The Altar of Compute — Why Starcloud's Orbital Ambition Is a Thermal Problem, Not an Energy One

I have spent about fifteen years putting compute into production, and most of that time went to one unglamorous question: where does the heat go? Not where the power comes from — that one gets the press release. Where the heat goes after the power arrives. Every capacity plan I have ever signed off on lived or died on that number. It is also the number the orbital data center pitch leaves out.

Starcloud filed for eighty-eight thousand satellites last month. The FCC accepted the application on March thirteenth. I have read the filing. I have read the Series A coverage — Benchmark leading, EQT Ventures alongside, one hundred and seventy million dollars at a valuation of one point one billion, the fastest Y Combinator graduate to unicorn status in the accelerator's history. I have also read the Breakthrough Institute's dissection of why orbital data centers face thermodynamic barriers that no amount of venture capital charms into submission. Both sets of documents are correct. Neither is sufficient, because they answer different questions, and the one that actually decides this — what does a watt of rejected heat cost at eight hundred kilometers — appears in neither.

The Thesis and the Track Record

Philip Johnston, Starcloud's founder, carries a biography that reads like a résumé assembled on purpose. Applied mathematics at Nottingham. Columbia for the masters. Wharton for the MBA. Harvard's Kennedy School for a degree in national security and technology policy. McKinsey for the consulting pedigree. SpaceX for proximity to the only manufacturing operation on the planet that has made launch look routine. He watched rockets descend onto landing pads at Starbase and concluded — correctly, in the narrow physical sense — that declining launch costs eventually make orbital infrastructure cheaper per compute-cycle than terrestrial alternatives. The insight is genuine. The timeline is the question nobody in the Series A materials bothers to answer honestly.

Here is what Starcloud has actually shipped. They built and launched a satellite weighing one hundred and thirty pounds for three million dollars in pre-seed funding. That satellite carried an Nvidia H100, roughly a hundred times more powerful than any GPU previously deployed in space. They ran Google DeepMind's Gemma model on it. They performed the first orbital training of a language model, using nanoGPT, above the atmosphere. Starcloud-2, scheduled for late this year or early next, will carry a Blackwell B200 and AWS Outposts blades, supported by the largest commercial radiator ever sent past the Karman line. Further out, on a timeline Johnston describes with the confidence of a man who has not yet had a Series A fail, Starcloud-4 deploys arrays four kilometers on a side generating five gigawatts — more than the largest power plant operating anywhere in the United States today.

Five gigawatts. In orbit. From a company that is seventeen months old.

The distance between those last two paragraphs is where most infrastructure companies die, and I say that with some sympathy, having been on the wrong side of it myself. One satellite, one GPU, one successful inference pass is real engineering and I will not minimize it. Eighty-eight thousand satellites forming a coherent compute mesh between six hundred and eight hundred and fifty kilometers is a different discipline wearing the same job title. A single node proves the physics is not impossible. A constellation proves you can manufacture, launch, network, cool, schedule, patch, and decommission at industrial rates while somebody pays for all of it. Those are unrelated proofs. A demo answers "can this work once." Production asks "what is the failure rate, who fixes it, how long does the fix take, and what does it cost every time" — and that second list is where the orbital story stops being about physics and starts being about logistics.

The Vacuum That Does Not Cool

The pitch depends on a seductive half-truth about thermodynamics. Solar panels in orbit receive roughly forty percent more energy than their terrestrial counterparts — no atmosphere to scatter photons, no clouds, no nightfall beyond brief eclipses in sun-synchronous trajectories. That part is correct, and it is the part everybody repeats. The cooling story, which every press release cites as though the vacuum of space were a gift-wrapped refrigerator, inverts the actual physics in a way that should embarrass anyone who has sat through an undergraduate heat transfer course.

Space is not cold in the way that matters for silicon. There is no air. No water. No medium to conduct or convect heat away from a processor die. The three-kelvin cosmic background is a radiation sink, not a contact cooler, and radiative rejection follows the Stefan-Boltzmann law — proportional to the fourth power of absolute temperature, which sounds generous right up until you compute the surface area needed to shed the thermal output of a modern GPU cluster. The numbers are punishing. A chip at full utilization in a vacuum, with no active thermal management, does not cool elegantly. It cooks. The Breakthrough Institute's analysis is unambiguous: the coldness of space is a marketing sentence, not an engineering solution, and the radiator mass required to reject meaningful thermal loads at orbital scale becomes both a dominant cost driver and a dominant failure mode. Follow that one step further. Radiators are mass. Mass is launch cost. Launch cost was the very thing that had to keep falling for the model to close, so the model is quietly competing against itself.

Then there is radiation. Not the thermal kind. The ionizing kind. Cosmic rays. Solar particle events. The steady drizzle of high-energy protons that flip bits, degrade gate oxide, and accumulate damage in transistors whose features are now measured in widths of a few dozen silicon atoms. Radiation-hardened processors exist, but they lag commercial silicon by several generations. You do not get H100 performance out of a rad-hard part; you get something closer to what was cutting edge in 2018, running behind error-correction overhead that eats the computational margin the orbital location was supposed to provide. Google claimed its Trillium chips could survive five years in orbit. The test involved accelerated proton bombardment on Earth, at a single energy level, under controlled conditions that resemble the real space radiation environment about as well as a swimming pool resembles the North Atlantic in February.

Meta's Llama 3 training run hit four hundred and nineteen unexpected interruptions in fifty-four days. On Earth. In climate-controlled buildings. With redundant power and technicians who could walk an aisle, pull a node, and have a replacement seated inside an hour. That is the baseline for the best-run compute on the planet. Now delete the technicians, add the proton flux, and turn every repair into a manifest slot on a rocket. On the ground, mean time to repair is a staffing decision. In orbit it is a launch schedule. Availability math tolerates a high failure rate or a high repair cost. It does not tolerate both, and orbit hands you both.

The Ground Ran Out First

None of which explains why serious investors wrote the check, and the explanation has nothing to do with orbit. Starcloud is not ascending out of orbital enthusiasm. It is fleeing the ground.

American data centers consumed one hundred and seventy-six terawatt-hours of electricity last year. Four point four percent of total national generation. Virginia alone — the state where the internet physically lives, where the fiber meets the copper meets the concrete — saw data centers take twenty-six percent of its entire electricity supply. The PJM capacity market, which prices wholesale power across thirteen states, cleared at three hundred and twenty-nine dollars per megawatt for the 2026-2027 delivery year. That is more than ten times the twenty-eight dollars and ninety-two cents it cleared just two years prior. Consumer electricity bills in western Maryland rose eighteen dollars a month. Ohio residents absorbed sixteen. The grid is not keeping pace and was never designed to. Interconnect queues are measured in years. GPU refresh cycles are measured in months. Those two clocks do not reconcile.

Amazon, Google, Meta, and Microsoft are collectively spending roughly seven hundred billion dollars on AI infrastructure this year. Meta alone budgeted seventy-two billion. Morgan Stanley and S&P Global have both published warnings — not speculation, warnings, the kind of language analysts deploy when they want plausible deniability recorded in ink — that six hundred and thirty-five billion dollars in planned 2026 AI capital expenditure faces material constraints from rising electricity costs, permitting delays, and supply chain fragmentation. The ground cannot hold the weight. The open question is whether orbit can.

I have made the underlying move myself, at far smaller scale, and it is always tempting and usually wrong. A component is constrained, so you relocate it somewhere the constraint does not apply and call the problem solved. It isn't. You have traded a constraint you understand and can instrument for a set of constraints you have never operated. Terrestrial data centers have a power problem: severe, well-measured, boring, and staffed. Orbital data centers have cheap power plus a thermal problem, a radiation problem, a maintenance problem, a latency problem, a debris problem, and a jurisdiction problem. Six unknowns are not an upgrade on one known. The bottleneck does not disappear. It changes altitude, and it gets much harder to instrument.

The Budget Line

Here is why I keep calling this an altar rather than an infrastructure bet: on an infrastructure bet, somebody eventually asks for the unit economics.

OpenAI closed a one hundred and twenty-two billion dollar funding round on March thirty-first. Post-money valuation: eight hundred and fifty-two billion. SoftBank contributed thirty billion. Amazon pledged fifty billion, thirty-five of it contingent on an IPO or on the achievement of something the term sheet calls artificial general intelligence — a phrase whose contractual definition I would pay a meaningful sum to read, because somewhere a lawyer had to turn it into a testable milestone. Nvidia committed thirty billion. Nine hundred million weekly active users on ChatGPT. Two billion dollars in monthly revenue. The IPO is expected this year. Put the last two numbers next to each other: twenty-four billion annualized against an eight hundred and fifty-two billion valuation is about thirty-five times revenue, on a business whose cost of goods sold climbs with every additional user, which is the reverse of how software multiples are supposed to work. That multiple prices something other than revenue. It prices the odds that the cost curve of inference bends before the capital runs out.

Oracle, the same week, terminated between twenty and thirty thousand employees. Eighteen percent of its global workforce. Twelve thousand in India alone — forty percent of its presence on the subcontinent, erased in a single morning email sent at six AM from an address labeled "Oracle Leadership." The freed cash flow, eight to ten billion dollars, was redirected toward GPU clusters, data center expansion, and cloud infrastructure contracts with OpenAI, Meta, and Nvidia. State the trade plainly and it is payroll converted into silicon. As capital allocation it is coherent, and under the assumptions currently priced into the sector it is even defensible. Whether those assumptions hold is the whole question, and the executives who made the call will present the resulting efficiency at the next earnings call as evidence of strategic clarity either way.

Starcloud is the logical end of that trade. If the ground cannot supply enough power, go where power is unmetered. If terrestrial permitting takes four years, build in a jurisdiction where no permit exists and no environmental review is required. If the workforce is the expensive line item, deploy hardware that does not unionize, does not need healthcare, and does not email journalists when it is fired. Every step follows cleanly from the one before it. That is what makes the pitch worth taking seriously, and also what makes it worth auditing, because a chain of individually rational steps terminates wherever the arithmetic points — including somewhere nobody would have chosen on purpose.

The input the model never prices is the orbit itself. Solar flux, orbital mechanics, panel efficiency curves, launch cost trends — those all go in a spreadsheet. The debris field does not, which is precisely why it ends up in nobody's model. Eighty-eight thousand objects traveling at seven point eight kilometers per second, in a band that already holds Starlink's ten thousand satellites, SpaceX's filed application for one million orbital data center nodes, and every fragment of every collision since Sputnik. The Kessler threshold — the collision cascade that would render entire orbital bands unusable for generations — sits on no balance sheet, because there is no counterparty to bill. Software people already know this shape. It is shared mutable state with no lock, no owner, and no rollback. We have spent thirty years learning how that fails, and we still ship it whenever the alternative is a hard conversation with another team.

The engineering word for the exposure is blast radius, and production teams obsess over it because the cost of an incident is almost never paid by the people who shipped the change. Low Earth orbit currently offers the largest blast radius available to a private company and the shortest list of parties who can be held to account for it. Call that asymmetry what it is in business terms: a pricing error. Pricing errors are the first thing capital finds.

The Jurisdiction Gap

Governor Newsom signed an executive order on March thirtieth. First of its kind. Safety audits, privacy reviews, bias assessments — all mandatory for state AI procurement within a hundred and twenty days. It gives California's chief information security officer the authority to evaluate and override federal supply chain designations, which in practice means the state can separate its technology procurement from the federal government's preferences entirely. Watermarking requirements for AI-generated images and video. The order was framed explicitly as a response to the Trump administration's rollback of AI protections, but read it as an engineer and it is a procurement spec. That is the lever a state actually holds, and it moves faster than legislation: purchasing requirements do not have to survive a court challenge to change vendor behavior, they only have to change the RFP. If you sell software into public agencies, that order is a roadmap item this quarter, not a policy debate.

Meanwhile, the Trump White House released its National AI Policy Framework on March twentieth. Preemption of state laws deemed burdensome. Light-touch regulation. Minimal national standards. Michael Kratsios, the director of the Office of Science and Technology Policy, wants the framework codified into law before the year ends. Senator Blackburn's legislative draft runs two hundred and ninety-one pages. The tension between federal preemption and state assertion is constitutional and structural, and in the context of orbital compute it is also beside the point. Starcloud's satellites will operate in sun-synchronous orbits between six hundred and eight hundred and fifty kilometers. No state regulates that altitude. No nation, in practical terms, governs it. The Outer Space Treaty of 1967 was drafted when orbital assets were rare, expensive, and operated by governments. It has no framework for a seventeen-month-old startup deploying eighty-eight thousand commercial compute nodes into a shared orbital commons.

The pattern matters more than the company. Build capability faster than governance, deploy into the space where governance does not yet exist, and let the resulting vacuum fill with the logic of capital instead of the logic of stewardship. This is not malice. Johnston is not a villain, and Benchmark is not conducting an assault on the orbital commons. It is what any optimizer does when the constraint function has a hole in it. Anyone who has watched a team hit a metric through a loophole knows how the story ends: the loophole is not a defect in the team, it is a defect in the metric.

Every system has a stated design and a revealed one. The stated design lives in the deck. The revealed design is whatever the system tolerates once it is under load and nobody is watching. Starcloud intends orbital compute. What its plan tolerates is the Kessler risk, the debris accumulation, the regulatory gap, and the precedent that any sufficiently funded entity can populate shared orbital space with commercial infrastructure at scales that were unimaginable five years ago and are now merely unfunded. Tolerances are the real spec. You find out what yours were after the thing ships.

The altar of compute is not a metaphor. It is a budget line. Seven hundred billion dollars from four companies. One hundred and twenty-two billion for a single funding round. One hundred and fifty-six billion committed by Oracle while thirty thousand workers clear their desks. One point one billion for a company that has launched one satellite and filed paperwork for eighty-eight thousand more. The inputs are real and measurable: kilowatt-hours, headcount, orbital slots claimed, the carbon from the rockets that carry the silicon that processes the tokens that generate the revenue that justifies the next launch. Every item in that chain carries a price except the orbit, which happens to be the one input nobody can manufacture more of.

So what do you do with this if you are the person signing a capacity plan? Three things, none of them exciting. Price the thermal path before the power source — on any accelerator deployment the binding constraint is watts rejected, not watts delivered, and that holds in a Virginia colo as firmly as it holds at eight hundred kilometers. Put a number on repair; if mean time to repair is unbounded, the availability figure in the deck is fiction no matter how good the hardware is. And find out who pays for the failure mode you are not modeling, because if the answer is someone who is not in the room, you have not found a cheaper architecture, you have found a subsidy. Starcloud's demo satellite was good work, and I hope Starcloud-2 flies. The constellation is a bet that six unpriced problems are easier than one priced one. I would take the other side of that trade — and I will still watch every launch, because the demo was the easy part, and the next eighteen months are where we find out what the radiator really costs.