Issue 5: Five weeks in — and AI4Space goes space-borne
Last week, the orbital data center category showed up as hardware; this week it got a launch date and a potential government customer, and the AI4Space side stopped describing models and started flying them.
On the AI4Space side, Loft Orbital's YAM-9 was reported to have run Google DeepMind's Gemma 3 vision-language model in orbit via NASA JPL's NAVI-Orbital stack - the first publicly reported in-orbit VLM deployment. Around that flight, ESA's Φ-lab took Best Paper at EarthVision 2026 for a reliability-aware geospatial foundation model, PiLogic signed on with AFRL to test AI satellite anomaly detection, and an arXiv review mapped the state of ML for solar energetic particle prediction.
On the Space4AI side, SpaceX used its Nasdaq IPO day to put a late-2027 date on the AI1 orbital data-center satellites, with compute first riding Starlink V3 as a canary. Days later, a bipartisan Senate bill proposed letting national security missions buy capacity on commercial orbital data centers - turning the category from a thesis into a procurement line.
The hardware getting a delivery date and the models actually running in orbit are converging on the same year.
Specifics below.
AI4Space
Loft Orbital's YAM-9 runs Google DeepMind Gemma 3 VLM in orbit via NASA JPL NAVI-Orbital — first publicly reported in-orbit VLM deployment
On June 15, TechCrunch reported that Loft Orbital's YAM-9 satellite ran Google DeepMind's Gemma 3 — a vision-language model purpose-built for edge hardware — in orbit during an April 2026 demonstration, executed via NASA JPL's NAVI-Orbital software. The flight is described as the first publicly reported in-orbit VLM deployment.
The model ran on an NVIDIA Jetson Orin AGX aboard YAM-9 at 500 km altitude, with LangGraph handling onboard agent orchestration. Asked to do things like flag where natural environment meets human development or pick out infrastructure around railway hubs, the satellite classified scenes and answered the queries onboard without downlinking the raw imagery. Loft frames it as a step toward "always-on patrol layers in space" — monitor a border and report what looks suspicious — and says global real-time coverage would take 50 to 100 such satellites, against the 12 it flies today.
SHRUG-FM geospatial foundation model wins Best Paper at EarthVision 2026
On June 16, ESA's Φ-lab announced that SHRUG-FM — Systematic Handling of Real-world Uncertainty for Geospatial Foundation Models — won the Best Paper Award at the EarthVision 2026 workshop, held at CVPR 2026, where it was presented on June 4. The work originated in the 2025 FDL Earth Systems Lab (ESL) "Foundation Models for Extreme Environments" sprint.
The problem it targets is foundation models that are "confidently wrong" in conditions underrepresented during training — dangerous in time-sensitive uses like disaster response. SHRUG-FM lets a model issue a prediction, raise a warning, or "shrug" and abstain when it is highly uncertain, separating data-driven from model-driven uncertainty, and showed consistently lower prediction risk on burn-scar segmentation, flood mapping, and landslide detection. Its win signals an EO/ML community shifting focus from raw capability to trust.
Source: ESA Φ-lab
PiLogic partners with AFRL on AI satellite anomaly detection
Also on June 16, Payload reported that LA-based PiLogic has partnered with the Air Force Research Laboratory (AFRL) under a no-cost Cooperative Research and Development Agreement to test AI-based anomaly detection for satellites — giving the startup access to test equipment at Kirtland Air Force Base, a path to demo for military customers, and sponsorship for a TS/SCI clearance.
PiLogic builds a model for each specific satellite, then ingests its telemetry to catch failures from internal faults, electronic warfare, cyberattacks, or space weather — and can respond autonomously, alert an operator, or recommend an action for a human to approve. CEO Johannes Waldstein said the system narrows down causes by process of elimination but stops short of naming a culprit satellite or nation: "It's very smart, but it's not magic." The work currently runs on two terrestrial flat sats — ground replicas of spacecraft — with the partners aiming to extend it into orbit.
arXiv review surveys ML models for solar energetic particle prediction
On June 17, a large multi-institution review led by NASA and Princeton's Spiridon Kasapis posted to arXiv (2606.19539) surveyed machine learning models developed for solar energetic particle (SEP) event prediction, framing the work around radiation hazards to aviation, spacecraft electronics, and crewed missions beyond Earth's magnetosphere.
It categorizes 24 ML models by architecture, inputs, and outputs — spanning from a two-parameter gradient-boosting model to a 285-million-parameter neural network, with 21 of the 24 trained on NOAA GOES data — and flags the field's recurring obstacles: SEP events are rare, the resulting class imbalance is severe, and the input data is heterogeneous and sparse. Most efforts remain proofs of concept, and the authors lay out practices meant to push SEP forecasting toward operational use.
Space4AI
SpaceX confirms AI1 orbital data-center satellites for late 2027, with Starlink V3 compute canary
On June 12 — the day SpaceX debuted on the Nasdaq under the ticker SPCX in the largest IPO in Wall Street history, pricing 556 million shares at $135 to value the company at $1.77 trillion — COO Gwynne Shotwell told CNBC that the company's AI1 orbital data-center satellites will launch in late 2027, with AI compute first flown on Starlink V3 broadband and mobile satellites as a canary deployment. "We love doing kind of canary sets and canary work before we fly the real thing," she said.
Each AI1 satellite is specified at 120 kW average and 150 kW peak compute across a 70-meter wingspan. SpaceX plans to rent that capacity out as it does with its terrestrial data centers — it has struck recent compute deals with Anthropic and Google, and folded in Elon Musk's xAI in February — and Shotwell said the company "100 percent" sees itself as a competitor to the neoclouds. It has already filed for a megaconstellation of up to one million satellites, with the Starlink V3 canary as the on-orbit proving ground ahead of the dedicated AI1 bus.
Senators introduce bipartisan NEW HORIZON Act to allow national security missions to use commercial orbital data centers
On June 16, Sens. Ted Cruz (R-TX) and John Hickenlooper (D-CO) — both on the Senate Commerce Committee, which Cruz chairs — introduced the NEW HORIZON Act, a bill that would let national security missions tap into commercial orbital data center capacity. The backronym unpacks as Nodes, Enterprise Workloads, and Hybrid Operations, Resilience, Integration, Zero-Trust, Orbital Networks.
Rather than buy capacity outright, the bill directs the Pentagon to stand up a pilot program at the Defense Innovation Unit within a year of passage — assessing how useful in-space storage and processing would be, whether it integrates with existing comms and intelligence systems, and how resilient (or vulnerable) orbital data centers are. Cruz framed it as a way to cut latency and exploit data "generated in space [that] goes underutilized because of network bandwidth issues." A likely vehicle is the FY27 NDAA, with findings due to Congress by the end of 2028.
Till next time,
Meta-beat Column of this week
Read also about the AI Pipeline that sits at the core, producing this Newsletter, including its ups and downs of this week: