Issue 007 / 2026.05.26 / 9 min read / Compute & Energy

Earth's grid, the asteroid belt, and the maths in between.

Apollo strapped three men to a silo of liquid fuel. The current race straps three trillion parameters to a silicon wafer — and burns roughly Japan's worth of electricity to find out whether the maths is right about off-world rocks.

A working analysis of the energy bill behind the next decade of compute. Numbers, sources, and one usable takeaway at the bottom.

01 · Ignition

Apollo was an engineering problem. This isn't.

The Apollo Guidance Computer ran at roughly 40 kHz with 4 KB of RAM. It got us to the Moon because the maths was solvable on paper — Newtonian, two-body, slide-rule territory. Three days, three astronauts, a known orbit.

The blockers between us and the asteroid belt aren't mechanical. They're mathematical: N-body trajectory optimisation across decades, autonomous proximity operations on rotating irregular bodies, spectral inversion against unknown mineral mixes. None of that fits on a slide rule.

One next-generation training cluster now does more arithmetic per second than every computer on Earth combined did in 2010. That is the actual story.

Figure 01 · Relative compute, log scale

Apollo Guidance Computer vs. a 2026 training cluster

One is a four-function calculator with extra steps. The other is a synthetic cortex. The axis is logarithmic because linear made the bar invisible.

02 · The Toll

The cooling water is the tell.

Humanity is constructing a planetary-scale intelligence to solve global problems, and powering it requires roughly the energy output of a G7 nation. The grid is groaning under matrix multiplications. None of this is a metaphor.

Projected · 2027
1,250TWh

Global AI data-centre demand. Within a rounding error of Japan's entire annual electricity consumption.

Source · IEA Energy & AI, 2024
Growth · since 2023
4.5×

Demand multiplier in twenty-four months. A curve that makes the 2021 crypto-mining panic look fiscally restrained.

Source · Goldman Sachs Carbonomics, 2025
Overhead · hyperscale
85%

Of net energy draw goes to cooling, not compute. We are boiling rivers to stop silicon melting itself.

Source · Uptime Institute Global Survey 2025
Figure 02 · TWh, 2018–2030

The trajectory most utilities are quietly planning around

If the curve holds, by 2030 a single global compute layer will consume more electricity than the United Kingdom, France and Italy combined.

Hyperscalers stopped being software companies in late 2024. They are sovereign energy buyers now, with a side hustle in matrix multiplication.
Figure 03 · USD billions, committed 2024–2026

Where the capex actually went

Solar and storage still dominates, but the SMR line is the one to watch — it didn't exist on a corporate balance sheet two years ago.

03 · Reactors

Three companies, one quiet nuclear arms race.

Municipal grids cannot deliver gigawatt-scale baseload on the timeline frontier-model training requires. Hyperscalers have responded by signing power-purchase agreements that would, ten years ago, have read as parody.

Microsoft has reopened Three Mile Island. Amazon has acquired a 960 MW data-centre campus directly behind a nuclear plant in Pennsylvania. Google has signed for 500 MW of Small Modular Reactor capacity from Kairos. None of these are software contracts.

The thing to track is not which lab publishes the next benchmark. It is which one secures stable baseload first.

04 · The Payoff

Why we'd burn a country to mine a rock.

The case for the energy bill rests on a single bet: that compute unlocks resource economics that physical engineering cannot. The maths is plausible. The timeline is the open question.

The minerals required to build the next decade of AI hardware — platinum-group metals, cobalt, neodymium, rhodium — are both terrestrially scarce and concentrated in politically inconvenient jurisdictions. Two near-Earth asteroids in our catalogue contain more accessible platinum than has been mined in the entirety of human history.

The blocker has always been the maths: optimal-transfer trajectories, autonomous landing on a rotating irregular body, in-situ spectral analysis. Solving these took decades of postdoc time and produced a single Hayabusa mission. A frontier model now redoes that targeting analysis in an afternoon, for a portfolio.

Figure 04 · Near-Earth asteroid mineral value vs. orbital accessibility

High value, low Delta-v: the targeting problem AI has already done

Bubble size encodes estimated mass. The lower-left quadrant is what an autonomous mining programme actually goes after first. Psyche is the outlier — enormous value, awful Delta-v.

Definition

Delta-v /ˈdɛltə viː/

The change in velocity, in km/s, required to move a spacecraft from one orbit to another. A useful proxy for "how much fuel — and therefore money — does this asteroid actually cost." Lower is better; it's the only metric mining economics genuinely care about.

Definition

N-body problem /ɛn ˈbɒdi/

Predicting the motion of three or more gravitationally interacting bodies. Famously unsolvable in closed form since Newton. The reason every Apollo trajectory was, technically, an approximation. The reason modern AI is so well-suited to long-horizon mission planning.

05 · The Loop

The argument runs on a feedback loop.

Stripped of its narrative furniture, the thesis is a four-step cycle. Each step depends on the previous one resolving. Each step also collapses the whole structure if it doesn't.

Step 01 Terrestrial energy is diverted, at scale, into training and inference.
Step 02 Compute solves astrodynamics, autonomous landing and spectral inversion.
Step 03 Autonomous fleets extract platinum-group metals from near-Earth bodies.
Step 04 Recovered minerals feed the next hardware generation. Loop closes.
Where it breaks

Step 03 is the open question.

The maths in Step 02 is, broadly, in hand. Autonomous extraction at commercial scale is not. The earliest credible delivery of platinum-group metals to Earth orbit sits in the early 2030s, contingent on three private launch programmes hitting nominal cadence and at least one regulatory framework that does not yet exist.

If Step 03 slips a decade, the energy bill from Step 01 still arrives on schedule. That is the actual risk.

Tactical Takeaway

Watch the energy contracts, not the model releases.

The next five years of AI capability will be decided by who secures stable baseload power — not whose benchmark wins this week. If you work in data, infrastructure or strategy, the more interesting filing is your power provider's grid-connect queue. Read it on a Tuesday.