For two decades SpaceX sold rides to orbit. In 2026 its fastest growing business is renting graphics chips. According to numbers compiled by analyst Jamin Ball and surfaced by the Latent Space AI News digest, SpaceX AI compute is already running at roughly $2.32 billion a month, an annualized rate near $28 billion. That figure would place Elon Musk's rocket company among the largest sellers of artificial intelligence computing power on Earth, ahead of every pure-play rental provider that went public in the past two years.
- SpaceX is renting GPU capacity at an annualized rate of about $28 billion, based on roughly $2.32 billion in monthly contracts.
- A new deal has Reflection AI paying $150 million a month from July 2026 through 2029 for Nvidia GB300 chips at the Colossus 2 site near Memphis.
- The run rate is roughly double CoreWeave's current annual revenue, reshaping who the biggest neocloud actually is.
- The capacity comes from xAI's Colossus supercomputer, folded into SpaceX after the two companies merged in early 2026.
The shift turns a launch provider into a landlord for the AI boom. The customers paying those bills are not small. Anthropic and Google, and now the open-source lab Reflection AI, have all signed multi-year contracts to run their training and inference workloads on hardware that physically sits inside data centers SpaceX controls in Tennessee. The story is less about rockets than about who owns the scarce resource every frontier lab is fighting over.
How SpaceX AI compute reached a $28 billion run rate
The $28 billion figure is a sum of disclosed contracts rather than an audited revenue line. Ball's breakdown, relayed through Latent Space, stacks three publicly reported agreements. Anthropic pays about $1.25 billion a month for capacity spanning the Colossus 1 and Colossus 2 sites, a pool that reportedly covers around 325,000 chips between the two facilities. Google pays roughly $920 million a month for xAI compute capacity. Reflection AI adds $150 million a month starting in July. Add those together and the monthly take lands near $2.32 billion, which annualizes to the $28 billion headline.
Ball also pointed to the implied pricing. At the contracted volumes, SpaceX is charging more than $10 per hour for Blackwell-class GPUs, a rate that reflects how tight supply remains for Nvidia's newest accelerators. When a single customer is willing to commit over a billion dollars a month on a multi-year basis, the per-hour math stops looking like a cloud price list and starts looking like a long-term lease on a power plant.
None of these contracts are short. The Anthropic and Google agreements both run through the middle of 2029, and the Reflection deal runs to the same horizon. That gives SpaceX a contracted backlog measured in tens of billions of dollars, the kind of visibility most infrastructure businesses spend years trying to build.
The Reflection AI deal, line by line
The newest piece is the most concrete. As TechCrunch reported, Reflection AI agreed to pay $150 million a month beginning July 1, 2026 and running through 2029, a commitment worth up to $6.3 billion across the full term. In exchange the lab gets immediate access to Nvidia's latest GB300 AI chips and the supporting hardware around them, hosted in the Colossus 2 data center near Memphis, Tennessee.
The contract includes an escape hatch. Either party can walk away with 90 days' notice once the first three months are complete, a clause that also appears in the other Colossus agreements. That flexibility matters in a market where chip generations turn over fast and a lab's compute needs can swing with a single model release.
Reflection AI is worth understanding here. The company was founded in 2024 by two former Google DeepMind researchers and positions itself as an open-source lab, releasing model weights rather than gating everything behind an API. A company spokesperson framed the deal in those terms, saying that recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs tied to depending only on closed models. Reflection called the arrangement one of the largest announced open AI infrastructure commitments to date.
There is an irony in an open lab leaning on the most capital-intensive infrastructure in the field. Open weights lower the barrier to using a model, but training a competitive one still requires the same enormous clusters that closed labs rent. Reflection's bet is that owning the model while renting the metal keeps it independent in the way that matters, even if the metal belongs to SpaceX.
From rocket company to neocloud
SpaceX did not build this compute business from a blank sheet. The capacity comes from xAI, the AI startup Musk launched in 2023. The two companies merged in early 2026, a transaction reported to value the combined entity at roughly $1.25 trillion, and xAI now operates as a subsidiary of SpaceX. The merger folded xAI's Colossus supercomputer project, and the contracts running on it, into the rocket company's balance sheet.
That corporate history is why the same hardware shows up under different names in different reports. The Google capacity is described as xAI compute, while the Anthropic and Reflection deals are framed as SpaceX agreements. They draw on the same physical clusters. The semantics changed; the silicon did not.
Anthropic's relationship with the facility is unusually deep. As Data Center Dynamics reported, Anthropic arranged to use the full output of the Colossus 1 site, an arrangement that over the life of the deal could total around $45 billion. CNBC noted that the Anthropic agreement even reaches into SpaceX's longer-term space ambitions, tying compute to the company's plans beyond Earth.
How the Colossus data centers were built
The physical anchor of all this sits in Memphis. xAI announced in June 2024 that it would build Colossus there, and the first cluster went live the following month. The speed of that buildout became part of the project's mythology, with the team standing up a supercomputer in a window that traditional operators would measure in years rather than weeks.
The real estate came together piece by piece. As Pulse 2.0 reported, an affiliate of SpaceX acquired a 785,000-square-foot property in Memphis for $185 million, a former Electrolux appliance factory sitting on 217 acres along Paul R. Lowry Road. That site became the nexus of a cluster of high-density data centers across greater Memphis that together form the Colossus supercomputer.
The scale of the hardware is what makes the contracts possible. Colossus 1 exceeds 300 megawatts of power draw and houses more than 220,000 Nvidia GPUs. Colossus 2 pushes further. Independent analysts at SemiAnalysis described Colossus 2 as the first gigawatt-scale data center in the world, a power footprint that rivals a mid-sized city. Capacity at that level is what lets SpaceX sign three separate billion-dollar-class tenants without running out of room.
Musk has signaled that Memphis is a step rather than a destination. He has talked about using the Colossus supercomputers to design orbital data centers, satellites launched by SpaceX and powered by xAI technology, with the first such facilities targeted for early 2028. Whether that timeline holds is an open question, but it explains why the compute business and the launch business are being run under one roof.
Where SpaceX sits against CoreWeave and the neocloud field
The term making the rounds for this category is neocloud, a label for companies that rent raw GPU capacity rather than the full managed stack of a hyperscaler. The poster child has been CoreWeave, which went public and built a brand around being the dedicated AI cloud. The SpaceX numbers reframe that pecking order.
CoreWeave's current revenue runs around $14.5 billion a year, against a post-IPO valuation that reached roughly $60 billion. SpaceX's annualized compute run rate of about $28 billion is therefore close to double CoreWeave's current revenue, and SpaceX got there without ever pitching itself as a cloud company. A business that exists to launch rockets quietly assembled one of the largest GPU rental operations in the industry as a side effect of building its own AI lab.
The $28 billion figure is an annualized run rate derived from disclosed monthly contracts, not a reported annual result. It assumes the current contract mix holds for a full year, which the 90-day exit clauses do not guarantee.
The structural advantage SpaceX holds is vertical integration most rivals cannot match. It controls the real estate. It has the capital to buy Nvidia allocations at scale. And through xAI it had an in-house tenant that justified the buildout before any external customer signed. A standalone neocloud has to win contracts to finance its clusters. SpaceX built the clusters for itself first, then sold the spare capacity at a premium.
The open-source bet Reflection is making
Reflection's framing deserves a second look because it captures a real fault line in the industry. The argument is that depending entirely on closed models from a handful of labs concentrates risk, both for nations worried about sovereignty and for enterprises worried about lock-in. Open weights are the hedge. A company that can download and self-host a capable model is not at the mercy of another firm's pricing, policy changes, or deprecation schedule.
The catch is that openness at the weights layer does not change the economics at the compute layer. To stay competitive, an open lab still has to train on the same gigawatt-class clusters that closed labs use, and that training is where the money goes. Reflection's $6.3 billion commitment to SpaceX is the tuition for staying in the frontier conversation. The model may be free to download, but the GPUs that produce it are anything but free.
That tension is now baked into the market structure. The labs compete on models and licenses, yet they increasingly share the same physical infrastructure and, in this case, the same landlord. Anthropic runs closed models, Reflection runs open ones, and both are tenants in the same Tennessee complex.
Why the GB300 chips matter to the deal
The hardware named in the Reflection contract is not incidental. Nvidia's GB300 is the latest step in the Blackwell line, a server-class accelerator built for exactly the kind of large-model training and high-volume inference that frontier labs run. Demand for these parts has outstripped supply since they began shipping, which is the reason a buyer will sign a multi-year lease rather than wait for spot availability that may never come.
That scarcity is what gives an operator pricing power. When Ball estimates SpaceX is charging north of $10 per hour for Blackwell-class GPUs, the number reflects a market where access matters more than the sticker price. A lab that cannot get chips cannot ship models, and missing a release window costs far more than a premium hourly rate. So the contracts get signed at terms that look expensive on paper and rational in context.
It also explains the structure of the agreements. Labs are not buying chips outright, partly because the next generation will arrive before the current one is fully depreciated. Renting at scale lets a customer ride each hardware wave without owning the depreciation risk, and it lets SpaceX keep the clusters full by rotating tenants. The 90-day exit clause is the pressure valve that makes the whole arrangement work for both sides.
How the financing math favors an integrated operator
Standing up a gigawatt-scale data center is one of the most capital-heavy projects in technology, and the financing usually comes before the revenue. A standalone neocloud has to convince lenders that future contracts will cover the cost of clusters it has not yet filled. That circular dependency is part of why the category has been volatile, with valuations swinging on each new customer announcement.
SpaceX inverted that order. Through xAI it had a guaranteed first tenant, its own lab, which justified the buildout before a single external contract existed. Once the capacity was online and partly idle, selling the surplus to Anthropic, Google, plus Reflection turned a cost center into a profit center. The company financed the hardest part of the business with internal demand, then monetized the slack at premium rates.
That advantage compounds. The revenue from external tenants helps fund the next cluster, which creates more surplus to sell, which funds the cluster after that. An operator that can self-finance the first turn of that wheel is in a stronger position than one that has to raise money against contracts it does not yet hold. It is the same flywheel logic that built SpaceX's launch business, applied to silicon instead of rockets.
The costs that do not appear on the invoice
A gigawatt of compute has consequences that reach past the contract terms. The Memphis buildout has drawn local opposition over its environmental footprint. As CNBC reported, the NAACP brought a lawsuit over air pollution tied to the xAI data centers in Memphis, focused on emissions from gas turbines used to power the site. Power at this scale has to come from somewhere, and when the grid cannot supply it fast enough, operators turn to on-site generation that brings its own emissions.
Those frictions are not unique to SpaceX. Across the industry, the race to stand up ever larger clusters is colliding with local power capacity, water for cooling, and community pushback. The economics that make a $28 billion run rate possible also concentrate enormous energy demand into a few square miles, and the people who live nearby do not get a line on the revenue.
For now the demand is winning. Every signed contract pulls more capacity online, and every new model release validates the next round of buildout. The labs need the chips, the chips need the power, and the power needs somewhere to land.
The compute landlord at the center of the AI race
What the SpaceX numbers reveal is how much of the AI economy now flows through infrastructure rather than software. The headlines go to model launches and benchmark wins, but the durable position belongs to whoever owns the buildings, the power contracts, plus the Nvidia allocations. SpaceX backed into that position by building for its own lab and discovering it had surplus to sell.
The open questions are about durability. The 90-day exit clauses mean the backlog is not as locked as the headline suggests, and chip generations move fast enough that today's GB300 advantage erodes on a predictable schedule. A customer paying $150 million a month will keep paying only as long as the hardware stays competitive and the alternatives stay scarce. If Nvidia supply loosens or a rival operator undercuts the per-hour rate, the math that produced $28 billion can move the other way.
Still, the direction is clear enough. The companies that will shape the next few years of AI are not only the ones training the models. They are the ones that control where those models get trained. On current numbers, SpaceX has quietly become one of them, and the contracts with Anthropic, Google, plus Reflection AI suggest the market already knows it.