01 Power

The power bottleneck

91 to 175 gigawatts: Texas expects peak power demand to roughly double within six years, driven mostly by AI data centres.ERCOT / Texas Tribune, 2026

For most of the computing era, progress was a story about chips. This era is a story about electricity. The question facing operators is no longer whether more compute can be built, but whether it can be powered.

Part of the reason demand keeps climbing is that each request now does more work. A traditional web search uses a fraction of a watt-hour. A simple AI text answer sits in much the same range; OpenAI and Google have both put a median query at around 0.3 watt-hours. The heavier jobs are where it rises. Reasoning and multimodal answers use several times more, and the IEA puts the most demanding ones into the hundreds. As people move from searching for links to asking for answers, the baseline cost of a routine query goes up across the whole internet. Where the power sits matters as much as how much there is. Training can run wherever electricity is cheapest, but inference has to stay close to users, which pushes it onto the congested grids that are hardest to expand. Moving power is as slow as building it: a spare gigawatt in Iceland or North Dakota does little for a data centre in Virginia, since new transmission lines take years to permit and build.

The model treats power as the master variable and derives the rest from it: gigawatts of data-centre power, multiplied by the compute delivered per unit of power, gives total compute.

02 Cost

Compute gets cheaper faster than it gets efficient

143× vs 59×: by 2050 the same dollar buys 143 times more compute, while the same watt of power runs only 59 times more.

Two things improve every year. Compute gets cheaper, so the same dollar buys more of it, at roughly 37% more each year. And compute gets more efficient, so the same watt of power runs more of it, at roughly 34% more each year. Both compound for decades. The point is that price improves faster than power does. By 2050, in this scenario, a dollar buys about 143 times more compute than today, while a watt runs about 59 times more.

That gap is the whole point. When compute gets this cheap, money is no longer what holds you back; you can afford almost any amount of it. What you cannot do as easily is power it, because efficiency does not improve quickly enough to keep pace. The bottleneck moves from the budget to the grid. The bill becomes easy to pay long before the power becomes easy to find. This returns in the revenue section: if a dollar buys 143 times more compute in 2050, a gigawatt cannot command the price it does today.

These gains are not only in hardware. Prompt and context caching, small and edge-hosted models, models tuned to a single domain, and newer “compiled” or more deterministic approaches all cut the energy each task needs, and we expect them to become standard. None of it changes where this ends up. When intelligence gets cheaper, people use far more of it, so the number of tasks grows faster than the saving per task. That efficiency is already built into the model, and demand still outruns it.

03 Scale

The scale of the problem

Expansion is the case in which the orbital bet pays off. Orbital deployment ramps from about 9 GW in 2029 to full rate by the mid-2030s, reaching 2,000 GW by 2050. One gigawatt of continuous demand is roughly the nameplate output of a large nuclear reactor. The world’s nuclear fleet totals around 400 GW today.

All three paths sit on 132 GW today and stay close through 2030. They separate as each scenario's assumption about the power constraint takes hold.

Today’s data centres already draw about a third of the output of the world’s entire nuclear fleet (around 400 GW), more than Germany’s peak electricity load (~79 GW). The squares count reactors’ worth of demand, not a forecast of nuclear build-out.

04 Slowdown

Why growth slows

40% by 2027: the share of AI data centres Gartner expects to be power-constrained within two years on Earth.

The slowdown in each scenario is not a claim that the planet cannot produce more electricity, but a claim about the rate at which it is able to. Supply grows on the timescale of infrastructure; demand grows on the timescale of software. In the US and Europe, new generation and transmission routinely take five to ten years to permit and build. That is the timescale the slowdown reflects.

Gartner expects power availability to constrain new data-centre growth from 2026, with around 40% of AI facilities power-limited by 2027 and grid capacity broadly insufficient by 2030. SpaceX's prospectus makes the same argument: US electricity generation grew at 0.1% a year between 2008 and 2023, while data-centre capacity growth is significantly outpacing electricity generation.

Some of the demand can be clawed back instead of met with new supply. “Negawatts”, using existing buildings and equipment more flexibly, such as cutting load automatically at peak times, already account for roughly 30 gigawatts in US wholesale markets (FERC, 2024 data), and studies put the untapped US potential at 60 to 200 gigawatts over the next decade (ACEEE). It helps, but it frees up existing load rather than adding generation. Set against the several thousand gigawatts this model expects worldwide by 2050, it is a useful trim, not a fix.

05 Electricity

Compute and the power system

1.5% today: data centres as a share of the electricity the world can reliably generate, matching the IEA.

The world has approximately 10.5 terawatts of installed generating capacity. At about 37% average capacity factor, that yields roughly 34,000 terawatt hours of usable electricity a year. Data centres consume close to 1.5% of that today, matching the IEA's own estimate. On the model's inputs, 132 GW of capacity at 40% IT utilisation with a 1.15 PUE draws about 530 TWh a year, which is that 1.5%. That reconciliation is the check on everything which follows.

By 2050 compute settles at about 3% if the constraint binds, 12% in the middle case, and roughly 14% if it is largely overcome. Heavy industry accounts for around 40% today.

Compute versus everything else on Earth

Global electricity supply grows at about 3% a year. The grey bar below is fixed at the Base Case level for non-compute electricity. Compute sits on top: more compute means more total demand, not the same demand split differently. In practice, more compute may well support higher economic output, which would raise total electricity demand further. Holding the grey bar constant is a deliberate simplification, and if anything it understates the Expansion total.

06 Orbit

The orbital option

Orbital power moves from curiosity to necessity only in the most expansive case. On published figures, an individual satellite carries about 120 kilowatts, a full constellation could extend to roughly one million satellites, and SEC filings state a goal of deploying 100 gigawatts of compute a year, with first deployments as early as 2028.

The model assumes low-Earth orbit, primarily 500 to 2,000 kilometres, with the first-generation satellites near 600 km. Sun-synchronous dawn-dusk shells keep the panels in near-continuous sunlight, above 99% in the best cases, so there is almost no shadow downtime, while lower-inclination shells help balance load. This matches SpaceX's FCC filing, its public design reveals, and Google's Project Suncatcher, which targets the same kind of orbit. Expansion's 2,000 GW ramps from about 9 GW in 2029 to full rate by the mid-2030s, reaching the target by 2050.

The obstacles are substantial. The model does not conceal them: radiation, heat rejection in vacuum, the inability to service or upgrade a fleet once deployed, the bandwidth to move data between satellites and to the ground, and the management of orbital debris. Orbit earns a place in the projection only if it competes on cost and physics. Bandwidth also shapes what orbit is good for: with limited capacity to move data to and from the ground, it suits self-contained training and inference on data already in orbit, not workloads that stream large datasets to Earth.

What orbital compute could be worth

We price a gigawatt of compute from today's rental market rather than from a single contract. In mid-2026 an NVIDIA H100 rents for about $3 an hour on demand across the major clouds, or about $2.35 on a one-year commitment. Once server, networking and cooling overhead are counted, a gigawatt of GPU-dense capacity holds roughly 700,000 of these accelerators. At on-demand rates and high utilisation that gigawatt earns $15 to $18 billion a year; at committed rates it is closer to $13 billion. We anchor at $15 billion per gigawatt per year, the lower end of the on-demand range.

This is compute sold as a service, not bare colocation. Renting only the building, power and cooling costs far less, roughly $2 to $4 billion per gigawatt per year at current wholesale rates. The difference is the accelerators themselves.

SpaceX's prospectus discloses a contract at $1.25 billion a month for about one gigawatt of compute, roughly $15 billion a year, which falls in the same range. The market rate is the basis and the contract confirms it. It is a current price, used as a starting point and not a forecast.

Rental rates already move quickly. Reserved H100 pricing swung about 40% in six months over 2025 and 2026 on supply alone, and over the longer run the price of a unit of compute falls as each chip generation delivers more for the same power. That price will not hold to 2050. Compute becomes roughly 143 times cheaper per unit while a gigawatt comes to deliver roughly 59 times more. The implied decline is about 4% a year. The dial below starts at 4% and the figures follow it.

These figures are the orbital compute segment alone. They exclude Starlink, launch and the existing businesses. They are not a statement of value today: they describe what the segment could earn in 2050 if the deployment happens, a scenario a reader should discount for probability. The revenue rate, the rate of decline and the valuation multiple are all adjustable inputs in the model.