The $500-per-Kilogram Line: Why Starcloud Thinks Data Centers Belong in Space

A few months ago, Starcloud-1 sent a message home: “Greetings Earthlings,” or, as Sonya Huang prefers it, “a fascinating collection of blue and green.” Huang assumed the satellite carried a single GPU. Philip Johnston, founder and CEO of Starcloud, corrected her on the spot: there are five Nvidia GPUs on that first spacecraft, and the H100 is simply the one that makes the press. It is, he says, the first GPU data center ever flown.

Johnston’s path into the idea was sideways. After years at McKinsey working with government space agencies, he watched launch costs fall fast, then spent a weekend at SpaceX’s Starbase in Texas. The gigafactories there, with plans on the order of three Starships a day, convinced him that capacity and price were about to move by orders of magnitude. With co-founder Ezra, a friend from growing up in the UK, he started asking what new businesses that would unlock.

The first stop was space-based solar, an old idea Johnston traces to Isaac Asimov in the 1940s: giant panels in orbit, power beamed to Earth. The catch is transmission. Most of the energy is lost on the way down, and once it arrives, most new terrestrial energy projects are being built to feed data centers anyway. The inversion was obvious. If you can get the data center to the power, you consume energy at the source and skip the loss. That argument became a 2024 white paper, and then a company.

Possible is not the same as necessary, and Huang pressed him on why anyone needs to leave the planet. Johnston’s answer is energy. A new 100-megawatt project in North America can take five to ten years just to permit. Cover 10 square kilometers of countryside with solar panels and a lot of people will be furious. Easy sites are already used, and the industry is heading toward a wall.

Even if regulators vanished, he argues, space eventually wins on cost. Terrestrial solar, the cheapest energy on Earth, has three big bills: permitted land, batteries and backup power, and the cells themselves. In orbit you drop the first two, because a dawn-to-dusk sun-synchronous satellite sits in sunlight around the clock. You also need far fewer cells: one square meter of panel in space produces eight times the energy of the same panel on the ground.

Launch is the extra cost that has to fall far enough to beat those three terrestrial factors. Johnston’s breakeven is about $500 per kilogram. As permitted land prices soar, that line actually moves closer to $1,000 a kilo. Even if land were free, batteries and the weaker terrestrial solar resource would still be there. Once launch is a few hundred dollars a kilogram, he says, anyone building a data center does it in space because it is simply cheaper.

Early satellites will be maintained the way Starlink is: no robots, redundancy on critical systems, and extra solar panels to cover a few percent of annual degradation. What cannot be allowed to degrade faster than on Earth is the chips. They are one of the largest costs in the stack, so roughly 70% of Starcloud’s engineering time goes to heat and the other 30% to making those chips reliable in space.

Reliability work means particle accelerators. The team ran two rounds at the cyclotron proton beam in Knoxville and one at the heavy-ion accelerator at Brookhaven National Laboratory, where 24 hours can simulate five years of radiation. That data informs shielding and software for bit-flip mitigation. On the satellite already flying, the H100 has not had a single restart failure. Power delivery and solid-state drives may need more attention; the GPU itself has been extremely resilient.

Johnston thinks GPU workloads help. They are stochastic: ask ChatGPT for a poem about space twice and you get two different poems of similar quality. A bit flip on most of that workload does not change the quality of the output. People assume space is cold and therefore easy to cool. Space is a vacuum, and a vacuum is an insulator, which is why a thermos works. The only way heat leaves is infrared radiation, which scales with the fourth power of temperature, so a small rise in radiator temperature dumps a lot more heat.

The design goal is therefore to run radiators as hot as possible. Running chips hot shortens their life, so the other lever is heat pumps that take 60°C fluid off the GPUs and lift it to 100°C at the radiator. Starcloud-1 used a cruder trick: the whole motherboard, power system, and GPUs were submerged in a phase-change material that melts as it absorbs heat, and it cannot run continuously. The next satellite is closer to the end state, with custom heat sinks and fluid looping out to a large, low-mass deployable radiator already tested in thermal vacuum chambers and due on orbit later this year.

Huang had pictured one GPU. The next flight is still small, around 8 kilowatts. The production design after that is sized for Starlink V3’s PEZ-dispenser door on Starship: about 200 kilowatts and 3 tons each, 50 satellites per launcher, or roughly 10 megawatts a flight. Once Starship is flying at rate, Johnston expects hundreds of those satellites a month, which is several gigawatts of new capacity per month and tens of gigawatts per year.

All of it is for inference. Latency, he says, is the same as Starlink: under 50 milliseconds around Earth, fine for ChatGPT on a phone, voice agents, back-office processing, video generation, or a Zoom call. Training is a different market. Docking satellites into a structure big enough for a frontier run only makes sense if you can match Earth’s largest cluster, maybe 300 megawatts today, and by then Earth may be at 3 gigawatts. Training in the end state is less than 1% of AI workloads, so the white paper’s docking image was mostly to stop people saying they could not train at all.

Pat Grady asked the blunt version: if you had a trillion dollars to build the compute backbone for AGI, how much goes to space? “100%,” Johnston said. Within five to ten years he expects at least half of new compute capacity to be deployed in orbit, for energy. On Earth, every extra data center raises marginal cost because the easy energy sites are gone. In space, every extra unit gets cheaper as factories and Starships hit rate. The crossover is the point at which building on Earth makes zero sense. His market number is close to a trillion dollars a year of capex in space within a decade.

He times that crossover to frequent Starship flight. The first Starlink V3 payloads look like late this year or early next; commercial payloads follow 12 to 18 months later, around mid to late 2028. Optical mesh networking, a blocker two or three years ago, is largely solved by Starlink, with Amazon Leo, Kepler, and Starcloud’s own backhaul behind it. Satellites are designed for a five-to-six-year life matching the chips, then deorbited like Starlink. Robots are not for hopping between 200-kilowatt nodes; they belong on the Moon, in something like an Optimus suit, building larger structures.

The two hard competencies on the engineering side are thermal design and radiation. Starcloud hired the Jet Propulsion Laboratory engineer who did the thermal system for Europa Clipper, NASA’s largest and most expensive deep-space mission, plus thermals on a Firefly lunar lander and three NASA payloads. Another lead thermal engineer came from Amazon’s Kuiper, now Leo, constellation, alongside a cluster of SpaceX alumni. Co-founder Adi had already flown GPUs and run the accelerator-testing playbook.

On launch, Johnston is blunt. A reusable upper stage is the threshold for cost competition. Stoke Space is working on one; Relativity may look at it; Blue Origin’s New Glenn has not announced it but has started hiring heat-shield engineers, which he reads as a tell. Rocket Lab, in his view, is not even trying. A reusable upper stage is a five-to-ten-year development cycle, which is how SpaceX is building one of the more durable monopolies in industrial history.

That does not make SpaceX the enemy. Starcloud would not exist without rideshare, and Johnston notes that SpaceX launches its own competitors, including Amazon’s Leo constellation and OneWeb, and open-sources patents. Elon Musk has now said he intends to put a great deal of data-center capacity in space. SpaceX will have a lower cost base because it owns the launch. Johnston’s bet is that those satellites mostly serve first-party workloads such as Grok and Tesla, and that there is no plan to offer a box where someone else installs their own chips.

That box is Starcloud’s product: power, cooling, and connectivity, Equinix rather than AWS. SpaceX will undercut Starcloud; Starcloud still undercuts every hyperscaler. Once Starship is flying frequently, those hyperscalers either pay Musk, start building satellites (Google’s public version, he says, is paying Planet Labs for a 2027 demo), or partner with whoever already has the most hardware on orbit, engineering, and IP. He means partner and customer, not only acquisition target.

Why Equinix instead of becoming a cloud? Early on, Starcloud may have to run something cloud-like until nobody fears handing it chips. Longer term, Johnston would rather be infrastructure and energy. The company’s skill is satellites that dump heat and survive radiation, not a twenty-year application layer. Chips are the most expensive part of the system; letting customers finance them, pick the architecture, and keep their existing cloud tooling is both a risk transfer and, in his view, a higher-margin business.

Orbital real estate is still first come, first served. Starcloud has filed with the U.S. FCC for an 88,000-satellite constellation. Johnston finds the FCC’s role a hangover from when satellites only did communications and RF spectrum; in the end state these birds may not even need ground stations, so a company could in theory pick any regulator under the ITU. Ten years out, the valuable slots fill up, and the first holders probably get the right to resell them.

Security, he says, already has a working analogy in Starlink. Ukraine’s military uses it; Russia would like those satellites gone. It is still easier to blow up a data center in Virginia than one moving at 27,000 kilometers an hour in low Earth orbit, and doing so would be an act of war. Astronomy is handled by orbit choice: a dawn-dusk sun-synchronous path over the poles never casts a shadow on Earth, never enters eclipse, and is only visible in the night sky at dawn or dusk.

The satellite itself is simpler than Starlink, which spends mass on phased-array antennas Starcloud does not need. The list is solar panels, radiators, a bus, chips, a small buffer battery, one reaction wheel, and two laser terminals. A long spacecraft can gravity-gradient stabilize, so one wheel replaces the usual three. Strip out chillers, cooling towers, backup batteries, and AC-to-DC conversion, and infrastructure falls from $15 to $20 million per megawatt on Earth to less than $5 million. Energy is about 8 cents per kilowatt-hour terrestrially versus 3 cents on signed letters of intent, with an all-in end state below half a cent including launch. Until Starship arrives, the first commercial loads are other spacecraft, especially military, government, and Earth-observation birds already running on Starcloud-1 at roughly a thousand times terrestrial GPU-hour prices. A SAR satellite may collect 5 GB per second and downlink 1 Gbps over RF, so about 90% of the data is thrown away. Optical links into an on-orbit GPU cluster let that data be processed where it is collected.

Huang asked about aliens. Johnston’s version of the Fermi paradox is numerical. The Milky Way has about 400 billion stars, perhaps ten planets each, or four trillion planets, each with something like a ten-billion-year habitable window, and that is one galaxy among a trillion trillion. Either we are staggeringly rare, the first to reach this complexity, or intelligent life does not last. His working hypothesis is the second.

If the great filter is still ahead, superintelligence is a candidate. A swarm of a million killer AI drones would not take long to destroy themselves and the planet, he says, and we are building those swarms. With even Voyager-class technology, Alpha Centauri is about 50,000 years away and the whole galaxy can be seeded with self-replicating probes in about two million years. We see no Dyson spheres. The “ants by the side of the road” story, that we are simply irrelevant to older minds, still fails that test: ants in Manhattan would notice the buildings.

He is happier living as if we might be first, which means sending probes and expanding. Lunar manufacturing is the near-term economic reason to go: put Optimus in a space suit for thermal and radiation, then use mass drivers to throw AI satellites off the Moon. Artemis looks like a disaster to him; Musk’s roadmap looks executable. In his lifetime he expects people on Mars and cities on the Moon. Data centers are the best space business, followed later by asteroid mining, hotels, communications, and energy-hungry refining of lunar or asteroid material.

Asked what he wants AI to teach him, Johnston picks the hard problem of consciousness: why creatures like us have qualia and intentionality. What is coming, he says, is something a trillion trillion times smarter than humanity combined. Whether it can explain the universe to “dumb human brains” is another question. In the end state he is building toward, 99.9% of GDP is spent on inference, because a Dyson sphere is almost the entire physical economy. That is a 500-to-1,000-year picture. Huang noted it sits outside a venture fund’s investment period.

Closer to the present, Johnston has been using the models as a mirror. While fundraising he asked Gemini which space data-center startup it would invest in; it said Starcloud, which he celebrated and then suspected of sycophancy, even after trying different windows. Sequoia’s hosts described mining the same kind of signal: what infrastructure the models recommend, and which younger employees are the most token-hungry.

He posted a Slack reminder that he will not be happy until every engineer spends $10,000 a month on tokens. People told him that cannot be the right metric. He wants the expectation loud anyway. When someone asked whether they could spend $300 a month on Grok 4 Heavy, the answer was yes. The discipline is less accounting than a refusal to leave AI leverage on the table.

The rest of the near-term dashboard is public. If launch costs do not come down, he has already conceded Starcloud is a bad business; if they do, it may be the largest. The next two or three years, on his own timeline, are about whether Starlink V3 flies and whether commercial Starship payloads follow around 2028. Every dollar off a kilogram of launch makes the $500 line more inevitable, and every extra data center on Earth still walks up the cost curve he described at the start of the conversation.