Starcloud. Starcloud.
Starcloud has brilliant people and a brilliant paper. But the transistor is still the transistor.
Jerome Milton sent me a paper this morning. He has been doing data center work for more than thirty-five years, and he does not send links idly. He was pointing back to a piece I wrote a few days ago called Hyperscalers Are RCA. The thesis, briefly, was this. Across the past spring, county commissions from Box Elder, Utah to Fort Worth to Warrenton to Manassas to several towns in Maine and Ohio have been processing the same conflict — a hyperscaler arrives, asks for water rights or transmission capacity or a property tax exemption, and the people who actually live where the facility will sit push back. The pattern repeats because the logic element repeats. The logic element is the transistor. The transistor is exhausted.
The earlier piece argued that today’s hyperscalers are RCA — not because they will disappear, since RCA persisted as a brand into the 1980s, but because they are pouring extraordinary capital into a logic element that has already been replaced. Bell Labs announced the transistor in December 1947. RCA, Sylvania, and Western Electric were expanding vacuum tube production in that month. They went on expanding vacuum tube production for another fifteen years.
The paper Jerome sent is what happens when an industry committed to the obsolete logic element runs that logic element to its physical limit and proposes — with serious math, with European Commission backing, with the chief technology officer of Thales Alenia Space on the record — to launch the entire facility into orbit.
The paper Jerome sent
It is twelve pages. Dated September 2024. Published by a company then called Lumen Orbit and now called Starcloud. The title is Why we should train AI in space. The authors are Ezra Feilden, who holds a doctorate, along with Adi Oltean and Philip Johnston.
The opening pages cite, in order, Sam Altman saying we will need fusion or radically cheaper solar plus storage to power what is coming, Elon Musk saying we have a silicon shortage today and an electricity shortage on the near horizon, Mark Zuckerberg saying he would build bigger clusters than he currently can if he could find the power, Tom Mueller — the first employee at SpaceX — saying that the compute required by 2045 will exceed the base power of the planet, and Christophe Valorge, the chief technology officer at Thales Alenia Space, citing a European Commission feasibility study concluding that orbital data centers are the more environmentally responsible option.
That is the table of authorities the paper opens with. It is a serious table. It is also, read carefully, a list of senior people saying — without quite saying — that the transistor has hit a wall.
The argument that follows is serious as well. Terrestrial solar farms in the United States, the paper notes, achieve a median capacity factor of twenty-four percent; in temperate Europe that figure drops below ten. A photovoltaic array in low Earth orbit, perpendicular to the sun and unobstructed by atmosphere, achieves better than ninety-five percent. Peak irradiance is roughly forty percent higher. The same panel produces more than five times the energy in space that it produces on the ground. The paper does the arithmetic at five million dollars per launch with one hundred tons to low Earth orbit, and arrives at an equivalent energy cost of two-tenths of a cent per kilowatt-hour, against United States wholesale of about four and a half cents, the United Kingdom of six, and Japan of seventeen. Twenty-two times cheaper than what households and industries now pay.
Then cooling. Deep space, the paper observes, is at 2.7 Kelvin — the temperature of the cosmic microwave background. A black plate held at twenty degrees Celsius will radiate about 838 watts of waste heat per square meter of surface, dissipating to a sink so cold that the math works without chillers, without water, without any of the apparatus that defines a modern terrestrial data center. The paper walks through the Stefan-Boltzmann calculation, the emissivity coefficient of zero point nine two, the absorptivity of zero point zero nine against Earth’s albedo of zero point three, the solar irradiance of 1366 watts per square meter, the view factor of zero point two five. They arrive at a net radiated power of 633 watts per square meter at twenty degrees, and they show the curve up to forty degrees, where it reaches over a thousand.
The orbit is a dawn-dusk sun-synchronous orbit — roughly five hundred miles above the surface of the Earth, with the plane of the orbit precessing around the planet at the same rate the planet circles the sun, which means the spacecraft sits permanently along the terminator and the solar arrays never enter Earth’s shadow. The paper has thought about radiation, and proposes shielding at roughly one kilogram per kilowatt of compute. It has thought about debris, about decommissioning, about demisable burnup, about the International Space Station’s fifteen-year design life as a reasonable analog. It has thought about laser inter-satellite links with Starlink, Kuiper, and Kepler, and about Jim Gray’s 1990s observation that physical shipping of disks can sometimes beat networking — which is why their architecture includes what they call data shuttles, capsules launched from the ground carrying petabytes or exabytes of training data, docking with the orbital cluster like cargo. Amazon already proved the concept, the paper notes, by sending a Snowcone to the International Space Station with seven gigabytes aboard.
For the five-gigawatt cluster the paper argues will be required by 2027 to train the next generation of frontier models, the solar array is four kilometers by four kilometers. Four thousand meters on a side. The radiator field is roughly half that. The whole assembly deploys using Z-fold or roll-out structures, with the largest radiators ever flown in space, all positioned in-line with the solar arrays so one side of each radiator faces the sun and the other faces the void. The paper walks through the modularity of the compute containers, the high-voltage DC bus, the docking ports, the redundancy at the system level, the eventual salvage or controlled re-entry. They cite Microsoft’s Project Natick — the sealed underwater data centers Microsoft operated for years — as evidence that environmentally isolated, sealed compute can outlast its terrestrial equivalent.
It is a beautiful paper. It is the kind of paper a serious team writes when serious money is on the table and serious people are paying attention. I have nothing to say against the math, the orbital mechanics, the launch economics, or the thermodynamics. Every number in the document is, as far as I can tell, defensible.
What the paper never argues for
Read it twice and notice what is missing.
Nowhere in twelve pages — through the energy section, the cooling section, the launch section, the orbital selection, the maintenance section — does anyone defend the transistor. The compute containers in the diagrams are racks of GB200-class NVL72 units at one hundred and twenty kilowatts apiece. The cooling infrastructure is direct-to-chip liquid or two-phase immersion. The thermal load that the radiators must dissipate is the thermal load that transistors generate when they switch billions of times per second to perform Boolean logic. The four-kilometer solar array exists because that thermal load exists. The radiator field exists because that thermal load exists. The launch cadence exists because that thermal load — which is to say, the power draw — exists in volumes the planet’s grid cannot supply.
The transistor is held constant. The entire paper is what falls out of that single assumption.
Once you hold the transistor constant, you are committed. You are committed to a power budget that exceeds the largest power plant in the United States. You are committed to a cooling architecture that requires either fresh water on the ground or deployable radiators in orbit. You are committed to a permitting and grid-interconnect timeline that — as the paper accurately notes about xAI’s Memphis cluster — already forces operators to truck in mobile natural gas generators because the utility cannot deliver. You are committed to a path where, at some point, the cheapest remaining solution is to launch the data center into space.
Get Derek Wayne Bailey’s stories in your inbox
Join Medium for free to get updates from this writer.
This is the through-line from Box Elder County to low Earth orbit. The Stratos developers and the Starcloud authors are not opponents. They are not even on different teams. They are running the same problem at different points along its curve. Stratos is still asking the Bear River watershed for the water. Starcloud has already concluded that the watershed will not give it, and the grid will not give it the power, and the planning commissions will not give it the permits, and so the only place left to put the facility is above the planning commissions altogether — in low Earth orbit, at the dawn-dusk line, at seventeen thousand miles per hour.
Given the transistor, the conclusion is correct. If the transistor is fixed, Starcloud’s paper is more right than the people on the ground stringing transmission lines through endangered species habitat. They have done the arithmetic the rest of the industry is avoiding. The two-tenths of a cent per kilowatt-hour is real. The 633 watts per square meter is real. The dawn-dusk orbit is real. The launch economics, if the next generation of heavy-lift vehicles arrives on schedule, are real.
What is not examined is the logic element that produced the arithmetic.
And the other piece applies too
Some weeks ago, before Hyperscalers Are RCA, I wrote about brilliant people breeding faster horses. The argument was that in the late nineteenth century, when the horse-drawn cities of the world were drowning in manure and the streets of New York and London had become so foul that public health authorities were running out of vocabulary, the best minds in animal husbandry produced extraordinary work on equine genetics, gait optimization, feed efficiency, harness design. They were brilliant. They were also, in the matter that mattered, irrelevant. The problem was not solved by a faster horse. The problem was solved by a different animal — the internal combustion engine, which was not an improved horse but the absence of a horse.
The Starcloud paper is that kind of brilliance, layered on top of the RCA pattern. The people who wrote it are not foolish. The investors who funded it are not foolish. The European Commission that ran the ASCEND feasibility study is not foolish. The chief technology officer of Thales Alenia Space — a company that has built spacecraft for fifty years — is not foolish. Tom Mueller, who built the engines that put humans back into orbit, is not foolish. None of them are foolish. They are working very hard on a problem they have correctly identified, with an assumption they have not.
The assumption is that the logic element of computing — the device that stores, switches, and copies the bit — must remain the transistor. Once you accept that, the path leads, eventually, to four kilometers of solar panels in low Earth orbit. The path is not wrong. The starting point is.
Replace the logic element, and the entire paper collapses gracefully. The five-gigawatt power draw becomes a five-hundred-megawatt power draw, which fits on terrestrial solar with reasonable acreage and no launches. The cooling infrastructure becomes negligible because the switching event itself does not dissipate the way a transistor’s switching dissipates. The four-kilometer radiator field becomes a building. The dawn-dusk orbit becomes a parcel of ground in Louisiana or California or Cabo Verde. The hundred launches become zero launches. The Snowcones full of training data become a fiber-optic link.
This is not theoretical. The Poovey Switch was validated externally at the Technion Israel Institute of Technology at one hundred fifty to two hundred femtoseconds — against a prior literature best of six hundred — The switch is on a new substrate: graphene-boron nitride on sapphire, which conducts heat at two thousand watts per meter-kelvin, an order of magnitude better than the substrate transistors are built on. The Boolean logic family is complete. The cooling-water requirement at scale is not reduced; it is near zero. The capital expenditure for a Clean Compute Center delivering compute equivalent to a two-gigawatt facility built on transistors runs to twenty-five to forty million dollars, against an industry per-gigawatt figure of thirty to fifty billion. Three orders of magnitude. Not three percent. Three orders.
Why Jerome sent it
Jerome is not naive. He has been around long enough to recognize the shape of a wrong assumption when it is dressed in real engineering. He sent the paper because it was the cleanest specimen he had seen recently — twelve pages of credentialed, careful, mathematically defensible work, all of it downstream of a choice no one in the document acknowledges making.
RCA did not lack engineers in 1947. RCA had some of the best engineers in the country. What RCA lacked was the willingness to retire the vacuum tube. Shockley, Bardeen, and Brattain had announced the transistor that December. The path was visible. RCA’s response was to build better vacuum tube factories, and they kept building them for two more decades. The factories ran. The tubes worked. The logic element of computing had already changed underneath them, and RCA went on perfecting what was no longer the question.
The Starcloud paper is what the same impulse looked like in 2024, with twenty-first-century engineering applied. The math is sound. The orbit is correct. The radiators will radiate. The launches, eventually, will launch. The four-kilometer solar arrays will unfurl in low Earth orbit, perpendicular to the sun, year-round.
And five hundred miles up, the transistor will still be the transistor.
Jerome knew that when he sent it. The rest of the industry has not yet noticed that the most ambitious space project of the century is, in the end, the most expensive vacuum tube factory ever proposed — built above the planning commissions that would not approve it on the ground, organized around a switching element that did not have to be the switching element.
Derek out.
Derek W. Bailey is the Founder of True Photonic, Inc. and author of Keep Computing: How Light Solves Computing’s Impossible Problem (2026). TPI is developing the Logic-in-Light™ architecture centered on the Poovey Switch, validated at 150–200 femtoseconds at the Technion Israel Institute of Technology.
Contact: dbailey@truephotonic.com
Disclosure: The author has a financial interest in the success of photonic computing technology. Performance projections are based on preliminary research, early-stage testing, and technical simulations. Actual results may vary.
Forward-Looking Statements: This article contains forward-looking statements regarding future technological capabilities and market developments. These statements involve risks and uncertainties that may cause actual results to differ materially. Readers should not place undue reliance on forward-looking statements, which speak only as of the publication date.
