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

Building AI inference hardware for orbit means navigating brutal constraints: severe Size, Weight, and Power (SWaP) limits, passive thermal dissipation in a vacuum, and cumulative ionizing radiation that causes single-event upsets (SEUs).

The market has shifted rapidly from using years-old radiation-hardened microcontrollers to adapting high-performance commercial architectures with fault-tolerant wrappers, FPGAs, and dedicated space modules.

The key players and technology providers to track are organized below by category:

Major Silicon & Hardware Giants

• NVIDIA – Launched formal NVIDIA Space Computing platforms, bringing modules like the Jetson Orin Go to product viewer dialog for this item., IGX Thor Go to product viewer dialog for this item. , and the space-grade Space-1 Vera Rubin Module directly to orbit for high-throughput edge inference and orbital data centers.
• Intel – Introduced space-grade offerings like the Starfire processor line, which adapts commercial system-on-chip (SoC) architectures with integrated neural processing and built-in radiation mitigation features designed for decade-long missions.
• Microchip Technology – A legacy heavyweight in aerospace silicon, continually releasing high-reliability and radiation-tolerant timing, processing, and control components (such as their Space CSAC line) critical for resilient satellite buses.

Systems Integrators & Edge Platforms

• Voyager Technologies – Developed Space Edge , a modular, radiation-tolerant edge architecture that packages components like NVIDIA Jetson modules alongside rad-hard monitors and FPGAs into SOSA-aligned single-board computers. It allows satellite builders to "compose rather than commission" custom compute.
• Hewlett Packard Enterprise (HPE) – Known for pushing commercial-off-the-shelf (COTS) edge supercomputing into orbit via their Spaceborne Computer initiatives on the International Space Station, proving software-based mitigation for radiation.

Emerging Orbital Cloud & Data Center Startups

• Starcloud – An orbital infrastructure startup closely tied to the ecosystem, focusing on heavy in-space compute and public-cloud architectures partnered with energy providers like Crusoe.
• Aetherflux & Planet Labs – While operating primarily as constellation and satellite data operators, they are key early adopters actively integrating high-performance edge inference to process Earth-observation data locally before downlinking.

Are you looking at this from a component supplier level (chips, packaging, IP), an avionics/board design level , or looking to deploy full orbital data center infrastructure ? Let me know so I can tailor the technical focus.

If the requirement is actual inference in orbit, with radiation, thermal, SWaP, and flight-heritage constraints, I’d track the market in tiers rather than just looking at conventional AI-chip companies.

Highest-priority companies

• AMD — probably the most important silicon vendor to track. Its radiation-tolerant Versal XQR AI Core combines programmable logic, CPU, DSP and AI engines, explicitly targeting low-latency inference in space. Its Kintex UltraScale XQR is also aimed at ML/on-board processing.
• Frontgrade Technologies — one of the strongest candidates if radiation assurance and mission qualification matter more than raw TOPS. It builds complete space processing systems around rad-hard/rad-tolerant FPGAs, processors, memory and fault-tolerant infrastructure.
• RAMON.SPACE — particularly interesting for high-performance AI/ML compute designed specifically for space. Its architecture includes an in-house ML/AI processor, radiation-error correction, FDIR and operation from LEO through deep space.
• NVIDIA — the obvious company to watch for the high-end/COTS-derived end of orbital AI. Jetson Orin targets compact onboard inference, while NVIDIA is now pushing IGX Thor and its Space-1 Vera Rubin platform for substantially larger orbital compute.
• KP Labs — worth watching for the integrated flight-computer/DPU layer rather than silicon itself. NASA's 2026 survey lists its Leopard DPU at up to ~10 W for deep-learning inference, using a Zynq UltraScale+ MPSoC with SEE mitigation.
• NOVI — interesting for compact, flight-proven AI compute. Its SP240 is AMD Versal-based, radiation tolerant and designed for CubeSats through larger payloads, with a stated 100-TOPS ML capability.

Also put on the watchlist

• Ibeos — EDGE-1100/Genie uses AMD Ryzen SoCs; NASA lists it at 6–35 W and with 30-krad TID / SEE characterization. This is interesting if you're willing to trade some radiation margin for much higher conventional compute.
• Innoflight — its CFC-510P uses an AMD Ryzen GPGPU, while CFC-600P uses Versal AI Edge; NASA lists these at roughly 12–40 W and 10–70 W respectively.
• EnduroSat — its GPC uses Jetson Orin and is an interesting example of pushing relatively powerful commercial AI compute into a spacecraft form factor. NASA lists ~130 W peak, so thermal/power architecture becomes a significant part of the trade.
• GomSpace — NanoMind HP MK3 provides a lower-power Zynq-based alternative, useful as a benchmark for the conventional onboard-compute end of the market.
• Resilient Computing — interesting if fault tolerance through redundancy/reconfiguration is more important than accelerator density. NASA lists its RISC-V/Artix-7 RadPC architecture at ~1.5 W.
• SkyLabs — very interesting for extremely constrained platforms: NASA lists its RISC-V/PolarFire-based NanoHPC OBC at ~3 W.

The strategic split I'd use

One important observation

Don't benchmark this market on TOPS alone. For orbit, I'd normalize candidates around:

inference throughput / watt / kg / cm³ × radiation tolerance × thermal rejection capability × flight heritage.

NASA's current avionics survey makes the trade quite stark: examples range from ~1.5–3 W RISC-V/FPGA systems to 130 W peak Jetson Orin, with radiation assurance varying dramatically.

And the industry is now bifurcating: AMD/Frontgrade/RAMON.SPACE are pursuing increasingly space-specific resilient compute, while NVIDIA is pushing terrestrial-class accelerated computing into orbit. NVIDIA explicitly positions Orin/Thor/Space-1 across this spectrum.

If you're evaluating this for a specific spacecraft, I can also build a 10–15 company landscape scored on TOPS/W, W, mass, volume, TID/SEE, thermal architecture, orbit, flight heritage, and expected availability.

Designing inference hardware for orbit means balancing brutal trade-offs: you need high teraflops per watt, but you are tightly restricted by mass budgets , passive-only or specialized radiative thermal dissipation in a vacuum, and single-event upsets (SEUs) caused by cosmic radiation.

The ecosystem is splitting into legacy defense primes building ultra-reliable, slow-to-evolve silicon, and aggressive venture-backed startups trying to fly commercial high-performance accelerators (like NVIDIA GPUs or custom TPUs) using clever packaging, shielding, or architectural fault tolerance.

Here are the key companies and entities you should track across the stack:

1. Commercial Orbital Data Center & Accelerator Pioneers

• Starcloud : A high-profile startup focusing explicitly on orbital AI infrastructure. They made waves by launching proof-of-concept hardware (including testing terrestrial-grade GPUs like the NVIDIA H100 and working on specialized modules like the Space-1 Vera Rubin Module) to handle real inference and fine-tuning in orbit. Track them for end-to-end space data center designs, including deployable low-mass radiative cooling systems.
• Voyager Technologies : Known for initiatives like Space Edge , they build active, flying processing architectures designed to bring modern, high-load AI computing substrates directly into space without requiring custom-redesigned computers for every single bus or mission.
• U.S. Space Force / Defense Innovation Unit (DIU) ecosystem : Programs like Victus Haze are pushing rapid on-orbit responsiveness. Keep an eye on commercial prime integrations—such as deployments utilizing AMD/Ryzen-based edge architectures (e.g., the Genie space edge processor achieving flight heritage)—which prove high-performance commercial silicon can survive LEO if packaged correctly.

2. Hyperscalers & Internal Silicon Initiatives

• Google (Project Suncatcher / Custom TPUs) : Google has been actively developing and testing space-compute concepts, including radiation-tolerant/hardened variants of its Tensor Processing Units (TPUs). Because they control the software stack and the silicon architecture, they are a major entity to watch regarding how they handle the surrounding bottlenecks like high-bandwidth memory (HBM) in high-radiation fields.

3. Traditional Radiation-Hardened (Rad-Hard) Semiconductor Giants

If your constraints demand components that are radiation-resistant from the ground up rather than relying purely on software redundancy or heavy external shielding, the trusted defense-grade vendors are mandatory tracking:

• Microchip Technology : A dominant force in space-grade microcontrollers and radiation-tolerant power management devices tailored for smaller satellite constellations.
• BAE Systems : Famous for its space-processor lineage (like the RAD5545 system-on-chip). They use "radiation-hardening-by-design" to ensure extreme reliability where a single flipped bit could mean mission failure.
• Infineon Technologies (IR HiRel) : Specializes in radiation-hardened power discretes, GaN transistors, and robust semiconductor architectures that withstand deep-space particle hits without relying on passive shielding.
• Renesas Electronics : Provides space-qualified and military-standard integrated circuits embedded across critical infrastructure, flight computers, and deep-space exploration capsules.

To help narrow this down, what is your target orbit/environment (e.g., LEO vs. deep space/GEO) and your expected compute throughput requirement (TOPS/Watts) ? I can tailor this list further toward commercial off-the-shelf (COTS) approaches vs. strict rad-hard suppliers.

For in-orbit AI inference hardware, I would track companies across four layers: rad-hard silicon, high-performance edge AI modules, space-qualified compute platforms, and orbital compute operators. The winning architecture is likely to be a mix of radiation-tolerant commercial silicon (with shielding/error mitigation) and traditional rad-hard components depending on mission lifetime and orbit.

1. Radiation-hardened processors and FPGA leaders

• BAE Systems — Track for heritage rad-hard CPUs such as RAD750/RAD5545-class systems. These are not the fastest AI accelerators, but they remain relevant for high-reliability missions where survival matters more than TOPS.
• AMD (including Xilinx space products) — Probably one of the most important names for space AI because FPGAs provide reconfigurability, radiation mitigation techniques, and high-performance onboard processing. Their space portfolio includes Versal XQR and RT Kintex-class devices aimed at space applications.
• Microchip Technology — Long-time space semiconductor supplier (FPGAs, MCUs, radiation-tolerant parts). Worth watching for low-power autonomous spacecraft compute.
• Teledyne Technologies — Strong position in space electronics, sensors, data converters, and radiation-tolerant components.
• Vorago Technologies — Interesting for rad-hard ARM-class control processors and supervisory compute.

2. AI accelerator / GPU-class compute for orbit

• NVIDIA — The most important commercial AI silicon company to watch. Their space-compute strategy spans:

• Jetson-class modules for small satellites

• IGX-class rugged edge AI

• higher-end orbital compute modules NVIDIA is explicitly targeting onboard inference, sensor fusion, and orbital data centers.

• Aitech Systems — One of the more direct “AI computer for spacecraft” vendors, offering radiation-aware systems with GPU/CPU/FPGA options. NASA’s small spacecraft avionics survey lists Aitech systems among high-performance onboard computing platforms.

• Unibap — Swedish company focused on intelligent space systems and onboard processing.

• KP Labs — Focused on AI-enabled satellite data processing, especially Earth observation workloads.

• NOVI — Developing radiation-tolerant AI-capable onboard computers based on AMD/Xilinx architectures, including Versal-based systems.

3. Orbital compute / AI satellite operators

These companies matter because they will determine what hardware actually gets flown.

• Starcloud — Building orbital data center concepts using high-performance compute hardware.
• Orbital — Focused specifically on AI inference infrastructure in low Earth orbit, including thermal and radiation challenges.
• Axiom Space — Relevant as a platform/integration partner for commercial orbital compute.
• Kepler Communications — Interesting because distributed orbital compute depends heavily on optical links and networking.
• SpaceX — Worth tracking because vertical integration (launch + satellite bus + networking + compute) could create a unique path for large-scale orbital AI infrastructure.

4. Thermal, packaging, and enabling technologies

The bottleneck may not be FLOPS — it may be watts-to-heat rejection ratio.

Watch:

• 3D PLUS — High-density radiation-tolerant packaging, memory, and system-in-package technology.
• Texas Instruments — Radiation-tolerant power electronics are critical because AI accelerators are power-limited in orbit.
• Ramon.Space — Interesting for radiation-tolerant storage and compute systems.

My shortlist to put on a serious watchlist

If I were building an investment/technology landscape, I would prioritize:

1. NVIDIA — highest upside for AI performance.
2. AMD — likely winner for adaptable, space-qualified compute.
3. BAE Systems — reliability benchmark.
4. Microchip Technology — rad-hard ecosystem.
5. Aitech Systems — direct spacecraft AI hardware.
6. NOVI — emerging AI spacecraft computers.
7. Unibap / KP Labs — application-driven onboard AI.
8. Starcloud / Orbital — if orbital data centers become viable.

The key technical split to watch: LEO Earth-observation inference (low latency, modest power) is likely to favor Jetson/FPGA-class systems, while orbital data centers (large AI models) will require breakthroughs in thermal radiators, shielding strategy, power generation, and fault-tolerant commercial silicon.

Building AI inference hardware for orbit means navigating a brutal engineering trade-off: the sub-nanometer transistors required for dense AI compute are inherently vulnerable to single-event upsets (bit flips) and total ionizing dose (TID) degradation from cosmic rays and solar protons.

The industry is splitting into two paths: radiation-hardened (rad-hard) by design (uncompromising reliability, slower compute cycles) and Commercial Off-The-Shelf (COTS) adaptation (high performance using shielding or fault-tolerant software tricks).

The key companies, silicon providers, and infrastructure innovators shaping the space-inference market to track include:

1. Major Silicon & Systems Giants

• NVIDIA : A dominant force in edge and data-center space compute, offering the NVIDIA Space Computing Platform ecosystem. They feature space-adapted iterations of the Jetson Orin and IGX Thor modules, alongside the Space-1 Vera Rubin Module engineered specifically to bring data-center-class AI inference and substantial TOPS-per-watt performance to low-Earth orbit (LEO) constellations and orbital data centers.
• AMD / Xilinx : Known for their space-grade adaptive SoCs. Their radiation-tolerant XQR Versal AI Core (like the VC1902) devices provide on-board, reconfigurable hardware acceleration and tensor-like processing, making them a primary choice for missions needing reprogrammable, high-throughput geospatial and radar data processing in flight.
• Intel : Has targeted the sector with specialized LEO-bound silicon such as the Starfire processor , which leverages advanced multi-tile architectures (utilizing Intel 18A process nodes and Foveros packaging) to deliver high neural processing unit (NPU) performance under harsh thermal profiles.

2. Rad-Hard Specialists & Foundries

• Microchip Technology : Famous for partnering with NASA on the High-Performance Spaceflight Computing (HPSC) program. They build extreme-reliability, fault-tolerant multi-core processors targeting multi-hundred-fold performance leaps over legacy space computers while prioritizing deterministic safety and radiation resilience.
• VORAGO Technologies : A leader in ruggedized microcontrollers and processors utilizing their patented HARDSIL technology, which hardens standard high-volume silicon manufacturing processes against extreme temperatures and radiation damage without sacrificing commercial pricing economics.
• GlobalFoundries : Acts as a critical manufacturing partner for defense and aerospace prime contractors (like BAE Systems and Northrop Grumman), fabricating specialized rad-hard semiconductors designed to survive prolonged exposure in high-radiation orbital regimes.
• Lattice Semiconductor : Provides ultra-low-power, small-form-factor FPGAs (such as their Nexus and Avant platforms) featuring built-in hardware scrubbers and error-correction coding (ECC) designed to resist single-event latch-ups in space constraints.

3. Edge AI Accelerators & Subsystems Startups

• EdgeCortix : Specializes in highly energy-efficient neural acceleration software and hardware IP (like their SAKURA architecture) designed for extreme edge environments where power and thermal dissipation budgets are tightly capped.
• Cosmic Shielding Corporation : Rather than changing the chip physics, they engineer advanced material shielding (such as their Plasteel composite) designed to wrap standard COTS hardware (like regular NVIDIA Jetson modules), mitigating radiation impact and cutting down launch-weight penalties for high-performance commercial payloads.
• Sophia Space : Focuses on modular "tile" architectures for orbital computing. Each self-contained tile manages its own distributed compute, power generation, and radiative thermal cooling linked via high-speed fiber optics, helping solve the massive thermal dissipation walls encountered in a vacuum.

Are you targeting Low Earth Orbit (LEO) constellations (where size, weight, and power/SWaP are tightly restricted) or orbital/deep-space data centers (where thermal dissipation and massive cooling arrays are the primary bottlenecks)? Sharing your compute density (TOPS) or power budget goals can help narrow down which of these architectures fits best.

Yes. If the requirement is actual inference in orbit, with radiation, thermal, SWaP, and long mission life treated as first-class constraints, I’d track a fairly different set of companies than the usual terrestrial AI-chip list.

My priority watchlist

One company I'd put especially high on the list

EdgeCortix deserves a close look. The distinction is that this isn't merely a conventional terrestrial accelerator being marketed for space: NASA's NEPP actually subjected SAKURA hardware to proton/heavy-ion testing. SAKURA-II's January 2026 results reportedly showed no destructive events and relatively few transient effects.

That's unusually relevant to your requirements because the real problem isn't simply TOPS/W. It's:

TOPS/W × radiation survivability × memory behavior × thermal density × mass × qualification risk.

A spectacular accelerator that requires a giant radiator and radiation shielding isn't necessarily competitive with a slower accelerator that can run at 2–5 W.

Don't overlook the system architecture

I'd divide the market into three layers:

1. Flight-qualified programmable compute

• AMD Versal XQR
• Microchip RT PolarFire / RTG4
• Frontgrade implementations

2. Dedicated low-power AI acceleration

• EdgeCortix
• Aqylon
• emerging neuromorphic/in-memory/spintronic approaches

NASA is actively evaluating a radiation-hardened AI inference accelerator targeting <0.4 W and 50 TOPS/W, which illustrates where the most interesting architecture race is heading.

3. High-performance orbital compute

• NVIDIA
• AMD
• Starcloud / Orbital / Exlumina-type operators

For example, Exlumina is pursuing GPU-based orbital inference and explicitly addresses radiator-based thermal management and radiation recovery mechanisms.

If I were building the competitive landscape

I'd rank the near-term technology bets roughly:

Tier 1 — investigate now

1. AMD
2. EdgeCortix
3. Microchip
4. Frontgrade
5. NVIDIA

Tier 2 — strategically interesting 6. Spacechips 7. VORAGO 8. NOVI 9. Aqylon

Tier 3 — longer-horizon architecture bets 10. Rocket One 11. Other neuromorphic/in-memory/spintronic accelerators

The key question I'd use to separate them is not peak TOPS. I'd request four numbers from every vendor:

• inference TOPS/W at the actual deployed precision (INT8/INT4/etc.)
• radiation dose + SEE/SEU data at the intended orbit
• W/cm² or W/package thermal density at sustained inference
• complete compute-module mass and volume, including memory, power conversion, shielding and thermal hardware

That fourth metric is where many seemingly attractive "space AI" solutions will fall apart. Current industry discussion increasingly recognizes that orbital AI is constrained by radiators, mass, radiation and power—not merely availability of compute.

If you tell me your orbit (LEO/GEO/cislunar), inference workload (vision/transformer/RF), power budget, and approximate mass ceiling, I can turn this into a 10–15 company competitive matrix with actual TOPS/W, radiation rating, temperature range, package mass, memory, and flight heritage.

Deploying AI inference hardware in orbit means balancing three unforgiving pillars: Size, Weight, and Power (SWaP) constraints, thermal dissipation in a vacuum (where convection doesn't exist), and radiation resilience (mitigating Single Event Upsets and Total Ionizing Dose).

The market is split between traditional aerospace radiation-hardened legacy providers and aggressive new entrants adapting high-performance edge/data-center silicon for orbit.

1. Heavyweight Tech Giants & Accelerated Computing

• NVIDIA: Expanded aggressively into the aerospace sector with dedicated space-ready platforms. Track their Space-1 Vera Rubin Module (built for orbital data centers offering dense AI compute in a SWaP envelope), alongside space-flight-adapted Jetson Orin and IGX Thor modules utilized heavily by commercial constellation builders.
• AMD: A major player via its Milpitas/Aerospace division, offering radiation-tolerant adaptive System-on-Chips (SoCs). Their XQR Versal AI Core Go to product viewer dialog for this item. (like the XQRVC1902)Go to product viewer dialog for this item. brings dynamically reconfigurable logic and dedicated AI engines directly into space-grade qualifications.
• Intel: Operates through Intel Government Technologies , pushing commercial-to-space pipelines. They have showcased space-grade functional SoCs like Starfire —aiming to bring high TOPS (Tera Operations Per Second) and dedicated neural processing units (NPUs) into a domestic manufacturing and packaging roadmap.

2. Specialized Edge AI & Space Processors

• EdgeCortix: Known for energy-efficient neural processing architectures, they offer the SAKURA-II AI Accelerator , which has undergone NASA validation testing for extreme environments, targeting local hazard detection and lunar/orbital autonomous navigation.
• Microchip Technology: Partnered closely with NASA on the High-Performance Spaceflight Computing (HPSC) processor initiative. While not a traditional deep-learning accelerator startup, it is the benchmark for high-throughput, fault-tolerant, radiation-hardened processing in deep space and orbital autonomy.
• Aitech Systems: Specializes in ruggedized, space-qualified AI supercomputers (such as the S-A2300 / Lyra series) designed around cold-plate conduction cooling and heavy in-orbit video/sensor analytics.
• Vorago Technologies: A pioneer in HARDSIL technology, which hardens standard commercial manufacturing processes against radiation. They provide microcontrollers and processor tech that let commercial off-the-shelf (COTS) logic survive extreme orbital radiation profiles.

3. Orbital Data Center & Infrastructure Integrators

• Starcloud: A prominent deep-tech player building dedicated orbital data center infrastructure designed to leverage continuous 24/7 solar power and passive radiative cooling in space.
• Orbits Edge: Focuses on the physical "chassis" problem—building ruggedized, shielded enclosure systems (like Sabir) that allow high-powered commercial edge servers to handle thermal regulation and radiation mitigation inside a satellite frame.
• Tycho Space Technologies: European developer designing specialized radiation-tolerant edge computing modules (utilizing architectures like AMD/Xilinx logic) aimed at real-time in-orbit AI evaluation and Earth observation.

Are you leaning toward using Commercial Off-The-Shelf (COTS) chips with software/shielding redundancy or do you need strictly space-qualified/radiation-hardened (Rad-Hard) ASICs for a high-radiation orbit (like MEO or GEO)? Let me know and I can narrow down the performance-per-watt tradeoffs.

If the requirement is actual onboard inference in orbit, with radiation, thermal dissipation, and SWaP constraints—not merely terrestrial AI chips adapted for a spacecraft—I’d track a fairly focused set of companies.

Highest-priority companies

Also worth tracking

GomSpace — particularly if you want integrated spacecraft computing rather than a bare accelerator. Its NanoMind family is already flying, although the performance ceiling is considerably below the AMD/KP Labs class. NASA lists NanoMind HP MK3 using Xilinx Zynq 7030/7045 with >20 krad tolerance.

VORAGO — more interesting as a radiation-hard processor technology supplier than as an AI accelerator today. Its collaboration with Frontgrade is explicitly aimed at autonomous, data-intensive space computing and AI.

BrainChip — worth tracking because of its Akida neuromorphic architecture. The Frontgrade GRAIN relationship is particularly interesting: neuromorphic/event-based inference could be a compelling answer when the real constraint is watts and heat per inference, rather than maximum throughput.

AMD's broader ecosystem — don't just watch AMD silicon. Companies such as Frontgrade are packaging Versal into space-grade processing cards, which may be much closer to something you can actually integrate into a spacecraft than the bare device.

My shortlist by use case

• Maximum inference performance: AMD Versal XQR
• Best combination of radiation + low power + programmable acceleration: Microchip RT PolarFire SoC
• Turnkey onboard AI/DPU: KP Labs Leopard
• Very low-power inference / event-driven workloads: Frontgrade Gaisler + BrainChip GRAIN
• Commercially packaged high-performance compute: Ibeos
• Long-term general-purpose space compute: Microchip HPSC
• Space-qualified processing ecosystem / integration: Frontgrade

One important distinction: don't compare these on TOPS alone. For spacecraft, I'd rank candidates on inference TOPS/W at the actual model precision, rad tolerance at your orbit, memory bandwidth, external-memory radiation strategy, peak/average thermal load, board mass/volume, fault recovery, and whether the vendor has demonstrated the specific workload on orbit. AMD's XQR, for example, is explicitly characterized for radiation and has AI engines, while Microchip is emphasizing substantially lower power and configuration-upset immunity.

If you tell me orbit (LEO/MEO/GEO/cislunar), target inference workload (EO vision, hyperspectral, RF, autonomy, etc.), and roughly your power budget, I can narrow this to a 5–8 company competitive landscape with estimated W, mass, radiation tolerance, compute, flight heritage, and maturity.

Yes. I’d track this as a space-edge inference hardware market, not just “space CPUs.” The interesting companies fall into three tiers:

My shortlist

If I were building a competitive landscape today, I'd put five companies in the first diligence wave:

1. Frontgrade Gaisler — best fit for radiation + low power + inference at the silicon level.
2. Ramon.Space — strongest “space-native compute platform” angle.
3. Microchip — compelling if FPGA acceleration/determinism is acceptable.
4. Unibap — compelling integrated hardware/software option.
5. NVIDIA/Aitech — compelling if you're willing to accept COTS-derived silicon and solve radiation at the system level.

The fundamental strategic split is important:

Rad-hard silicon → lower performance, excellent radiation pedigree, much easier thermal/SWaP story.

COTS/high-end GPU → dramatically better TOPS/W and model flexibility, but radiation, memory reliability, qualification, and thermal design become system-level problems.

That's increasingly becoming a real architectural choice. Recent research argues that LEO inference can be viable while frontier-scale orbital training is much less attractive because of networking, power, radiation and cost constraints. arxiv.org And current work on radiation-tolerant AI is explicitly exploring algorithmic fault tolerance rather than relying exclusively on expensive rad-hard silicon.

One company I'd watch especially closely: Frontgrade Gaisler. Its GR801/GRAIN approach—rad-hard RISC-V + neuromorphic inference—is unusually well aligned with your radiation + thermal + mass constraints rather than simply putting a terrestrial GPU in a spacecraft.

If you tell me your approximate TOPS, watts, mass, orbit (LEO/GEO/cislunar), and mission life, I can turn this into a much more useful vendor × radiation × TOPS/W × mass × qualification × maturity matrix.

Building AI inference hardware for orbit means navigating brutal Size, Weight, and Power (SWaP) constraints, extreme thermal management in a vacuum , and radiation-induced soft errors (bit flips) or latch-ups.

The market has shifted rapidly from traditional, slow-to-develop rad-hard chips to high-performance commercial-off-the-shelf (COTS) adaptation and architectural fault tolerance.

The key players and emerging companies in the on-orbit AI and space compute ecosystem worth tracking fall into distinct categories:

1. High-Performance Silicon & GPU Giants

• NVIDIA: A major force in the space compute conversation with the NVIDIA Space Computing initiative. They provide energy-efficient edge platforms like Jetson Orin and IGX Thor , alongside the Space-1 Vera Rubin Module , which brings data-center-class inference performance directly to orbital deployments.
• AMD: Offers space-grade adaptive SoCs and radiation-tolerant FPGAs (such as the XQR Versal AI Core series), enabling robust on-board ML inferencing and data preprocessing for aerospace and defense missions.

2. Radiation-Hardening & Specialized Space Processors

• VORAGO Technologies: Known for its patented HARDSIL® technology, VORAGO provides extreme radiation-hardened/tolerant microcontrollers. Watch their VA7230 Space Edge Processor , which pairs an Arm architecture with an embedded GPU designed specifically for real-time space AI and computer vision.
• Microchip Technology: A staple in aerospace silicon, partnering heavily with NASA on ultra-reliable, fault-tolerant processing architectures like the High-Performance Spacecomputing (HPSC) initiative to bring massive compute leaps over legacy space processors.
• Lattice Semiconductor: Provides small form-factor, highly power-efficient FPGAs (utilizing FD-SOI and FinFET nodes) that feature built-in hardware error correction and high immunity to single-event latch-ups in LEO.

3. Subsystems, Chassis, and Thermal/Radiation Shielding

• OrbitsEdge: Builds ruggedized orbital edge platforms (like Edge1 ) that feature integrated neural processing units alongside custom physical radiation shielding and thermal dissipation frameworks designed to house commercial compute safely in LEO and GEO.
• Cosmic Shielding Corporation: Specializes in advanced materials science (such as their "Plasteel" shielding technology) designed to wrap around high-performance commercial edge GPUs (like NVIDIA Jetson Go to product viewer dialog for this item.) to block radiation without adding prohibitive mass penalties.
• Sophia Space: Develops modular orbital data center architectures, radiation-hardened edge servers (like the Sophia Blade and Sophia TILE platforms), and specific in-orbit operating software tailored for localized AI workloads.

4. Orbital Infrastructure & Systems Pioneers

• Starcloud: Focuses directly on deploying orbital data center clusters powered by persistent solar energy and cooled via deep-space infrared radiation, utilizing high-end commercial GPUs for on-orbit processing, fine-tuning, and inference.

If you're targeting a specific orbit (like LEO vs. GEO ) or a particular processing throughput requirement, let me know and I can help narrow down which architecture fits your payload profile best.