The Token Factory Squeeze: Why An AI Slowdown Makes SpaceX A Profit Machine

How could a slowdown in AI data center construction make SpaceX the most profitable company in the world? It sounds backwards. Fewer data centers going up looks like bad news for anyone betting on AI, and most people still picture SpaceX as a rocket company that sells internet from the sky.

But look at what SpaceX owns now. My argument is that agents are multiplying demand for AI work, the power grid can't keep up, and SpaceX is one of the few companies that can both build compute fast and run its own frontier model on it. That combination turns a slowdown into a windfall.

The rocket company that bought an AI stack

SpaceX rebuilt itself during 2026. In February it completed the buyout of the AI company Elon Musk founded, which had already absorbed the social network X. In June it went public on the Nasdaq, raising about $85.7 billion in net cash after fees. High-vote shares left Musk with about 82% of the voting power. In August, SpaceX closed the buyout of the company behind the AI coding platform Cursor at a $60 billion equity valuation.

Its filings now report three operating segments. Space covers rockets. Connectivity covers Starlink. The third is called Super Intelligence, or SI, formerly the AI segment, and it holds the data centers, the company's frontier AI model, Cursor and X.

The quarter that flipped the story

The third-quarter report flips the rocket story upside down. In the quarter ending in September, SpaceX generated about $14.6 billion in revenue and roughly $2.8 billion in operating profit. Rockets brought in about $1.8 billion and ran near break even. Starlink brought in about $5.2 billion with roughly $2.4 billion in operating profit, the steady cash engine.

SI brought in nearly $7.6 billion, more than rockets and Starlink combined. Roughly $7 billion came from AI solutions and infrastructure, the rest from advertising. A year earlier that division lost about $1.7 billion on operations. This quarter it made $431 million in operating profit.

In that one quarter, roughly $16.4 billion of capital went into SI, about 84% of every capital dollar SpaceX invested. Over the first nine months, nearly $40 billion went into the AI buildout. So why is a rocket company spending like that?

Every data center is a token factory

Start with the basic unit of AI work. When a model reads a prompt or writes code, it breaks text into small pieces called tokens. A token is roughly three quarters of an English word. Tokens go in, tokens come out. So an AI data center is really a token factory, turning electricity and chips into intelligence.

In the first chatbot wave, people ordered those tokens. Somebody typed a question, read a short answer and closed the tab. Plenty of people assumed smarter, more efficient models would need less compute.

That missed who the main user became. It's an AI agent now, software that takes a big goal, breaks it into dozens of steps, uses tools, checks its own work and loops until the job is done. Asking a carpenter a quick question in a doorway takes seconds. Hiring that carpenter to build a house takes thousands of measurements, cuts and inspections.

Cheap tokens, expensive chips

The numbers say this is already underway. Menlo Ventures estimates enterprise spending on deployed AI hit $176.8 billion by September 2026, up 5.4 times in a year. OpenRouter, a Menlo portfolio company that routes developers to hundreds of AI models, says its weekly volume went from about half a trillion tokens in mid-January 2025 to 128.4 trillion in mid-September 2026. Menlo found agents generate 80% of all tokens there. After agents overtook human traffic in February, agent token use multiplied 37 times while human use grew five times.

Token prices and chip rents moved in opposite directions. Andreessen Horowitz says average token prices have fallen 67% since May. From mid-July to late October, its figures show rental prices up 106% for Nvidia's B300, 66% for the B200 and 17% for the H100. The firm cautions that token price data come from small, biased samples, and its rough math shows total token spending still rose more than 120% since May.

The plug is the problem

Buying chips is the easy step, because a chip in a warehouse box makes zero tokens. Nvidia's chief financial officer said on an August earnings call that production shipments of next-generation Vera Rubin chips had started, and that Nvidia expects supply to remain a bottleneck at least through the end of its fiscal 2028, which ends in early 2028. A working factory also needs a concrete hall, liquid cooling, high-voltage substations and steady power. A gigawatt is a billion watts, roughly the output of a full-size nuclear reactor.

CBRE, the commercial real estate firm, says vacancy across the eight primary North American data center markets held at 1.4% in the first half of 2026. In Northern Virginia, the largest of those markets at about 4.5 gigawatts of inventory, it fell to 0.2%. A record 7.5 gigawatts is under construction, and more than 80% of it is already pre-leased. Less than 1.5 gigawatts of future capacity is left to lease, about six months of demand.

Research firm GlobalData points to the clock mismatch. A data center shell takes one to three years. New grid infrastructure takes up to 15. Goldman Sachs banker Jason Tofsky put it this way in an October analysis: "Power is probably the forever bottleneck." Goldman says environmental reviews can take three to four years, grid interconnection now takes a median of about five years, and substation and transformer lead times run roughly three to five years.

Then the neighbors show up. Goldman counted 86 projects cancelled or extensively delayed since January 2026, 76 of them directly linked to local opposition. In May, Goldman researchers forecast that about 60% of US data center capacity scheduled for the next year would come online on time, dropping to roughly 50% over two years. And at a global summit session in September, Musk pointed to analyst expectations of at least a 15 gigawatt power shortfall for AI chips next year, because chip output is rising up to 50% a year while power outside China grows much more slowly.

The token factory squeeze

Put demand and supply together and you get what I call the token factory squeeze. I think every business has to adopt agents or get crushed by competitors that do. If construction stalls while agent use compounds, there won't be enough working factories, and the price of intelligence gets pushed sharply up.

The winner in that shortage does two things at once. It builds compute fast, and it owns a frontier model. That gives it two ways to get paid. Path one rents scarce compute to rivals at a premium. Path two runs its own models on its own compute at raw cost, with no landlord markup. SpaceX wrote this into its June prospectus as a "dual monetization strategy," saying it can lease capacity to third parties while keeping the right to reallocate it to its own AI applications later.

Rivals paying the rent

Path one is already live. SpaceX's filings describe its cloud contracts as stand-ready promises of reserved capacity for fixed monthly fees, so it sells the reserved GPU month instead of billing by the token. A GPU is the kind of chip, in these deals made by Nvidia, that runs AI models.

In May, Anthropic, the maker of Claude, signed agreements for capacity across SpaceX's Colossus and Colossus 2 data centers. Per the prospectus, SpaceX provides roughly 325,000 Nvidia GPUs along with CPUs, storage and networking, and Anthropic pays $1.25 billion a month through May 2029, about $3,846 per GPU per month if you divide by GPUs alone. In June, Google signed for roughly 110,000 Nvidia GPUs plus CPUs and memory at $920 million a month starting in October, about $8,364 per GPU on the same math. And in the third quarter, a single unnamed customer tied to the SI segment accounted for 34.5% of SpaceX's total revenue, about $5 billion.

Path two runs in the same halls. On its developer platform, SpaceX's Grok 4.7 model lists at $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens. A rival lab short on buildings pays a landlord markup on every chip. SpaceX pays for electricity, site operations and chip depreciation.

The 90-day lever

The sharpest term in both the Anthropic and Google deals is the cancellation clause. Either side can exit the Anthropic agreements on 90 days' notice after an initial three months, and the Google deal on 90 days' notice after December 31, 2026. So rivals help pay for SpaceX's power gear and chips while it builds. If token prices surge, SpaceX can give notice, reclaim the GPUs and run its own models at cost.

On SpaceX's October earnings call, Musk said one of its AI products was in early deployment with trusted customers, expected wide enterprise deployment within weeks, and that he thinks it may end up using most of SpaceX's superintelligence compute. He said that implies SpaceX annual recurring revenue between $500 billion and $1 trillion by the end of 2027. At the November 6 Baron Investment Conference, he called ARR above $500 billion in December 2027 his best guess, not a promise, and put $1 trillion at maybe 2%. Treat those as Musk's projections. I see a toll booth where rivals fund the factory and can be evicted in 90 days.

Google and Meta, the only other players

In my view, only two other major companies pair a frontier model with massive company-owned compute. Alphabet is far bigger today. In the third quarter it brought in $127.6 billion in revenue and $40.6 billion in operating profit. It owns Gemini, a cloud division with $30.7 billion in quarterly revenue, and its own AI chips called TPUs, which it also sells to outside businesses. It plans roughly $200 billion of capital spending this year.

Yet on Alphabet's October earnings call, CFO Anat Ashkenazi said that because of a supply-constrained environment, Google started using third-party compute at the end of the third quarter and would ramp it in the fourth, creating near-term margin headwinds. Put that next to SpaceX's filings and my take is that even the richest AI giant on Earth couldn't energize its own buildings fast enough, so it became a tenant in SpaceX's halls.

Meta builds its own AI models and expanded its Hyperion campus in Louisiana into a 5 gigawatt project projected to cost more than $50 billion, alongside a multi-gigawatt chip deal with AMD. But Hyperion targets 2 gigawatts by the end of the decade, and a Meta spokesperson told CNBC the full 5 gigawatts should be complete around 2032.

SpaceX tracks nameplate compute, the GPUs installed multiplied by their all-in power rating. It went from 0.6 gigawatts at the end of September a year earlier to 1.7 gigawatts twelve months later, nearly triple.

Built like a rocket program

How does a rocket builder do that? SpaceX grew up on hard physical problems, reusable Falcon and Starship rockets and more than 11,100 Starlink satellites in low Earth orbit. That meant mastering high-pressure fluids, heat rejection, heavy power electronics and fast factory construction. A giant GPU cluster is the same kind of problem: push massive power into a hall and pull hundreds of megawatts of heat back out through liquid cooling.

So for Colossus in Memphis, Tennessee, and Colossus 2 across Memphis and a site in Mississippi, SpaceX skipped the long utility queue and brought in on-site mobile natural gas turbines. To buffer power swings, it bought $480 million of Megapack batteries from Tesla in the first nine months of the year. Tesla is still a separate public company. Both have formed special board committees to evaluate a potential combination, and Musk said in October that no terms had been agreed.

A November 3 SemiAnalysis report says SpaceX's first Colossus retrofit took 122 days and Colossus 2 took 91, figures that match SpaceX's prospectus. The firm said it has identified more than 9 gigawatts of SpaceX capacity and wrote: "We now believe that SpaceX may be the fastest datacenter builder in the world, even more so than Chinese peers."

On the October call, Musk said he'd be surprised if SpaceX brought on less than 13 gigawatts of compute by the end of next year, and called over 20 gigawatts next year a small but nonzero possibility. SemiAnalysis says SpaceX has raised its 2027 estimate to about 13 gigawatts, which the firm says would roughly match all of Google's data center capacity built from 1999 to 2026, in about 1.5 years.

Brakes on the ground

Speed has costs. In April, the NAACP and its Mississippi State Conference sued two SpaceX AI entities in federal court in northern Mississippi, alleging that the mobile gas turbines powering Colossus 2, which run with state regulators' permission, are stationary sources lacking required Clean Air Act permits. The companies and the State of Mississippi oppose a preliminary injunction, and the suit was pending as of SpaceX's October filing. Separately, in August, SpaceX AI companies sued a contractor and related parties for more than $500 million over Memphis-area construction work. The defendants dispute the claims and have filed liens and counterclaims seeking about $569 million.

Token factories in orbit

That's why SpaceX is looking up. Speaking at the Baron Investment Conference on November 6, Musk said SpaceX had launched one satellite carrying an Nvidia GPU and another carrying a Google TPU, and both are working in space. Alphabet CEO Sundar Pichai said in October that Google launched its TPUs into space with Project Suncatcher. Musk also said, while noting he was guessing, that the first orbital computers co-designed with Nvidia could launch around the third quarter of next year. He argued there isn't enough power on the ground to reach 200 gigawatts a year, let alone a terawatt.

In January, SpaceX applied to the FCC for an orbital data center network of up to 1 million satellites. The FCC accepted it for filing and public comment on February 4, which is not a final license. SpaceX's June prospectus targets orbital AI compute satellites as early as 2028 and warns the timeline depends on Starship, which flew three times in the first nine months of the year.

Why a slowdown favors SpaceX

So here's the chain. Better models mean more agents, and agents multiply token demand. Grid queues, transformer waits and local opposition stall new data centers, so compute gets scarce and pricier. SpaceX holds both halves of the machine in that squeeze: the engineering to bring gigawatts online faster than anyone, and its own models to run on them. I think that's how SpaceX turns an AI construction slowdown into the most profitable company in the world.