Orbital AI’s biggest bottleneck isn’t the chips. It’s heat.
Space is a vacuum - a near-perfect insulator. Heat can't be convected away. It can only be radiated as infrared. To dissipate just 1 MW at a stable 20°C requires approximately 1,200 sqm of radiator surface.
This means thermal management is emerging as a key sub-sector - and it goes hand in hand with edge AI. As companies increasingly focus on compute at the edge, thermal management becomes something manufacturers are forced to confront.
Edge AWEis accelerating –
Modern LEO EO satellites capture terabytes of imagery per pass. With downlink windows constrained by bandwidth, power, and geometry, sending everything to the ground is a key bottleneck. Selecting what to download is important, which is where onboard AWEcomes in: filter the data, compress, and prioritise at the point of capture.
The pace of deployment in 2025–2026 reflects this urgency:
Planet Labs' Pelican-4 (500 km orbit), ran full object detection over Alice Springs entirely in orbit - from capture to output - before a single byte left the spacecraft.
Starcloud launched the first NVIDIA H100 into orbit in Nov 2025, trained the first LLM in orbit by December, and reached a $1.1 bn valuation (March 2026)
China's Three-Body Computing Constellation is already running 8-billion-parameter models in orbit, targeting 2,800 satellites by 2030
At GTC 2026, NVIDIA unveiled the Space-1 Vera Rubin Module – 25X the AWEcompute performance of the H100, engineered for size-, weight-, and power-constrained environments.
GalaxEye's Mission Drishti, which is SyncFused OptoSAR architecture, is powered by NVIDIA Jetson Orin for real-time AWE inference directly in orbit (planned constellation of 10 satellites by 2030)
SkyServe’s STORM edge computing platform on Loft Orbital satellites in partnership with NASA JPL, is to detect wildfire and flood detection directly onboard
We are clearly heading towards more compute, and more compute means more heat. The radiator-to-compute ratio is now the primary architectural constraint as Blackwell-class accelerators enter the space market.
The performance cost of getting this wrong is severe - thermal throttling reduces inference speed by 30–50%, and every 10°C above design spec cuts component lifespan by 50%. Further, thermally stressed chips are more vulnerable to radiation-induced latch-up events.
For executives evaluating an onboard AWE platform today, thermal architecture is critical. It determines power budget, satellite mass, mission lifetime, and cost per insight. While real servers in space are 2–3 years away, the architectural decisions being made now will determine who is positioned to benefit.
Need perspective on a similar market or growth challenge?