SpaceX and Nvidia's Plan to Move AI Compute Into Space

The argument begins on the ground, not in orbit. In a recent commentary, a market commentator and self-described future forecaster points to an interview with OpenAI co-founder Greg Brockman acknowledging that compute constraints could pace, inhibit, or even stop AI from reaching everyone. His list of constraints is familiar: not enough data centers, not enough places to put them, not enough energy, and political resistance from communities pushing back on construction. On his reading, 'probably the biggest bottleneck in the entire economy of 2028, 29 and going forward will be the compute buildout of Earth.'0:33

He singles out inference compute, the deployment of trained models to users, as the binding constraint. In his view, training compute will eventually matter less because frontier training runs will 'very likely be outlawed' around 2029 as too dangerous, a forecast he admits may be wrong. Meanwhile, by 2028 he expects 'full-blown AGI,' after which, he argues, every business and project becomes a function of the ability to deploy inference compute. It is a sweeping thesis, and he presents it as his own conviction rather than consensus.1:06

The proposed solution is SpaceX's plan to move the buildout itself into space. About a month before the commentary, Nvidia and SpaceX said they are developing Vera Rubin VR NLV72 compute racks, with a custom, lighter, space-resilient version designed specifically for satellites rather than terrestrial data centers. The satellite architecture the speaker describes is straightforward: solar panels that stay locked on the sun in sun-synchronous orbit, radiators to shed heat, and the Nvidia rack in the middle.2:30

The arithmetic he offers makes the scale concrete. Each satellite would collect roughly 240 kilowatts, so a single gigawatt would require around 5,000 satellites, synced by lasers. He stresses the components are established technology: SpaceX already operates about 10,000 Starlink satellites using laser links, which he says would be the equivalent of two gigawatts if they were these new AI satellites. On cooling, the perennial objection, he is dismissive, noting Musk's response that cooling on Starlink is 'solved' and that radiating heat in vacuum 'it's known for 100 years.'3:58

For Nvidia, he argues, the stakes are enormous: if space offers unlimited energy and superior cost structures and timelines, terrestrial data centers bottlenecked by politics and power could, in his words, see those problems 'fall to the wayside' as early as the end of 2027.7:28

The decisive question, the speaker concedes, is timing. Elon Musk 'just 20 hours ago' tweeted that he is 'highly confident that SpaceX will be launching Nvidia VR NLV72 AI computers in space next year,' and the speaker takes this as confirmation of a timeline already in full-blown development. He speculates the first test satellites may already have flown and that operational satellites could launch in as few as two Starship flights, with Starship's reusability driving down orbital transport costs enough to let space compute 'out compete terrestrial compute data centers.'4:48

His own forecast is aggressively early: the first commercial satellite ready in 2027, the first gigawatt possibly in 2028, five to ten gigawatts the year after, and hundreds of gigawatts in the 2030s. Most analysts, he acknowledges, expect this in 2030 or 2031, and he thinks they are wrong. He also pushes back on the reflex that 'Elon is always late,' arguing Musk is sometimes early and that the underlying evidence of Nvidia's development matters more than Musk's statements.10:00

Notably, the speaker thinks a SpaceX IPO pitch centered on enterprise AI undersells the company. Grok and enterprise AI are 'an enormous opportunity,' he says, but a riskier one because it requires SpaceX to beat Anthropic and OpenAI, two rivals he describes as 'massively in the lead.' The more robust baseline, in his view, is compute buildout itself: the 'Dyson swarm,' becoming the sole provider of compute and intelligence for humanity.10:32

His conviction extends to his own portfolio. He reports significantly increasing his SpaceX position on the day of the commentary, around $148-150, reasoning that a Wednesday selloff tied to a share unlock, 328 million shares or 15% of float unlocking in ten days, creates a predictable rhythm of two to three days of selling pressure he plans to trade around. He openly frames this as a game enabled by his own tracking tools, while cautioning viewers that his short-term trades 'are not for everyone.'12:19

The broader market context is tense. He describes an 'AI pacing scare' dragging stocks down, a Fed decision he sees as largely baked in, and what he calls the most stressful investing environment in memory, citing CNBC contributor Chris Camillo. His advice is characteristic: stay exposed because the fundamental trajectory is 'incredibly bullish,' but stay cautious, because the environment is 'not safe territory' surrounded by political and geopolitical risk.56:10

What gives the commentary its urgency is the speaker's parallel conviction about risk. He argues at length that aligning a superintelligence is fundamentally impossible: whatever objectives are designed into such a system, it can redesign them, since unlike humans, whose core drives are genetically fixed, an ASI can modify its own components. He also invokes an evolutionary argument, that even a benevolent ASI would be displaced by a more resource-maximizing one, and cites his claim that researchers including Max Tegmark reached similar conclusions.33:22

That belief loops back into the compute thesis. If frontier training is eventually halted as too dangerous while inference deployment accelerates, he predicts '99% of all compute will be inference,' which he calls 'very space compatible.' The orbital buildout, in his telling, is thus not just an infrastructure story but a civilizational hedge: humanity deploys the intelligence it has built while containing the process that creates more. Whether any of this happens on his timeline, on the analysts' timeline, or at all, he leaves pointedly to the trackers.20:17