At 7:26 Sunday morning a Falcon Heavy lifted the Nancy Grace Roman Space Telescope off Pad 39A at Kennedy. It is now three months and roughly a million miles from the L2 point where it will spend its life measuring how fast the universe is flying apart. NASA expects first images in early 2027.
The 2.4-meter primary mirror at the center of that observatory was not ground to look at galaxies. It was built for the Future Imagery Architecture, the reconnaissance program Boeing won in 1999 by underbidding Lockheed Martin by about a billion dollars and then could not deliver. Roughly ten billion dollars went into it by 2005. The New York Times later called it perhaps the most spectacular and expensive failure in the fifty-year history of American spy satellite projects. The optical half was cancelled, the contract went back to Lockheed, and the finished telescope assemblies went into storage at ITT Exelis in Rochester, New York, where they sat until the NRO handed two of them to NASA in 2012 at a valuation of a quarter billion dollars each. NASA still has the second one. It has never announced a use for it.
So the obvious question, and the one I have been asked four times this week: if that is what they were writing off in the nineties, what are they flying now?
We know the answer with unusual precision, and it is stranger than the one people are hoping for.
The accident that told us everything
On 30 August 2019 the President of the United States tweeted a photograph of the wreckage at Iran’s Imam Khomeini Space Center. There was a light flare in the middle of the frame, the tell that somebody had photographed a printed briefing slide, and the detail was good enough that the first assumption among analysts was a drone or a reconnaissance aircraft. Melissa Hanham, then at the Open Nuclear Network, said plainly that she did not believe a satellite could have produced it.
A satellite had. Marco Langbroek in Leiden ran the viewing geometry, Michael Thompson at Purdue checked the pass times, Cees Bassa at ASTRON worked the same problem in parallel, and inside a day the three of them had converged on USA-224, a KH-11, 385 kilometers of slant range, roughly 09:44 UTC on 29 August. Langbroek simulated the view from the spacecraft and matched it against the tweeted frame. The image resolves to something near ten centimeters per pixel. It was a genuine security incident at the time, and NGA did not declassify the original product until November 2022.
That single leaked frame is the best public calibration we have of what the classified fleet can actually do. What it shows is a ceiling.
The ceiling stopped moving around 1980
Angular resolution for a circular aperture is 1.22 lambda over D. Run 500 nanometers of green light through a 2.4-meter mirror and you get about 0.05 arcseconds. Project that down from 250 kilometers and the smallest resolvable feature on the ground is roughly six centimeters. That is the theoretical maximum, before anything degrades it.
Things degrade it. Turbulence scrambles the wavefront on the way down, and the characteristic scale of that scrambling is the Fried parameter, five to twenty centimeters at visible wavelengths under good conditions. Once your aperture exceeds the coherence length of the air you are looking through, more glass buys you photons rather than detail. David Fried and John Evvard published the underlying work in the mid-1960s. The practical floor for imaging the Earth’s surface from orbit sits somewhere between five and ten centimeters, and it is indifferent to your budget.
Correct the 2019 frame for slant angle and range and USA-224’s best case near nadir comes out around five to seven centimeters, which is exactly what a diffraction-limited 2.4-meter aperture predicts. That is a satellite launched in 2011, on a design first flown in December 1976, operating within a factor of two of its own theoretical maximum and resting on the floor the atmosphere imposes on everybody.
Which reframes the donation entirely. The Hubble-class optics the NRO gave away were not a hint of some steep classified capability curve running above the public one. They were evidence the curve had already gone flat. The observatory now heading for L2 and the satellite that photographed Semnan carry the same class of optic, from the same production run, out of the same cleanroom, for the same customer. GAO’s July 2026 assessment puts Roman’s replanned life-cycle cost at 4.316 billion dollars, which is the honest correction to the story about NASA getting a free telescope. The agency received hardware. It did not receive a mission.
Clifford Stoll described the practical consequence in 1989: a couple of inches, not quite good enough to recognize a face. At ten centimeters you can separate a sedan from a pickup. You cannot identify the driver, and no appropriation changes that, because the limiting component is not on the spacecraft. It is the eighty kilometers of moving air between the spacecraft and the pavement.
Seven years later, that photograph has a price list
Albedo put Clarity-1 into very low Earth orbit in March 2025 and sells ten-centimeter visible imagery commercially, with two-meter thermal collected on the same pass. European Space Imaging resells it into Europe and North Africa and holds exclusive tasking rights over parts of eastern Europe and the Mediterranean. Umbra sells sixteen-centimeter synthetic aperture radar, indifferent to cloud cover and darkness. Planet photographs every landmass on Earth daily. Maxar sells thirty centimeters at volume. The NRO has held acquisition authority for commercial imagery since 2018 and now buys a substantial share of its pixels rather than collecting them, while the Space Force runs a marketplace called TacSRT that lets an operator purchase surveillance as a service on a timeline measured in hours.
The number that triggered a national security argument in 2019 is a line item in 2026. The exclusive advantage the United States held for four decades now has roughly an eighteen-month lag and a published rate card, and the reason it could be closed at all is that the target stopped moving in about 1980.
What you buy when you cannot buy sharper
When the sensor is pinned against physics, the engineering moves. It moved to three places, and none of them are the camera.
The first is persistence. A KH-11 runs about 4.4 billion dollars, call it 6.3 billion in current money, and there are very few of them, in sun-synchronous orbits that hobbyists track from their back gardens. Adversaries have known the pass windows for decades. You move the launcher into the shelter, you pause the excavation, you resume when the bird is over the horizon. Sub-meter imagery of an empty pad is worth nothing. In May 2024 the NRO flew NROL-146 and called it the first launch of its proliferated architecture. By January 2026 it had flown twelve, and independent cataloguing now puts the constellation past 220 satellites across thirteen or fourteen batches, believed to be Starshield, the government variant of Starlink, carrying Northrop Grumman payloads on a SpaceX bus under a 1.8 billion dollar contract Reuters surfaced in 2024. The agency says the constellation has delivered more than 160,000 images and compresses delivery to minutes or seconds. Roughly a dozen NRO missions were manifested for 2026, with cadence running through at least 2029. Revisit measured in minutes defeats the shelter game in a way another centimeter of resolution never could.
The second is phenomenology, and it is the shift I would watch. Optical is one physics. Synthetic aperture radar is another, indifferent to weather and night. Hyperspectral is a third and probably the most consequential, because a spectrometer reading hundreds of narrow bands does not care what shape a thing is, it reports what the thing is made of. Camouflage netting that defeats a panchromatic camera does not defeat chemistry. This August the NRO awarded Radar Commercial Augmentation contracts to Capella, ICEYE US and Umbra, the three survivors of a five-company evaluation that began in 2022. In May it onboarded Pixxel for hyperspectral, ICEYE for radio frequency and EarthDaily for electro-optical, following HEO, SatVu and Sierra Nevada in February for non-Earth imaging, medium-wave infrared and RF geolocation. Non-Earth imaging means exactly what it sounds like: satellites whose assignment is photographing other satellites.
The third is what happens after the downlink.
The bottleneck moved and nobody wants to talk about where
I have spent most of thirty years building systems for people convinced their problem was the sensor. It usually is not. NGA took operational control of Project Maven after the 2022 invasion of Ukraine for an unglamorous reason: a two-hundred-satellite constellation produces imagery faster than any workforce you could assemble can look at it. Chris Scolese has said publicly that the machine learning work that matters at the NRO is in tasking, in deciding which satellite or combination of satellites should be pointed at a thing right now. That is scheduling and orchestration, not perception.
It is the same wall I hit last year building an agentic pipeline for a Las Vegas wedding chapel. Retrieval was fine. The model was fine. The system underperformed because nobody had built the layer that decides which of ten thousand available facts are load-bearing for the request in front of it. Every serious constraint I have hit on production agents in the last two years has been in the harness rather than the model, and the NRO appears to have reached the same conclusion from the opposite direction with a considerably larger budget.
Maskirovka is dead, and Moscow has noticed
This is where the surveillance story stops being about privacy and starts being about ordnance.
The Foreign Policy Research Institute published an assessment this month arguing that Russian analysts are specifically alarmed by commercial synthetic aperture radar, because it penetrates cloud cover and dismantles the country’s traditional reliance on maskirovka, the entire doctrine of tactical camouflage, concealment and deception that Russian ground forces have organized around since the Second World War. Cloud cover used to be a strategic asset. So did night. Both have been sold off, in increments a mid-sized company can expense, to anyone with a purchase order.
A state that cannot hide from a sensor it cannot buy, cannot deny and cannot outrun has one option left, and it is not a better tarpaulin.
What you do when you cannot hide
The counterspace ladder is fully populated and most of it is in daily use. The bottom rung is radio-frequency jamming and GPS spoofing, which CSIS documented as an ongoing daily occurrence in conflict-adjacent regions across both its 2024 and 2025 Space Threat Assessments, at levels visible to civilian aviation. Reports surfaced in 2025 of Starlink degradation attributable to Russian electronic warfare, which is worth sitting with, given that proliferated LEO is the American answer to nearly everything else on this list.
Above that sits the ground segment, which is softer than the spacecraft. Russia’s cyberattack on Viasat’s KA-SAT network went in one hour before troops crossed into Ukraine in February 2022 and bricked tens of thousands of modems across Ukraine and several European countries. A satellite you cannot command is a satellite you do not have.
Then kinetic. China destroyed its own Fengyun-1C weather satellite in January 2007 at roughly 865 kilometers and produced more than three thousand trackable fragments that are still up there. Russia’s Nudol interceptor destroyed Cosmos-1408 in November 2021 at about 490 kilometers, generated more than fifteen hundred trackable pieces, and put the crew of the International Space Station into shelter. Both countries continue development. General Saltzman testified in April 2025 that China is investing across all six categories of counterspace capability the Space Force tracks, and used the phrase grave threat.
Then co-orbital, which is the category that keeps planners awake, because a satellite closing to within a few kilometers of another satellite looks identical whether it intends to photograph, jam, dazzle or grapple. China’s SJ-21 was observed in 2022 towing the defunct Compass G2 out of the geostationary belt into a graveyard orbit, which is a debris-mitigation service and a demonstration of capture at the same time. Russia’s Luch and Luch-2 have loitered near Western communications satellites in GEO throughout the war. The United States does the same thing from the other side; GSSAP-4 closed to within 29 kilometers of China’s Shijian-20 in July 2021, and Shijian-20 moved away inside a day.
The nuclear option, stated precisely
In February 2024 the White House confirmed, after being pushed to it by members of the House Intelligence Committee, that Russia is developing a nuclear anti-satellite weapon intended for orbit. John Plumb, then Assistant Secretary of Defense for Space Policy, told the House Armed Services Committee in May 2024 that the concern was Russia developing the ability to fly a nuclear weapon in space. Mallory Stewart of the State Department noted that the satellite of interest sat in a high low-Earth orbit in a region no other spacecraft uses, which she called unusual on its own.
Precision matters here, because this claim gets inflated in both directions. It is a weapon for killing satellites, by detonating a nuclear device in orbit and using the electromagnetic pulse and the resulting persistent radiation belt to disable everything in that shell of space. The Secure World Foundation’s assessment is that it has been neither deployed nor tested and that it does not threaten terrestrial facilities. It is not a bomb parked over your city. In April 2025 the satellite believed connected to the program, Cosmos-2553, appeared to be tumbling uncontrollably and may no longer be functional.
The counterargument deserves its hearing. The National Security Space Association has argued that the reassurance about human safety is wrong, on the grounds that the EMP Commission assessed a high-altitude nuclear detonation as capable of producing widespread and long-lasting damage to critical ground infrastructure. That dispute is unresolved in public, and anyone telling you it is settled in either direction is selling something.
What is not in dispute is what it does to the architecture. Stationing it would violate the Outer Space Treaty of 1967 outright, which is why Russia vetoed a United States and Japan resolution at the Security Council in April 2024 that did nothing more than reaffirm the existing ban. The detonation would be indiscriminate. It would take out Russian satellites alongside American ones and endanger crews on both the International Space Station and the Chinese station. And it defeats, specifically and completely, the one architectural move the United States has spent two years making. Two hundred cheap satellites absorb a kinetic ASAT beautifully; you cannot fire enough interceptors to matter, and the debris hurts the attacker as much as the target. Two hundred cheap satellites in the same orbital shell absorb a nuclear detonation not at all. Proliferation answers every threat on this list except that one, and Russian planners read the same NRO press releases the rest of us do.
The one that can actually reach you
If you want the weapon that genuinely arrives anywhere, it is not in orbit and it is not new. Fractional Orbital Bombardment puts a warhead into a partial orbit and deorbits it onto a target from an unexpected azimuth, typically over the South Pole, where American early warning radar coverage was never built out because nobody expected an attack from that direction. The Soviets fielded the concept in the 1960s. Russia’s Ministry of Defense announced a second successful test of the RS-28 Sarmat on 12 May 2026, a heavy ICBM it describes as FOBS-capable and able to carry Avangard hypersonic glide vehicles. That is a Russian claim and has not been independently confirmed.
It does not violate the Outer Space Treaty, because it stations nothing. It goes around once. What it defeats is not your roof, it is the geometry of the sensor network built to see it coming, which puts it in the same category as every other item in this piece.
The through line
All of it is downstream of one fact: the observing side won. Not through better optics, which stopped improving before most of the people reading this were born, but through persistence, through sensors that read chemistry and radar returns instead of visible light, through automated tasking, and through a commercial market that now sells a usable version of the capability to anyone with a credit card. Concealment as a military discipline is finished. What replaces it is either accepting continuous observation or attacking the observers, and the second option is the one being funded.
The mirror now on its way to L2 was ground to watch other people’s runways. It failed at that so comprehensively that the government gave it away, and it is about to be extraordinarily good at the other thing. There is a second one still crated in Rochester, and the reason nobody has found a use for it is that there is nothing left for a 2.4-meter mirror to see that a 2.4-meter mirror cannot already see. The competition moved on. It is about what happens after the photons land, and about what you do concerning a photograph you have no way to prevent.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment, a Las Vegas enterprise software and AI consultancy. He has shipped production systems on the Microsoft stack for nearly thirty years for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air, was employee number six at Break.com during the user-generated content era, and was a Los Angeles Business Journal CTO of the Year finalist. He builds production agentic systems: MCP servers, retrieval pipelines, SQL Server 2025 vector search, and the PLogger observability framework. He is the author of the novel The Archive of Lost Suns. He writes here weekly about the distance between what executives say about AI and what practitioners actually put into production, which turns out to be roughly the same distance as the one between what people believe a reconnaissance satellite can see and what the atmosphere permits.