Spy Satellites · 2026-07-27

A Spy Satellite Can Already See Individual People From 480 km Up. A New Chip Just Gave It a Brain.

Intel just shipped a chip designed to run artificial intelligence in orbit. The announcement dropped quietly this week. Most people scrolled past it.

They probably shouldn't have.

What's Already Up There

The United States operates a fleet of reconnaissance satellites under the National Reconnaissance Office — the NRO. Most of them are classified. Their orbits, exact capabilities, and true numbers don't appear in any public catalog.

What we do know: in 2011, the NRO quietly donated two spare telescope mirrors to NASA. Each one was described as "at least as capable as Hubble." They were spares. The NRO had better ones already in orbit.

~10 cmEstimated resolution per pixel — defense analysts' figure

~480 kmTypical orbital altitude for recon satellites

DozensEstimated NRO optical, radar, and SIGINT satellites

At 10 centimeters per pixel — the widely cited estimate based on declassified budget documents and the donated mirror disclosure — you can see a person walking on a rooftop. You can count vehicles in a parking lot. You can watch a naval vessel enter a harbor and identify its class.

To be fair: The NRO has never officially confirmed resolution figures. The 10 cm estimate comes from defense analysts and indirect evidence, not an official statement. The actual capability could be sharper — or different in ways that aren't publicly understood.

But here's what these satellites have never been able to do: understand what they were seeing.

The Bottleneck Nobody Talks About

A single pass of a modern Earth observation satellite over a dense urban area produces terabytes of imagery. The satellite records it all. Then it has to send it home.

That's the chokepoint.

Radio downlinks from orbit are constrained by physics — bandwidth, antenna size, the brief windows when a ground station is in line of sight. A satellite crossing a target area has maybe eight minutes to beam data down before it's over the horizon. You can't send everything. You send compressed strips. You miss things.

8 min

Average window a low-orbit satellite has to downlink data to a single ground station before it disappears over the horizon

Intelligence analysts have raised this problem for years. By the time imagery reaches a screen, processed and cleared, the convoy has moved. The missile transporter is back in its hangar. The meeting has ended.

What if the satellite could do the analysis itself — before it ever sent a single byte home?

The Starfire Chip

That's exactly what Intel's new chip is designed for. Announced this week and named Starfire, it's built specifically for onboard data processing and AI workloads in space environments.

Space is brutal on electronics. Cosmic rays and charged particles tear through anything in orbit — flipping memory bits, corrupting sensor data, crashing processors mid-operation. Normal silicon fails. Space-grade chips are hardened against this radiation, but historically that came at a steep cost: you got reliability by sacrificing performance. You got a chip that survived, not one that thought.

What Starfire is trying to solve: Close the gap between radiation-hardened reliability and enough raw processing power to run actual AI — the kind that can look at an image and flag: that structure wasn't there last Tuesday, or that vehicle is on a prohibited route.

Intel hasn't named specific customers or programs. But the defense and intelligence communities have been explicit for years about wanting exactly this capability: satellites that process, filter, and decide — autonomously, in orbit — before anything reaches the ground.

What an Intelligent Eye in Orbit Actually Changes

Here's the current sequence: spy satellite images a target, stores raw data, downlinks during a ground station pass, analysts receive it hours or days later, flag what matters.

Here's the new sequence: satellite images a target, AI model runs onboard, compares against a baseline, flags the anomaly, sends only the relevant tile — in near real time.

Hours → SecondsPotential latency improvement for time-sensitive intelligence

~95%Data volume reduction possible with onboard AI filtering

16,119Satellites publicly tracked in orbit right now

For military applications, the difference is not incremental. It's operational. A target that moves during the old system's downlink window disappears. In a system with onboard AI, the satellite detects that movement in the moment — and a commander gets an alert before the target reaches the next location.

You can explore what's overhead right now — commercial, civilian, and the publicly catalogued ones — on the SkyLens live tracker. The classified ones don't appear. That's the point.

It's Not Just America Racing Here

China's Jilin-1 commercial imaging constellation already provides near-daily coverage of major global locations. Its military Yaogan program — optical, synthetic aperture radar, and signals intelligence variants — operates in orbits that analysts match to specific surveillance tasks.

830+

Chinese satellites tracked in public catalogs — the actual number, including classified platforms, is higher

Russia, France, Israel, and India all field dedicated reconnaissance satellite programs. What's in the public TLE catalog is a floor, not a ceiling. Amateur astronomers using optical telescopes and radar track objects that don't appear in any official list — because orbital mechanics doesn't care about classification level.

The public catalog gap: Some intelligence satellites are deliberately omitted from public orbital databases. Hobbyist trackers find them anyway by watching the sky. Here's the physics of why you can't fully hide a satellite.

The Part That Should Make You Pause

There's a natural rate limit on surveillance when humans have to analyze the imagery. Analysts can only review so many images per day. There are only so many watching hours. Targets get prioritized. Most of the Earth, most of the time, is simply not being looked at.

Onboard AI removes that limit.

When a satellite processes its own imagery — flagging vehicles, tracking individuals across multiple passes, detecting construction activity, identifying changes in land use — the cost of surveillance per target approaches zero. The question shifts from which targets can we afford to watch to which targets do we want the system to ignore.

Commercial operators are already moving this direction. Planet Labs images the entire Earth's landmass daily with over 200 satellites. AI startups are selling automated change-detection tools that run on that imagery. The civilian and military pipelines are converging.

The real shift: Intelligence satellites used to answer questions humans already had. The next generation may start surfacing patterns nobody thought to look for yet — operating at a scale and speed that has no human equivalent.

What Happens Next

Intel hasn't released launch timelines or confirmed which programs Starfire is being integrated into. The defense procurement pipeline between chip announcement and operational satellite typically runs years.

But the trajectory is clear. Chips get smaller, faster, more power-efficient. AI models shrink to fit constrained hardware. Each hardware generation gets closer to running serious computer vision — object identification, motion tracking, behavioral pattern analysis — entirely in orbit, entirely autonomously.

The satellites are already watching. They're about to start understanding.

That's a different world from the one where a grainy downlinked image sits in an analyst's queue for 36 hours.

See what's in orbit right nowOpen live tracker

More on reconnaissance satellites, orbital surveillance, and the intelligence race above Earth — in the SkyLens archive.

SkyLens editorial — live CelesTrak + NASA/JPL data (16119 objects)