FPGAs have played a critical role in space exploration for more than 40 years, from the first satellites to deep-space and long-duration missions, and the meteoric rise of the New Space market. The next chapter is orbital compute, and it has already begun. On November 2, 2025, a Falcon 9 rideshare dropped a 60 kg satellite into orbit with a data center GPU inside. It was an NVIDIA H100, the same part that fills terrestrial AI racks. Five weeks later, it trained a small language model over the South Pacific. It was a small experiment, but it proved that data center silicon can work in orbit.

It also was not the first proof. In 2017, HPE bolted two off-the-shelf servers into a locker on the International Space Station and expected them to last days. They ran 615 days and crossed the South Atlantic Anomaly, the place where the inner radiation belt dips toward Earth, more than 6,800 times. Nine years and a dozen flights later, commercial hardware keeps surviving orbit.

The reason compute must move beyond Earth is power. Demand is surging: McKinsey sees roughly 125 GW of additional AI capacity demand by 2030, and SemiAnalysis projects that US grid headroom turns negative in 2027. Supply cannot keep pace, with interconnect queues running five to seven years. Orbit sidesteps both problems. A solar panel in a dawn-dusk orbit produces up to 5 to 8X the energy of the same panel on Earth, with no night, no weather, no land, and no cooling water. Elon Musk, Jeff Bezos, Sundar Pichai, and Jensen Huang have each said some version of the same thing this year: the next wave of AI compute will be built where the power is.

That shift has already begun, and early customers are paying a premium for it. Optical relay constellations carry data between satellites at 100+ Gbps without touching a ground station. Earth-observation and signals-intelligence satellites run inference onboard and downlink detections instead of pixels, cutting raw data 10 to 100X and delivering alerts in minutes. Sovereign cloud nodes hold government records where no one can physically reach them. Defense planners need missile tracking and guidance fused in orbit, and command and control that survives the loss of the ground segment.

None of this arrives without hard engineering, and we should be clear-eyed about what it takes. The economics depend on launch cost, which depends on Starship or other transports flying often and cheaply. Falcon 9 took a kilogram to orbit from roughly $10,000 to under $3,000, yet Google's own analysis says the case closes below $200/kg, which it places in the mid-2030s. Radiators are heavy, and repair is impossible. Independent analysts put gigawatt-scale orbital compute in the 2030s. Even so, we believe the direction is settled, because every one of these constraints is an engineering and cost problem, and those get solved over time.

Whatever the timeline, the architecture of the first nodes is already clear. Look at what every one of those nodes is. It is a commercial grade server: LINUX®, PCIe®, NVMe®, an accelerator, and a boot chain that assumes the memory holding the next stage is reliable and secure. Put that server in orbit and three things change. Power is shared with the thruster that keeps the satellite from falling out of the sky, and every watt of waste heat becomes radiator area. Recovery from a bit flip or a brownout has to happen onboard, in microseconds or less. And the adversary is a nation-state, so the boot chain needs measured boot, attestation, and quantum-resistant signatures; CNSA 2.0 requires post-quantum cryptography on national security systems starting in 2030 and whatever is deployed must be crypto-agile to adapt to the latest encryption algorithms over the air.

Those three demands (low power, autonomous recovery, and a secure boot chain) are control and supervisory functions, a key part of the server Lattice already helps build. Lattice Nexus™ FPGAs are fabricated on 28 nm FD-SOI, which gives them a soft error rate of 1.3 FIT/Mb, up to 100X lower than bulk CMOS in their class, immunity to single-event latch-up to LET 80 MeV-cm²/mg, and 100 krad(Si) total dose tolerance. These figures are characterized in-house and attested independently across the space ecosystem, from partner qualification programs to devices already on orbit. Lattice MachXO5™-NX and Lattice CertusPro™-NX space-grade variants carry NASA PEM INST-001 and ESA ECSS-Q-ST-60-13C qualification, run at 100 mW to 600 mW typical, and ship in 14 x 14 mm and 19 x 19 mm ruggedized packages footprint-compatible with the commercial parts a designer already knows. They configure in milliseconds with hardened scrubbing and ECC built in. On the security side, MachXO5™-NX TDQ and Lattice Mach™-N2 implement CNSA 2.0 post-quantum algorithms in silicon, hardened and built in, not bolted on, alongside NIST SP 800-193 firmware resilience, DICE and SPDM attestation, and multiple stored images with anti-rollback, flight ready, and crypto-agile.

Moving AI servers into orbit is not simply a component-qualification exercise. It changes the failure model of the entire server. On Earth, the platform assumes stable power, convection or liquid cooling, replaceable hardware, continuous network access, and human intervention. In orbit, recovery has to be autonomous, deterministic, and executable from hardware that remains available when the main compute complex does not. Solving these problems took decades of evolving radiation standards, qualification processes, and flight heritage. Rad hard is table stakes; what matters is reliability engineered into every device from chip architecture to built-in redundancy and autonomous error recovery and final product screening . It starts with selection of the right process node, built-in redundancy of control functions, hardened cells and 100% auto-correctable SEU, verified through qualification, and backed by high reliability production flow, the same silicon ships by the millions into critical infrastructure, automation and data center systems, and that volume is what makes the reliability statistics measurable, tuned, and real.

Lattice devices are already flying in signal processing and compute, communications, and distributed sensing payloads. Customers who ask us about this are also asking a newer question: not just how a device behaves under radiation, but where its wafers were manufactured. The resilience we get from FD-SOI process technology or other mature and mainstream products we have shipped do not depend on a single geography or a single fab. Dual sourcing, including an onshore fabrication option, is coming to our space-grade portfolio, and you will hear more from us on it.

Orbital compute is coming. It will take longer than the optimists hope and arrive more surely than the skeptics expect. The accelerators that fly will change every two years. The low power, secure, resilient control plane that boots them, attests them, and recovers them should not. That is the part we have been building all along.

To learn more about how Lattice FPGAs and secure control solutions can help bring resilient compute to orbit, reach out to our team today.