← back

The AI Investment Economy

Everyone's watching the top of the stack.
The real story is at the bottom.

Where Are We in the Cycle?

Eight investment booms normalized to the same starting point. The X axis shows years from the start of each boom. Toggle curves on to compare velocity across eras. Historical & NASDAQ data from FRED. Crypto from CoinMarketCap. AI CapEx from earnings reports (2025–26 projected).

Modern
Historical

AI CapEx at Year 4 (2024) is tracking NASDAQ-like velocity. Years 0–4 are from earnings reports. The dot-com peaked at Year 5 ( 529 indexed) and crashed at Year 6. AI CapEx Year 5–6 are analyst projections — reaching 733 at Year 6, surpassing the dot-com peak by 39%. Note the 2023 dip (Year 3) — a real pause before the AI ramp.

The CapEx vs Revenue Gap

Hyperscaler capital expenditure vs AI-generated revenue. The gap between spending and revenue IS the story.

Hyperscaler CapEx ($B)
Gen AI Vendor Revenue ($B)

CapEx 2020–2024 from earnings reports & Platformonomics. 2025–2026 projected (analyst consensus, Futurum). Revenue line is estimated throughout — see sources.

Hyperscaler CapEx is 10–20x generative AI revenue. Not all CapEx is AI-specific, but the acceleration from 2024 onward is primarily AI-driven. Note the 2023 dip — a post-pandemic pullback before the AI ramp. Today's leaders are profitable: only 20% of tech companies are unprofitable vs 36% during dot-com (Bernstein). Note: the revenue line uses estimated figures — the gap is directionally right, not precise.

The Investment Stack — Beyond the Obvious

Six layers of AI infrastructure. Most attention goes to the top. The hardest bottlenecks — and the longest runways — are at the bottom.

Applications

The products end users actually touch

High crowding+
Models & Platforms

The foundation models and companies building them

Medium crowding+
Chips & Silicon

The processors and memory doing the thinking

Medium crowding+
Data Centers & Cooling

The buildings and thermal systems housing AI

Medium crowding+
Power & Energy

The electricity that feeds every GPU cycle

Low crowding

30% transformer shortfall

Key Insight

30% transformer shortfall (Wood Mackenzie). Lead times 128–144 weeks. Nuclear renaissance underway — Constellation signed a 20-year 1.1GW deal with Meta.

You can't train or run AI without electricity. Transformer shortfalls and grid constraints are the hardest bottleneck in the stack — money alone can't solve it.

Bottleneck

128–144 week lead times — grid is maxed out

Timeline2–3 years for new transformer capacity

By the Numbers

Transformer lead time

Wood Mackenzie

128–144 wk

DC electricity demand by 2030

IEA

2–3x

Nuclear deal (Meta + CEG)

Constellation

1.1 GW / 20 yr

Uranium Price

$69.71/lb

+183% since 2020

Nuclear renaissance. Constellation + Meta 20-year 1.1GW deal. BofA targets $130/lb by Q4 2026.

FRED 2026-01

US Electricity Price

19.2¢/kWh

+43% since 2020

Data center electricity demand projected 2–3x by 2030. (IEA, Goldman Sachs)

FRED 2026-01

Key Players

EatonETN

Grid Infrastructure

GE VernovaGEV

Power Generation

Constellation EnergyCEG

Nuclear

NextEra EnergyNEE

Utilities

Sprott Uranium MinersURNM

Uranium Basket

Raw Materials

The mines, metals, and minerals AI runs on

Low crowding+