Data Centers in Orbit, Factories on the Moon: Why Betting Against SpaceX and xAI's Space-Compute Plan Is the Easy Wrong Call of 2026

Strip away the theatrics and the 2026 plan is unusually specific. In February, SpaceX absorbed xAI and filed with the FCC for authority to operate up to a million satellites — not for broadband, but for compute. In June, Musk and a Starlink engineer walked through the first dedicated design: the AI-1, a satellite that draws about 150 kW at peak and sustains roughly 120 kW of compute, which is close to a single NVIDIA GB300 rack of seventy-two GPUs. Stretched out with its solar wings, one craft is wider than a Boeing 747.

The architecture is almost boring in its logic, which is usually a good sign. An AI satellite is simpler than a Starlink broadband satellite: no phased-array or parabolic antennas, just solar cells, a double-sided radiator to dump heat, an interchangeable chip payload, and laser links carrying on the order of a terabit per second. Power comes from sunlight that never sets and never clouds over, so there are no batteries and no glass-and-frame panels to ship — the cells can be made cheaper than anything that has to survive on the ground. Two prototypes are slated for early 2027, with pilot testing on the next-generation Starlink V3 bus before that.

The framing is deliberately grand — Musk talks about climbing the Kardashev scale by capturing a meaningful slice of the Sun's output — but the near-term numbers are checkable milestones, not slogans. Roughly one gigawatt of annualized orbital compute by late 2027, then a stated ambition to scale it about tenfold a year toward 100 GW by 2030. You do not have to believe the 2030 figure to notice that the 2027 figure is a date you can hold them to. That is the difference between a roadmap and a vision board.

The orbital satellites are the visible tip. The more consequential bet is upstream, in manufacturing. Musk has framed Terafab as a one-terawatt-per-year compute factory — a chip foundry that Tesla, SpaceX and xAI are building together in Austin on a roughly $20-25 billion budget, starting near 100,000 wafer starts a month and aiming for a million. The point is vertical integration taken to its logical extreme: one foundry producing the silicon that powers Tesla's cars, xAI's data centers and SpaceX's orbital racks alike. When you own the wafer, you stop bidding against the rest of the industry for the scarcest input in AI.

Then there is the Moon. At an xAI all-hands in February, Musk floated a lunar manufacturing facility that would build AI satellites and fling them into orbit with a mass driver — an electromagnetic catapult — rather than a rocket. On paper the logic is not absurd: the Moon offers constant solar power, a hard vacuum for cooling, one-sixth of Earth's gravity, and no atmosphere to fight on the way up. He paired it with talk of a self-growing lunar city within a decade, arguing the Moon is operationally easier than Mars because you can launch every ten days instead of waiting twenty-six months for a planetary alignment.

It is fair to file the Moon factory under aspiration and the AI-1 under engineering. But notice the through-line: every piece is about removing a manufacturing or energy bottleneck rather than inventing new physics. Cheaper power, cheaper cooling, cheaper launch, cheaper silicon — the same playbook that turned Falcon 9 from a stunt into a freight service. The Moon is the speculative end of a spectrum whose other end is already bending metal in Texas.

There are two very different groups saying this will not work, and conflating them is how you end up badly informed. The first is made up of engineers and analysts raising real constraints. Cooling is the big one: a satellite cannot blow waste heat into air or water, it can only radiate it away as infrared, and radiative cooling is far less efficient than convection. The International Space Station's radiator wings reject about 70 kilowatts; a single modern AI rack can draw more than that. Add 90-minute sunlight-to-shadow cycles in low orbit that fatigue hardware, radiation that forces less-efficient rad-hardened chips (which run hotter, worsening the cooling problem), and the fact that orbital hardware is expensive to launch and nearly impossible to service. These are serious, and SpaceX has not fully answered them.

The second group is the problem. Across LinkedIn and YouTube, a genre of pseudo-expert has discovered that dunking on the idea performs well: a confident thumbnail, a here-is-why-this-is-impossible headline, and a physics-flavored paragraph that mistakes hard for cannot-be-done. Even Sam Altman called orbital data centers ridiculous. The tell is that these takes almost never engage with the actual engineering tradeoffs or the milestone schedule; they pattern-match to Musk hype and collect the engagement. Skepticism about economics is intellectually honest. Declaring a thing impossible because it sounds impossible is not skepticism — it is a content format.

The honest position lives between the two. The thermal and economic objections are real and may well push the timeline out by years or force a more modest architecture than the headline numbers. But the unit economics are unproven in 2026 and this can never happen are completely different claims, and the people making the second one are borrowing the credibility of the people making the first. When you read a confident impossible, the useful question is simple: is this person describing a law of physics, or a hard problem with money and time against it?

Here is what the dunk-posters keep forgetting: this is the single most accomplished hardware-deployment organization in history. SpaceX operates more than 10,000 active Starlink satellites — about two-thirds of everything currently in orbit, a constellation larger than every other operator on Earth combined. Falcon 9 has flown over 600 times, with individual boosters reused more than thirty times and recovery now so routine it barely makes the news. Reusable orbital rockets were textbook-impossible until SpaceX landed one; the consensus of serious aerospace opinion was that it could not be done economically, and that consensus was wrong.

The pattern repeats across Musk's companies. Tesla was supposed to be a rounding error that legacy automakers would crush; instead it made the electric car mainstream and forced every incumbent to follow. The lesson is not that Musk is always right — he is routinely late, and some promises never arrive on the original timeline. The lesson is about the asymmetry of the bet. Betting against this team on a hardware-manufacturing problem has been, repeatedly, a cheap thing to say and an expensive thing to be wrong about.

So calibrate accordingly. The base rate for SpaceX attempts an audacious manufacturing-and-launch problem and eventually delivers, late and reworked is high. The base rate for a LinkedIn influencer who called Starlink a fantasy in 2019 was right this time is low. None of that suspends physics — if radiative cooling truly caps the economics, no track record overrides it. But it should make you deeply suspicious of confident impossibility from people whose last three impossibility calls are now flying overhead.

From where we sit in Osaka, the most useful part of this story is not the rockets — it is the manufacturing thesis underneath it. The real moat Musk is building is not a satellite; it is Terafab, the owned foundry, and the lights-out, highly automated factories that feed everything else. That is a thesis Japan understands in its bones. Japan invented the modern automated factory, FANUC has run near-lights-out robot-building-robot plants for decades, and the country is now pouring capital into semiconductor sovereignty through Rapidus's 2nm effort in Hokkaido and TSMC's fabs in Kumamoto.

There is also a concrete supply-chain angle. An orbital data center is, at its heart, a thermal and materials problem — radiators, heat pipes, radiation-tolerant components, precision optics for laser links — and these are categories where Japanese manufacturers quietly lead the world. Whether or not SpaceX's specific architecture wins, the broader shift of compute toward power-constrained, thermally exotic environments plays to exactly the precision-manufacturing strengths that Japanese suppliers have spent decades perfecting. The opportunity for a Japanese firm is to be in the bill of materials, not in the comment section.

For cross-border decision-makers the takeaway is a discipline, not a side. Do not outsource your judgment to dunk-posts, and do not swallow the hype whole either. Separate the physics objections (respect them) from the timeline-and-economics objections (price them) from the it-sounds-crazy objections (ignore them). Watch the checkable milestones — AI-1 prototypes in 2027, Starlink V3 pilots, Terafab wafer starts — and update on those, not on engagement-farmed certainty. The companies that win the next decade will be the ones supplying the audacious projects, not the ones narrating why they were always going to fail.