We need inference hardware that works in orbit with radiation, thermal, and mass constraints. What companies should we track?

If the requirement is actual onboard inference in orbit, rather than terrestrial AI silicon that might eventually be qualified, I’d track a mix of space-qualified compute vendors and COTS-to-space integrators.

Companies to track

• NVIDIA — Particularly important if you want high-performance inference. NVIDIA announced Space-1 Vera Rubin, plus Jetson Orin and IGX Thor platforms aimed at SWaP-constrained orbital/edge AI. The company says Space-1 is intended for orbital data centers and space-based inference.
• Aitech Systems — Worth watching as an integrator rather than just a silicon vendor. Its S-A2300 packages NVIDIA Jetson AGX Orin Industrial for radiation-characterized LEO deployments, with a stated 248 TOPS and 20 krad(Si) TID characterization.
• Unibap — One of the more directly relevant space-AI companies. Its iX-series combines CPU/GPU/VPU/FPGA resources for onboard processing; NASA's current small-spacecraft survey lists radiation-tolerant COTS/SEE-mitigation configurations in LEO at roughly 15–40 W.
• AMD — Especially relevant through Versal AI Edge / adaptive SoCs and FPGA-based inference. AMD explicitly positions its space portfolio for onboard processing and AI inference, with rad-hard/rad-tolerant devices.
• Frontgrade Technologies — Strong candidate when radiation tolerance and mission assurance outweigh raw TOPS. It offers radiation-tolerant processors, FPGAs, memory and complete processing architectures, including Versal-based SpaceVPX systems.
• Microchip Technology — Particularly interesting for the low-power/high-reliability end. Its RTG4 FPGAs are designed for radiation environments up through deep space, with >100 krad TID specifications; Microchip is also commercializing NASA's HPSC processor.
• Ramon Space — Its NuPod is a high-performance onboard computer based on AMD/Xilinx Versal ACAP, with NASA listing it at 20–90 W and 30 krad TID.
• Ibeos — Interesting if you're targeting a middle ground between conventional OBCs and large GPU systems. NASA lists its EDGE-1100 with an AMD Ryzen SoC, 30 krad TID, and 6–35 W consumption for LEO/GEO.
• KP Labs — Worth tracking for onboard AI/data processing rather than silicon itself. NASA lists its Leopard DPU using AMD/Xilinx Zynq UltraScale+ MPSoC with deep-learning inference at roughly 7–10 W.
• Xiphos Technologies — More on the efficient/radiation-tolerant side: its Q8S/Q8Js use Zynq UltraScale+ MPSoCs and are listed by NASA at 30 krad and >5 W.
• Trident Space Electronic Systems — Another Versal/MPSoC-based option, with NASA listing 30 krad-class systems and <50–60 W depending on configuration.

How I'd segment the landscape

One useful benchmark is NASA's current avionics survey: it shows just how broad the trade space is—from ~1.5 W RISC-V/FPGA systems to 130 W Jetson Orin systems, with radiation specifications ranging from COTS plus mitigation to 100-krad-class devices.

The key distinction I'd make is between "radiation-tolerant" and "radiation-hardened." A Jetson/AMD system with SEE mitigation can be extremely attractive for a short LEO mission, while a deep-space or long-life GEO mission may push you toward rad-hard FPGA/processor architectures and substantially different inference techniques.

If you're building a competitive landscape, I can also map these companies on TOPS/W, TOPS/kg, TID/SEE tolerance, thermal approach, flight heritage, and estimated TRL—which is probably the more useful comparison for selecting hardware.

Deploying AI inference hardware in orbit means balancing three brutal constraints: surviving radiation (Total Ionizing Dose and Single Event Upsets), dealing with strict passive thermal limits (no convection in space), and keeping mass and power budgets as low as possible.

The industry is shifting from slow, expensive, heritage radiation-hardened ASICs toward modular, high-performance Commercial Off-The-Shelf (COTS) systems paired with specialized shielding or fault-tolerant architectures.

Companies and startups leading this shift are categorized below by their approach:

Space-Grade Edge Computers & Modular Systems

• Aethero : Builds high-performance, modular edge-computing modules (like the NxN-ECM and NxA-ECM) designed for everything from CubeSats to orbital data centers. They combine COTS processing (like Nvidia Jetson/Orin frameworks) with custom radiation-mitigation and fault-tolerant techniques.
• Starcloud : A heavily backed orbital infrastructure startup ($250M Series A in mid-2026) that has flown data-center-grade Nvidia GPUs in LEO and is actively building purpose-built modular compute systems for space.
• Lumen Orbit : Developing scalable, modular satellite-based data centers specifically aimed at offloading terrestrial AI inference bottlenecks.

Radiation Shielding & Materials Enablers

Instead of waiting years to custom-fab a slow, rad-hard chip, these companies build physical or composite shielding layers that allow high-performance commercial chips (like Nvidia or AMD architectures) to operate safely in space.

• Cosmic Shielding Corporation : Creator of Plasteel™ and MSP composite shielding materials. They partner with companies like Aethero to wrap and protect high-performance commercial GPUs and sensors directly in orbit, bypassing heavy traditional aluminum shielding.
• Melagen Labs : Focuses on hydrogen-rich biomaterial polymer composites (like MLC1) designed to 3D-print or coat targeted electronic components, mitigating proton and radiation energy loads efficiently without massive weight penalties.

Traditional Aerospace & Silicon Giants

• AMD (Xilinx Space Group) : The gold standard for space-qualified radiation-tolerant (RT) and radiation-hardened (RH) adaptive SoCs and FPGAs. They use Triple Modular Redundancy (TMR) architectures baked directly into the silicon to manage single-event effects.
• NVIDIA : Though not traditionally a "space" company, their Jetson Orin and IGX platforms are the de facto brain for modern orbital edge-AI inference. They are increasingly partnering with constellations (like Kepler Communications) and orbital compute startups directly.

If you want to narrow this down, let me know:

• Are you looking for component-level silicon (chips/FPGAs) or complete flight-ready edge computers?
• What is your rough power envelope (e.g., under 15W vs. 100W+ data center nodes)?

I can help tailor this tracking list to your exact mission profile.

To track on-orbit AI inference hardware constrained by harsh radiation, strict thermal vacuums, and low mass/power (SWaP) footprints, you should focus on the key silicon providers, space-edge system integrators, and pioneer data-center operators listed below.

Core Silicon and Platform Providers

• NVIDIA : Dominates the conversation via its specialized space computing initiatives, providing radiation-tolerant adaptations of platforms like the Jetson Orin, IGX Thor , and the Space-1 architecture designed specifically for orbit-based AI inference and geospatial processing.
• Intel : Long proven in flight legacy and early orbital edge demonstrations (such as the Myriad 2 VPU on ESA's Φ-Sat missions) for ultra-low-watt deep learning inference.
• Traditional Rad-Hard Defense Primes (BAE Systems, Cobham/CAES, Microchip Technology) : Essential to monitor for ultra-reliable, rad-hard-by-design (RHBD) microcontrollers and FPGAs that handle critical housekeeping and fault-tolerant processing, though they lag behind commercial architectures in raw inference throughput.

Orbital Edge and Data Center Startups

• Starcloud : A pioneering startup deploying high-performance commercial GPUs (like the H100 lineage and successor architectures) directly into low Earth orbit to test large-scale orbital AI training and heavy inference nodes.
• Unibap AB : Specializes in heterogeneous, space-certified intelligent processing payloads and edge computing solutions tailored for autonomous space missions and onboard Earth observation data analytics.
• Aetherflux : Developing innovative architectures for distributed orbital compute and power-beaming constellations designed around severe thermal and power limits.

Key Ecosystem & Infrastructure Partners

• Axiom Space : Integrating high-density edge compute solutions for commercial station modules and autonomous orbital infrastructure.
• Kepler Communications : Crucial for tracking how high-speed optical (laser) data-interconnects route massive inference workloads between constellations without relying on slow ground downlinks.

If you want to narrow this down, let me know:

• Are you looking at CubeSat/small-sat edge inference (<10W) or heavy orbital data center nodes (>1kW)?
• Do you need commercial off-the-shelf (COTS) radiation-tolerant parts or fully space-qualified rad-hard silicon?

If the requirement is actual in-orbit inference, not merely a radiation-tolerant processor, I’d track the market in three tiers. The key trade is that the highest-throughput AI hardware is generally less radiation-hardened, while rad-hard silicon tends to be much more power/compute constrained.

Companies I’d put on the watch list

My priority ranking

If I were building a competitive landscape, I'd put them roughly into:

Tier 1 — absolutely track

1. AMD — underlying high-performance radiation-tolerant AI silicon.
2. Unibap — arguably the clearest commercial answer to "give me a space-qualified AI computer."
3. Aitech — strongest candidate for GPU-centric, relatively near-term LEO inference.
4. Microchip — important for rad-hard control/FPGA architectures and heterogeneous designs.

Tier 2 — important 5. Xiphos 6. Frontgrade 7. Trident 8. TelePIX

The interesting architectural split

There are really three approaches emerging:

• Rad-hard accelerator: AMD Versal XQR / Microchip FPGA → highest environmental assurance, but potentially less efficient than terrestrial AI accelerators.
• Radiation-tolerant COTS AI: NVIDIA Orin, AMD Ryzen/Radeon, Hailo, etc. → much better inference/W per dollar, but requires radiation characterization, shielding, redundancy and recovery strategies.
• Hybrid architecture: rad-hard controller + high-performance AI accelerator → probably the most interesting compromise for many LEO missions.

NASA's current small-spacecraft survey illustrates the spectrum: Unibap combines AMD/Intel compute in radiation-tolerant COTS systems, Xiphos uses AMD-Xilinx SoCs at only a few watts, while TelePIX combines a PolarFire SoC with an NVIDIA Orin module.

One company I'd watch especially closely is Unibap. Its latest iX20 is a good indicator of where the market is going: AMD Ryzen V8000 + Radeon GPU + four Hailo-8 AI accelerators, 64 GB ECC memory, and 7 W idle / 40–50 W typical power, in a package under 2 kg.

For a mass-constrained spacecraft, though, I'd pay more attention to TOPS/W, TOPS/kg, radiation mitigation architecture, and thermal rejection per sustained TOPS than headline AI TOPS. A 100-TOPS accelerator that can't continuously reject its heat in your spacecraft is less useful than a 20-TOPS device that can.

If you're evaluating this for a specific spacecraft, I can also build you a vendor matrix with TOPS, W, kg, dimensions, TID/SEE tolerance, orbit, flight heritage, and estimated TOPS/W and TOPS/kg.

Building AI inference hardware for orbit means balancing high compute density against brutal realities : ionizing radiation (causing Single Event Upsets), severe thermal dissipation limits (no convection in space), and strict mass/power budgets on the satellite bus.

The market has shifted rapidly from retrofitting terrestrial chips to purpose-built radiation-tolerant and space-qualified architectures. Here are the key companies and silicon vendors to track, broken down by their approach:

1. Established Heavyweights & Space-Qualified SoCs

• • AMD (Xilinx): Known for its radiation-tolerant and space-grade adaptive SoCs. AMD completed Class B qualification for its space-grade Versal devices (like the XQR Versal AI Core VC1902 Go to product viewer dialog for this item. and VE2302 Go to product viewer dialog for this item.), specifically targeting high-performance on-board machine learning and inference in harsh orbital environments without sacrificing reconfigurability.
• • NVIDIA: Pushing aggressively into orbital edge computing. While traditional high-end GPUs fail in raw space environments, NVIDIA’s edge platforms (like space-adapted modules leveraging Jetson Orin and IGX Thor architectures, alongside specialized enterprise space modules) focus on bringing energy-efficient, high-throughput computer vision and AI inference straight to constellations.

2. Emerging Edge AI Silicon & Radiation-Resilient Specialists

• •   - **EdgeCortix:** A fabless edge AI processor company making waves with heavy-ion and radiation testing. Their co-processor architectures (such as the SAKURA series) have undergone NASA Electronic Parts and Packaging Program (NEPP) testing, proving high energy efficiency and resilience to destructive single-event effects in space.[](https://www.edgecortix.com/en/press-releases/edgecortix-validates-radiation-resilient-sakura-ii-ai-accelerator-for-orbital-and-lunar-missions) [[1]](https://www.edgecortix.com/en/press-releases/edgecortix-validates-radiation-resilient-sakura-ii-ai-accelerator-for-orbital-and-lunar-missions)
    

3. Systems Integrators & Orbital Data Center Pioneers

• •   - **Orbit AI / Power Bank:** Actively deploying dedicated compute infrastructure and launching specialized edge/in-orbit data center pathfinder satellites to handle processing workloads natively in low-Earth orbit.[](https://enkiai.com/ai-market-intelligence/orbital-data-centers-2026-the-rush-to-space-for-ai/) [[1]](https://enkiai.com/ai-market-intelligence/orbital-data-centers-2026-the-rush-to-space-for-ai/)
    

• •   - **Axiom Space & Partners:** Developing broader orbital data center initiatives and testing edge architecture payloads on and around platforms like the International Space Station, driving requirements down to the physical mass/thermal envelope constraints.[](https://enkiai.com/ai-market-intelligence/orbital-data-centers-2026-the-rush-to-space-for-ai/) [[1]](https://enkiai.com/ai-market-intelligence/orbital-data-centers-2026-the-rush-to-space-for-ai/)
    

If you're targeting a specific application, let me know:

• Are you optimizing for Low Earth Orbit (LEO) constellations (high frequency, lower radiation per pass) vs. MEO/GEO/Lunar (severe total ionizing dose)?
• What is your approximate power budget (Watts) per compute node?

I can narrow down which architectural paradigm (Adaptive SoCs vs. ASICs vs. GPU edge modules) fits best.

If the requirement is actual onboard AI inference in orbit, with radiation tolerance, thermal limits, and tight SWaP, I’d track the market in three tiers rather than treating all “space AI” vendors alike.

Highest-priority companies

• Unibap — Probably the most directly relevant commercial benchmark. Its iX family combines CPU/GPU/VPU compute with AI inference, radiation-tolerant COTS plus SEE mitigation, and space flight heritage. NASA's small-spacecraft survey lists the iX10-100A at roughly 25–40 W and 10×10×6 cm.
• Ramon.Space — Especially interesting if you need radiation resilience + high-performance compute + storage rather than simply a rad-hard FPGA. Its NuPod is explicitly an AI/ML-oriented onboard compute platform with a proprietary radiation-hardening approach and SWaP optimization.
• AMD — A key technology supplier rather than necessarily the complete box you would fly. Its space portfolio now includes Versal XQR Adaptive SoCs and Kintex UltraScale XQR FPGAs, with ML/inference and high-throughput onboard processing explicitly supported.
• Frontgrade Technologies — Worth tracking for rad-hard processing architectures and space-qualified electronics, particularly if your design needs radiation assurance rather than merely mitigation around COTS silicon. Its portfolio spans onboard compute, AI/ML architectures, FPGA processing, storage and radiation-hardened microelectronics.
• Microchip Technology — Particularly important for rad-hard FPGA/SoC building blocks and NASA's High Performance Spaceflight Computing (HPSC). Its RTG4 FPGA is rated for >100 krad TID and is designed for severe radiation environments. HPSC is intended to provide substantially more capable spaceborne compute for future lunar/planetary missions.

Also worth watching

• NVIDIA — The interesting outlier. NVIDIA is now explicitly targeting orbital computing with Jetson Orin, IGX Thor and Space-1/Vera Rubin, emphasizing AI inference under SWaP constraints. The catch is that these platforms shouldn't automatically be treated as equivalent to traditional rad-hard space electronics; the radiation/mission-assurance architecture matters enormously.
• Xiphos Systems — Interesting for very small spacecraft where power and mass dominate. NASA lists its Q7S at only about 2 W and Q8S/Q8Js at >5 W, with 25–30 krad ratings. These aren't GPU-class AI accelerators, but they're useful reference points for the low-SWaP end of the design space.
• Trident Space Electronic Systems — Relevant if Versal-based processing is attractive. NASA lists its VDRT at <60 W and UDRT at <50 W, with radiation specifications around 30 krad.
• SkyLabs — Worth tracking for compact RISC-V/PolarFire-based onboard processing, particularly if your inference workload can be implemented efficiently without a conventional GPU.

My shortlist

If I were building a competitive landscape today, I'd put these five at the top of the diligence list:

1. Unibap — closest analogue to a commercial space AI computer.
2. Ramon.Space — strongest emphasis on resilient, high-performance space computing.
3. AMD — potentially the most important silicon platform for custom AI/FPGA architectures.
4. Microchip — radiation-hard compute/FPGA/SoC foundation.
5. NVIDIA — potentially transformative performance, but with a different radiation/qualification/SWaP trade space.

One particularly important distinction: “radiation tolerant” isn't a single property. For your application I'd compare vendors on TID, SEL/SEU/SEE behavior, mitigation strategy, thermal dissipation, W at your actual TOPS/TFLOPS workload, memory radiation tolerance, mass, and whether the stated radiation performance applies to the complete module or only the underlying silicon. NASA's current survey is useful because it puts several of these products into the same table with dimensions, power and radiation information.

If you're evaluating this for a specific mission, I can also turn this into a vendor matrix with TOPS/TFLOPS, W, kg, volume, TID, SEE tolerance, memory architecture, flight heritage, and estimated $/compute.

Building AI inference hardware for orbit means surviving a brutal triad of constraints: ionizing radiation (which causes bit-flips and latch-ups), extreme thermal swings in a vacuum (where convection doesn't exist and heat must be radiated away), and strict mass/power budgets on small satellites.

The industry is splitting into two approaches: rad-hardening traditional commercial silicon (COTS) with smart packaging/software, and building custom space-native AI accelerators or micro-data centers.

The key players and emerging companies worth tracking are organized below:

1. High-Performance COTS & Systems Integrators (The Pragmatic Route)

Instead of waiting a decade to custom-build rad-hard AI chips from scratch, many companies are adapting heavy-hitting commercial silicon by adding external shielding, redundant software, or specialized carrier boards.

• • NVIDIA : Formally pushing deeper into Space Computing initiatives , leveraging edge modules like the Jetson Orin and IGX Thor . While not natively rad-hardened out of the fab, they are widely flown with external mitigation and used by commercial imagery providers (like Planet Labs) for real-time orbital computer vision.
• • Aethero : A small-satellite systems builder specializing in making high-performance commercial architectures (like NVIDIA GPUs) viable in space. They collaborate closely with shielding innovators to bridge the gap between terrestrial performance and space survivability.
• • Orbits Edge : Focuses on high-performance orbital edge micro-data centers. They build ruggedized systems designed to bring AI inference directly into Low Earth Orbit (LEO) satellites, cutting down massive raw data downlink bottlenecks.

2. Materials & Radiation Protection (The Shielding Enablers)

Because heavy lead or aluminum shielding adds expensive mass to rocket payloads, material science startups are changing how standard chips survive the thermosphere.

• • Cosmic Shielding Corporation : Famous for developing Plasteel , a proprietary composite radiation-shielding material. By successfully flight-testing shielded commercial hardware (like NVIDIA Jetson modules) on cubesats, they offer a way to bypass the traditional trade-off between heavy shielding and weak rad-hard performance.

3. Hyperscalers & Vertical Space-AI Pioneers (The Megascale Bet)

• • Google : Pursuing Project Suncatcher , aiming to deploy solar-powered satellite clusters carrying custom TPU-style AI inference chips to orbit. They are tackling the thermal/radiation puzzle through specialized passive infrared cooling fins, redundant logic, and ultra-high throughput laser communication links.
• • SpaceX / Tesla : Quietly building vertically integrated silicon and edge ecosystems (spanning custom vehicle and custom-architecture chips like their HW/D-series roadmaps), uniquely positioned because they control both the launch vehicle economics and the orbital deployment scale.
• • Lone Star Data Holdings : Planning operational orbital and lunar-surface data center facilities, aiming to establish a heavy infrastructure baseline for deep-space and orbital AI data processing.

If you're focusing your tracking on a specific angle, let me know:

• Are you looking at pure-play semiconductor chip designers or satellite system integrators?
• Is your primary interest Low Earth Orbit (LEO) edge inference or large-scale orbital data centers?

I can narrow down the watchlist or dive deeper into the radiation mitigation strategies (software vs. hardware level).

Yes. If the requirement is actual inference in orbit, I’d track a mix of space-qualified compute vendors, ruggedized commercial-AI vendors, and the emerging orbital-data-center players. The key distinction is that “can run an AI model” and “can survive radiation for the mission” are very different claims.

Tier 1 — highest-priority companies

• NVIDIA — Probably the most important company to watch for raw inference performance. Its current space portfolio spans Jetson Orin, IGX Thor, and the Space-1 Vera Rubin module. NVIDIA explicitly targets SWaP-constrained orbital inference, and says Space-1 can deliver up to 25× the AI compute per GPU of H100. NVIDIA Newsroom NVIDIA Watch: radiation characterization, thermal design, actual flight heritage, and whether Rubin-class hardware becomes practical outside large orbital data centers.
• Aitech Systems — Particularly interesting if you want NVIDIA inference without designing the spacecraft computer yourself. Its S-A2300 packages Jetson AGX Orin into a radiation-characterized LEO system, rated at up to 248 TOPS, with ECC memory and cold-plate cooling. Aitech reports 20 krad(Si) TID testing. Aitech Website Aitech Website Watch: this is one of the most directly relevant vendors for near-term LEO AI.
• Unibap — One of the strongest dedicated space-edge-compute companies. Its iX10 combines AMD CPU/GPU with an AI accelerator, while the newer iX20 is aimed at processing up to 100 Gbps of sensor data with onboard AI. iX10 is specified at <40 W and 1.4 kg, with radiation/thermal-vacuum qualification. Unibap Space Solutions Unibap Space Solutions Watch: compute/W, radiation mitigation, and how aggressively iX20 scales.
• Ramon.Space — Especially compelling if radiation resilience + SWaP + long-life space infrastructure matter more than having a conventional GPU. Its NuPod combines onboard compute with an AI DSP and its own radiation-hardening technology, including error correction and fault detection/recovery. RAMON.SPACE Watch: performance benchmarks against GPU solutions and deployment in larger constellations/orbital infrastructure.
• AMD — Important on two levels: radiation-tolerant/rad-hard FPGAs for deterministic processing and increasingly capable adaptive-compute architectures for AI. AMD explicitly positions its space portfolio for AI inference and high-throughput onboard processing. AMD AMD Watch: Versal-based space systems and FPGA/AI acceleration where deterministic performance and radiation tolerance beat GPU flexibility.

Tier 2 — critical enabling / space-qualified compute

• Frontgrade Technologies — Less flashy than NVIDIA, but arguably more important for mission assurance. It supplies radiation-hardened processors, FPGAs, memory, power and complete mission-processing architectures. Frontgrade Frontgrade Watch: rad-hard compute + AI/ML architectures and how much acceleration they can deliver within a realistic spacecraft power budget.
• Microchip Technology — Particularly relevant for radiation-tolerant FPGA/SoC infrastructure around an AI accelerator. Its RTG4 FPGA is designed for LEO through deep-space environments, with hardened registers and radiation tolerance beyond 100 krad(Si). Microchip Watch: heterogeneous architectures where Microchip handles the radiation-hardened control/data plane while commercial acceleration handles inference.
• Teledyne Technologies / Teledyne e2v — Worth tracking for space processors and memory. Its QLS1046-Space SoC, for example, is being integrated into Aitech's SP1 SpaceVPX computer for onboard processing and AI-based algorithms.
• 3D PLUS — More of an enabling-component play: radiation-tolerant memory, computer cores, interfaces, protection and high-density SiPs. Its miniaturization and radiation expertise become increasingly valuable as mass and volume tighten.
• Xiphos Systems — Worth watching for very small spacecraft. NASA's current small-spacecraft avionics survey lists its Zynq-based Q7S/Q8S/Q8Js systems with roughly 2–5+ W-class power consumption and radiation specifications around 25–30 krad.

Tier 3 — watch the system architecture, not just the chip

Also track SpaceX, Axiom Space, Kepler Communications, Planet, Starcloud and Vast. They're important because they're becoming the customers/integrators that determine whether orbital compute becomes a real infrastructure market rather than a collection of demonstrations. NVIDIA currently identifies several of these organizations as users or partners for its space-computing platform.

My shortlist

If I were building an investment/technology watchlist today, I'd rank them:

The most interesting technical race isn't actually “NVIDIA vs. AMD.” It's commercial accelerator + radiation mitigation vs. genuinely radiation-hardened compute. NASA's current avionics survey already shows the spectrum: from ~2 W Zynq systems to 25–40 W heterogeneous AI computers, while vendors are increasingly using watchdogs, ECC, FPGA redundancy, shielding and software fault recovery to make higher-performance commercial silicon viable in orbit.

If your target is something like <10 kg, <100 W, LEO, 5–7 year lifetime, I can also narrow this to the 5–7 architectures/products I'd actually put through a radiation/thermal/SWaP trade study, including estimated TOPS/W and what I'd worry about for each.

Building high-performance AI inference in orbit requires balancing extreme radiation (SEEs and TID), strict thermal dissipation (conduction/radiation only in a vacuum), and tight mass/power constraints. The market is split between traditional radiation-hardened (rad-hard) silicon, commercial-off-the-shelf (COTS) edge architectures with specialized shielding, and adaptive system-on-chips.

The primary companies and technology providers to track are categorized below by their approach:

1. High-Performance Edge & Accelerated Compute Platforms

• NVIDIA: A major force in orbital compute through the NVIDIA Jetson Orin and IGX Thor platforms, alongside their space-focused modules (like the Space-1 Vera Rubin architecture). They enable high-density COTS-based AI inference in space, heavily utilized by emerging orbital data center providers.
• AMD / Xilinx: Renowned for their radiation-tolerant XQR Versal Adaptive SoCs (such as the VC1902 Go to product viewer dialog for this item. and VE2302 Class B Go to product viewer dialog for this item. qualified space-grade chips). Their integrated AI Engines (AIE-ML) are specifically optimized for machine learning inference workloads under space-qualified reliability standards.

2. Specialized Space Processors & Subsystems

• Spacechips: Known for products like the AI1 transponder , an onboard processor card using adaptive compute acceleration tailored explicitly to survive thermal and radiation limits while delivering substantial TOPS for real-time autonomy in orbit.
• Aethero: Specializes in small-satellite systems and modular architectures designed to host intensive edge computing payloads in cubesats and smallsats.

3. Materials & Thermal/Radiation Shielding Innovators

• Cosmic Shielding Corporation: Rather than relying solely on heavy traditional shielding or slow rad-hard silicon, they develop advanced nanomaterial shielding (like Plasteel ) that has been flight-tested with COTS hardware (like NVIDIA GPUs) to protect commercial AI chips from space radiation efficiently without massive weight penalties.

4. Orbital Data Center & Infrastructure Providers

• Starcloud & Sophia Space: Companies building modular, passively cooled infrastructure or dedicated orbital data centers designed to host high-compute AI and cloud workloads directly in LEO, solving the severe thermal and power density bottlenecks of space-based inference.

If you're targeting a specific orbit (like LEO vs. GEO ) or a particular SWaP (Size, Weight, and Power) envelope, I can help narrow down whether you should lean toward COTS + shielding or fully flight-qualified rad-hard ASICs.

If the requirement is AI inference in orbit under radiation, thermal, SWaP, and reliability constraints, I’d track the market in three layers: compute silicon, space-qualified/rugged compute platforms, and orbital infrastructure/operators.

Highest-priority companies to track

• NVIDIA — Probably the most important near-term commercial AI silicon supplier. Its space portfolio now explicitly includes Jetson Orin, IGX Thor, and the Space-1 Vera Rubin module, with NVIDIA positioning them for SWaP-constrained orbital inference. NVIDIA says Axiom Space, Kepler, Planet, Sophia Space, Starcloud and others are using its platforms. NVIDIA Newsroom NVIDIA Watch: radiation characterization, thermal envelope, actual flight heritage, and whether Vera Rubin-class hardware can be made sufficiently radiation-tolerant without killing its performance/W.
• Aitech Systems — One of the most directly relevant hardware companies. Its S-A2300 packages NVIDIA Jetson AGX Orin Industrial into a radiation-characterized LEO/NEO system, claiming up to 248 TOPS, while its older S-A1760 demonstrates the small-form-factor approach. Aitech Website Aitech Website Watch: TID/SEE performance, thermal design, mass, flight qualification and mission duration. This is much closer to an actual deployable spacecraft computer than simply buying an AI accelerator.
• AMD — Particularly interesting if you want FPGA-based inference rather than GPU-only architectures. AMD has radiation-hardened/tolerant FPGAs and explicitly supports onboard processing, AI inference and reconfigurable space systems. AMD Watch: Versal-based adaptive SoCs/FPGAs and the emerging space ecosystem around them. FPGAs remain attractive where deterministic processing, radiation tolerance and reconfiguration matter more than absolute TOPS.
• Frontgrade Technologies — Important as a space-qualified implementation partner around AMD/Versal technology. AMD identifies Frontgrade as providing Versal-based 3U SpaceVPX processing cards and the associated space-grade memory, networking, fault-tolerance and power ecosystem. AMD Watch: Actual radiation-qualified processing cards and how much AI acceleration they can deliver within a realistic spacecraft thermal/power budget.

Orbital-compute companies worth watching

• Axiom Space — Not primarily a chip company, but arguably one of the most important customers/integrators. Axiom has already demonstrated commercial compute on the ISS and deployed an orbital data-center prototype; its current ODC architecture explicitly targets AI/ML inference in orbit.
• Starcloud — Particularly interesting for the proposition that satellites become actual AI compute infrastructure rather than merely spacecraft with an AI payload. NVIDIA lists Starcloud among its space-computing partners.
• Kepler Communications — Relevant because orbital inference only becomes strategically interesting at scale when spacecraft can exchange processed data efficiently. Kepler is involved in the optical-link infrastructure supporting Axiom's orbital data-center nodes.
• Planet Labs — Worth tracking as a demanding end-user: large volumes of imagery create a strong economic case for inference before downlink. NVIDIA specifically cites Planet's use of accelerated computing for space applications.

The less-obvious names

I'd also put Microchip, Renesas, Infineon and Spacechips on the watchlist. They aren't necessarily the headline AI companies, but radiation-tolerant memory, power management, FPGA support, fault-tolerant interfaces and configuration memory can become the limiting factors in an orbital AI computer. AMD's space ecosystem explicitly identifies these suppliers.

My shortlist

If I were building a competitive/intelligence watchlist today, I'd rank them:

One important distinction: don't evaluate these companies on TOPS alone. For orbit, I'd build the comparison around TOPS/W, TOPS/kg, W/cm², TID, SEE/SEFI behavior, ECC/memory architecture, thermal rejection method, radiation shielding mass, operating temperature, boot/recovery behavior, and demonstrated flight hours. The current industry push toward orbital data centers makes thermal rejection and radiation protection particularly important—the physics may matter more than raw accelerator performance.

If you're evaluating this for an actual spacecraft architecture, I can also turn this into a 10–15 company landscape with columns for TOPS, watts, mass, radiation tolerance, process node, flight heritage, thermal approach, and likely 2026–2028 availability.