Radiation-Hardened AI Chips Reach Orbit: EdgeCortix SAKURA-II Passes NASA, DIU, and Air Force Validation

60 TOPS at 8 watts. In space.

For decades, putting a computer in space meant accepting performance roughly equivalent to a 1990s desktop PC. BAE Systems' RAD750, the processor that powered the Mars rovers and a generation of deep-space probes, delivers about 0.2 TOPS at several hundred watts. It was designed for telemetry, not inference.

The gap between what terrestrial AI chips can do and what orbit can tolerate has been the single hardest engineering constraint on space-based AI. A satellite generates far more data than it can economically downlink. The only way to process that data in real time is to run inference onboard. But the same nanometer-scale transistors that make modern AI accelerators fast also make them vulnerable to radiation: high-energy protons and heavy ions that flip bits, latch circuits, and degrade silicon over time.

As we wrote in July, in-space semiconductor manufacturing is gaining policy backing as orbital fab tests succeed. But building a chip in orbit solves only half the problem. The chip itself also has to survive the environment it operates in.

The Three Validation Gates

EdgeCortix, a Tokyo-based fabless semiconductor company founded in 2019, put its SAKURA-II AI co-processor through three distinct qualification streams in parallel over twelve months. Its CEO, Dr Sakyasingha Dasgupta, described the process as "a single qualification cycle." The same hardware passed all three without redesign.

Gate 1: NASA Heavy Ion Testing

In January 2026, NASA's Electronic Parts and Packaging Program (NEPP) published a report confirming that SAKURA-II withstood heavy ion bombardment at the Texas A&M Cyclotron with no destructive events and relatively few transitory radiation effects. The testing covered conditions from low Earth orbit through geosynchronous orbit to lunar surface operations.

SAKURA-II is the second chip from EdgeCortix to pass this test. Its predecessor, SAKURA-I, was tested in 2024 and outperformed comparable AI processors in radiation resilience. The second-generation chip matched that performance. Each generation pushes transistor geometries smaller and more vulnerable, so matching the prior result is itself a data point.

Gate 2: DIU Success Memorandum

The Defense Innovation Unit awarded EdgeCortix a Success Memorandum after it met all prototype project objectives. The DIU evaluation included independent AI benchmarking and performance validation by Carnegie Mellon University's Software Engineering Institute. That was an external check that the chip's advertised inference speeds held up under real conditions, not just in the lab.

Gate 3: Airborne Flight Test

The chip was integrated into an Advanced Intelligent Gateway System, a prototype designed to provide resilient multi-domain battlefield connectivity, and flown aboard a KC-135 during a large-force exercise. The Air Force confirmed that it executed AI workloads in flight under operationally relevant conditions.

"The U.S. 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," said Lt Col Spencer Liedl, KC-135 Operational Test Director at the Air National Guard Air Force Reserve Command Test Center.

How It Works: The DNA Architecture

SAKURA-II is built around EdgeCortix's Dynamic Neural Accelerator (DNA), a run-time reconfigurable neural network IP core. The architecture uses digital in-memory computing, storing and processing data in SRAM arrays rather than shuttling information between separate memory and compute circuits. This limits data movement, which is the primary driver of power consumption in AI inference.

Key specifications, verified by independent testing:

The chip supports sparse computing, skipping zero-weight operations rather than treating them as computations, and arbitrary activation function approximation in hardware. Both features matter for space deployment, where every millijoule of energy must be justified against the thermal budget of a spacecraft that cannot rely on convection for cooling.

The run-time reconfigurability of the DNA core is a separate advantage for space applications. A satellite in low Earth orbit might run computer vision models during a pass over a target area, switch to signal processing during communications windows, and run anomaly detection on telemetry during silent periods. All within a single orbit. A rad-hard CPU can handle these sequentially but inefficiently. A fixed ASIC cannot handle them at all. SAKURA-II's reconfigurable fabric allows the same silicon to shift between workload types without the latency penalty of loading new bitstreams onto an FPGA.

EdgeCortix's Voyager SDK and Model Zoo provide pre-optimized model libraries that abstract away the hardware complexity for application developers. This is a software ecosystem the traditional rad-hard market has never invested in. The company's software-first approach, described as "an OS for Edge AI," means that qualification for space does not require rewriting inference pipelines from scratch.

In terrestrial edge AI deployments, a 10W chip dissipates heat through ambient air or a small fan. In a vacuum, the same chip requires a conducted path to a radiator surface, and every watt of dissipation demands roughly 0.5–1 kg of thermal management hardware depending on orbit altitude and attitude. An 8W chip that delivers 60 TOPS changes whether the thermal budget for AI inference fits inside a 12U CubeSat or requires a dedicated ESPA-class platform. For constellation operators launching 200+ satellites, the difference is multiplicative across the entire fleet.

Three Approaches to Space-Grade AI Compute

EdgeCortix's SAKURA-II belongs to the first category, but it does not own the space alone. Three competing approaches have emerged, each with distinct cost, performance, and reliability profiles.

Microchip Technology's PIC64-HPSC, a 10-core RISC-V processor in Moog's Cascade computer, represents the latest in rad-hard-by-design for general-purpose space computing. It targets a 100x gain over the RAD750. But it is a general-purpose CPU, not an AI accelerator. For neural network inference, EdgeCortix's SAKURA-II offers a different trade: less general-purpose flexibility, dramatically higher TOPS-per-watt for the specific workloads that matter in orbit — computer vision, signal processing, and on-board classification.

The shielding approach, championed by Atlanta-based Cosmic Shielding Corporation with its AFWERX-backed Plasteel nanocomposite, wraps commercial chips in protective enclosures. Starcloud launched an NVIDIA H100 to orbit inside a shielded container as proof of concept. The advantage is access to current-generation silicon. The disadvantage is mass: every kilogram of shielding is a kilogram not available for sensors, fuel, or payload.

Radiation-aware training, where fault injection teaches neural networks to tolerate corrupted weights, remains a research topic. Results in simulation are promising. Real-world validation in actual radiation environments is still sparse. Bits Atoms Brains noted in May 2026 that most deployed AI pipelines do not include confidence bounding, where the system refuses to act on inference results that fall outside expected statistical ranges.

The Market Gap EdgeCortix Fills

EdgeCortix raised more than $110 million in an oversubscribed Series B that closed 30% above target, with participation from TDK Ventures, Jane Street Global Trading, and CDIB Cross Border Innovation Fund II, alongside existing investors SBI Investment and Global Hands-On VC. Mizuho Bank extended a ¥1.5 billion unsecured credit facility in parallel.

The company also secured a ¥3 billion project from Japan's New Energy and Industrial Technology Development Organization (NEDO) to develop a next-generation low-power AI inference and learning chiplet platform called SAKURA-X.

Total disclosed funding, including equity, debt, and government grants, exceeds $140 million. That places EdgeCortix in a small group of well-capitalized edge AI chip startups alongside Axelera AI ($250M+), Hailo, and Untether AI. None of those competitors have pursued the space and defense qualification path EdgeCortix has completed.

Limits and Open Questions

Three caveats prevent this from being a clean victory lap.

First, SAKURA-II's 60 TOPS at 8W applies to INT8 precision. BF16 throughput is half that. For FP32, still required for certain scientific and training workloads, the effective throughput is lower. The chip is an inference accelerator, not a training platform.

Second, the Air Force flight test validated one specific integration: the Advanced Intelligent Gateway System. Generalizing from a KC-135 gateways demo to broad orbital deployment across multiple satellite bus architectures requires additional integration work that no customer has yet funded.

Third, Nvidia's Vera Rubin Space-1, announced at GTC 2026, targets 25x the AI performance of H100 for orbital data centers. It is not a competitor to SAKURA-II at the chip level. Space-1 is a full module designed for orbital data centers, not for sub-10W edge inference. But it signals that the largest AI hardware company on Earth now considers space a viable market. That changes the competitive landscape for every startup in this space.

There is also the question of procurement timelines. The defense and space markets operate on acquisition cycles measured in years, not quarters. A DIU Success Memorandum allows a program to transition from prototype to procurement without restarting the competition. It is not a production contract. EdgeCortix has cleared the technical gates. The commercial gates are still ahead: volume orders, integration into satellite bus architectures, and long-term supply agreements.

The counter-argument is that terrestrial cloud inference is improving fast enough to make orbital compute unnecessary. If Starlink-class low-latency satellite communications can downlink raw data to ground stations at 10 Gbps or more, and if terrestrial inference costs continue to fall, the argument for on-orbit processing weakens. But the physics of signal propagation from deep space, where round-trip latency to a lunar mission is 2.5 seconds and to Mars 8–40 minutes, means that for planetary exploration and cislunar operations, onboard AI is not optional. And in the defense domain, contested electromagnetic environments may deny downlink entirely, forcing autonomous decision-making at the edge regardless of terrestrial compute costs.

Why Now: The Orbital Compute Inflection

SAKURA-II's validation coincides with three structural changes that did not exist in commercial form two years ago.

First, satellite sensor resolution has outpaced downlink capacity. A single modern Earth-observation satellite generates 10–50 TB of imagery per day. Downlink budgets, constrained by antenna size, power, and regulatory spectrum, typically allow 1–5 TB per day under ideal conditions. The satellite collects more data than it can send home, and the gap widens every year.

Second, the cost of launch has fallen enough that the payload architecture itself can be reconsidered. At $7,000/kg on a Falcon 9, a 5 kg AI accelerator module costs $35,000 to put in orbit — within budget for a commercial small-sat operator. At $25,000/kg on the Shuttle-era pricing, the same module would have been $125,000, and the conversation would not have happened.

Third, defense doctrine is shifting toward distributed, autonomous edge processing. The Pentagon's Combined Joint All-Domain Command and Control (CJADC2) concept explicitly requires sensors and shooters to share processed intelligence without a terrestrial relay. That means AI inference on orbit, in the air, and at the tactical edge, all three domains SAKURA-II was tested in.

The market numbers reflect this. Global space-based AI compute, estimated at roughly $1.2 billion in 2025, is projected by multiple analysts to exceed $8 billion by 2030, driven by constellation-scale operators who need onboard processing to make their business models viable. Planet Labs, which operates the largest constellation of Earth-imaging satellites, is already building a GPU-native AI engine with NVIDIA. Kepler Communications is deploying Jetson Orin for intelligent data routing across its optical constellation.

These are commercial deployments, not research experiments. AI will run in orbit. The open question is whose silicon will run it.

Sources