The insurance adjuster's nightmare scenario isn't the hurricane itself. It's the three days after, when clouds still blanket the coast and optical satellites—those billion-dollar eyes in the sky—can't see a thing.
This happens more often than you'd think. At any given moment, something like 70% of Earth's surface sits under cloud cover. Throw in nighttime, and the monitoring windows for traditional satellites narrow to the point where calling them "continuous" becomes generous. For industries that live and die by real-time intelligence—insurers racing to process claims, defense analysts tracking troop movements, supply chain managers watching cargo ships—it's a reliability problem that compounds with every passing hour.
Synthetic Aperture Radar satellites don't care about clouds. Or smoke. Or darkness. They've been the all-weather workhorse for decades, punching through atmospheric interference that grounds their optical cousins. The catch? SAR data is nearly unintelligible unless you happen to be a radar specialist. What you get isn't a crisp photograph—it's backscatter intensity, a kind of digital static that requires serious expertise to decode. A flooded neighborhood reads like noise. A port full of shipping containers looks like visual static.
Enter AxionOrbital Space, a two-person startup from Y Combinator's Winter 2026 batch that thinks AI can fix this. Their pitch is refreshingly direct: train foundation models to translate SAR's radar gibberish into something that looks like Google Earth. Make it interpretable, make it fast, and suddenly radar becomes the default for anyone who can't afford downtime.
Whether they can deliver on that promise at scale remains an open question. But the timing suggests they're onto something.
The Physics Problem That Won't Go Away
CEO Dhenenjay Yadav spent time in ISRO's machine learning group before this. CTO Atharva Peshkar holds a PhD from CU Boulder and logged hours at Harvard's Visual Computing Group. They're building what they describe as "foundation models for all-weather Earth observation"—systems designed to ingest raw SAR data and spit out imagery that looks optical, that analysts can actually work with.
AxionOrbital markets two models publicly. Orion, their restricted high-res system, operates around 0.5 meters. Hubble, released as open-source, works at roughly 10 meters. According to the company's Y Combinator materials, Orion scored 30.24 on the Fréchet Inception Distance benchmark using the MSAW dataset—a standard test for SAR-to-optical translation. That reportedly beat the previous state-of-the-art model, C-DiffSET, by about 19 points. They also cite an SSIM score of 0.60.
The company claims their models use "deterministic one-step diffusion" for real-time inference, with processing speeds under 0.1 seconds mentioned in a March 2026 LinkedIn post. Independent verification of these performance numbers isn't publicly available yet, which is typical for early-stage startups but worth noting.
Their go-to-market strategy targets sectors where weather-independent monitoring commands premium pricing: defense contractors, financial traders watching commodity flows, disaster response teams, agricultural monitors. Deployment comes in three flavors: API access, on-premises installations, or edge computing for latency-sensitive applications.
Radar's Coming-of-Age Moment
Here's why the timing matters. SAR is no longer some niche government capability tucked away in classified budgets. It's becoming a real commercial business, and the revenue numbers are starting to look serious.
ICEYE, the Finnish operator that's become something of a bellwether for the sector, reported around €250 million in revenue for 2025. More striking: they disclosed a €1.5 billion backlog during a March briefing this year. The company pulled in €150 million last December at a €2.4 billion valuation and is openly targeting €1 billion in annual revenue within the next few years. They're scaling their fourth-generation constellation and cutting sovereign deals—including a partnership with Japan's IHI—that suggest national governments see SAR as strategic infrastructure.
Umbra, which made waves in August 2023 with a 16-centimeter SAR image (still the sharpest commercial radar shot on record), opened a new manufacturing facility in 2025. In February they announced a $6.75 million expansion into Virginia. They also deepened ties with e-GEOS late last year, signaling a push toward integrated service offerings.
Capella Space secured an AFWERX STRATFI contract back in September 2024 to accelerate next-gen SAR development, building on $60 million in growth equity from mid-2023. When the Department of Defense starts writing checks for commercial radar tech, it's usually a sign the sector's crossed from experimental to essential.
On the government side, the infrastructure is maturing fast. The European Space Agency's Sentinel-1C went operational in December 2024; Sentinel-1D launched this past November. Together they ensure C-band SAR continuity through the 2040s under the Copernicus program. NASA and ISRO's NISAR mission, which launched last July, adds dual-frequency L-band and S-band coverage for tracking everything from ground deformation to forest biomass. First-light data activities kicked off late last year.
The pieces are falling into place for SAR to become default infrastructure rather than specialty capability. Maybe more than the industry anticipated just a few years back.
The Foundation Model Gold Rush
AxionOrbital isn't working in a vacuum. Earth observation is having its foundation model moment, and the pace of releases has been relentless.
NASA and IBM dropped Prithvi-EO-2.0 in December 2024—a multi-temporal transformer pretrained on harmonized Landsat-Sentinel data. The model claimed state-of-the-art performance across multiple remote sensing tasks and landed on Hugging Face for anyone to download. NASA Goddard followed with SatVision TOA, a cloud-focused foundation model, in late 2024. Researchers introduced TerraTorch in March 2025, a toolkit specifically for fine-tuning and benchmarking geospatial foundation models.
A survey paper in Nature Reviews Electrical Engineering last August laid out the emerging landscape of multimodal remote sensing models. The authors emphasized physical grounding, uncertainty quantification, and evaluation methods that go beyond simple pixel-level metrics. NASA and ESA have been running joint workshops on multimodality and physical principles in model design, with discussions edging toward "agentic" AI systems that can reason about Earth observation tasks autonomously.
The training infrastructure is getting better, too. The Copernicus Data Space Ecosystem—launched in 2023 and continuously updated—provides petabyte-scale access to free satellite data. That's critical for pretraining large models; you can't build these systems without massive datasets. The Committee on Earth Observation Satellites established Analysis Ready Data specs, and NASA's OPERA program certified Sentinel-1 radar products with standardized metadata last November, making it easier to plug SAR into optical workflows.
Some of this is even heading into orbit. ESA's Φ-Sat-2 began its science phase mid-2025 to test onboard AI processing. Planet announced plans last June to fly NVIDIA Jetson processors on its Pelican constellation. Cosmic Shielding and Aethero demonstrated radiation-tolerant Jetson Orin hardware in July 2024. Academic teams reported the first on-orbit demo of a geospatial foundation model late last year.
The infrastructure is there. The question is whether the models can deliver on operational promises.
Stress-Tested in the Storm Zone
The value proposition for all-weather monitoring isn't hypothetical. Disasters have a way of clarifying what works.
When Hurricane Helene tore through the Southeast in September 2024, NASA's OPERA SAR products mapped flooding while clouds still grounded optical satellites. ICEYE delivered building-level inundation analysis for both Milton and Helene within roughly 24 hours—providing wind and flood intelligence to insurers while conditions were still too dangerous for aerial surveys. They launched a "Flood Rapid Impact" product in July 2025 designed specifically for near-real-time flood intelligence. Partnerships followed: MAPFRE RE last August, Munich Re's location risk platform this January.
Even Planet, which built its business on optical satellites, introduced "Fusion with SAR" back in October 2021 to plug visibility gaps. That's a tacit admission that optical alone can't meet enterprise reliability standards, especially in the tropics and humid regions where cloud cover is persistent and supply chains run dense.
These aren't pilot projects anymore. They're operational workflows with money changing hands.
The Devil in the Evaluation Details
But there's a technical asterisk worth acknowledging. SAR-to-optical translation isn't a solved problem, and the evaluation methods have blind spots.
The field relies heavily on pixel-level metrics—PSNR, SSIM, LPIPS—borrowed from image generation. Multiple papers from the past two years caution that strong pixel scores don't necessarily translate to semantic correctness for downstream tasks. You can generate an image that looks sharp but misrepresents actual ground conditions. Registration errors and temporal mismatches between SAR captures and optical reference images compound the challenge.
Benchmarks are improving. The SpaceNet-6 MSAW dataset, released back in 2020, remains widely used. A newer dataset called SkyCap, released in 2025, provides high-resolution SAR-optical image sets for change detection and model evaluation. Still, practitioners worry about "hallucination" risk—if users treat generated optical-like outputs as ground truth without independent verification, errors creep into critical decisions.
NASA and ESA workshop discussions over the past two years have pushed for task-based evaluation, physical consistency checks, and explicit uncertainty quantification. The shift matters because these models are leaving the lab. Operational contexts don't tolerate elegant failures.
Market Forces and Export Headaches
The broader Earth observation market is on track to roughly double by the mid-2030s. Fortune Business Insights pegged the 2025 market around $7.0 to $7.1 billion in reports published this February, projecting growth to somewhere between $14.5 and $15.9 billion by 2034 or 2035—call it an 8.3% compound annual growth rate. Grand View Research, in a February refresh, projected growth from a $5.1 billion baseline in 2024 to $7.24 billion by 2030.
The variance reflects different definitions of scope—space segment versus data versus analytics—but the direction is consistent. Demand is being driven by climate accountability mandates, defense modernization budgets, and the maturation of cloud-based analytics infrastructure. Sovereign constellation deals, like ICEYE's partnership with IHI, suggest a shift toward localized capacity and data sovereignty concerns.
Export controls remain messy. A 2020 overhaul streamlined US commercial remote sensing licensing and removed many legacy SAR restrictions. But industry feedback in 2024 argued that proposed ITAR and EAR changes for high-resolution SAR remained too restrictive relative to foreign availability. Legal analyses from 2025 flagged potential tightening around AI models and training data, which could complicate cross-border deployments.
The Copernicus program's free, full, open data policy creates an interesting contrast. It provides a stable, unrestricted training corpus and has become the de facto foundation for many geospatial AI efforts. The policy asymmetry creates winners and losers in who can build these systems at scale.
When the Default Shifts
The signals point in one direction. All-weather monitoring is moving from specialty capability to baseline expectation.
With Sentinel-1C and Sentinel-1D ensuring C-band continuity through the 2040s, NISAR adding L-band and S-band science coverage, and commercial operators like ICEYE projecting billion-euro revenue within reach, SAR isn't filling gaps anymore. It's infrastructure. The question is whether foundation models can solve the usability problem—whether they can make radar data immediately legible to analysts who don't have radar engineering PhDs.
AxionOrbital's bet is that their models build the "interpretability bridge," translating radar physics into human-readable intelligence. Whether that bridge holds under operational stress—across diverse geographies, seasonal variations, and sensor configurations—still needs demonstrating at scale. Early metrics look promising, but early metrics often do.
The broader industry trajectory, though, seems clear enough. Insurance companies are baking SAR into catastrophe workflows. Defense budgets are funding next-gen radar constellations. Satellite operators are flying AI processors to process data in orbit, cutting latency and bandwidth costs. And foundation model research is migrating from academic preprints to production deployments faster than many expected.
The 70% visibility gap isn't going anywhere. Cloud cover is atmospheric physics; you can't negotiate with thermodynamics. What's changing is the technology stack built around that constraint.
Perhaps the real question isn't whether SAR becomes the default for critical monitoring. It's how quickly the translation layer gets good enough that users stop noticing they're looking at radar in the first place.
