Point a camera at Earth from orbit and, most of the time, you get a photograph of a cloud. Weather and nightfall block conventional optical satellites for roughly 70% of any given stretch - a fact the imagery industry has largely accepted as the cost of doing business. AxionOrbital Space, a two-founder company in San Francisco, decided that number was a market rather than a limitation.
The company builds foundation models for Earth observation. Its first, called ORION, does something specific and slightly counterintuitive: it takes radar data - the kind that already passes through clouds - and translates it into optical imagery a human can read at a glance. Radar has always seen through weather. The problem was that its raw output, Synthetic Aperture Radar backscatter, looks like static to anyone who isn't a remote-sensing specialist. AxionOrbital's bet is that the breakthrough isn't a new sensor in space. It's a translator for the sensors already up there.
The problemThe two-thirds of the time nobody can see
Consider who cares when the ground disappears. A commodities trader wants to know how full an oil terminal is before the market does. A defense analyst is tracking vehicles that don't stop moving when it gets cloudy. A disaster team needs a flood map precisely when the storm is still overhead. All three run into the same wall, and it's usually made of weather.
The blind majority. Standard optical imaging is useful only in the thin slice of time when the sky cooperates. Radar doesn't care about the sky - it just needed a translator.
The conventional fix is to launch more satellites and wait for a clear pass. That is expensive and, on a cloudy week, futile. AxionOrbital's fix is to stop waiting - convert the radar feed that's already coming down into an optical picture in real time.
The productWhat ORION actually does
ORION reads radar and writes photographs. Under the hood it uses what the company describes as a deterministic one-step diffusion architecture - a mouthful whose practical payoff is speed and consistency. It renders in about 0.06 seconds, roughly a blink, and it does so without inventing detail. The output is what AxionOrbital calls physically anchored: every feature in the generated image is meant to trace back to the actual radar signal, not to a model's imagination. For customers making million-dollar or life-and-death calls, that distinction between "grounded" and "plausible-looking" is the entire product.
Radar in, picture out. The sensor stays the same. What changes is that the signal arrives as something a person can interpret without a PhD in remote sensing.
The proofBenchmarks as the opening argument
Deep-tech companies rarely get to lead with a customer logo. They lead with numbers. On the MSAW benchmark for SAR-to-optical translation, AxionOrbital reports ORION at a 30.24 FID score - about 19% better than the prior published state of the art, a method called C-DiffSET - along with a new-best SSIM of 0.60. FID measures how close generated images are to real ones; lower is better. SSIM measures structural similarity; higher is better. Read together, they're the company's first sales pitch, delivered before a single contract.
C-DiffSET
Lower is better, so shorter would normally win. Here the bar shows relative standing on the benchmark - ORION reports the leading FID, with C-DiffSET trailing by about 19%.
The customersThree rooms, one blind spot
AxionOrbital's early markets don't obviously belong together, which is part of the point. A capability that removes a shared constraint tends to find buyers in rooms that never speak to each other.
Situational awareness
Continuous tracking of movement that doesn't pause for weather or nightfall.
Market alpha
Hedge funds and commodity traders reading supply chains - terminals, ports, fields - in any conditions.
Disaster mapping
Flood and crop monitoring exactly when the storm is overhead and optical is useless.
There's a fourth audience worth noting: the satellite data providers and analytics firms who could fold an all-weather optical layer into what they already sell. In that sense AxionOrbital is less a competitor to the imagery industry than a missing utility for it.
The teamFrom a national space agency to a two-person startup
AxionOrbital was founded in 2025 by Dhenenjay Yadav and Atharva Peshkar. Yadav, the CEO, previously worked as a machine-learning engineer at ISRO, India's national space agency, and as a reinforcement-learning researcher at IIM Ahmedabad - someone who watched the blind-satellite problem from the inside. Peshkar, the CTO, holds a computer-science PhD from CU Boulder and researched computer vision at Harvard's Visual Computing Group. Space-agency engineering on one side, academic vision research on the other; the model sits at the seam between them.
The marketA software layer on top of orbit
Most of the well-known names in Earth observation either operate satellites or sell the pictures those satellites take. AxionOrbital positions itself one rung up the stack: a software model that makes existing radar feeds more valuable, sold to defense, finance, agriculture, and the data providers themselves. It competes, loosely, with SAR specialists and analytics layers, and it competes philosophically with the assumption that better vision means launching more hardware. Backed by Y Combinator's Winter 2026 batch and reportedly around $500K in early funding, it's still a very small team making a very large claim. The claim happens to come with a benchmark attached.
Whether "replace passive optical satellites" turns out to be a mission statement or marketing depends on customers the company hasn't named yet. But the framing is disciplined: AxionOrbital doesn't sell diffusion architectures. It sells the removal of a constraint everyone else priced in. The radar-to-optical machinery is how. Seeing the ground on a cloudy Tuesday is why.