The Thesis

100,000× effective compute by 2027

Three compounding vectors of progress, each contributing ~0.5 OOMs/year. Compounded: ~5 orders of magnitude from 2023 to 2027.

Vector 01

Compute Scaling

+0.5 OOM/year
+2 OOMs by 2027

GPT-2 (2019) used 4e21 FLOP; GPT-4 used 8e24-4e25 FLOP — 3,000-10,000× in 4 years.

Vector 02

Algorithmic Efficiency

+0.5 OOM/year
+2 OOMs by 2027

GPT-4o costs 6×/4× less than GPT-4. Gemini 1.5 Flash costs 85×/57× less.

Vector 03

Unhobbling

RLHF · CoT · Scaffolding
+1 OOM equivalent

RLHF'd small model = non-RLHF'd model 100× larger. Context 2K → 1M+ tokens.

The Trillion-Dollar Cluster

From GPT-4 at 10MW to 100GW clusters by 2030 — greater than 20% of all US electricity consumption.

2022
GPT-4
H100 equiv~10K
Cost~$500M
Power~10 MW
Equivalent to
10,000 homes
2024
Current frontier
H100 equiv~100K
Cost$Billions
Power~100 MW
Equivalent to
100,000 homes
2026
Million-GPU cluster
H100 equiv~1M
Cost$10s of B
Power~1 GW
Equivalent to
Hoover Dam
2028
$100B cluster
H100 equiv~10M
Cost$100s of B
Power~10 GW
Equivalent to
Small US state
2030
Trillion-dollar cluster
H100 equiv~100M
Cost$1T+
Power~100 GW
Equivalent to
>20% of US electricity

“Power is the binding constraint, not chips or investment.”

— Leopold Aschenbrenner, Situational Awareness, June 2024
Timeline

Path to superintelligence

Proto-engineer by 2026 · AGI 2027 · intelligence explosion 2028-29 · superintelligence by 2030.

2019
GPT-2
4e21 FLOP baseline
2023
GPT-4
8e24–4e25 FLOP — 3,000× to 10,000× GPT-2
2024
Essay published
Leopold publishes 'Situational Awareness: The Decade Ahead'
2025
Fund scales
Situational Awareness LP goes from $254M to $5.52B exposure
2026
Proto-engineer
AI engineer that speeds up coding 1.5–2×
2027
AGI arrives
GPT-4-level model trainable in one minute. ~100,000× effective compute scaleup from 2023.
2028
Proto-researcher
>90% automation of AI research. Intelligence explosion underway.
2029
10× pace
Decade of algorithmic progress compressed to ~1 year.
2030
Superintelligence
100M automated researchers. Economic growth 30%+/year.
AI industry capex
$150B → $8T by 2030
Economic growth
30%/yr +

“Multiple doublings a year” possible — a regime shift comparable to hunting → farming → industrialization.

Automated researchers
100M+

Each working at 100× human speed. A decade of algorithmic progress compressed to one year.

Power draw by 2030
~100 GW

Greater than 20% of all US electricity. The binding constraint is gas — not chips.