A Software Play for Satellites: Why This Founder Left SpaceX to Build AI Infrastructure in Orbit

From Internal Pitch to Seed Round

Rama Afullo spent years inside two of Silicon Valley's most ambitious infrastructure companies watching the same pattern repeat: engineers would propose running computation in orbit, and executives would decline. At Google's cloud division and later at SpaceX's Starlink programme, Afullo observed that satellite operators were leaving compute capability on the table, relying instead on ground-based controllers to process data and resolve anomalies. When both organisations passed on his proposals, he left to co-found Satlyt with headquarters in Sunnyvale, California, and Nairobi.

The company announced an USD 8 million seed round led by Houston-based Non Sibi Ventures, where partner Bernard Harris, a former NASA astronaut, backed the thesis that satellites need software infrastructure before they need data centres. Satlyt is building a platform to run AI models across multiple spacecraft, positioning itself as middleware rather than hardware operator. The firm has already flown its software on two demonstration missions and expects a third deployment this week aboard a spacecraft built by Indian startup TakeMe2Space.

The Middleware Bet

Satlyt's approach diverges from the vertically integrated strategies of SpaceX, Google and emerging startups such as Starcloud and Cowboy Space Company, all of which are building their own orbital data centre hardware. Afullo compares his model to VMware and Snowflake: platforms that abstract away infrastructure complexity and let users focus on workloads. The company does not plan to manufacture satellites or launch them; it sells software that spacecraft operators install to run AI inference on existing compute modules.

The analogy Afullo uses is operating-system level: if SpaceX's orbital data centres are the iPhone, Satlyt aims to be Android, an open ecosystem that works across hardware from multiple vendors. The strategy assumes that satellite builders will converge on advanced processors and GPUs over the next four years, creating demand for a common software layer. Afullo projects that Satlyt could be live on 20 per cent of satellites by 2030, a target that depends on GPU adoption becoming standard practice in spacecraft design.

The economics hinge on downlink costs. Transmitting data from orbit to Earth remains expensive and bandwidth-constrained, so processing sensor readings or resolving software errors onboard can yield significant savings. Earlier in 2026, Satlyt deployed a version of Google DeepMind's Gemma model on a spacecraft operated by Momentus, a US-based space logistics company. The model compressed telemetry about software errors by more than 60 per cent, a reduction Afullo estimates can save hundreds of thousands of dollars per satellite annually.

Three Customers, One Demonstration

This week's launch carries a more complex test. Satlyt's software will run aboard a TakeMe2Space satellite serving three distinct customers: NASA, which is funding trials of cloud computing protocols in orbit; Stellerian, a startup focused on space surveillance that wants to process imagery onboard rather than downlink raw frames; and TakeMe2Space itself, demonstrating that its hardware can host third-party workloads. The mission is designed to show that a single spacecraft can handle heterogeneous compute tasks from multiple tenants, a prerequisite for the distributed cloud Satlyt plans to build.

The company's next milestone is to link two satellites into a shared computing system, effectively creating a multi-node cloud in orbit. Satlyt expects to attempt this in 2027. Success would allow the firm to offer compute-as-a-service to spacecraft operators, turning satellites into revenue-generating infrastructure rather than single-purpose assets. The pitch to satellite builders is straightforward: install Satlyt's software, and your spacecraft becomes a managed service that can sell spare compute capacity.

At Opentechwire, we have tracked the gradual shift in satellite design from bespoke, mission-specific hardware towards modular, software-defined architectures. Satlyt's middleware model mirrors the cloud transition that occurred in terrestrial data centres a decade ago, when abstraction layers enabled multi-tenancy and workload portability. The question is whether the orbital market will mature quickly enough to support a third-party software provider before vertically integrated players capture the value chain.

Why Investors Backed a Pre-Revenue Software Startup

Non Sibi Ventures led the seed round with a thesis that Satlyt does not require the most ambitious version of orbital data centres to succeed. Partner Kent Lucas noted that the company's revenue potential scales with the number of satellites launched, not with the realisation of large-scale space-based computation. The firm's conviction rests on two trends: the falling cost of launch, which is increasing the satellite population, and the adoption of AI accelerators in spacecraft, which creates demand for software to manage those chips.

Bernard Harris, the Non Sibi partner and former astronaut, brought operational credibility to the investment. His experience with NASA's mission control and spacecraft systems informed the firm's view that onboard autonomy will become a requirement, not an option, as constellations grow and ground-station bandwidth becomes a bottleneck. The fund's assessment is that Satlyt is positioned to capture value from incremental efficiency gains today while retaining optionality on the orbital cloud thesis tomorrow.

The seed capital will fund engineering hires in both Sunnyvale and Nairobi, where Afullo is building a distributed team. The company is prioritising integrations with spacecraft manufacturers and securing contracts with satellite operators who are already deploying GPUs. Satlyt's challenge is to prove that its software can deliver measurable cost savings across diverse hardware platforms before competitors with deeper pockets bundle similar functionality into their own systems.

The GPU Adoption Curve

Afullo's timeline assumes that GPU deployment in satellites will accelerate sharply over the next four years. Right now, few high-performance AI accelerators are in orbit; most spacecraft rely on radiation-hardened processors optimised for reliability rather than throughput. The shift towards GPUs is being driven by two use cases: onboard image processing for Earth observation and remote sensing, and autonomous decision-making for constellation management.

Stellerian's participation in this week's launch illustrates the first category. The startup tracks objects in orbit to monitor space debris and prevent collisions. Processing imagery onboard reduces the volume of data that must be downlinked, cutting costs and latency. NASA's involvement signals interest in the second category: protocols that allow satellites to coordinate workloads and share compute resources without relying on ground control.

The broader question is whether satellite operators will standardise on a common software layer or fragment across proprietary stacks. Satlyt is betting on the former, arguing that the economics of shared infrastructure will outweigh the control benefits of vertical integration. The precedent is terrestrial cloud computing, where AWS, Azure and Google Cloud emerged as dominant platforms despite early scepticism that enterprises would cede control of their data centres.

What Comes After Efficiency

Satlyt's near-term revenue comes from cost reduction: helping satellite operators spend less on downlink and ground operations. The longer-term opportunity is monetisation: enabling spacecraft owners to sell compute capacity to third parties. This shifts satellites from cost centres into revenue-generating assets, a model that could reshape the economics of constellation operators.

The analogy Afullo uses is mobile networks. Telecommunications companies built infrastructure to serve their own subscribers, then opened wholesale capacity to mobile virtual network operators. Satlyt envisions a similar evolution in orbit, where constellation operators lease compute time to customers who do not own satellites. The market for this service depends on latency-sensitive applications that benefit from processing data in space rather than on Earth, a category that remains speculative.

Google's Project Suncatcher, which is launching a prototype alongside Satlyt's software this week, represents the vertically integrated alternative. Google is building both the hardware and the software for its orbital data centre, aiming to serve its own cloud customers rather than third-party satellite operators. SpaceX has also committed to orbital data centres, leveraging Starlink's existing constellation and launch capacity. Satlyt's success depends on convincing enough spacecraft builders that a neutral, multi-vendor platform offers more value than locking into a single provider's ecosystem.

The company is among a cohort of startups competing at TechCrunch Disrupt's Battlefield event in San Francisco later this month, a showcase that will test investor appetite for space infrastructure plays that do not involve building rockets or satellites. Afullo's pitch is that software margins are higher than hardware margins, and that abstraction layers capture value across the entire market rather than within a single vertical. Whether that thesis holds in orbit remains to be seen.