Why Orbital AI Data Centers Face Massive Engineering Hurdles

I've seen the pitch before — uninterrupted solar, vacuum cooling, no local permitting fights. I've also debugged enough 3 a.m. pages to know that physics doesn't take feedback.

Moving compute to orbit doesn't fix the bottleneck that wakes me up. It usually creates new ones.

The orbital pitch, stress-tested

The appeal is straightforward on paper. Solar power is uninterrupted above the atmosphere. Cooling radiates into hard vacuum. No land permits, no community pushback on water usage.

The appeal ends the moment you sit down with the spec sheet.

Round-trip latency to LEO is a physics floor — not a tunable knob in anyone's control plane. For inference serving, that's borderline. For training synchronization across orbital clusters, it's a non-starter without aggressive co-location on the same pass. Bandwidth to ground stations is limited, contended, and weather-dependent. Hardware lives in a radiation environment that degrades memory and storage faster than any terrestrial failure mode I've debugged. Maintenance means another rocket, and another budget line.

Throughput, in theory, is fine. Reliability is the part I can't sell to a CTO.

What the ground-side numbers actually say

Meanwhile, back in the data center, capital is already committed to terrestrial builds. A recent Simply Wall St screen of AI infrastructure names surfaced Western Digital and Digital Realty Trust — players sitting on revenue bases in the billions and backlogs tied directly to AI workload demand. Western Digital is reporting capacity effectively pre-sold well into the next several years as hyperscale customers lock in high-capacity drives. Digital Realty Trust carries a record US$1.9b leasing backlog and rising funds-from-operations guidance tied to power-rich campuses under development.

That isn't speculation money. That's signed leases and pre-sold hardware. The ground game is funded. The orbital game is a pitch deck.

My read: deploy or discard

I'm not saying don't research it. Long-horizon bets deserve napkin math, and orbital compute has genuine appeal for batch workloads that don't care about round-trip latency. But if your roadmap has an orbital compute line item before you have a working multi-region failover on the ground, your priorities are inverted.

Discard the orbital DC idea for production training and real-time inference today. Revisit when someone publishes a latency, MTBF, and TCO comparison I can actually run against a colocation contract. Until then, the frontier is the same boring, reliable aisle it was last quarter — and that's exactly where the capital is going.