One Million Compute Satellites? Morgan Stanley Just Handed SpaceX the AI Power Crown — and the Grid Can’t Compete

Posted on by

(SeaPRwire) –   By: Reginald Vance

The AI buildout has a dirty secret, and it is not chips. It is electricity. Morgan Stanley’s Stephen Byrd projects AI power demand will approach 100 gigawatts by 2028, roughly 8% of total U.S. generating capacity, and utilities physically cannot build that fast. That bottleneck is now the market’s central panic. Hyperscalers are staring at interconnection queues measured in years. Into that gap walks SpaceX, fresh off a mid-June 2026 IPO and a roughly 45% stock run since August, despite slipping 0.6% to $151.85 on Monday while the Nasdaq rallied 2.3%. Byrd’s argument is blunt. SpaceX has proven it can scale time-to-power solutions faster and cheaper than rivals, and he calls that the key competitive advantage against a projected national power shortfall. This is a hardware scaling story dressed as a software story. Whoever gets electrons to silicon first wins the next four years of AI economics. Everyone else pays waitlist prices.

Now catalog the actual hardware commitments. SpaceX closed Q2 2026 with 1.4 gigawatts of AI computing capacity and is targeting 10 GW by the end of 2027, with Byrd modeling terrestrial compute peaking near 15 GW in 2031. That is a sevenfold capacity jump in eighteen months, a pace no conventional data center developer is matching. Then comes the orbital layer. Musk says he is highly confident SpaceX will begin launching NVIDIA’s most advanced AI computers into orbit as early as 2027, adapting Starlink satellites with NVIDIA chips and solar panels, scaling to larger deployments in 2028, and floating a long-term constellation of up to one million compute satellites. NVIDIA is the exclusive chip supplier and already holds an equity stake in SpaceX, so this is vertical alignment, not a purchase order. Starship underpins the entire cost curve, with its 14th test flight set for September 28. The physics, however, bites hard. Cooling in vacuum demands radiators spanning millions of square feet. Radiation shielding, downlink logistics, zero repair access, and unresolved regulation all stack against the concept. Analysts see no meaningful orbital revenue before 2029, treating 2027 and 2028 as feasibility runs.

Follow the cash, and the endgame sharpens. The terrestrial 10 GW build is the real business, funded by a public market that just handed SpaceX a 45% post-IPO re-rating, while the orbital plan functions as a call option on infinite solar power and zero permitting fights. NVIDIA’s dual role as supplier and shareholder means every gigawatt SpaceX energizes feeds chip demand on top of existing earthbound appetite, locking one vendor into both sides of the transaction. My read on consolidation: the AI infrastructure race collapses into whoever controls power-first compute, and the orbital experiment, revenue-deferred past 2029, becomes the moat that keeps every terrestrial rival permanently behind.