FundingRounds, leads & terms
Satlyt runs Google's Gemma on orbiting satellites to diagnose faults and shrink downlinked data after $8M seed
Satlyt, a software-only startup that recently raised an $8M seed, runs a quantized Google Gemma model on satellites already in orbit to diagnose onboard faults and shrink downlinked data, a first step toward pooling many operators' spacecraft into orbital compute.
From the This Week in Startups video of November 12, 2026 and cited sources: A quantized Gemma 3 1B model, run through llama.cpp, reads system logs and stack traces from onboard image-processing jobs. [2] [1]
SOURCES[2] How Satlyt is building local AI for space with Gemma –… deepmind.google[1] Running Google’s Gemma on Satlyt’s orbital AI… youtube.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- A quantized Gemma 3 1B model, run through llama.cpp, reads system logs and stack traces from onboard image-processing jobs. [2] [1]
- Original source
- This Week in Startups video, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 12, 2026 video published
- Covered
- Nov 13 edition
- Evidence
- 11 sources across 10 sites; transcript sourced
- 1Running Google’s Gemma on Satlyt’s orbital AI…youtube.com
- 2How Satlyt is building local AI for space with Gemma –…deepmind.google
- 3NASA award 80NSSC25C0340 – Federal Compassfedconsult.federalcompass.com
- 4Satlyt raises $8M seed – Payloadpayloadspace.com
- 5Satellite software provider Satlyt closes $8M – SiliconANGLEsiliconangle.com
- 6Satlyt Secures $8M in Seed Financing – Via Satellitesatellitetoday.com
RisingHow it’s measured
YouTube video · 1.9K views · 16 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
A quantized Gemma 3 1B model, run through llama.cpp, reads system logs and stack traces from onboard image-processing jobs. In two injected-error scenarios it cut diagnostic payloads by about 64% (for example, 1,319 to 469 bytes) at roughly 23–26 tokens per second. These figures come from a case study co-published by Satlyt and Google, which supplies the model. [2] [1]
Satlyt builds no hardware. Its product is a Kubernetes-style layer that places containerised services such as telemetry analytics, compression and encryption across satellites with different designs, within tight power budgets (15 W per node in the demo). The next model under test, Gemma 4 E2B on an 8 GB Jetson Orin Nano, peaked near 4 GB of RAM and about 11 W. [1] [2]
The $8M seed round, led by non sibi ventures, is reported consistently by three outlets, and its valuation was not disclosed. The size of Satlyt's NASA work is disputed: one outlet describes a $4M STTR contract, but the federal award record shows a Phase I contract worth up to $179.5k. [4] [5] [6] [3]
So far, the evidence shows smaller diagnostic payloads on one class of edge hardware.
Read the full assessment
If that holds up, ground-station fees and power limits could make small onboard models a practical pattern, and a software-only layer could earn revenue from other operators' satellites. Neither implication is proven yet.
Executive brief
Satlyt is running a quantized 1-billion-parameter Google model on a satellite that is already in orbit, and Satlyt does not build any satellite hardware. Its software runs on other companies' spacecraft. Google DeepMind's case study, written with the company, says Gemma 3 1B reads onboard error logs and shrinks the diagnostic data sent to Earth by about 64% (DeepMind). In a 12 November episode of This Week in Startups' AI Basics series, which Google for Startups sponsors, CEO Rama Afullo presented this as the first step toward "virtual" orbital data centres built from many operators' satellites (video, 21:24). The company announced an $8M seed round on 1 October.
What changed and event timeline
NASA STTR Phase I award
NASA gave Satlyt a Phase I contract worth up to $179.5k for "delay-tolerant mobile edge cloud" work (). Payload describes a $4M STTR contract, which does not match that record ().
Satellite with Satlyt software launches on Momentus
SiliconANGLE lists Momentus as a customer whose satellite launched in March ().
Gemma inference in orbit
Payload reports a May demonstration that ran Gemma on an orbiting satellite, processed imagery, produced encrypted downlink output and diagnosed software faults ().
$8M seed
The company announced the round, led by non sibi ventures. Valuation was not disclosed (;).
Workload-scheduling test
Under NASA-funded work with the University of Houston, one spacecraft computer assigned tasks across a simulated network of four hosts ().
Podcast demo
Afullo walked through the mission-operations software and said the Gemma experiment has run on "a couple of satellites" ().
Capabilities and access
- Deployed: Gemma 3 1B, quantized and run through llama.cpp. It analyses system logs, software errors and stack traces from onboard image-processing jobs (DeepMind).
- Being evaluated: Gemma 4 E2B on an NVIDIA Jetson Orin Nano with 8 GB of RAM (DeepMind).
- Other managed services shown in the demo: a container registry, telemetry analytics, compression, encryption, and placing services on specific nodes (video, 18:26).
Read the full section
- Deployed: Gemma 3 1B, quantized and run through llama.cpp. It analyses system logs, software errors and stack traces from onboard image-processing jobs (DeepMind).
- Being evaluated: Gemma 4 E2B on an NVIDIA Jetson Orin Nano with 8 GB of RAM (DeepMind).
- Other managed services shown in the demo: a container registry, telemetry analytics, compression, encryption, and placing services on specific nodes (video, 18:26).
- Access: commercial "packages" booked through satlyt.ai (video, 31:55).
Technical analysis
- How it is built: a Kubernetes-style orchestration layer that runs containers across satellites with different hardware designs, within power limits (15 W per node in the demo) (video, 19:57).
- How it was evaluated: common software errors were deliberately injected into an image pipeline.
- Gemma 4 E2B profile (4-bit Q4_K_M): about 4 GB peak RAM, about 11 W during inference, 3–5 °C temperature rise, 19.08 tokens per second (DeepMind).
Read the full section
- How it is built: a Kubernetes-style orchestration layer that runs containers across satellites with different hardware designs, within power limits (15 W per node in the demo) (video, 19:57).
- How it was evaluated: common software errors were deliberately injected into an image pipeline. The model's summaries cut the diagnostic payload from 1,319 to 469 bytes (64.4%) and from 1,318 to 464 bytes (64.8%), generating 22.7–25.5 tokens per second (DeepMind).
- Gemma 4 E2B profile (4-bit Q4_K_M): about 4 GB peak RAM, about 11 W during inference, 3–5 °C temperature rise, 19.08 tokens per second (DeepMind).
- What is missing for reproducibility: no published code, error set or diagnostic-accuracy metric. The two test scenarios measure compression, not whether the diagnoses were correct.
Claims and evidence
- About 64% reduction in downlinked telemetry: this comes from the company's own case study, co-published with Google, which also supplies the model (DeepMind).
- "Only 11.7% of data was useful": this number appeared in the podcast demo interface and is not a published result (video, 22:02).
- $8M seed led by non sibi ventures: announced by the company and reported consistently by three outlets (Payload; Via Satellite; SiliconANGLE).
Read the full section
- About 64% reduction in downlinked telemetry: this comes from the company's own case study, co-published with Google, which also supplies the model (DeepMind). SiliconANGLE repeats it but did not test it (SiliconANGLE).
- "Only 11.7% of data was useful": this number appeared in the podcast demo interface and is not a published result (video, 22:02).
- $8M seed led by non sibi ventures: announced by the company and reported consistently by three outlets (Payload; Via Satellite; SiliconANGLE). No regulatory filing was found to confirm it. Total funding "well past $10M" is the company's own figure.
- Goals of 50 spacecraft deployments in 2026 and 30% of non-Starlink satellites by 2030: company targets only (DeepMind; Payload).
Context and prior work
- Earlier onboard AI: other satellites have recently detected targets on their own using vision-language models (TechCrunch).
- Hardware-first rivals: Google's own orbital data-centre satellite (TechRepublic) and the startup Orbital, which raised a $5M pre-seed (TNW).
- Satlyt's positioning: it says it routes data and does not compete with companies building orbital data-centre hardware (Payload).
Read the full section
- Earlier onboard AI: other satellites have recently detected targets on their own using vision-language models (TechCrunch).
- Hardware-first rivals: Google's own orbital data-centre satellite (TechRepublic) and the startup Orbital, which raised a $5M pre-seed (TNW).
- Satlyt's positioning: it says it routes data and does not compete with companies building orbital data-centre hardware (Payload).
- Networking base: the delay-tolerant networking work builds on NASA Glenn's High-Rate DTN project (NASA NTRS).
Limitations, safety and contested findings
- NASA contract size is disputed: Payload says $4M, while the federal record shows a $179.5k Phase I award (Payload; Federal Compass).
- Multi-satellite claims are early: the scheduling test used one physical computer and four simulated hosts, and NASA has not selected Satlyt for a mission (TechCrunch).
- Hardware is not radiation-hardened: Afullo says the Jetson units fly with spinning fans, and that longer lifetimes need radiation shielding and costly testing (video, 30:33).
Read the full section
- NASA contract size is disputed: Payload says $4M, while the federal record shows a $179.5k Phase I award (Payload; Federal Compass).
- Multi-satellite claims are early: the scheduling test used one physical computer and four simulated hosts, and NASA has not selected Satlyt for a mission (TechCrunch).
- Hardware is not radiation-hardened: Afullo says the Jetson units fly with spinning fans, and that longer lifetimes need radiation shielding and costly testing (video, 30:33).
- Disclosure: Google for Startups sponsors the episode (video, 32:39).
Business and practitioner implications
- Cost case: every byte not sent down saves money. Afullo puts ground-station fees at $10–20 per GB (video, 22:09).
- Launch economics: he cites rideshare launches at about $15,000/kg against roughly $100/kg needed for orbital compute to match ground data centres (video, 16:29).
- For developers: the practical pattern is small quantized models on Jetson-class hardware, packaged in containers and kept within 11–15 W.
Read the full section
- Cost case: every byte not sent down saves money. Afullo puts ground-station fees at $10–20 per GB (video, 22:09).
- Launch economics: he cites rideshare launches at about $15,000/kg against roughly $100/kg needed for orbital compute to match ground data centres (video, 16:29). These figures are his and were not independently checked.
- For developers: the practical pattern is small quantized models on Jetson-class hardware, packaged in containers and kept within 11–15 W.
- For investors: Satlyt is a software-only bet that depends on other operators' satellites. Valuation and revenue are undisclosed.
The source trail.
Sources (11)
Running Google’s Gemma on Satlyt’s orbital AI (w/ CEO Rama Afullo)
Transcript retrieved via youtube_auto_captions; language en. Timestamped text, not direct audiovisual review. Source: https://www.youtube.com/watch?v=fEwH9Z-wwQQ. Automatic captions/transcription may contain errors.
www.youtube.com