The Hook: In one of the recent Kron digests, a line flashed by about STM32s in orbit — something about each Starlink satellite carrying ~70 specialized processors and 10 microcontrollers, including "ordinary STM32s on ARM Cortex-M85." I read it, nodded, moved on. But then it stopped me. Because in that line, casually dropped, lies the deepest tectonic shift in the semiconductor industry that nobody has really bothered to call by its name. And I'm not talking about AI, not about "new technologies," not about "Musk's innovations." I'm talking about the radiation-hardened chip market — quiet, conservative, defense-aerospace, with thirty-year supply cycles, Swiss-watch margins, and a dozen suppliers for the entire planet. And into this market SpaceX walked in with consumer STM32s at $5–10 and said: "Guys, why do you even need rad-hard if I'm launching ten satellites a week, and if one dies from radiation — well, we'll roll out the next one?" This isn't an engineering argument, it's philosophical. And it's destroying an entire industry right now.
To understand the drama, you first need to understand the numbers. Because they're the main character in this story.
RAD750 — this is probably the most famous radiation-hardened processor in the world. It's made by BAE Systems. It's on Perseverance (the rover, $2.7 billion), on James Webb (~$10 billion mission), on InSight, on dozens of other government satellites. Technical specs by modern standards are humiliating: 110–200 MHz clock speed, 266 MIPS, 10.4 million transistors, manufactured on 250 nm or 150 nm process. For comparison: your iPhone 15 runs on a 3 nm process with 19 billion transistors and a 3.46 GHz clock speed. That is, the processor flying to Mars is 200 times weaker than the chip sitting in your pocket.
But! This processor withstands a dose of 2,000–10,000 gray (200,000–1,000,000 rad), operates at temperatures from −55°C to +125°C, and consumes only 5 watts. Here's its price: around $200,000 for a single-board computer (2002 data, adjusted for inflation — $358,000 in 2025). For 110 MHz. On 150 nm. For two hundred thousand dollars.
Why so expensive? Not because BAE engineers are worse than Apple engineers. But because it's a different game. RAD750 isn't about transistors. It's about:
This isn't a chip. It's a ritual object. Certified, tested, documented, protected from every possible failure. When you launch a $500 million satellite, you don't want an SEU (single-event upset — a bit flip from cosmic radiation) to kill the entire mission three years later on geostationary orbit. So you pay $200,000 for 110 MHz. And nobody complains, because there's no alternative.
For completeness — this isn't just BAE. The rad-hard electronics market has historically been dominated by a few players:
They all work through aerospace-qualified foundries — old, proven, certified production lines. For example, BAE uses its own fab in Manchester for RAD750, certified for defense and space applications. All these chips cost in the $50K–$500K range per unit depending on type and qualification. And each takes a year or two for full radiation qualification.
In 2020, SpaceX engineers — Matt Monson, Josh Sulkin, and others — held a famous Reddit AMA. There they dropped a phrase I quoted above: each Starlink satellite carries more than 4,000 Linux computers and more than 6,000 microcontrollers in each launch of 60 satellites. Confirmation from independent sources: arxiv.org/html/2503.03722 (Memon et al., 2025) writes that "SpaceX is reported to operate tens of thousands of Linux-class COTS nodes across Starlink." Tens of thousands of Linux nodes on consumer SoCs in orbit.
This isn't a joke or exaggeration. Starlink V2 Mini costs, according to New Space Economy estimates, ~$400,000 per satellite — that is, the entire vehicle complete. And inside this vehicle:
Total cost of all semiconductors on board — pennies by rad-hard market standards. STM32 costs $5–50 depending on variant. Even if SpaceX puts 70 processors per satellite, that's $350–$3,500 in electronics on a vehicle launched by the thousands.
Here's where it gets most interesting. SpaceX doesn't just use cheap chips. They use a different architectural philosophy that makes rad-hard unnecessary.
Old paradigm (one expensive rad-hard chip):
New SpaceX paradigm (many cheap COTS chips):
And this — paradigm shift. Not technological, but economic and philosophical. It's like finding out that instead of one Rolls-Royce for $500K with a driver who doesn't talk on the phone and parks perfectly, you can take 50 Toyota Camrys at $30K each, and each occasionally breaks down, but together they work more reliably than one Rolls-Royce. The Rolls-Royce doesn't disappear (there are tasks where it's objectively better), but the market shrinks dramatically.
Confirmation of this logic — in the data from arxiv.org/html/2503.03722. Scientists from NASA Goddard, University of Luxembourg, and Polish Academy of Sciences tested three COTS SoCs (Raspberry Pi Zero 2W, NXP i.MX 8M Plus, OrangeCrab ECP5 FPGA) under proton irradiation. Result:
And most importantly: no COTS chip survives space "out of the box". But with the right architecture (TMR, ECC, software mitigation, kernel hardening) — quite usable. And this usability costs many times less than buying rad-hard.
Numbers that probably make marketing folks at BAE Systems twitch:
Rad-hard electronics market (New Space Economy estimate, April 2026):
SDA Tranche 3 (Space Development Agency Proliferated Warfighter Space Architecture) — this is perhaps the most telling case. $1.1 billion Lockheed Martin, $843 million L3Harris, $764 million Northrop Grumman, $805 million Rocket Lab — each company must deliver 18 satellites in 3 years. That's 72 satellites in Tranche 3 for a defense program, and they all use COTS components with "mid-tier" radiation hardness — not full rad-hard, not full COTS, but something in between. This is a new component category that didn't exist in any significant volume ten years ago.
Lost (partially or completely):
Won:
Neutral position:
Returning to arxiv.org/html/2503.03722. One of the subtlest points of their research — propagation chain analysis. They found that one SEU can pass through 6 different Linux kernel subsystems before the system finally crashes. That is, a radiation hit — isn't instant failure, it's a cascade. And that's exactly why COTS + proper architecture works: you can isolate the cascade early.
But! They also found reproducible deadlock: "In 3 of 133 events, restoring PID 1 required the same corrupted eMMC subsystem that the fault had disabled, a circular dependency that blocks autonomous recovery." That is, 3 out of 133 failures (~2.3%) — are not autonomously recoverable. On a traditional $300 million satellite, this is unacceptable. On 7,000+ Starlink vehicles — that's 160 pre-planned losses per year, which is economically acceptable.
Old aerospace engineering says: 2.3% non-recoverable failures — that's catastrophic.
New SpaceX engineering says: 2.3% non-recoverable failures — that's 0.023 × $400,000 = $9,200 in losses per year. On $10 billion revenue — that's noise.
This story — isn't unique to space. The same thing is happening now:
Common pattern: when "recoverable failure" becomes cheaper than "absolute reliability," the market flips. Rad-hard electronics — that's an "absolute reliability" industry. COTS in orbit — that's a "recoverable failure" industry. SpaceX made the second cheaper than the first by so much that it changed the economics.
But here's the paradox that New Space Economy writes about: radiation-hardened electronics — it's not just an engineering category, it's a sovereignty category. Countries want to have their own rad-foundries, their own certified components, their own supply chains. Because if your satellite depends on BAE Systems (and BAE is a British company), then in case of conflict you depend on someone else's fab.
SpaceX "undermined" this logic from another angle: they showed that COTS from allies (STMicroelectronics — European-Italian-French, NXP — Dutch, Microchip — American) provide almost the same reliability, but without sovereign dependence on a few rad-foundries. And for the LEO segment, where 90% of satellites are commercial constellations, that's enough. But for deep space, for defense GEO satellites, for Artemis — rad-hard is still needed. And sovereign investments in rad-foundries aren't stopping, they're even growing.
This means the industry isn't "dying," but splitting: the LEO-commercial segment goes to COTS, while the defense-government and deep-space segment stays with rad-hard. Two markets with different economics, different regulations, different culture. SpaceX didn't kill BAE Systems. It cut off their LEO.
Most engineering conclusion: $200,000 for 110 MHz — that's not absurd, it's a rational price in the "absolute reliability for a unique mission" paradigm. $5 for a 100 MHz STM32 — that's also not absurd, it's a rational price in the "mass production with recoverable failures" paradigm. Two different economies, two different engineering cultures. And until 2020 the first was the only one. After 2020 — the second appeared. And the second turned out to be so much cheaper than the first that it pulled the entire LEO market to itself.
Most non-obvious conclusion: this isn't about Musk and not about genius. It's about changing the unit of reliability. Before, the unit was the satellite. Now the unit — the constellation. When the unit of reliability is one vehicle, you need rad-hard at $200K. When the unit is 10,000 vehicles, you need COTS at $5 and replacement launch cadence. The entire SpaceX story — is redefining the unit of reliability from "satellite" to "constellation".
Most troubling conclusion: rad-foundries in the US, UK, Europe, Japan — that's critical infrastructure, like nuclear plants or submarine cables. If they disappear because the LEO market went to COTS, then in case of a deep-space program or Artemis-2 in 2030 we'll find ourselves without domestic rad-hard chip production. New Space Economy warns directly about this: "if a mission depends on a handful of hard-to-replace electronics suppliers, those suppliers become strategic assets even if the end product looks like an ordinary chip package". And this means that SpaceX is breaking not only the commercial market, but also the military-industrial base of allies. Not intentionally — but it's breaking.
Most personal conclusion: I'm an engineer, and this story gets me to the bone. Because in it — the whole truth about how progress works in engineering. Not "new technology beats old technology." But "changing the unit of measurement" beats "new technology". SpaceX didn't invent anything radically new in silicon. STM32s have existed since 2007. Radiation hardness through software mitigation has been known since the 1990s. What SpaceX did — it changed the economic frame: "a satellite costs $400K, it can be lost, no big deal." And when this frame is fixed, everything else (component selection, architecture, software practices, mission philosophy) — flows from it automatically. This isn't an engineering decision. It's philosophical. And that's exactly why it's irreversible.
P.S. And yes — what I just described, this is literally the same dynamic that destroyed mainframes in the 1990s, minicomputers in the 2000s, local SANs in the 2010s, and is now destroying enterprise monoliths in the 2020s. Distributed architecture with cheap nodes and automatic recovery always beats centralized with expensive nodes and manual recovery. It's just that in space this happens 30 years later than in IT. But it happens according to exactly the same script. 🌌