benefit - because the benefit shows up as consumer surplus that nobody notices and the disruption shows up as specific people losing specific jobs in specific places.

It is least discussed for a structural reason, not an accidental one: neither pole of the public argument has an incentive to describe it. The accelerationist case is funded and told by people whose valuations require the high branch; the catastrophist case is organized around tails; and "large real gains, politically illegible, badly distributed" recruits no coalition and sells no book. That silence is itself a forecast input - it means the modal world arrives *unprepared for*, with institutions tuned for either boom or doom encountering something that is neither. The corpus treats the thinness of middle-case planning as evidence the scenario is under-priced, not evidence it is unlikely; the failure mode of that reasoning is survivorship in the analogy set, since the same argument could have been made for nuclear power's middle case and the poles turned out closer to right there.

That middle case is also what [prices](../09-macro/prices.md) and [assets](../09-macro/assets.md) describe from the macro side: the two-economy split fully mature, wealth concentrated in the inelastic complements, fiscal systems still chasing a labor-income base that has thinned.

## What determines which branch

### Primary: robotics cost curves and automated experimentation

These are the two places where cheap cognition converts into physical output.

**Robotics.** The binding constraint remains [data](../03-domains/physical/robotics/data-problem.md), not intelligence - there is still no internet of manipulation. The economic threshold is delivered cost per hour against local wages and utilization, not unit price. Who wins is decided by [manufacturing capacity](../03-domains/physical/robotics/supply-chain.md) (actuators, magnets, cells, assembly skill), not model quality - and that capacity is not distributed the way AI research talent is. China is stronger here than in frontier models. The [humanoid form](../03-domains/physical/robotics/form-factor.md) is a bet on general-purpose data collection; it may be right for reasons other than mechanical efficiency.

> General-purpose humanoids at meaningful economic scale is a **2032–2040** story. Structured environments ran 5+ years ahead and should already be commercial. Unstructured - homes, hospitals, construction sites - is what this period decides.

**Science.** [Automated experimentation](../03-domains/cognitive/science.md) is the other conversion path: closed-loop labs that design, run, and interpret experiments without a human in every cycle. Where ground truth is instrumented (materials, chemistry, parts of biology), this compounds. Where it is not, it doesn't. The same master asymmetry that ordered [Part III](../03-domains/) still orders which sciences accelerate.

### Secondary: energy

The high-growth branch requires enormous amounts of power, and the permitting question from [Uncertainty 2](../06-uncertainties/power-permitting.md) still constrains the ceiling. By this period the question is less "can we generate" than "can we interconnect and deliver" - transmission and substations, not nameplate capacity.

[Demography](../09-macro/demography.md) pushes adoption pressure *up*: labor-scarce economies pull automation in rather than fighting it. Japan is the natural experiment. The US, still the advanced economy with a growing labor force, is the least representative case for political economy of adoption - a fact the corpus underweighted until round 7.

### Wide error bars: recursive research acceleration

If [Uncertainty 1](../06-uncertainties/recursive-self-improvement.md) lands hard, this entire period compresses forward. The three governors (verification, physical supply chain, financial feedback through the neutral rate) still argue that the observable consequence is more likely a decade of unusually fast progress than a discontinuity - but a decade of unusually fast progress *is* this period arriving early, which is enough to scramble every date below.

## What the period looks like under the modal (middle) case

### Labor: composition, not aggregate

The [Game 4](../02-games/4-labor.md) aggregate displacement estimate was always an upper bound once demography is included. By this period the acute shortage is in **care and skilled trades**, not a labor surplus in cognitive work. Real wages in both outpace credentialed cognitive professions - the same claim [assets](../09-macro/assets.md) makes from the ownership side.

The [apprenticeship gap](../06-uncertainties/apprenticeship-gap.md), if unmitigated through the 2020s, is now a structural expert shortage. Professions that cut junior hiring in 2025–30 discover in 2035 that they cannot staff senior roles at any price. Immigration is the arithmetic response and the political non-response in the countries with the sharpest shortage.

### Macro: the feedback loops mature

- **Rates.** If productivity stepped up, the real neutral rate should have risen - success making the next round of capex more expensive. Demographics still pull the other way. The most likely world remains one where AI works well enough to matter and real rates stay lower than the pure productivity story implies. → [Rates](../09-macro/rates.md)
- **Fiscal.** Tax systems built on labor income meet an economy using proportionally less of it, *and* a demographic dependency ratio that does the same thing independently. The two are the same problem arriving twice. → [Fiscal](../09-macro/fiscal.md)
- **Assets.** Six of the seven inelastic complements are owned by capital. Consumer-surplus gains coexist with wealth concentration. That is not a bug in the framework; it is the framework's distributional prediction.

### Geopolitics: partial bifurcation, not a winner

The [most likely 2030s configuration](../03-domains/contested/geopolitics/bipolar.md): two capability ecosystems closer than controls intend; diverging capacity favoring whoever built power; third countries choosing per-layer not per-bloc; **robotics as the mirror image of AI**, with the dependency arrow reversed. A Western reshoring program executed with imported robots is a weaker strategic position than the announcements say.

- **India** has either pivoted toward domestic demand and energy, or is managing a permanent shortfall in the export-led model. → [India](../03-domains/contested/geopolitics/india.md)
- **Europe** has either converted rule-setting into industrial capacity or accepted permanent dependence on foreign producers while taxing its own adopters. → [Europe](../03-domains/contested/geopolitics/europe.md)
- **The Gulf** knows whether compute was a real diversification or a depreciating reserve asset with a five-year half-life. → [Gulf](../03-domains/contested/geopolitics/gulf.md)
- **Global South** outcomes split hard: welfare up where complements exist, development strategy failed where they don't. → [Global South](../03-domains/contested/geopolitics/global-south.md)

### Governance: living inside the post-incident architecture

Whatever was written in the 18 months after the first salient incident is now the regime. [Part V](../05-probabilities/) row 4 - binding international agreement with real verification - remains the low-probability row precisely *because* the competitive use of export controls and the only future verification mechanism are the same object. That trade was likely made implicitly in the late 2020s. → [Bipolar](../03-domains/contested/geopolitics/bipolar.md) on the compute-governance trap

[Insurance](../03-domains/cognitive/insurance.md) either matured into a material line with published rates or forced a state backstop after a correlated event. Either way, the deployment frontier is a financial object as much as a capability one.

## The Taiwan inheritance

Every projection above still assumes Taiwanese leading-edge fabrication continued through the critical window. If that assumption failed earlier, this period is unrecognizable - capacity bifurcation is more extreme, timelines for everything compute-gated slip by years, and robotics supply chains reorder around whoever holds the remaining fabs and the magnet/actuator stack. The ~90% modelling assumption was never a geopolitical forecast; it was an input the whole period rests on. → [Uncertainty 4](../06-uncertainties/taiwan.md)

## What to watch

| Signal | Meaning |
|---|---|
| Delivered $/hour of robot work vs. local wages, by environment class | Which branch of the fork |
| Teleoperation-to-autonomy ratio on unstructured tasks | Whether general-purpose is real |
| Global real GDP growth sustained above/below 3% | Scoring the 40/45/15 split |
| Care-worker real wages vs. cognitive professional wages | Demography + Game 4 compositional claim |
| Share of new frontier capacity on owned/contracted generation | Whether energy ceiling loosened |
| Internal research cycle time at frontier labs | Whether RSI compressed this period forward |

## Failure modes

- **If humanoids hit >1M units/yr deployed early in the period with rising autonomy share**, the middle case was too cautious and the high-growth branch probability should have been higher. → [Part V row 5](../05-probabilities/)
- **If TFP never clears 2%/yr sustained despite visible deployment**, either the J-curve is longer than assumed, measurement is broken in the way [prices](../09-macro/prices.md) describes, or the organizational-redesign constraint was even harder than claimed.
- **If learned verification retires the master asymmetry**, the domain ordering of [Part III](../03-domains/) collapses and physical-world lag was a temporary artifact of 2020s training methods - the document's central claim fails. → [Uncertainty 5](../06-uncertainties/learned-verification.md), [Steelman](../08-method/steelman.md) §1

### Fork on delivered work, not form factor

The growth fork is [B12](../07-indicators/diffusion/capability.md) and delivered $/hour in enough physical tasks - warehouses, yards, fields, then messier environments. Humanoid unit sales without autonomy and utilization are a form-factor fashion cycle ([form factor](../03-domains/physical/robotics/form-factor.md)). Score the period by whether aggregate growth and labor share move with physical automation, not by whether bipeds make keynotes.

**Evidence family: physical.** Delivered work, autonomy ratios, wafer share outside TW - not financial Q&A tone. Branches diverge too far for a single narrative past this period; conditionals beat calendars ([timelines hub](README.md)). Full joint worlds (best, base, worst, named branches): [scenarios](scenarios.md).

---

**Related:** [Robotics](../03-domains/physical/robotics/) · [Demography](../09-macro/demography.md) · [Prices](../09-macro/prices.md) · [Geopolitics](../03-domains/contested/geopolitics/) · [Uncertainty 1](../06-uncertainties/recursive-self-improvement.md) · [Part V](../05-probabilities/) · [Scenarios](scenarios.md)

**Previous:** [2028–2032](2028-2032.md) · **Next:** [Scenarios](scenarios.md)