The Situation For twenty years, the SpaceX story has been simple enough for anyone to follow: build rockets, land rockets, reuse rockets, launch satellites, get humans to Mars someday. Starlink came along and turned into the company’s obvious cash engine, internet beamed down from orbit to homes, boats, and battlefields in 167 countries. That’s the SpaceX most people think they know. Rockets. Satellites. Musk talking about Mars. It’s also, according to Musk himself, already out of date. The Struggle Here’s the problem every AI company on Earth is running into right now, whether it’s Microsoft, Google, or the smaller players building the tools you and I use every day: AI needs an almost unthinkable amount of electricity and an almost unthinkable amount of cooling, and Earth is running out of cheap ways to supply both. Data centers on the ground are hitting real walls, power grids that can’t keep up, water rights fights over cooling, land leases, taxes, permitting delays, lawsuits. Musk has personally run into this: at the Colossus data center in Memphis (built by xAI, which has since been folded into SpaceXAI), the local grid couldn’t supply power fast enough, so the company trucked in dozens of mobile natural gas turbines to keep the AI training clusters running. It worked, but it also became a legal headache. Regulators later ruled that many of those turbines were operating without the required Clean Air Act permits, drawing lawsuits from the NAACP and environmental groups, and the company has committed to fully removing the unpermitted units by mid-2027 as it transitions to a permanent power plant. That’s not a hypothetical bottleneck, that’s Musk racing the current AI boom against Earth’s own infrastructure, permitting, and regulation, and hitting friction on all three. Meanwhile, the appetite for compute isn’t slowing down. In his own all-hands talk, Musk cited a Cloudflare estimate directly to his employees, saying AI traffic could be a thousand times greater than human internet traffic within five years. Sit with that number for a second. Not double. Not ten times. A thousand times. So the struggle is this: the world wants exponentially more AI, and the ground can’t supply the power and cooling fast enough to keep pace. The Shift Musk’s answer, laid out plainly to his own employees, is to stop trying to solve that problem on the ground at all, and move the data centers into orbit. He’s calling the project Starmind, and the pitch is almost absurdly simple once you hear it: * No cooling costs. Space is cold. You don’t pay for water, chillers, or climate systems. * No power bill, ever. Once solar panels unfold in orbit, you get free sunlight 24 hours a day, 365 days a year, no clouds, no night, no grid. * No land, no lawyers, no local permitting fights. Just the cost of building and launching the hardware. The economics, according to figures laid out in a JCristina breakdown of the situation, are stark: ground-based AI data centers currently run somewhere around $35–40 million per megawatt in ongoing costs, while orbital compute, once built, could drop to roughly $8–10 million per megawatt, because after the satellite is up, the power and cooling are essentially free forever. Those specific figures are estimates from that analysis rather than numbers Musk stated directly, so treat them as illustrative of the direction of the cost advantage rather than an official SpaceX projection. Musk’s own words to his team made the ambition concrete. He said SpaceX is targeting 10 gigawatts of AI compute online by the end of 2027, and that at an estimated $30–50 in value per watt, that scale of compute translates to somewhere between $300 and $500 billion a year in revenue, from AI alone. That first 10 GW is just the starting line, too; both source videos describe SpaceX scaling well past it once the orbital buildout is proven out. And then the line that should stop anyone paying attention: Musk told his own employees that within four to five years, 99% of SpaceX’s total value will come from AI, not rockets, not Starlink. Not “might.” He said it twice, for emphasis, in front of the room. He even said SpaceX’s AI revenue is expected to pass all of SpaceX’s other revenue combined by this September, and significantly exceed it by the fourth quarter of this year. Let that land: the company famous for reusable rockets is telling its own workforce that rockets are becoming the delivery truck, not the destination. Starship’s real job, in this new framing, isn’t Mars tourism, it’s hauling satellite-sized AI data centers into orbit, sometimes 60 at a time, on a cadence Musk wants pushed to daily, then hourly, launches. Why This Isn’t Just an Investor Story It’s tempting to file this under “stock news” and move on, and yes, analysts have already started revising SpaceX price targets upward on the back of this shift, with at least one firm quoted setting a new target around $160. But that’s not really the point for most of us. The point is what this signals about where AI is actually headed: A quick clarification before the takeaways: the 10 GW and $300 to $500 billion numbers Musk gave are for SpaceX’s overall AI compute buildout, which is happening on the ground right now, not exclusively in orbit. Starmind, the orbital piece, is the longer-term solution he’s aiming at: training stays on Earth, but the day-to-day running of AI models (what’s called inference) is planned to shift into orbit aboard dedicated Starmind satellites, with launches targeted to begin next year. So think of it as two overlapping timelines: SpaceX is racing to scale AI compute on the ground right now, while building toward Starmind as the move that removes the ground’s power and cooling constraints entirely. 1. The bottleneck for AI was never really intelligence, it was power. For the last few years, the AI conversation has been dominated by model size, parameters, benchmarks. Musk’s bet says the next constraint isn’t smarter models, it’s finding enough cheap, clean energy to run them at scale. Solving that in orbit is a bet that whoever solves the power problem first controls the next decade of AI. 2. This is happening faster than most people expect. Musk isn’t describing a 20-year moonshot. He’s talking about revenue crossovers by this quarter and 99% value shift within five years. If that pace is even directionally right, the AI landscape most of us are adjusting to right now will look completely different by 2030. 3. AI infrastructure is becoming its own industry, separate from the AI models people actually talk to. Most people think about AI as ChatGPT, Grok, Claude, the chatbot in front of them. But underneath every one of those tools is a hardware and energy story, and that story is now being rewritten in orbit. For anyone building a business, a skill set, or just a plan for the next five years, that third point is the one worth sitting with. The AI conversation isn’t just “which chatbot is smartest.” It’s who controls the compute, and increasingly, that answer is being written above the atmosphere. The Part Worth Staying Skeptical About None of this is a done deal, and it’s worth saying so plainly. Launching thousands, eventually potentially up to a million, of AI satellites on a daily or hourly cadence is a manufacturing and logistics problem nobody has solved at that scale before. Hardening chips against radiation, keeping laser communication links reliable across a massive constellation, and securing enough GPU supply from partners like NVIDIA are all real, unresolved challenges, not settled facts. Musk’s own track record includes plenty of ambitious timelines that slipped by years, not months. So hold this vision with genuine excitement, but also with the understanding that “SpaceX said it will happen” and “it happened” are two different things, and the gap between them has swallowed plenty of bold predictions before. What do you think, is orbital compute the inevitable next step for AI, or is this the kind of bet that sounds obvious right up until it isn’t? I’d love to hear where you land. Sources: Elon Musk’s SpaceX all-hands address, posted to X on August 11, 2026, as reported by Teslarati and 24/7 Wall St.; commentary and cost breakdown from Joseph Cristina’s YouTube channel. Get full access to AI for Humans at henryaitech.substack.com/subscribe
