Technology · Issue 01
Five times more compute, every year
Epoch AI: training compute for frontier language models has grown 5× per year since 2020 — a doubling every 5.2 months. That is not Moore’s law. That is a wartime production curve wearing a research-lab badge.
- Frontier training compute since 2020
- 5×/yr
- Doubling time
- 5.2 mo
- Orders of magnitude per year
- 0.7
- The line the 2025–26 frontier crossed
- 10²⁶
A slope that should be illegal
Epoch AI’s February 2026 trends page is the print: training compute for frontier language models has grown at 5× per year since 2020, doubling every 5.2 months, 0.7 orders of magnitude a year. Moore’s law was 2× every 24 months. This is a mobilisation. The chart is that slope applied to a 2020 GPT-3-class anchor (~3×10²³ FLOP). It is a trend, not a claim that a particular 2026 model used 4.7×10²⁷ FLOP. The trend is the news. Hardware, cluster size, and money all have to rise together for a 5× to happen. That is why five firms now outspend the oilfield.
The unit is watts and wafers
Pair this with The power behind intelligence, The $1.5 trillion bottleneck, and 2,500 gigawatts waiting for a wire. A 5× compute curve is a 5× claim on HBM, on CoWoS, on substations, on water. It does not continue because a blog says so. It continues until a bottleneck — power, packaging, or a board — says no. The investing question is which layer rations first.
The long view is a slope that must break
Exponential curves on a linear planet break. The break can be a pause, an efficiency miracle, or a wall. Own the rationed inputs on the way up. Do not own a 2030 model that needs another four 5×. Size the labs as a call option on the slope remaining intact for one more doubling, not four.
Investing lens
Horizon 2–6 years · Educational, not advice
A 5× compute slope is a claim on power, memory and packaging. Own the rationed layer. Do not underwrite four more doublings.
Where the map points
- HBM, foundry, advanced packaging
- Power and interconnect with contracted offtake
- The labs, sized as slope-risk
- A pause sleeve: 5× is not a law
What can break it
- Efficiency that bends FLOP demand
- A power or packaging wall
- The trend was always a small-n fit
CHART does not recommend securities, funds or trades. Figures can be revised by their publishers. Do your own research and consider regulated advice before allocating capital.
Sources
Every headline number traces to a named publisher. Contextual sources inform the essay, not the key stat.
- 01PrimaryEpoch AI2026-02-05Trends in Artificial Intelligence
Training compute for frontier language models has been growing at 5× per year since 2020 (doubling every 5.2 months; 0.7 OOM/year).
- 02PrimaryInternational Energy Agency2026Key Questions on Energy and AI — Executive summary
Largest tech companies’ capex exceeded $400bn in 2025, expected to jump another 75% in 2026. Five firms now spend more than global oil and gas production investment. AI factories more than tripled in 18 months.
- 03PrimaryWorld Semiconductor Trade Statistics2026Spring 2026 semiconductor market forecast
2026 market projected at $1.51 trillion, +90%; memory >$800 billion.
Keep reading
Five firms, more than the oilfield
The largest technology companies spent more than $400 billion of capex in 2025. The IEA expects another 75% jump in 2026. That is more than the world spends drilling oil and gas. Distribution of intelligence is a capital-stock event.
$400bn+
Big-tech capex, 2025 (IEA)
The power behind intelligence
Data centres used 485 TWh in 2025. The IEA’s base case still nearly doubles that by 2030 — and AI sites are growing three times as fast as the rest.
485 TWh
Data-centre electricity, 2025
The chip supercycle
Semiconductors did $792 billion in 2025. WSTS’s spring 2026 forecast put 2026 at $1.51 trillion — a 90% leap, almost all memory and AI.
$1.51 tn
WSTS spring-2026 forecast for 2026 sales