On June 12, 2026, SpaceX comes to public markets at $135 a share, a valuation near $1.75 trillion and about 107 times sales, among the most expensive equities ever brought to investors. Some of that price is artificial intelligence. Goldman Sachs, the lead underwriter, projects $322 billion of AI revenue in 2030 inside a $474 billion consolidated total, and it published that single segment line with nothing beneath it. The question worth asking before the roadshow talking points harden is not whether to trust the number. It is what the number is built from, and how much of it rides on Grok, the model Elon Musk names most often.
Break the aggregate into its parts and most of it has little to do with how good Grok is.
Start with the one piece that carries a documented price today: compute, which SpaceX earns as a landlord rather than as a model-maker. SpaceX rents capacity on Colossus 1, its original Memphis cluster, to rival labs. Anthropic pays about $1.25 billion a month under terms scoped through May 2029, and Google about $920 million a month from October 2026 through June 2029, each running its own models on SpaceX hardware. Those two contracts alone run near $26 billion a year, and not one dollar of it depends on Grok’s quality. The tenant’s model does the work, and SpaceX collects the rent.
Goldman broke out nothing below its total, so the rest has to be reconstructed from the disclosed contracts and published pricing. Built from the bottom up, the 2030 projection sorts into four blocks. Operational AI, the compute leasing plus the autonomy embedded in vehicles and the network, comes to roughly $92 billion, about 29% of the projection, and depends on Grok not at all. Grok itself, the language-and-reasoning model, comes to roughly $45 billion, about 14%. X-platform advertising adds roughly $12 billion, about 4%. The largest block by far, roughly $173 billion or about 54%, is a pure residual: the gap between what can be built from documented sources and Goldman’s much larger aggregate. Recon Analytics reads that residual as further compute and infrastructure growth, but nothing that we can see documents it, so the conclusion that compute and operational AI dominate the projection holds only if that reading is right. All four sizes are Recon Analytics estimates. Goldman published none of them.
That leaves Grok carrying about a seventh of the projection, roughly $45 billion by 2030, against revenue well under $2 billion in 2025, inside an AI segment the S-1 reports at $3.2 billion that year. The climb the projection assumes, from under $2 billion to $45 billion in four years, is the part of the SpaceX AI story that actually turns on Grok. At about 107 times sales, every dollar of projected revenue carries a steep premium, and the 14% exposed to Grok carries it like the rest.
That climb is not a given, and the people closest to Grok are the reason. Among AI users who name a primary tool, only 1.7% name Grok, and that base is the most male-skewed of any major assistant at 77.2% against 54.2% for the field. They adopt AI at work less than the field does and bring Grok to work least of any major tool, they are the least willing to pay more, naming $6.50 a month to upgrade against $11.70 for Claude’s users, and they name bias a concern most often, at 23.0% against 13.5% for the field. A revenue line built to reach $45 billion has to convert a base that is, for now, the hardest of the majors to convert. The counterweight sits in the same data. Grok’s accuracy rating climbed from clearly negative into positive territory in April and May 2026, and the users who already pay report good value for the money. That is early, directional evidence that the rebuild Musk describes may be reaching the product, even as the trust and bias gaps have not moved.
So the disaggregation points an investor toward a cleaner way to price the AI line. Pay for what Grok does today, which is real but modest: bounded jobs across X, Tesla, Starlink support, and SpaceX’s own engineering, none of them in a loop where a wrong answer costs a life or a mission. Then treat the run to $45 billion as an option on a rebuild that Musk says is underway, priced as a probability rather than a promise. Most of the $322 billion is landlord compute and operational autonomy that will earn or fail to earn on their own merits no matter what happens to Grok.
One sensitivity sits above the others, and it cuts the other way. The Anthropic and Google leases carry 90-day cancellation clauses, and Musk has described the Anthropic arrangement in public as limited to six months, which sits in tension with the headline term running to 2029. If those leases shorten, the operational base that anchors this whole reconstruction shrinks with them. That is the single largest swing in the picture, and it works against the compute revenue, not against Grok.
What to watch from here is narrow. During and after the roadshow, listen for what SpaceX and Musk claim about Grok’s role inside that AI number, because the gap between Grok’s small current revenue and the large role the selling story implies for it is the whole tension. Watch whether the compute leases hold their headline terms or shrink toward Musk’s six-month characterization. And watch whether the rebuild produces a materially better model rather than a better name. Price the parts, not the headline, and most of SpaceX’s AI number sits clear of Grok.
Source: Recon Analytics AI Pulse, US named-primary AI users, March through May 2026 (Grok n=744; field n=44,541).