Challenges of Orbital Data Centers: Heat, Cost, and Launch

The challenges of orbital data centers are the part of the story that separates a real infrastructure thesis from a brochure. Dylan Jovine’s “Elon’s 100X AI Stock” pitch treats orbital AI compute as nearly inevitable, quoting Musk’s “The cheapest place to put AI will be in space within 36 months or less.” We read the pitch and the physics claims behind it, and here is the honest engineering view: what checks out, what is genuinely hard, and what the marketing glosses over.

The Case For Orbit, Stated Fairly

Start with what is verifiable. Terrestrial AI is hitting a power wall: Microsoft, Amazon, Google, and Meta are pouring a combined $725 billion into AI infrastructure in 2026, while Morgan Stanley projects the U.S. will be short 49 gigawatts of electricity for data centers by 2028. Google, the pitch says, now waits up to 12 years to connect a new data center to the grid. Denver, Oklahoma City, and Tulsa have paused new data center construction, and Maine’s legislature voted to ban them outright.

Orbit offers three things no earthbound site can: “a solar panel in space generates roughly eight times the energy of the same panel on the ground” (no night, no clouds, no atmosphere), continuous sun, and “Space is 250 degrees below zero, and it never warms up.” The concrete artifact in the pitch is AI1, a satellite whose “solar wings stretch 230 feet from tip to tip. That’s longer than the wingspan of a Boeing 747.” The race is real, not invented: the Nvidia-backed Starcloud trained the first AI model in space, Bezos is putting $1 billion a year into Blue Origin, and Google’s Project Suncatcher is underway. We mapped orbital data centers as an investment theme earlier this year.

Challenge One: Heat Has Nowhere to Go

Here is the counterintuitive part. A vacuum is an excellent insulator. On Earth, a data center can blow air past its chips; in orbit there is no air, so waste heat can only leave by radiation, through radiator panels that have to be enormous relative to the compute they serve. The pitch itself acknowledges this when it says radiators must be “enormous,” and the engineering logic flows straight from it: the less power a chip wastes as heat, the less radiator mass you have to fly. That is why Jovine’s investment angle centers on power-efficient silicon, “the only chip that can run AI in orbit without cooking itself,” in the pitch’s words, supplied by an unnamed licensor that “gets paid” on every AI1 satellite. The identification of that company is unconfirmed; the thermal logic behind why efficient chips matter in orbit is solid engineering.

Challenge Two: Launch Mass and Cost

A data center is the heaviest kind of computing facility you can build, and every kilogram has to ride a rocket. The pitch’s answer is scale: SpaceX has “about 10,000 satellites in orbit” today and, per the January 30, 2026 filing it cites, “SpaceX asks the FCC for permission to launch one million satellites.” Mass production of launch capacity is Musk’s core competency, so the argument is coherent when applied to SpaceX specifically. The honest caveat: a filed request is a request, not a built fleet, and the gap between 10,000 units and one million is the entire bet. The 100X framing in Jovine’s pitch (“he could soon face a 100 times bigger bill to this company”) is an inference from that request, not a forecast with engineering milestones behind it.

Challenge Three: Repair, Degradation, and Radiation

Nobody fixes a server in orbit. Terrestrial data centers swap failed drives and GPUs constantly; an orbital facility must carry redundancy for every component or accept degraded capacity. Radiation also stresses electronics over time in ways data centers on the ground never face. None of this makes the thesis unworkable, and the pitch’s free-cooling and free-power claims remain true, but these are the operating realities that determine whether the “36 months or less” timeline is credible or aspirational. The honest answer is that the first orbital facilities are experiments, and Starcloud’s in-space training run is exactly that: a proof of concept, not a hyperscaler replacement.

What This Means for the Investment Angle

Jovine’s pitch lands on a supplier thesis: “Elon doesn’t own that chip. He licenses it. That means, every time a Tesla rolls off the line… Every time SpaceX launches an AI1 satellite… This company gets paid.” The royalty model is real, and the switching-cost argument (rewriting “every line of software in every Tesla and every satellite he owns”) is economically sound. The unverified part is everything specific: the 2012 supplier relationship, the doubling data center royalties, the blocked NVIDIA acquisition. We treat the teased company as unidentified; the clues point to a chip-IP licensor, and no more than that.

The urgency device is the December 8th SpaceX lockup expiration, when “the last block of roughly 800 million insider shares comes loose, all at once.” We covered that event in our SpaceX lockup expiration analysis, and Jovine’s Superdeal-era supplier thinking in our Superdeal stock pick coverage.

Bottom Line

The challenges of orbital data centers are real but tractable, and the direction of travel is genuinely toward orbit as the grid bottleneck deepens. The pitch’s physics is honest; its timeline and 100X math are the presenter’s inference, hedged as such. If you want to evaluate the full argument, the presentation is here, and the $49 subscription carries a six-month full-refund guarantee.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.