Satlyt Raises $8M for AI on Satellites: Smart, Essential Bet

Satlyt, a start-up with headquarters in Sunnyvale, California and Nairobi, Kenya, has raised an $8 million seed round to build software that runs AI models on satellites. Its chief executive, Rama Afullo, told TechCrunch he pitched orbital computing inside both Google and SpaceX before leaving to do it himself. “When I was at SpaceX, I tried to pitch this internally. They said no. When I was at Google, I tried to pitch this internally. They said no,” he said.

The timing is striking. On 1 October 2026, the same day the round was announced, Satlyt’s software flew on a SpaceX rideshare rocket that also carried the first prototype of Google’s Project Suncatcher space data centre. Unlike Google, SpaceX or Starcloud, Satlyt is not building spacecraft. It wants to be the software layer that lets many operators’ satellites run AI and, eventually, share computing work.

This article explains what Satlyt builds, what its Gemma experiment in orbit actually showed, what flew this week, who invested, how it compares with the big orbital computing projects, and what the plan means for anyone who buys cloud or edge computing on the ground.

Table of contents

What Satlyt Builds
Why Run AI on a Satellite at All
What Satlyt’s Gemma Experiment in Orbit Showed
What Flew on Transporter-18
Why Small AI Models Suit Satellites
Who Invested in Satlyt
How Satlyt Compares With the Orbital Computing Race
Satlyt’s Road to a Cloud in Orbit
What the Coverage Got Wrong
What Satlyt Means for Businesses on the Ground
Satlyt FAQ
References and Further Reading

What Satlyt Builds

Satlyt describes itself as “the above cloud service provider”. In practice it makes software for computers already on board satellites.

Software, not spacecraft

Afullo compares the company to VMware and Snowflake, whose platforms let customers run complex software without managing the underlying machines. Satlyt’s software is meant to sit on satellites built by other companies, decide what runs where, and keep AI workloads within the tight power, heat and memory budgets of a spacecraft.

“If they are the iPhone, we’ll build Android”

Afullo frames the market as a platform contest. “The folks like SpaceX who are doing orbital data centers, if they are the iPhone, we’ll build Android as a horizontally integrated, open ecosystem,” he told TechCrunch. The bet is that most satellite makers will not build their own computing stack and will want one that works across many fleets.

A managed service on someone else’s satellite

“The way to think about it is that we turn your satellite into a revenue-generating managed service,” Afullo said. A satellite operator with spare onboard computing could rent it to other customers through Satlyt’s layer, much as cloud providers rent spare server time on the ground.

Why Run AI on a Satellite at All

The case for onboard AI starts with a bottleneck: getting data down to Earth.

Downlink is slow and expensive

A satellite in low Earth orbit is usually within range of its operator’s ground antenna for only a few minutes each day, SiliconANGLE notes. Sending data down, known as downlink, is costly and slow, so satellites often collect far more than they can send. Processing data on board and sending only the useful result eases that squeeze.

Fixing faults without waiting for the ground

Spacecraft usually depend on flight controllers on the ground to diagnose problems. TechCrunch reports that Satlyt’s first focus is operational efficiency: letting the satellite analyse its own errors and sensor readings, so operators get a short diagnosis instead of raw logs. Afullo says that kind of saving can be worth hundreds of thousands of dollars per satellite each year.

The same logic as edge computing

On the ground this is called edge computing: process data where it is created when the link back to a data centre is the constraint. It is the same reasoning behind running machine learning on factory sensors or in shops rather than in the cloud. Our IoT solutions team applies it to devices on Earth; Satlyt applies it in orbit.

What Satlyt's Gemma Experiment in Orbit Showed

Satlyt’s most concrete evidence comes from a case study Google DeepMind published about its work with Gemma, Google’s family of small open AI models.

A one-billion-parameter model in orbit

Satlyt deployed a quantised Gemma 3 1B model on a satellite operated by Momentus, running through the open-source llama.cpp runtime. SiliconANGLE says that satellite launched in March 2026. The model read system logs, software errors and stack traces from onboard image-processing jobs and wrote a short root-cause diagnosis with a recommended fix.

Two fault-injection tests

Satlyt deliberately broke an image-processing pipeline and measured the result. In one test, the telemetry showed 99.6% memory use and an out-of-memory error from loading a 12,000 by 12,000 pixel image. In the other, a camera capture failed with an input/output error. The diagnosis Gemma wrote was much smaller than the raw data it summarised.

Diagnostic payload before and after onboard summarisation, in bytes (bar length relative to 1,319)

Test 1, raw diagnostic data: 1,319

Test 1, Gemma summary: 469

Test 2, raw diagnostic data: 1,318

Test 2, Gemma summary: 464

The reductions are (1,319 − 469) ÷ 1,319 = 64.4% and (1,318 − 464) ÷ 1,318 = 64.8%, the figures in Google DeepMind’s case study. TechCrunch rounds this to “more than 60%”. The model generated text at 22.71 and 25.48 tokens per second in the two tests.

Keep the scale in perspective

These are payloads of about 1.3 kilobytes. The test shows the method works on real flight hardware, not that Satlyt has already cut a large satellite’s downlink bill by two thirds. Larger savings would come from applying the same idea to images and sensor data, which is exactly what this week’s mission sets out to try.

Next: smaller models, bigger jobs

The case study says Satlyt is now testing Gemma 4 E2B on an Nvidia Jetson Orin Nano board. For some diagnostic tasks it is aiming for a more than 90% reduction in model memory, and an 85% to 90% cut in the memory each managed AI service needs, through pruning and distillation. It also says Satlyt has software packages prepared for up to 50 spacecraft deployments in 2026.

What Flew on Transporter-18

Satlyt’s second mission rode on SpaceX’s Transporter-18 rideshare flight, which lifted off from Vandenberg Space Force Base in California on 1 October 2026.

A crowded rocket

TechCrunch says the Falcon 9 carried more than 100 payloads; CNBC-TV18 put the count at 130. Alongside Satlyt’s software were Google’s first Project Suncatcher satellite, built on a Planet Labs platform, and a mission from Cowboy Space Company. We covered Google’s prototype in our look at Project Suncatcher’s first test of AI chips in space.

TakeMe2Space’s MOI-1A

Satlyt’s software runs on MOI-1A, a satellite from Hyderabad-based TakeMe2Space, an Indian start-up that builds computing hardware for satellites. Indian media describe it as the country’s first orbital computing satellite. It is TakeMe2Space’s second attempt: its first satellite was lost on an ISRO PSLV launch in January, ThePrint reported. The Next Web says MOI-1A carries Nvidia chips and has 23 customers.

Three customers on one satellite

TechCrunch reports three customers on the mission. NASA is paying Satlyt to test protocols for cloud computing in space; iAfrica says this is tied to Satlyt’s NASA Glenn small business research work with the University of Houston. Stellerian, a space surveillance start-up that tracks objects in orbit, wants to test computer vision workloads that process images on the satellite. TakeMe2Space wants to prove its satellite can host other companies’ software.

When	Milestone	Source
2024	Company founded by Rama Afullo after a stint at Starlink	SiliconANGLE, TechCrunch
March 2026	Gemma 3 1B runs on a Momentus satellite	SiliconANGLE, Google DeepMind
1 October 2026	Software flies on TakeMe2Space’s MOI-1A aboard Transporter-18; $8M seed announced	TechCrunch
2026	Packages prepared for up to 50 spacecraft deployments	Google DeepMind
2027	Attempt at a shared compute cloud spanning two satellites	TechCrunch
End of the decade	Goal: live on 20% of satellites	TechCrunch

Why Small AI Models Suit Satellites

Running AI in orbit is not like running it in a data centre. Google DeepMind’s case study lists the constraints that shape every design choice.

Every watt, byte and second counts

“Space systems are one of the clearest examples of why local AI matters,” the case study says, because operators cannot rely on cloud connectivity or afford to send every raw log home. A satellite’s computer shares a small power budget with its radios, sensors and pointing systems, has limited memory, and must shed heat without air to carry it away.

What quantisation, pruning and distillation do

Quantisation stores a model’s numbers at lower precision, which shrinks the memory it needs and speeds it up, at some cost in accuracy. That is how a one-billion-parameter model fits on a small board alongside the rest of the flight software. Pruning removes parts of a model a task does not need, and distillation trains a smaller model to copy a larger one. The case study says the company is using all three, and is removing “modality-specific components” a given job does not need.

Many small specialists, not one big model

The goal, according to the case study, is to run several specialised AI workloads at once within one constrained computer, for tasks such as software diagnostics, anomaly detection and summarisation. That runs against the trend on the ground towards ever-larger general models.

People stay in charge

The case study is explicit that spacecraft operators “retain command authority”. The onboard model recommends; it does not take control of the spacecraft. That matters for safety and for insurers, and it will matter to any operator deciding whether to let third-party software run on its flight computer.

Who Invested in Satlyt

Non Sibi Ventures leads

The seed round was led by Houston-based Non Sibi Ventures, where former NASA astronaut Bernard Harris is a partner. Partner Kent Lucas told TechCrunch the firm liked that the company does not depend on the most ambitious vision of space data centres: “We don’t need data centers in space for Rama to be wildly successful, right? It can just be driven by the number of satellites going up.”

A long list of African and US backers

iAfrica lists the other investors, several of them Africa-focused funds.

Investor	Role in the round
Non Sibi Ventures	Lead
TLCOM, Launch Africa Ventures, Enza Capital, Axian Investment, Askya Investment Partners	Participants (Africa-focused, per iAfrica)
Antler, Slauson & Co., Demos, BAG Collective, Gaingels	Participants
Existing backers	Follow-on

What the money is for

According to iAfrica, the funding will go to product development, hiring and deploying the software on more third-party satellites. The company keeps significant engineering in Nairobi, iAfrica notes, rather than moving it to the US.

The founder

Afullo is a Kenyan American engineer. He worked in Google’s cloud computing business, which SiliconANGLE describes as a cloud architect role, then spent a short time at SpaceX’s Starlink in 2024, where SiliconANGLE says he was a product manager. Satlyt is also one of the start-ups in TechCrunch’s Startup Battlefield at Disrupt, held on 13 to 15 October in San Francisco.

How Satlyt Compares With the Orbital Computing Race

Several far bigger players want to put AI computing in orbit. Satlyt’s approach is different from all of them.

Google’s long-term moonshot

Google’s Suncatcher prototype will run its Tensor Processing Units in 15-minute bursts to protect the satellite’s power and cooling, TechCrunch reports. A two-satellite demonstration linked by lasers is planned for next year, and Google’s long-term picture is a formation of 81 satellites working in parallel. Travis Beals, who runs the project, calls it a “long-term moonshot”. Our earlier piece on Google’s teensy-tiny space data centre puts that first satellite’s one-kilowatt power budget in context.

Launch is the bottleneck for big data centres

Google’s newly peer-reviewed paper, due in the journal Joule, expects launch prices near $200 per kilogram by 2035. TechCrunch’s reading of it is that getting there would need Starship to fly about 1,800 times over ten years, carrying 200 tonnes each time.

Starship launches per year: needed under Google’s cost curve versus the most flown so far (bar length relative to 180)

Needed: 1,800 launches over 10 years: 180 a year

Most Starship flights in any year so far: 5

The arithmetic is 1,800 ÷ 10 = 180 a year, and 5 ÷ 180 = 2.8%. That gap is why Non Sibi likes a company that earns money from ordinary satellites already flying, rather than waiting for orbital data centres.

A side-by-side view

Player	Approach	Builds its own spacecraft?
Satlyt	Software to run and share AI workloads across other companies’ satellites	No
Google Project Suncatcher	TPU clusters in close formation; prototype built with Planet Labs	With a partner
SpaceX	Orbital data centres alongside its launch business	Yes
Starcloud, Cowboy Space Company	Start-ups building dedicated computing spacecraft	Yes
TakeMe2Space	Computing hardware for satellites, hosting others’ software	Yes

Satlyt's Road to a Cloud in Orbit

Step one: useful work on single satellites

For now, few powerful GPUs are in orbit, TechCrunch notes, so the near-term business is the kind of work Satlyt already does: fault diagnosis, anomaly handling and processing sensor data on one spacecraft.

Step two: two satellites, one cloud

The next project, planned for next year, is a shared computing system spanning two different satellites. If that works, Satlyt could offer a true compute cloud in orbit as a third-party provider to spacecraft builders.

Step three: 20% of satellites

Afullo hopes to be live on 20% of satellites by the end of the decade, when he expects nearly every spacecraft builder to fit GPUs and similar processors. “If you’re putting up a satellite without putting up a GPU on it, at the very least, you’re doing yourself a disservice, right?” he said.

What could go wrong

The open-platform plan depends on satellite makers agreeing to run third-party software on flight computers, which raises security, certification and liability questions. Radiation and heat limit what a chip can do. And if SpaceX’s vertically integrated approach wins, the “Android” layer could be squeezed out, as happens to independent platforms in many markets.

A small round for a big idea

iAfrica points out that onboard satellite computing serves operators rather than enterprises, and that $8 million is a seed round rather than infrastructure capital. The plan works only if the software spreads across many fleets before larger players build their own.

What the Coverage Got Wrong

A few details differ between reports, so here is what the primary sources say.

Gemma, not Gemini

SiliconANGLE says Satlyt’s platform “used Gemini” to collect error logs. Google DeepMind’s own case study and TechCrunch both say the model was Gemma, Google’s small open model, which is what makes running it on a satellite practical.

The launch date

SiliconANGLE describes Google’s first AI satellite as launched “last week”, and iAfrica says no launch date was given. TechCrunch and launch coverage confirm Transporter-18 flew on 1 October 2026, with Satlyt’s software and Google’s prototype aboard.

A corrected headline

TechCrunch’s companion story on Google’s paper first said Starship would need 1,600 launches. It later corrected the headline to 1,800.

What Satlyt Means for Businesses on the Ground

Edge AI is getting smaller

The most transferable lesson is about model size. A one-billion-parameter model on a small board did useful diagnostic work in orbit, and Satlyt is now aiming to cut model memory by around 90%. The same small-model approach suits factories, vehicles and remote sites on Earth where connectivity is limited.

Cloud capacity may one day include orbit

Space computing will not replace ground data centres soon. But if third-party layers like Satlyt mature, “cloud” could one day include capacity on satellites, especially for Earth observation and communications work. Our cloud computing services team tracks where workloads are best placed.

Watch the next 12 months

The signals to watch are the results from MOI-1A, the two-satellite cloud attempt next year, and whether satellite builders beyond TakeMe2Space sign up.

Satlyt FAQ

What is Satlyt?

Satlyt is a start-up based in Sunnyvale, California and Nairobi that builds software to run AI models on satellites and, later, to share computing work across satellites from different operators.

How much did Satlyt raise?

It raised an $8 million seed round led by Non Sibi Ventures, with participation from investors including TLCOM, Antler, Launch Africa Ventures and Enza Capital.

Who founded Satlyt?

Rama Afullo, a Kenyan American engineer who worked in Google’s cloud business and at SpaceX’s Starlink, co-founded the company in 2024.

Has Satlyt’s software been to space?

Yes. It has flown on two missions: a Momentus satellite that ran Google’s Gemma 3 1B model, and TakeMe2Space’s MOI-1A, launched on SpaceX’s Transporter-18 on 1 October 2026.