FPGA and RF engineering services or a new hire: How to decide
Engineering is the hardest role in aerospace and defense to fill. Three questions show whether your FPGA or RF gap needs a new hire or FPGA engineering services.
Back to top
The short answer: The satellite payload processing figures that hold up come from seven publishers, including Novaspace, BryceTech, ESA, NASA and AMD. Earth observation launches rise from 1,864 in the decade to 2023 to a forecast 5,401 for 2024 to 2033, and flight-qualified AI silicon now reaches 184 TOPS. Six widely quoted figures failed verification.
Novaspace counts 1,864 Earth observation satellites launched in the decade to 2023, and forecasts 5,401 for 2024 to 2033. Six other numbers in wide circulation about satellite payload processing turned out not to exist at the publishers they are credited to, including a thermal range, a processing deadline and a beam-test result. Seven institutions publish the figures that do hold up: two space agencies, a weather agency, a launch-data firm, a market-intelligence firm, a chip vendor and an analyst newsletter. Every figure below was read on the publisher’s own page, and each row names the publisher and the class of source.
The sensing fleet is growing faster than the downlinks that carry what it produces.
| Figure | Source, and what class of source it is |
|---|---|
| 1,864 EO satellites launched in the decade to 2023; 5,401 forecast for 2024 to 2033 | Novaspace, Earth Observation Satellite Systems, 2024 edition, July 2024 (market-intelligence firm, own series) |
| More than 6,500 EO satellites expected to launch by 2035, generating $155.9B in manufacturing revenue | Novaspace, current edition, August 2026 (same series, revised upward) |
| 325 orbital launches and 4,544 spacecraft deployed in 2025, up 25% and 54% on 2024 | BryceTech, 2025 Year in Review, April 2026 (primary launch data) |
| SpaceX launched 85% of all spacecraft in 2025, a total of 3,854 | BryceTech, same report (primary launch data) |
| Spacecraft under 1,200 kg were 98% of everything launched in 2025 and 85% of total upmass | BryceTech, same report (primary launch data) |
The fleet is growing and the spacecraft are shrinking, so a compute board competes for volume, mass and radiating area inside a bus that got smaller while the processing requirement got larger. Novaspace reports the same shift from the market side, with satellite buses standardizing while value moves “toward the integration of advanced payloads.”
ESA flew the first case in 2020 on Φ-sat-1, where an AI chip screened cloudy imagery before transmission so that, in ESA’s own wording, “only usable data are returned.” The hardware below is now flying on commercial constellations.
| Figure | Source, and what class of source it is |
|---|---|
| EDGX systems rated at up to 157 TOPS of in-orbit compute | Kratos, Constellations, January 2026 (named publisher, vendor-adjacent editorial) |
| Aitech flying NVIDIA Orin processors in space at 248 TOPS | Kratos, same article (same) |
| Ubotica has flown 11 missions with over 30 AI models in orbit | TerraWatch Space, February 2026 (analyst newsletter, free portion only) |
| Starcloud put an NVIDIA H100 in orbit in November 2025; Axiom Space deployed its first orbital data center nodes in January 2026 | TerraWatch Space, same (same) |
| China’s Three-Body Computing Constellation has demonstrated 8-billion-parameter models running in orbit | TerraWatch Space, same (same) |
The TerraWatch deep dive is a subscription piece whose opening section is all that renders publicly, so those three rows come from the free portion. No published count exists of how many spacecraft carry onboard processing at all. Operators either do not disclose the capability or count it differently, and every search for that census returns a market forecast instead.
Qualification status sets the device list before any architecture trade begins, and the list is short.
| Figure | Source, and what class of source it is |
|---|---|
| Up to 184 TOPS of AI acceleration, over 500K LUTs and up to 200K DMIPS in a radiation-tolerant, flight-qualified device | AMD, November 2025 (vendor, own product) |
| Class Y packaging targeted at missions of up to 15 years; Class B and Class Y both derive from MIL-PRF-38535 | AMD, same (vendor) |
| Versal AI Core XQRVC1902 sampling in 2026, flight-qualified units in 2027; space-qualified RF and AI Edge Gen 2 parts expected in 2029 | AMD, same (vendor) |
| Custom ASIC development at a cost “exceeding $100M” | AMD, same (vendor claim about the alternative) |
None of the three TOPS figures on this page states the numeric precision it was measured at. That omission belongs to the publishers, and an unqualified TOPS number gets discounted accordingly. The ASIC cost row is a vendor’s figure for the path a customer takes instead of buying from that vendor, and it is quoted on that understanding. A part that samples in 2026 and reaches flight qualification in 2027 cannot carry a payload that has to ship sooner, which makes device selection partly a calendar decision, and the calendar belongs to the vendor.
A payload compute board has to shed every watt it draws, and in orbit there is no air to carry the heat away.
| Figure | Source, and what class of source it is |
|---|---|
| The exterior of a spacecraft in low Earth orbit alternates between about +121°C in sunlight and -157°C in darkness | ESA, July 2026 (primary; stated about the ISS, a large crewed platform, so read it as the order of magnitude of the environment and not a spec limit for a board) |
| “In a vacuum, heat is transferred only by radiation and conduction with no convection“ | NASA, State-of-the-Art of Small Spacecraft Technology, chapter 7, May 2026 (primary) |
| Thermal management on small spacecraft is limited by mass, surface area, volume and power | NASA, same chapter (primary) |
| NOAA’s JPSS polar satellites circle the globe 14 times each day at about 512 miles | NOAA NESDIS (primary, operator) |
Every watt the processor draws has to leave by conduction to a radiating surface. That surface sits on a structure moving between the two temperatures ESA measured on the ISS, and NOAA’s JPSS satellites circle the Earth 14 times a day. A power delivery network that closes at average load and opens under worst-case switching will throttle in orbit. The board throttles, the processing does not finish, and the pass is wasted. Closing that budget is hardware and PCB layout work: layer stack, plane assignment, decoupling, simultaneous switching noise and the thermal path, all validated against each other in one pass.
Three firms sizing the space-based edge computing market for roughly the same year put it at $2.1B (Stratistics MRC), $3.8B (Dataintelo) and $182.86B (Fortune Business Insights). None of those three appears below, because a market whose published estimates run from $2.1B to $182.86B has three definitions and one label.
| Figure | Source, and what class of source it is |
|---|---|
| Satellite payload market estimated at $12.79B in 2026, projected to $27.23B by 2031, a CAGR of 16.31% | Mordor Intelligence (forecaster, proprietary estimation framework) |
| Software-defined payloads forecast as the fastest-growing type at 21.92% CAGR to 2031; communication payloads held 49.20% share in 2025 | Mordor Intelligence, same (same) |
| A record 4,466 smallsats launched in 2025, from an operator base of 305 organizations including 133 first-time operators | BryceTech, Smallsats by the Numbers 2026, August 2026 (primary launch data) |
| More than 300 launches took place in 2025 | ESA, Space Environment Report 2026 (primary) |
Mordor states on the page that its figures come from a proprietary estimation framework updated with the latest available data, which is the disclosure the three reports above do not make.
Six numbers were checked at the institution each was attributed to, and refused. They circulate widely enough that a reader will meet them elsewhere.
| Figure as it circulates | What the check found |
|---|---|
| “Roughly 3x more EO satellites than the prior decade,” attributed to TerraWatch Space | Not in the accessible text, and the remainder is paywalled. The claim belongs to Novaspace, which publishes the satellite-count series. Re-sourced, and it survives. |
| “A 4x increase in spacecraft carrying onboard edge computing” | No publisher anywhere. Not in TerraWatch, Kratos, Novaspace or BryceTech. The only results are the three market estimates that run from $2.1B to $182.86B. |
| “Processing windows as tight as 250 ms,” attributed to Kratos | Not in the article, which carries no latency or millisecond figure of any kind. |
| “LEO thermal swings of +150°C to -150°C, cycling thousands of times a year,” attributed to Military Aerospace | Not at that publisher. The article carries no temperature figure and no degree symbol anywhere in its body. ESA’s +121°C to -157°C stands in above, with NOAA supplying the cycling rate. |
| “Versal XQR beam-tested with no SEL or SEFI above 80 MeV·cm²/mg LET,” attributed to an AMD blog post | Not in that post, which contains no single-event, LET or dose figure. The claim traces to radiation test reporting and belongs cited there. |
| Φ-sat-1 “saved 30% of bandwidth” | On no ESA page. It is carried by three trade outlets writing about ESA. ESA’s own qualitative wording is quoted above instead. |
Five of those began as a real measurement and arrived attached to the wrong publisher. The 4x figure is the worse case, because no publisher stands behind it at all and there is nothing to go and read. A payload program that budgets power or radiation margin against a figure nobody published is carrying a risk it cannot see, and the check is cheap: open the publisher’s page and find the number in the form quoted.
One payload compute board carries a fixed power allocation, a conduction path through a structure that cycles between sunlight and eclipse for the life of the mission, and a device list set by qualification schedules. A downlink budget decides how much sensor output ever reaches the ground. Fidus designs payload compute hardware across FPGA, embedded software, signal and power integrity, and thermal and mechanical work in one team, so those budgets close against each other while the board is still on the screen.
Contact Fidus to work the numbers against a specific mission.
Novaspace forecasts 5,401 Earth observation satellites for 2024 to 2033, up from 1,864 in the decade to 2023. Its August 2026 edition expects more than 6,500 by 2035.
AMD lists a radiation-tolerant, flight-qualified device with up to 184 TOPS of AI acceleration. Kratos reports Aitech flying NVIDIA Orin processors at 248 TOPS. None of these figures states the numeric precision it was measured at.
ESA puts the ISS exterior at about +121°C in sunlight and -157°C in darkness. In a vacuum, heat leaves only by radiation and conduction, so every watt the processor draws needs a conduction path to a radiating surface.
Mordor Intelligence estimates $12.79B in 2026, growing to $27.23B by 2031 at a 16.31% CAGR. Software-defined payloads are forecast to grow fastest, at 21.92% a year.
Six widely quoted figures could not be found at their attributed publishers, including a “4x” rise in onboard edge computing, 250 ms processing windows and a +150°C to -150°C thermal swing. The 4x figure has no publisher at all.
Engineering is the hardest role in aerospace and defense to fill. Three questions show whether your FPGA or RF gap needs a new hire or FPGA engineering services.
Earth observation launches are nearly tripling, but downlink capacity per spacecraft hasn't kept up. Deciding what to process in orbit is now a hardware decision made years before launch.
When should you outsource heterogeneous computing design? Keep the CPU, FPGA, GPU and NPU architecture call in-house, and get outside validation before board layout, while changing the partition is still cheap.
Trust us to deliver on time. That’s why 95% of our customers come back.