Moving AI inference to space to escape Earth's grid constraints sounds visionary until you remember you still have to get the answers back down.
The Summary
- Orbital Inc., a Los Angeles startup backed by A16z, plans to launch space-based data centers for AI inference starting with a 2027 prototype
- Their pitch: bypass Earth's strained power grids by tapping "free" solar energy in orbit, using a mesh constellation of up to 10,000 satellites with 100kW each
- The real constraint isn't power generation in space—it's latency, launch economics, and whether anyone actually needs inference results beamed from orbit
The Signal
Orbital wants to put GPUs in low Earth orbit because data centers are eating the grid. Fair premise. Power demand for AI inference is real, and utilities are scrambling. But space-based compute has a physics problem that "abundant solar energy" doesn't solve: the speed of light still matters when your user is asking a chatbot a question and waiting for an answer.
The technical plan is straightforward enough. Each satellite gets solar panels the size of a tennis court, radiative cooling panels of similar scale, and a fridge-sized GPU rack pulling 100 kilowatts. Launch them on SpaceX Falcon 9s—cheaper per kilogram than ever, but still not cheap—and build out to 10,000 satellites forming a distributed inference cloud. The 2027 prototype will validate whether GPUs can actually run reliably in orbit under those thermal and radiation conditions.
"There simply isn't enough capacity here [on Earth], and the only way is up."
Here's what Orbital's pitch glosses over: latency and link budget. Low Earth orbit is 340 to 1,200 kilometers up. Round-trip signal time to a satellite at 550km altitude—Starlink's ballpark—is roughly 7-8 milliseconds at the speed of light, before you account for ground station handoffs, atmospheric delays, or routing through multiple satellites in a mesh. Compare that to a terrestrial cloud region where inference latency is measured in tens of milliseconds end-to-end. For chatbots, agents, and real-time applications, adding orbital hops kills responsiveness.
Then there's bandwidth. Beaming training data up and inference results down requires serious RF infrastructure. Starlink's user terminals pull 100-200 watts and cost $600 retail for consumer broadband. Orbital's satellites would need bidirectional links capable of moving model weights and constant inference traffic—gigabits per second, minimum—which means either larger ground stations or denser satellite constellations to maintain coverage. Both cost money and power, Earth-side.
Launch economics don't favor compute-in-space yet:
- Falcon 9 launches cost ~$67 million, lifting 22,800kg to LEO. That's roughly $2,900/kg.
- A "fridge-sized" satellite at 100kg per unit plus solar arrays means you're spending ~$290,000 per satellite just on launch, before hardware, insurance, or operations.
- To reach 10,000 satellites, you're looking at $2.9 billion in launch costs alone—and that assumes no failures, no refurbishment, no replacement cycles.
The most coherent use case for orbital inference isn't replacing terrestrial data centers. It's edge inference for applications already in space or remote locations with no grid access—satellite imaging analysis, autonomous spacecraft decision-making, maritime or polar operations where latency to a distant cloud is already high. Orbital mentions matching "the solution with a problem," and that problem is likely narrower than "AI inference in general."
The Implication
If you're building agents that need to respond in real-time, orbital inference is a non-starter. But if you're processing sensor data from satellites, ships, or Antarctic research stations—places where the nearest GPU cluster is already thousands of kilometers and multiple network hops away—space-based inference starts to make sense. Watch whether Orbital's 2027 prototype focuses on consumer AI workloads or specialized edge cases. The former is a grid-escape fantasy. The latter might actually have a business model.
For now, the real signal is this: A16z is betting that launch costs will keep falling and that someone will figure out how to monetize orbital compute before the satellite replacement cycle eats all the margin. If they're right, the bottleneck for AI won't be energy—it'll be where you can physically put the chips.