On-Orbit AI Inference for Satellite Autonomy

The case for running AI inference in space is not about latency. The round trip to geostationary orbit takes half a second. That is more than fast enough for most applications. The case is about what happens when a satellite captures more data in one pass than it can transmit in a week. And about what happens when the satellite has to make a decision before the next ground station comes into view.

Why satellites cannot send everything home

A single Earth observation satellite generates terabytes of imagery per orbit. The downlink budget is measured in megabytes per pass. Most of what a satellite sees is never seen by anyone.

This is the bottleneck that on-orbit AI inference addresses. Instead of transmitting all raw sensor data to a ground station, the satellite runs inference onboard and sends back only what matters: a change detection alert, a classified object, an anomaly signature. The downlink load drops by orders of magnitude compared to sending everything. The insight arrives in minutes instead of hours.

As we wrote in July, EdgeCortix SAKURA-II had already passed NASA heavy-ion testing for orbital radiation environments with zero destructive events. What changed in the months since is the scope: from component validation to operational deployment, from lab to airborne mission systems, from a single NASA report to a DIU procurement pathway.

Three ways to use AI on a satellite

On-orbit inference clusters into three use cases, each with its own performance requirements and revenue model. All three are already in production or late-stage prototyping.

Earth observation filtering. The largest addressable market. A satellite images the same swath of land on every pass. Most frames show no change. Running a change-detection model onboard means the satellite transmits only the frames where something happened: a ship in a new port, a construction site that was bare earth last week, a crop stress pattern that emerged overnight. Planet Labs already uses NVIDIA Jetson to convert satellite imagery into real-time intelligence in orbit. The bandwidth savings are a factor of 100 to 1,000 depending on the scene complexity.

Autonomous navigation and collision avoidance. Satellites in low Earth orbit share space with tens of thousands of debris objects and operational spacecraft. Today, conjunction alerts are computed on the ground and uploaded. The round trip takes time. On-orbit inference lets the satellite run its own collision-assessment model and execute an avoidance maneuver without waiting for a ground controller. The same capability applies to lunar and deep-space missions where light-speed lag makes ground-based control impractical for time-sensitive decisions.

Maritime and border awareness. Over ocean regions, satellites are the only persistent surveillance layer. Synthetic aperture radar captures ship tracks regardless of cloud cover. Onboard AI classifies vessels by size, heading, and behavior pattern: an anomalous stop in international waters, a rendezvous at sea, a fishing vessel crossing a marine protected area boundary. ESA's AI-eXpress mission already demonstrates this for maritime domain awareness, with three satellites processing radar and optical data in orbit and downlinking only the classified results. The program counts Eni, IBM, and Ubotica among its partners.

60 TOPS at 8 watts

SAKURA-II delivers 60 trillion operations per second at a typical power envelope of 8 watts. That is not a server-grade GPU in the traditional sense. It is an edge accelerator the size of an M.2 module that fits inside a satellite's avionics bay and runs on the same power budget as a sensor.

The numbers matter because power is the real constraint on orbit, not raw compute throughput. A typical small satellite generates a few hundred watts from its solar panels. Every watt spent on compute is a watt not spent on transmission, thermal management, or propulsion. Its efficiency means a satellite can run continuous inference without redesigning its power system.

Flight test over the United States

In June 2026, EdgeCortix integrated SAKURA-II into an Advanced Intelligent Gateway System and flew it aboard a US Air Force KC-135 during a large-force exercise. The system ran AI inference in flight, under operational conditions, as part of a prototype project backed by the Defense Innovation Unit. The DIU issued a Success Memorandum, its formal certification that the technology met all prototype objectives and is eligible for fast-track procurement across all US military branches.

The flight test validated something the laboratory environment could not: the platform works under vibration, thermal cycling, and the electromagnetic environment of a military aircraft. The same environment overlaps substantially with the conditions inside a satellite in low Earth orbit. The vibration spectrum during launch, the thermal swings as the spacecraft passes from sunlight to shade, the radiation levels at the higher end of the aircraft's operational ceiling. Carnegie Mellon University's Software Engineering Institute independently verified the AI performance benchmarks, adding a third-party credibility layer that matters for defense and space procurement.

The DIU Success Memorandum is not a contract award. It is a procurement signal. It tells program managers across the Department of Defense that the technology has been vetted, tested in operationally relevant conditions, and certified as meeting its stated performance. For EdgeCortix, a Tokyo-based fabless semiconductor company with 62 employees, it is the single most important milestone since the company was founded in 2019. It opens a channel into US government acquisitions that most non-US chipmakers never access.

Lt Col Spencer Liedl, KC-135 Operational Test Director at the Air National Guard Air Force Reserve Command Test Center, said the demonstration validated SAKURA-II's ability to execute AI workloads aboard an aircraft, delivering low-power, high-performance computing for defense and aerospace missions.

The US Air Force and EdgeCortix worked together to integrate SAKURA-II into a relevant mission system and fly it in a large force exercise, validating AI inference in flight with a tactically relevant application in operationally relevant scenarios.— Lt Col Spencer Liedl, Air National Guard AATC

The orbital compute market

EdgeCortix is not the only company pursuing on-orbit AI inference, but it is the only one with a radiation-tested, flight-validated edge accelerator at this power efficiency. The broader orbital compute market includes three distinct layers.

At the component level, NVIDIA's Jetson Orin and the newly announced Space-1 Vera Rubin GPU target orbital data centers. These are full-scale compute nodes in space, designed for sustained AI workloads. The Space-1 module delivers up to 25 times more AI compute per GPU than the previous generation, according to NVIDIA's published specifications. Companies like Sophia Space package Jetson into modular orbital platforms. Firefly Aerospace is integrating Jetson on its Elytra spacecraft for real-time lunar imagery processing through its Ocula service. Planet Labs already runs NVIDIA-powered inference on orbit for real-time intelligence.

At the constellation level, Orbital raised a $5 million pre-seed round led by a16z Speedrun in June 2026 to build purpose-built AI inference satellites. Its first mission, Pathfinder, is slated for 2027 on a Falcon 9 rideshare. The company envisions a network of 100,000 satellites delivering 10 GW of orbital compute. This ambition depends heavily on Starship-class launch economics.

At the institutional level, ESA's AI-eXpress program has launched three satellites with onboard AI processing for maritime domain awareness and disaster response. China's GuoXing Aerospace deployed Alibaba's Qwen3 large language model on its space computing satellite constellation in January 2026. It achieved end-to-end reasoning entirely in orbit: questions transmitted from Earth, processed onboard, results returned within two minutes. The company plans 2,800 computing satellites by 2035.

The pattern repeats across the industry. What started as a single NASA validation report has become a procurement pathway, a flight demo pipeline, and a growing list of companies treating orbital compute as addressable infrastructure rather than science payload. The question is no longer whether inference can run on a satellite. It is which business model scales first: adding AI to existing satellites, building dedicated compute satellites, or deploying edge accelerators on orbital platforms operated by someone else.

Why this shifts the economics

The terrestrial data center industry spends roughly 40% of its energy on cooling. In space, radiative cooling is free. Solar power in low Earth orbit is continuous for large parts of the year. The unit economics of orbital compute could undercut terrestrial inference for workloads that do not require human-scale latency, especially once launch costs fall.

SAKURA-II sits at the edge of this shift. It is not designed for orbital data centers. It is designed for the satellites already in orbit. The thousands of Earth observation, communications, and surveillance platforms currently underutilize their sensor capacity because they cannot process what they capture. Upgrading an existing satellite with an edge AI module is not a theoretical exercise: the same PCIe Gen 3 interface that connects SAKURA-II to a Raspberry Pi 5 can connect it to a satellite's onboard computer. The cost to add inference to a satellite already on orbit is the cost of the module plus integration, rather than the cost of a satellite launch.

According to Tracxn, EdgeCortix has raised $138 million across six rounds from investors including SBI Investment, Renesas, and GHOVC. The company is headquartered in Kawasaki, Japan and operates additional offices in Arlington, Virginia and Hyderabad, India. Its technology readiness level advanced with the DIU flight test, opening a path to production defense and space contracts that would have been inaccessible for a non-US chipmaker without the DIU certification.

Dr. Sakyasingha Dasgupta, founder and CEO, described the milestone as validation that energy-efficient edge AI platforms can enable trusted autonomy across next-generation defense and space systems. The phrase matters: trusted autonomy implies the system can be certified for decisions that a human would otherwise make, which is the regulatory gating question for autonomous satellite operations. Certification pathways for onboard AI decision-making remain undefined at most space agencies, which is both a risk and a first-mover advantage for whatever company demonstrates compliance first.

What still limits on-orbit inference

Three constraints will define how fast this market develops and which companies capture the most value.

Radiation. SAKURA-II passed heavy-ion testing with zero destructive events and relatively few transitory effects. But long-duration exposure in geosynchronous orbit or beyond the Van Allen belts will require additional validation. The NASA NEPP testing covered LEO through cislunar regimes, enough for most commercial constellations but not yet for interplanetary or deep-space missions. SAKURA-I and SAKURA-II are now the only AI accelerators with two published NASA test reports on the public record. This gives EdgeCortix a documentation advantage in government procurement cycles that competitors without radiation data cannot match.

Downlink bandwidth. On-orbit inference reduces downlink demand but does not eliminate it. Processed insights still need to reach the ground. Optical intersatellite links, laser communication between satellites, are improving throughput but remain expensive and unproven at constellation scale. The bottleneck shifts from raw data transmission to the latency requirements of specific applications. A maritime surveillance alert can tolerate a 15-minute delay. A tactical defense alert cannot.

Launch cost. The economics of orbital compute depend on Starship or equivalent heavy-lift vehicles achieving their projected $100 per kilogram to low Earth orbit. At current Falcon 9 pricing of roughly $2,600 per kilogram, the business case for a 100,000-satellite constellation does not close. EdgeCortix's approach upgrades existing satellites, avoiding the launch dependency entirely. A module swap inside an existing satellite bus costs a fraction of deploying a new constellation. Sophia Space, by contrast, is building dedicated orbital compute satellites and raised $22 million to do it. Both models may work. They address different parts of the market.

What terrestrial edge AI teaches us

The pattern playing out in orbit is the one terrestrial edge AI converged on over the past decade. Perception ran at the device first. Coordination across devices followed, as gateways became capable enough to maintain state. The cloud receded into what it does well: longer-horizon analytics and model lifecycle management.

The orbital case runs the same arc, only faster, because the constraint at the far edge is severe enough to force the architectural choices that took terrestrial deployments years to implement. Ambarella, which produces edge AI vision silicon for security cameras, published an analysis in June 2026 drawing the same parallel: the processors that succeed in space and the processors that succeed inside a ceiling-mounted camera have more in common with each other than either has with a data center GPU.

EdgeCortix's SAKURA-II runs on the same Dynamic Neural Accelerator architecture whether the target is a drone, a security gateway, a lunar rover, or a communications satellite. The company's $138 million in funding and its partnerships with Renesas and SBI Investment reflect a strategy that treats space as one deployment environment among many, not a separate market with separate hardware.