Starfire’s Ascent: Intel’s 18A Chip and the Coming Orbital AI Shake-Up

In January 2026, SpaceX quietly filed an application with the FCC for an “orbital data center system” that could eventually put up to a million AI compute satellites in low Earth orbit. By mid‑year, Elon Musk’s company had already locked in roughly $28 billion in annualized compute orders from Anthropic and Google. That number, even if padded with termination clauses, would make most terrestrial data center operators choke on their coffee.

Then, on July 9, 2026, Intel published a product page for a chip that suddenly made those sky‑high ambitions feel a little less speculative. Code‑named Starfire, it’s the company’s first space‑grade system‑on‑chip, built on the 18A node—the same process that CEO Lip‑Bu Tan has called a “national treasure.” With 45 to 75 TOPS of AI performance inside a 35‑watt thermal envelope, and a 10‑year design lifetime, Starfire wasn’t just a press release; it was a signal that Intel intends to be the silicon backbone of the orbital AI economy.

Starfire isn’t some exotic, clean‑sheet design. Under the hood, it’s essentially a Panther Lake variant—four P‑cores, four low‑power E‑cores, and a four‑core Xe3 GPU—repackaged for a world without air, with temperatures swinging from -55°C to +125°C, and with high‑energy protons trying to flip your bits every few seconds. The chip comes in two flavors: a 10‑watt low‑power SKU that runs the P‑cores at 1.0 GHz and pumps out 45 TOPS (INT8), and a 35‑watt performance SKU that cranks the P‑cores to 3.1 GHz and delivers 75 TOPS. Both carry the same 4‑core GPU tile and a neural processing unit fabricated on Intel’s crown‑jewel 18A node.

That manufacturing split is one of Starfire’s neatest tricks. Using Foveros 3D packaging, Intel puts the CPU complex and NPU on 18A—where RibbonFET gate‑all‑around transistors and PowerVia backside power delivery should, in theory, provide better radiation tolerance and lower leakage—while the GPU tile sticks to the more mature Intel 3 process. The company hasn’t come out and said it, but the message is clear: the mission‑critical compute sits on the most advanced node Intel can offer, while the graphics block rides a proven, high‑yield process where a few extra picoseconds of latency won’t doom the satellite. As one Tom’s Hardware forum commenter put it, “Chip designed for the US government leverages Intel 3 for the GPU.” That’s not a bug; it’s a political and engineering feature.

Memory support is standard LPDDR5 and DDR5, and the I/O tops out at 12 lanes of PCIe Gen4. No Gen5—again, a pragmatic choice. Less power, less heat, and frankly, when you’re beaming processed SAR images to the ground, the extra bandwidth would be academic. First engineering samples are scheduled for Q3 2026.

The 18A Gambit: Radiation Hardening the Un‑hardenable

Space‑grade chips historically live a decade behind their terrestrial cousins. They use older, thicker‑oxide processes that are inherently less leaky and less susceptible to total ionizing dose (TID) and single‑event effects. Intel is trying to short‑circuit that lag by adapting a cutting‑edge commercial node for space—and doing it while that node is still climbing the yield curve. Morgan Stanley pegged 18A yields around 50% earlier this year, while BlueFin Research Partners saw numbers in the 55–70% range. By July 2026, Intel declared yield problems “solved” and was churning out roughly 30,000 wafers per month across its Arizona and Oregon fabs. That’s a remarkable turnaround, and it gave Starfire a manufacturing foundation that few space chip vendors can match.

Yet here’s the rub: radiation qualification is still in progress. Intel says the chip will be tested against TID, single‑event latch‑up (SEL), and single‑event effects (SEE), but it hasn’t published any thresholds—no krad figures, no LET numbers. “Radiation‑hardened” in a press release is not the same as “Rad‑hard by design” on a silicon‑on‑insulator substrate. The distinction matters a lot. True rad‑hard parts from companies like BAE Systems or Honeywell often use dedicated SOI processes that gracefully degrade over decades. Starfire, by contrast, is a commercial derivative being characterized after the fact. The Hacker News crowd has been skeptical: “Is this really space‑grade, or just space‑tolerant? Until we see the TID curves, it’s marketing.”

And there’s the cooling challenge. Space isn’t just cold—it’s a vacuum, which means all heat rejection must happen radiatively. Intel rates Starfire for junction temperatures up to 125°C, but a Foveros multi‑chip package under thermal cycling will test every material interface. Different coefficients of thermal expansion between the silicon, the organic interposer, and the solder bumps can induce fatigue that shows up only after thousands of thermal cycles. Intel’s claim of a 10‑year lifetime suggests the packaging team has done its homework, but without published reliability data, we’re left guessing.

The Market That’s Suddenly Very Real

If you’d asked a satellite operator five years ago about onboard AI, they’d have laughed. Today, the spaceborne AI processing unit market crossed $1.27 billion in 2025, with over 1,100 units operating in orbit—nearly double the 2024 number. China alone accounts for about 34% of shipments. Research firm IIM projects the market will hit $6.84 billion by 2030, a 32% compound annual growth rate driven by LEO constellations that need to process data at the edge instead of downlinking raw sensor streams. When a Starlink‑style satellite generates terabytes a day, moving compute to orbit isn’t just elegant; it’s an economic necessity.

SpaceX’s AI1 design gives a taste of what’s coming. Each satellite—20 meters tall with a 70‑meter wingspan—packs 120 kilowatts of average AI compute, roughly a single terrestrial server rack, and dissipates heat through 110 square meters of liquid radiators. The first prototypes are supposed to launch in early 2027, with volume production at a new Gigasat factory later that year. Morgan Stanley analysts have penciled in 2029 as the year when terrestrial AI power expansion starts to plateau and orbital compute begins taking the marginal load. Whether that timeline holds is anyone’s guess, but the contracts are real—even if they come with escape hatches.

Meanwhile, AMD has been quietly eating SpaceX’s existing satellite compute business. Roughly 70% of the Starlink V2 satellites launched after 2023 use AMD Versal AI Core adaptive SoCs for on‑board processing. Those chips are radiation‑tolerant, not fully rad‑hard, but they’ve already passed Class B flight qualification. And they’re field‑reprogrammable in orbit, which is a huge deal when you can’t swap out a bad board. Microchip’s PIC64‑HPSC, co‑developed with JPL on a $50 million fixed‑price contract, brings a 64‑bit RISC‑V core with AI acceleration to the table. Google’s Project Suncatcher aims to put TPU‑powered satellites with free‑space optical links into a dawn‑dusk sun‑synchronous orbit by 2027. And China’s newly formed Space Computing Committee, backed by the Ministry of Industry and IT, has started throwing 10‑million‑yuan grants at anti‑radiation chip startups. It’s getting crowded up there.

Where Starfire Fits—and Where It Doesn’t

Intel’s pitch for Starfire boils down to three things: American manufacturing, consumer‑volume economics, and enough TOPS to run useful AI models on orbit. The “Domestic US manufacturing” tag is a dog whistle for the Pentagon and the NRO, where ITAR restrictions make a TSMC‑fabbed chip a paperwork nightmare. If Starfire can clear rad‑testing, it becomes the default option for classified programs. Intel’s foundry team has already notched over 200 design wins on 18A, and while none of them are yet from a major space prime, the pipeline is filling.

But the performance gap is jarring. A single NVIDIA Blackwell GPU churns out north of 1,000 TOPS at 700 watts; Apple’s M4 hits 38 TOPS at 20 watts in a consumer laptop. Starfire’s 75 TOPS at 35 watts is impressive for a rad‑tolerant part, but compared to anything on Earth, as Extremetech bluntly put it, “it’s pedestrian.” The gap reflects the fundamental physics of radiation: you can’t just slap a heatsink on a Blackwell and launch it into LEO without the silicon slowly frying itself from single‑event burnout.

There’s also the question of who exactly is buying. Tom’s Hardware reported that Starfire is “designed for the US government,” but Intel’s own slides list “potential users are currently unclear.” That kind of ambiguity usually means the product is being shopped to multiple agencies that don’t want to be named. The real litmus test will be whether SpaceX or OneWeb ever puts Starfire into production satellites. SpaceX CFO has already signaled that first‑gen orbital data centers will run on NVIDIA hardware (likely radiation‑shielded COTS), with a long‑term shift to a custom “Terafab” rad‑hard chip developed jointly with Tesla and Intel. That suggests Starfire might be a stepping stone, not the final answer.

Community Scrutiny and the Economic Elephant in the Room

The comment sections of tech publications have been brutal and, in many ways, more honest than the corporate press releases. “If building a data center on the ground takes tens of billions of dollars, building one in space, where every launch costs tens to hundreds of millions of dollars, seems unlikely to offer much return on investment,” an Extremetech reader wrote. Others pointed out that radiative cooling in a vacuum is non‑trivial—your giant liquid radiators are going to glow in the infrared, potentially interfering with optical sensors on neighboring satellites. And then there’s the maintenance problem: if a server fails in a hyperscale data center, a technician swaps it out in ten minutes. In orbit, that server becomes space junk.

Still, the true believers counter that the economics aren’t static. Starship’s cadence is ramping, and if per‑kilogram launch costs drop by another factor of three, the math starts to work. Free 24/7 solar power at eight times the terrestrial efficiency, combined with zero property taxes and no NIMBY lawsuits, could tip the balance for latency‑insensitive training workloads. “The first space‑grade AI chips will be underpowered and expensive, just like the first smartphones,” a Reddit comment speculated. “The trajectory matters more than the starting point. If Intel can iterate Starfire every two to three years, the performance gap closes fast.”

What Happens Next

By the end of 2026, we’ll have the first engineering samples in the hands of unnamed customers. The radiation data will either validate Intel’s design choices or send the company back to the drawing board. Meanwhile, SpaceX’s AI1 prototype will be on the pad, and China’s space‑computing push will be distributing grants to homegrown chip firms. The LEO compute market is projected to grow from $2.15 billion in 2026 to nearly $5.9 billion by 2034, but that forecast assumes a lot of things go right—starting with Starfire’s rad‑tolerance numbers.

Intel has done something genuinely hard: it has dragged an advanced commercial node into an environment where failure is measured in microns of lattice displacement. Whether that translates into a sustainable business is a different question. The chip exists. The demand is plausible. The execution, as always, remains to be seen.

We’ll know in a few years whether Starfire was the first domino or just a very expensive science project.