Very good analysis 👀
I'd like to add that by launching an
@Nvidia H100 in space,
@Starcloud_Inc1 started the new space computing industrial revolution now embraced by
@SpaceX,
@Google and others.
Datacenters in space are now the new frontier.
100 GW of Compute Above Earth: The Hardware Leap Hidden in Elon’s Claim 🔥🧠⚙️
Elon says SpaceX could launch ~100 GW of energy, in the form of high-orbit compute, in ~five years. That implies very specific performance requirements for chips, radiators, and solar arrays if launch cadence is going to stay reasonable.
In space, compute is constrained by a triangular bottleneck:
1️⃣ Power generation (solar)
2️⃣ Heat rejection (radiators)
3️⃣ Compute density (processing power per kg)
All three must rise together. If one lags, the other two stop contributing, and the launch count to hit 100 GW explodes.
Today’s generally assumed baseline:
☀️ Solar ~0.8 kW/kg (rigid LEO-class PV)
🌡 Radiators ~1 kW/kg
🧠 Compute ~0.1 kW/kg (typical GPU rack)
At these levels, you’d need ≈ 3000 Starship launches to deploy 100 GW of space compute!
Compute is the dominant lever: every doubling of compute W/kg roughly halves required launches. Musk’s AI5 → AI8 roadmap points to ~50-100 % annual gains, far faster than the ~25 % GPU trend. Tesla’s AI ASICs are built for efficiency and power density, rather than flexibility; the exact silicon needed for mass-constrained data centers.
Crucially, those gains demand new satellite architectures. To sustain cadence, future SpaceX platforms would need to move beyond Starlink-style LEO buses toward larger, optimised HEO compute vehicles with thin-film arrays > 1.5 kW/kg and light, high-temperature radiators > 2 kW/kg: performance levels cited only in advanced NASA STMD and ISNPS studies.
Given the 100 GW / ~5 yr claim, Elon is effectively telling us he believes compute, solar, and thermal systems will all hit near-frontier performance, and that Tesla’s chips will deliver the compute-per-kilogram leap needed to make it physically and economically possible.
It’s ambitious but plausible: if those PV and radiator systems reach projected “advanced” specs, SpaceX could feasibly deploy ~100 GW of orbital compute per year with ~300 Starship launches.
100 GW matters: it’s roughly the entire terrestrial data-center load expected by 2030. If SpaceX hits that number in orbit, the next AI scaling curve won’t be built on Earth.
Link below for the full breakdown 🧐