Orbital computing still needs a cheaper ride
- Satlyt has demonstrated AI filtering satellite data before it is sent to Earth.
- Stoke is developing a reusable rocket, but orbital computing still depends on much cheaper launches.
A satellite collecting data may not need to send all of it back to Earth. Satlyt CEO Rama Afullo told This Week in Startups that the company has tested Google’s Gemma language model aboard satellites to pick out useful information before downlink. In one demonstration, just 11.7% of the data was worth transmitting; ground-station links can cost $10 to $20 per gigabyte.
Satlyt’s software is designed to coordinate computing jobs across satellites made by different manufacturers, even when their hardware varies. Afullo sees that as a practical first step toward an orbital cloud: process information close to where it is gathered, then send only what is needed. He also points to emergency response and experiments that could be run in orbit without sending equipment or people to the International Space Station.
The larger vision runs into a basic constraint: getting equipment into orbit. Afullo put typical satellite rideshare launches at about $15,000 per kilogram and said costs would need to fall to roughly $100 per kilogram for orbital data centers to match terrestrial economics.
He cited an estimated $1,000 per kilogram for Starship if it reaches that price. ARK Invest’s FYI discussed a different threshold: host Brett Winton estimated orbital computing could become competitive near $300 per kilogram. The estimates describe different economic tests, but both make cheaper access to space central to the opportunity.
That is where Stoke Space’s rocket plans meet the computing story, not as a partner to Satlyt, but as a potential answer to the broader launch-cost problem. On FYI, Stoke’s Devin Poppin Drew described a fully reusable rocket built to return and fly again.
Its upper stage uses liquid hydrogen to cool the surface during reentry, with the engine also serving as the heat shield. Stoke has completed more than 11,000 seconds of upper-stage engine testing, built both stages of its first vehicle and is preparing for static-fire tests; Drew said the company is targeting its first orbital flight for early 2027.
Neither a successful data-filtering demo nor a rocket under development proves that orbital data centers can compete with those on Earth. Afullo also flagged the engineering challenge of shedding heat in a vacuum, alongside the need to protect electronics from radiation. The near-term case is narrower: processing data in orbit could save transmission capacity and get useful information to the ground sooner. Whether that grows into a cloud above Earth depends on both the computing and the cost of the ride.
Source Intelligence
- Deep dive into what was said in the episodes
Running Google’s Gemma on Satlyt’s orbital AI (w/ CEO Rama Afullo) • Nov 12
- Rama Afullo says satellites from different manufacturers lack a standard way to pool compute for large jobs. Satlyt’s software uses containers and Kubernetes-style orchestration to place services across heterogeneous satellites with different hardware.
- Rama Afullo sees orbital computing as edge infrastructure: satellites can process data near where it is collected, reduce latency, and support emergency response in remote areas. He also cites local opposition to terrestrial data centers and limits on power and communications links.
- Rama Afullo says sun-synchronous orbits can give satellites more consistent solar access. Their speed also makes orbital compute a mobile cloud region, able to pass over any point on Earth within 90 minutes.
- Rama Afullo says the number of satellites matters less than their available compute and ability to relay data. He expects satellite count to grow sharply, creating more capacity for distributed orbital workloads.
- Rama Afullo says orbital compute could run experiments without sending equipment to the International Space Station or paying an astronaut to operate it. He also sees direct-to-device links as a way to deliver services without changing handset architecture.
- Rama Afullo says Satlyt has tested Google’s Gemma small language model on satellites, using onboard AI to analyze and compress data before downlink. In one demonstration, only 11.7% of the collected data was useful to send, against ground-station charges of $10 to $20 per gigabyte.
- Rama Afullo describes a five-satellite demo with Nvidia Jetson hardware, a 15-watt power ceiling, and onboard services for containers, telemetry, compression, storage, security, and encryption. Satlyt’s orchestration software places those services across satellite nodes.
- Rama Afullo says Satlyt’s costs are already workable because the company does not build its own spacecraft hardware. Broader parity with terrestrial data centers depends on cheaper launches and larger vehicles reaching full operation.
- Rama Afullo predicts direct-to-cell services could make ground-based telecom operators obsolete within a decade, while faster satellite data could improve emergency response. He says some insurers accessed information before the Los Angeles fires and changed policies before the fires hit.
- Rama Afullo says Satlyt is working with NASA on relaying 4K video from the Moon to Earth. Delay-tolerant networking can route data through lunar-orbit and low-Earth-orbit satellites while limiting packet loss and bit flips.
- Rama Afullo says orbital compute faces a real thermal problem: vacuum is cold, but heat must still be carried away, often with large radiators. He points to space-built compute modules designed for higher temperatures and power as another way to manage heat.
- Rama Afullo says ordinary electronics and moving parts can survive launch and operate in orbit, citing Planet Labs’ use of repurposed components and Nvidia Jetson Nano computers with fans. Longer service life still requires radiation shielding, rugged design, and costly testing.
Also discussed on this episode: (3)
Startups (1)
- Rama Afullo says his satellite-computing idea grew from a family steeped in the field: his father was an electrical-engineering professor focused on satellite communications, and his mother was a computer scientist. After roles at Tesla, Google, and SpaceX, he left SpaceX to start Satlyt.
Space (1)
- Rama Afullo expects space engineering to produce useful terrestrial technologies, as earlier space work did with solar panels and microwaves. He argues that companies building orbital systems should pair them with Earth-based work to transfer useful lessons and outputs.
Psychology (1)
- Jason and Rama Afullo describe ikigai through four overlaps: passion, mission, vocation, and profession, formed by what someone loves, does well, what the world needs, and what people will pay for.
Filling The Falcon 9-Shaped Hole: Stoke Space's Bet On Full Reuse • Nov 12
- Devin Poppin Drew says Stoke Space is building fully reusable rockets to replace the manufacturing bottleneck of one-use stages with an operations model built around frequent flights and routine returns from space.
- Devin says Stoke’s upper stage uses liquid hydrogen to cool the reentry surface, combining the engine and heat shield; he argues active cooling avoids fragile tiles and lengthy inspections.
- Devin says Stoke has tested the upper-stage engine across varied operating conditions, built both stages of its first vehicle, and is preparing for static fires before its first orbital flight.
- Devin says vertical integration helps Stoke control schedule, cost, and supplier risk, while tightly coupled stages make mix-and-match designs difficult; he expects more standardized interfaces only as launch matures.
- Devin argues full reuse is the next economic step after Falcon 9 proved partial reuse; higher flight rates and fewer first-flight vehicles can offset payload penalties while improving reliability.
- Devin says Nova Block 2 scales Stoke’s existing full-reuse architecture to Falcon 9-class payloads, using a traditional fairing and stainless-steel construction that needs no fixed tooling.
- Devin says launch demand extends beyond Starlink and defense: Amazon has satellites awaiting launch, while communications, Earth observation, alternative navigation, and signals-intelligence companies seek capacity to scale constellations.
- Devin argues independent launch providers are needed because major operators also serve their own constellations or government programs, leaving a short list of companies dedicated to outside customers; Stoke currently plans to stay payload-agnostic.
- Devin says Stoke’s traditional separating fairing lets customers keep existing satellite designs while benefiting from reuse; he argues a narrow dispenser can exclude payloads such as high-power optics.
- Devin says Stoke designs for a 24-hour turnaround, although not at launch; with a fleet of dozens of vehicles, that operating model could support hundreds or thousands of annual flights.
- Brett Winton argues that cheap launch could make orbital computing competitive, citing solar power availability and AI workloads; he estimates the economics could cross over near $300 per kilogram of launch cost.
- Devin says customers have contracted with Stoke years ahead because its technical progress and funding have strengthened confidence in its schedule; launch buyers are placing bets early to secure future constellation capacity.
- Devin says safe upper-stage return is essential for human and cargo missions, and that frequent debris over populated areas would be unacceptable; Brett raised hypothetical reentry failure rates to illustrate the risk.
Also discussed on this episode: (1)
Space (1)
- Devin sees orbital manufacturing as a large market because some biotech, pharmaceutical, and advanced-material products made in microgravity cannot be produced terrestrially, yet often lack a reliable ride back to Earth.