SpaceX's $1.75 Trillion IPO Funds the Most Vertically Integrated Data Center Stack in AI
AI's binding constraint has shifted from chips to power, and SpaceX's $1.75 trillion IPO funds the most complete response: on-site generation, an internalized xAI, a $55B chip fab, and solar-powered data centers in orbit. Until it ships, the revenue still lands on Nvidia, GE Vernova, and Vertiv.
Elon Musk now owns the launchpad, the power, the chips, and the orbit. The public-market revenue still lands on three companies he is trying to stop paying.
Disclosure: I hold a position in Tesla ($TSLA), which would absorb SpaceX under the merger this piece discusses, per the house rules. This is impersonal research, not personal investment advice (see the disclaimer).
In the spring of 1927, ore boats from Michigan's Upper Peninsula docked at the River Rouge, Henry Ford's mile-square plant outside Detroit, where iron and coal and rubber went in one end and a finished Model A rolled out the other. Ford owned the mines, the ore ships, the railroad between them, even a rubber plantation in the Amazon, because he refused to let a supplier set the pace of his factory.
A century later, the binding constraint in artificial intelligence has moved off the chip and onto the power to run it, the years-long wait for a grid connection, and the permit fights over where the electrons come from. Elon Musk has built the Rouge for that problem, and on June 12 he gave it a stock ticker. SpaceX priced its IPO at $135 a share and raised about $75 billion, the largest offering in market history, opened at $150 on Nasdaq under SPCX, and closed its first day up 19% at $160.95, a roughly $2.1 trillion company by the bell and the seventh most valuable on earth. The headline valuation sits near $1.75 trillion. The reason to read past the headline is the compute stack underneath it, most of which Musk now owns outright.
The Rouge, reassembled
Walk the stack. xAI's Colossus in Memphis is the power plant and the compute floor, one of the largest GPU clusters on earth, and since early February it sits inside SpaceX: Musk merged xAI into the rocket company in an all-stock deal that valued the combination near $1.25 trillion, folding Grok and Colossus under one roof before the IPO. Compute and launch now ride the same balance sheet.
Colossus runs partly on on-site gas turbines, because the grid interconnection queue moves too slowly to wait on (the bottleneck the power map on this site tracks). That is the short-term bypass: when the utility cannot hand you electrons on AI's timeline, you generate your own behind the meter. Starlink is the network and the cash engine, $11.4 billion of revenue in 2025 at high margins, the recurring money that funds everything upstream. SpaceX is the launch. And then there is the silicon.
In March, Musk announced Terafab, a chip plant near Giga Texas built with Tesla and the now-internal xAI. SpaceX's filings put the first phase at $55 billion and the full buildout as high as $119 billion, the most expensive factory ever proposed. Intel signed on in April as process partner, anchoring Terafab to its 14A node, the generation after 18A. It is a prototype first, design and fab and test under one roof, with volume years out, and roughly 80% of its planned output is earmarked for space-hardened "D3" chips for orbital satellites. Terafab is the Rouge's blast furnace, Musk making the silicon instead of ordering it from Nvidia. Which points at the strangest layer of the stack.
Compute leaves the planet
If on-site gas is the short-term bypass, orbit is the long bet, and it is further along than it sounds. Starcloud, a Redmond startup, put the first Nvidia H100 into orbit in November 2025 and raised $170 million at a $1.1 billion valuation to build solar-powered data centers in space. In May it ordered more than 50 Starlink Mini laser terminals to wire its satellites into a constellation. Starcloud is independent of Musk, yet it launches on SpaceX rockets and talks over SpaceX lasers, and its roadmap (Blackwell-class GPUs in orbit next) rhymes with Terafab's D3 plan. The pitch is plain physics: a data center in orbit draws continuous solar power, radiates its heat into space instead of evaporating water, and answers to no county permit board. Every constraint throttling the terrestrial buildout, power and cooling and water and siting, relaxes the moment the rack leaves the ground. Musk is building the heavy lift, the laser mesh, and the rad-hard chips meant to fly on it.
Who actually books the revenue
Here is the irony a vertically integrated empire creates for the investor. Ford's Rouge still bought its furnace linings and its glass lines from somebody, and that somebody booked the order. Musk's AI Rouge is the same, for now. Until Terafab ships at volume, the GPUs in Colossus and in orbit are Nvidia's. $NVDA reported $81.6 billion of revenue in the April quarter, up 85%, with data center alone at $75.2 billion, up 92%, and every new Musk cluster adds to that line. The turbines feeding Colossus while the interconnection queue grinds are the franchise of $GEV (GE Vernova), which carries a $163 billion backlog, booked $2.4 billion of data-center electrification orders in a single quarter (more than all of last year), and just pulled its $200 billion backlog target forward to 2027. The power distribution and liquid cooling inside the racks is $VRT (Vertiv) territory. Musk's whole vertical play is a bid to eventually stop ordering from those three. Every gigawatt he stands up in the meantime turns into their backlog.
The merger that closes the loop
The xAI half of the loop is already shut. The bigger one is Tesla. Within weeks of the IPO, Musk floated folding Tesla and SpaceX into a single roughly $3.4 trillion entity (Tesla near $1.6 trillion plus SpaceX's new public value). On IPO day, SpaceX president Gwynne Shotwell told CNBC a combination "might make Elon's life a little easier." Wedbush's Dan Ives puts the odds at 80 to 90% by early 2027; Polymarket bettors price closer to 38% for this year. A merger would put launch, power, compute, model, chip fab, and Tesla's energy-storage and grid assets inside one ticker, the most complete answer any public company has assembled to AI's power problem.
What the tape is not pricing
None of this is clean. The on-site gas bypass is already in court: the NAACP, the Southern Environmental Law Center, and Earthjustice are suing over the turbines powering Colossus, where xAI (now SpaceX) ran 27 of them without air permits at a site in Southaven, Mississippi, a plant the filings say can throw off more than 1,700 tons of smog-forming nitrogen oxides a year. That litigation is SpaceX's to defend now, the cost of importing xAI's operations onto a public balance sheet. Terafab is a prototype with volume years away, on an Intel node that has not run at scale. Orbital training clusters depend on a Starship cadence SpaceX has not yet demonstrated. And the cash engine under all of it, Starlink, booked $11.4 billion in 2025 against a market value north of $2 trillion, a price that already assumes the orbit bet pays.
Ford's Amazon rubber plantation was called Fordlandia, and it failed, the trees blighting and the workers walking off, because vertical integration has a limit, the point where the next input you insist on owning costs more than buying it would. The question for anyone pricing a $1.75 trillion stock, or the $3.4 trillion merged one Musk is dangling, is which layers of this stack genuinely break the power constraint (cheaper launch, solar in orbit, custom silicon) and which are expensive theater (a turbine farm in court, a fab that has not yet etched a wafer). Power is the limiter now. The first company to route around it at scale owns the decade. The chips that decide whether that company is SpaceX start coming off the Terafab line in 2027.