The most interesting number in the orbital data center debate isn’t about rockets. It’s 11.7%.
That’s the share of satellite-collected data that turned out to be useful in a recent Satlyt experiment, according to CEO Rama Afullo, speaking on This Week in Startups. The other 88.3% was junk — and at $10 to $20 per gigabyte to beam it down through a ground station, junk is expensive. “If you can send only what’s useful, it’s cheaper for you, much more efficient,” Afullo said. Satlyt’s answer: run Google’s Gemma language model onboard the satellite, filter the data in orbit, and downlink only the fraction worth paying for.
It sounds like a science project. It’s actually the first version of a business case for putting data centers in space — and it’s already making money for at least one company, even as the giants fight over whether the whole idea is realistic.
The economics: profitable for software, not for hardware
Afullo’s most important claim is that Satlyt doesn’t need launch costs to fall to be commercially viable, because it doesn’t build or launch anything. It sells orchestration software that runs on satellites owned by others. The brutal per-kilogram math that governs satellite operators simply doesn’t apply.
“For us, for Satlyt, it’s already cost effective because we don’t build our own hardware,” he said.
That math, though, is why the broader industry is still waiting. Getting mass into orbit remains extraordinarily expensive:
| Launch scenario | Cost per kg | Status per Afullo |
|---|---|---|
| Rideshare, typical satellite operator | ~$15,000 | Current market rate |
| SpaceX Starship estimate | ~$1,000 | Approaching, pending catch tests |
| Parity with terrestrial data centers | ~$100 | The threshold that matters |
The gap is a launch-capacity problem, not a physics problem, in Afullo’s framing. Starship and Blue Origin’s New Glenn both need to reach full operational tempo — “anyone who’s building a massive vehicle to transport more satellites into orbit” has to be running at scale. He noted that if Starship’s booster catch succeeds, “we’re off to the races.”
The demand-side argument is real estate and power. Low Earth orbit has a larger circumference than the planet, so vastly more available area than any terrestrial site. Sun-synchronous orbits give satellites continuous access to solar power without atmospheric loss — addressing what Afullo calls the biggest limiting factor for compute generally. And an orbital data center is flexible in a way a cloud region isn’t: a satellite moving at 17,500 mph can be positioned over any point on Earth within 90 minutes, while a terrestrial data center is stuck where it was built.
What Satlyt actually sells: Kubernetes for satellites
The product demo is the most concrete part of the conversation. Satlyt’s mission operations center lets an operator define a job across a constellation of what Afullo calls “systems with dissimilar architectures” — satellites from different makers — and the software places and orchestrates containerized workloads across them. The components include a container registry (so software can be ported without rewriting it), data compression, telemetry analysis, encryption, and storage. He credits Google’s Kubernetes work as the model.
In the demo, five satellites each run an NVIDIA Jetson Orin Nano — 8 GB of memory, six cores, a 15-watt power budget — and data must downlink through at least one node to a ground station. The software schedules services across nodes and enforces the power ceiling.
Then Gemma goes to work filtering. The flow looks like this:
Afullo says this exact experiment has already run on a couple of satellites, and that Google DeepMind published a case study on the work.
Heat: both camps are right, and the fix is to run chips hotter
The cac talk he keeps encountering: space is cold, so cooling should be free. The counterargument: space is a vacuum, so hardware should overheat. Afullo’s answer is that both are correct. Waste heat cannot conduct away in vacuum — it must be radiated, which is why the International Space Station carries those massive radiator panels. At the same time, the ambient environment really is cold.
The resolution is hardware designed for the environment. Afullo points to Intel’s Starfire and NVIDIA’s space modules as examples of compute built to run much hotter and dissipate around 3,000 watts. “They’re essentially just skipping the entire debate,” he said — instead of solving cooling for terrestrial chips, run space-grade chips hot and throttle as needed. He notes the scale of what’s changing: some compute now being deployed in orbit is roughly 10 times the total compute ever put across the roughly 20,000 satellites currently in space.
That last figure is striking, and it’s also unattributed — one of several places where Afullo’s claims run ahead of the public evidence.
Direct-to-device: the telco gets landlined
Afullo’s most aggressive forecast is that orbital capability erases the traditional telecom carrier within about a decade, reducing it to the status of the landline telephone. The mechanism is direct-to-cell service: if satellites can beam data straight to a handset regardless of the handset’s power efficiency, “you no longer actually have to have a telco on the ground. It essentially eradicates the typical telco in the next 10 years or so.”
He cites this as the reason there’s a race in direct-to-satellite services right now. Handsests wouldn’t need architectural changes to receive service — a detail that matters enormously for adoption.
He also raises a darker use case. During the Los Angeles wildfires, he says, insurance companies obtained data about what was about to happen faster than ordinary residents did, and several changed policies just before the fires hit. Host Jason Calacanis calls the example “a little bit dark”; Afullo agrees, but uses it to show how much control faster orbital data confers — and to whom.
Other applications he names: real-time scientific experiments that no longer require launching to the ISS and paying an astronaut to run them, and faster emergency and wildfire response for remote areas.
NASA wants 4K video from the Moon
Satlyt is working with NASA’s Glenn Research Center on delay-tolerant networking — a protocol family Afullo says was built by early internet creators including Vint Cerf for deep-space communication. The goal: relay 4K video from a lunar base through lunar-orbit satellites to low Earth orbit satellites and down to a ground station at low latency, while suppressing packet loss and bit flips — ones and zeros swapping — which Afullo calls a “showstopper.”
On extending this to Mars, he’s skeptical of daisy-chaining satellites across interplanetary distances. The geometry is too complex — planets rotate on their axes while satellites move — and the intermediate orbits would need to be populated first. He expects the industry to add orbital infrastructure incrementally rather than attempt a Mars relay chain soon.
The skeptics are not fringe
The contrast with the terrestrial cloud establishment is sharp. AWS chief executive Matt Garman said in a Nov. 10 interview that he puts the odds of orbital data centers being economically viable or practical within 10 to 20 years as very low. It’s not a problem one person can solve, he argued: a single GPU rack weighs a metric ton, radiating heat in a vacuum is extremely hard, and servers must be radiation-hardened at added cost. Among space, the ocean, and the desert, Garman called the desert the most realistic site for data centers.
Garman’s own business explains why the debate matters: AWS is adding one GPU to its data centers every 14 seconds and still can’t keep pace with demand, with roughly $220 billion in capital spending budgeted this year. “Demand is far outpacing supply right now,” he said. He thinks the inference market may be running at less than 1% of its eventual size.
Afullo’s answer to the physics objections is essentially: build for the environment, don’t fight it. Radiation-hardened or shielded hardware, thermally robust design, and efficiency are mandatory anyway — every kilogram and every bit matters. And the spillover thesis is that space-grade engineering pays back on Earth: microwaves and solar panels came from space programs, and hardware built to survive launch and vacuum tends to last longer and need replacing less often.
He’s not dogmatic about ruggedization. Planet Labs, he notes, ignored conventional wisdom — sending up smartphones with solar panels attached, and hardware with moving parts. NVIDIA Jetson Nanos with spinning fans are operating in orbit right now. But for longevity, shielding and expensive testing are still required. He also predicts the smartest space players will maintain a terrestrial counterpart to absorb the volume of data and outputs coming from orbit.
“If you can build it to survive in space, you’re going to have to have material science combined with the engineering to make it very rugged,” Calacanis summarized. “If you make it that efficient, now I kind of understand your earlier answer.”
From Starlink to Satlyt: a decade-long bet
Afullo’s path runs through family and a long-held conviction. His father is a professor of electrical engineering whose PhD focused on satellite and microwave technology; his mother is a computer scientist who taught him copy-paste. He’s Kenyan-American, raised partly in Botswana, where his older brother ran a CD-copying business — the family’s first lesson in spotting an unmet market.
He says he knew what his startup would be called about a decade before founding it, and bought the Satlyt domains — including satellite.ai — four to five years ago. The name is a backronym for “space and technology analytics.” Before Tesla he was the company’s third employee in Africa; he later worked at Google as a professional cloud architect and solutions consultant, where he pitched orbital compute at the start of the AI boom. Google’s priorities were terrestrial LLMs competing with ChatGPT, so he left for SpaceX and a product manager role at Starlink. Three months in, he was told in no uncertain terms that the last people to propose orbital compute were no longer at the company. He left and started Satlyt just over two years before this conversation.
He also mentions getting married the week of recording, and invokes the Japanese concept of ikigai — the overlap of what you love, what you’re good at, what the world needs, and what you’ll be paid for — as the frame for his career.
What the smart money is actually doing
The market isn’t waiting for the orbital debate to resolve. SpaceX, now public, has become a compute landlord first and a rocket company second: its Super Intelligence segment rented out $7.6 billion of capacity in the third quarter, up 830% year over year, with Anthropic accounting for more than a third of sales. Deutsche Bank estimates AI compute will account for more than half of fourth-quarter revenue. Musk has said SpaceX plans to begin launching AI-computing satellites in the third quarter of next year, and analysts are split on timing — GF Securities sees 5 gigawatts of space AI capacity by 2028, while Deutsche Bank sees nothing material until 2029 or 2030.
The near-term truth is simpler: the compute shortage on the ground is so severe that the ground is where the money is. Garman’s one-GPU-every-14-seconds line is the tell. Orbital data centers are a bet on a launch-cost curve that hasn’t bent yet, a thermal problem that’s solvable but expensive, and a regulatory environment on Earth that is getting harder — Goldman Sachs estimates new restrictions could affect around 45% of planned U.S. data center capacity expansion. That last point is the quiet bull case for space: if you can’t build on the ground, orbit starts looking less like science fiction.
Afullo’s forecast that telcos become relics within a decade is speculative and unproven. But the 11.7% number is real, the DeepMind case study exists, and the NASA Glenn work is underway. For investors, the question isn’t whether orbital compute replaces the cloud — it’s whether the launch-cost curve, not the physics, is the binding constraint. Watch Starship’s catch tests, watch direct-to-device spectrum rules, and watch whether Satlyt’s software layer becomes the Kubernetes of orbit before the hardware economics catch up.