data-centre power today, about 132 reactors
more compute by 2050, Base Case
of world electricity by 2050
orbital, if Musk is right
01 Power
The power bottleneck
For most of the computing era, progress was a story about chips. This era is a story about electricity. The mismatch: chips and data centres scale in months; power plants, grids and transformers scale in years. Once existing grid slack is absorbed, growth proceeds only as fast as new power can be built. Terrestrial build-out is near-identical across scenarios — the difference that matters is orbital.
Total data-centre power by scenario
Gigawatts, terrestrial + orbital, 2026–2050 · linear scale
The same buildout, counted in nuclear reactors
Each square ≈ one large nuclear reactor’s output (1 GW)
Today’s data centres already draw about a third of the output of the world’s entire nuclear fleet (~440 reactors, ~420 GW) — more than Germany’s peak electricity load (~79 GW). The squares count reactors’ worth of demand, not a forecast of nuclear build-out.
A dollar buys 37% more compute every year. A watt manages 34%. Epoch AI — and that gap is why money stops being the binding constraint.
Compute gets cheaper faster than it gets efficient
Cost falls faster than efficiency rises, and the gap widens across the whole period — capital buys ever more compute per unit spent, while physics does not concede at the same rate. Money stops being the binding constraint; what stays scarce is power. It also means a gigawatt cannot command today's price in 2050 — that returns as the decline dial in section 03.
Terrestrial vs orbital
Slowdown timing per Gartner: power limits new data-centre growth from 2026; 40% of AI data centres power-constrained by 2027; grid broadly insufficient by 2030. Gartner forecasts only to 2027 — the slowdown continuing to 2050 is our assumption, not their finding.
02 Energy reality
Compute and the power system
Global electricity supply grows roughly 3% a year (IEA), so the pie roughly doubles by 2050. Compute takes a growing share of a growing pie — it does not reduce what everyone else gets. Households, industry and transport keep growing in absolute terms in every scenario. Terrestrial only; orbital power is generated off-grid.
World electricity roughly doubles by 2050. Compute's slice: about 13%. Base scenario · all heavy industry combined today: ~40%.
Share of Earth's usable power
Data-centre electricity ÷ usable global generation · terrestrial only
Does compute take power from other consumers?
For scale: all heavy industry combined is roughly 40% of global electricity today. From 2028 the modelled data-centre electricity runs above the IEA's published path — deliberate: IEA counts announced projects; the model builds what power-constrained demand asks for. If the ~3%/yr of new global generation does not get built, compute and other consumers compete directly — that supply-speed limit is the whole argument for orbital.
03 The orbital option
What orbital compute could be worth
Disclosure: Redstone is an investor in SpaceX.$1.25bn a month — what Anthropic pays today for one gigawatt in orbit. SpaceX S-1 — a filed contract price, not an estimate.
The contract covers roughly 1.0 GW across COLOSSUS and COLOSSUS II (S-1) — about $15bn per gigawatt per year. For scale: close to Spotify’s entire annual revenue (€17.2bn in 2025) — from a single gigawatt, every year. Apply that price to each scenario's orbital build and dial in how fast it erodes.
Don't believe the price? Move the dials — everything below recomputes.
Orbital compute revenue
$ trillions per year · whole orbital market, SpaceX leading · linear scale
The counterweight
Morningstar puts 7% on the Moonshot — and 43% on "No Go".
- SpaceX's own S-1: orbital AI compute involves "unproven technologies or technologies that do not exist" and "may not achieve commercial viability"; the 100 GW/yr timeline "may be difficult or impossible to determine".
- Morningstar values SpaceX's entire AI segment at ≈$170bn — "more akin to the value of a call option" — assigning ≈7% probability to the Moonshot case and ≈43% to "No Go".
- Today SpaceX pays Google ~$920m/month for ~110,000 terrestrial GPUs while its orbital compute revenue is zero. Everything above is prospective.
- A gigawatt cannot sell for today's price in 2050 while compute gets ~37%/yr cheaper — hence the decline dial. At 0% those two claims contradict each other.
04 Founder simulator
Simulate your start-up's impact on the world
A projection of this kind can feel remote from the decisions any single company makes, so the model ends by reversing the question: pick the area you are building in, set the levers, and we send you how far it moves the system. Twelve areas, three levers: more efficient (more compute per watt), cheaper (more compute per dollar), more capacity (more overall compute).
Don't believe one start-up can move the curve? Choose your area, add your link, then set the sliders.
Stays on this page until you request your results below.
We send your results here — nothing else.
The results are not shown on this page — they go to your inbox: the chart with your company's band in it, the year-by-year data behind it (CSV), your headline figures, and a link that reopens the simulator exactly as you set it.
World compute, with and without this founder
Billion PFLOPS (1 PFLOPS ≈ one H100 GPU) · Base scenario as fixed reference, so results stay stable across scenarios
The tool conveys scale rather than precision. Two companies pursuing the same lever with the same assumptions see the same figure, because it models the category rather than the firm. Its value lies in distinguishing work that moves the curve modestly from work that moves it substantially. Default improvements are the team's working estimates.
05 Sources & method
Every number, accounted for
Fixed inputs
Held constant across scenarios; from the methodology table in the white paper
| Input | Value | Status |
|---|---|---|
| Data-centre power, 2026 | 132 GW | Sourced — Gartner |
| Compute per MW, 2026 | 114 PFLOPS/MW | Calibrated to Epoch (≈15M H100-equivalents) |
| Cost improvement, start | 37% a year | Sourced — Epoch (30–45%) |
| PUE | 1.15 | Sourced — hyperscale range 1.1–1.2 |
| Utilisation / load factor | 40% | Chosen — calibrated to IEA |
| World generating capacity | 9.6 TW | Sourced — IRENA |
| Fleet capacity factor | 37% | Sourced — Ember and IRENA |
| Global electricity supply growth | ~3% a year | Sourced — IEA |
Method: power first. Terrestrial capacity grows at Gartner's observed rate, then decays once the grid bites; compute per MW follows Epoch's efficiency curve with the same-shaped slowdown; orbital capacity ramps linearly from its start year to its 2050 endpoint; electricity = power × utilisation × PUE. Orbital revenue = orbital GW × a declining price per GW anchored to the Anthropic contract. Terrestrial electricity sits somewhat above the IEA's central projection from 2030 onwards, reflecting a higher assumed load factor — the IEA path is a conservative reference, not an upper bound. Morningstar figures accessed via secondary reporting — verify before publishing.
06 Tell us where we can improve
This model is a working instrument
The assumptions behind it are stated openly because we don't expect everything to be right, and we would rather find out from people closer to the work than we are. If you have better data, a different reading of the assumptions, or you are building in one of the twelve areas, we would like to hear from you.