|
The bottleneck for AI moved from chips to electricity
Picture the machine at the center of this. SpaceX’s first orbital data center satellite, called AI1, stands 20 meters tall with a wingspan of 70 meters, wider than the fuselage of a Boeing 747. It is designed to unfold in orbit, drink in constant sunlight, dump its heat into the vacuum of space, and run AI calculations far above the nearest power line. SpaceX has asked to launch up to a million of them. In its filing, the company argued that within a few years, freed from the limits of building on Earth, the cheapest place to make AI compute will be orbit rather than the ground.
Strip away the science-fiction sheen and this is a confession. The most capitalized companies in the world have looked at the American power grid and concluded it may be easier to leave the planet than to plug in. That is the signal worth taking from all of this, and it has almost nothing to do with rockets.
Why the grid gave out. Artificial intelligence is astonishingly hungry for electricity. A single AI query can use up to a thousand times the power of an ordinary web search, and the servers behind them are multiplying fast. PJM, the largest US grid operator, which serves 65 million people, now projects a six-gigawatt shortfall against its own reliability targets by 2027. Texas has paused new grid connections. High-voltage transformers that once took months to order now take three to five years. Gartner expects power shortages to restrict 40 percent of AI data centers by 2027. The industry spent years worrying about a shortage of chips. The shortage that arrived was electricity.
Why space is the pitch. In orbit, three of the grid’s hardest problems disappear. The sun shines constantly, so power is free and uninterrupted. The vacuum of space carries heat away, so there is no need for the rivers of cooling water that terrestrial data centers consume. And there is no land to buy, no county to lobby, no neighbors to placate. Google, running the same math, has estimated that once rockets are cheap and reusable enough, launching data centers to space could beat building them on the ground by around 2035. Nvidia is already selling the chips for it.
Why to keep a hand on your wallet. Now the cold water. SpaceX’s filing is not approved; regulators have only opened it for comment. The company has not named a single paying customer for orbital compute, and its own investor documents call the effort early-stage and unproven at commercial scale. A million satellites would multiply the number of objects in orbit many times over, raising real concerns about collisions and debris. And valuations have already run ahead of the hardware: one orbital startup is worth more than a billion dollars on the strength of a single 130-pound satellite and no disclosed revenue. When a story is this good, the price usually arrives before the proof.
Where the durable money actually is. Here is the part that matters for a long-term investor, and it sits firmly on the ground. Whether or not a single AI satellite ever turns a profit, the demand for electricity to run artificial intelligence is real, enormous, and growing now. That points to the least glamorous corner of the market: the companies that generate power, build and repair the grid, make the transformers and turbines, and supply the natural gas and nuclear capacity that data centers are signing decade-long contracts to secure. The satellites may or may not fly. The power bill is already here. And for the cautious, a Treasury bill still pays more than 4 percent while you decide what, if anything, to believe about any of it.
A million satellites may never leave the ground. The reason a company would try tells you everything: the AI boom has outgrown the planet’s ability to power it. Watch where the electricity comes from. That is the trade the rockets are pointing at.
|