SpaceX's xAI Seeks Massive AI Compute Deals

SpaceX built its reputation by making rockets reusable. Now, it's applying the same logic to compute infrastructure — build it big, use it repeatedly, and sell the excess to whoever needs it most.

That pivot became concrete when Anthropic signed a deal to lease the entire capacity of xAI's Colossus I data center for $1.25 billion per month over three years. Read that number again. Per month. For a single facility. This isn't a co-location agreement or a modest cloud partnership — it's a signal that SpaceX is actively repositioning xAI from an internal AI lab into something closer to a neocloud provider, competing in the same arena as CoreWeave and Lambda Labs, while sitting on rocket-fueled capital that those companies can only dream about.

CEO Elon Musk confirmed as much, stating plainly: "We are in discussions with other companies to do the same." The Anthropic deal wasn't a one-off. It was a proof of concept.

From Internal Tool to Compute Marketplace

xAI was supposed to be SpaceX's answer to OpenAI — an in-house generative AI operation built around the Grok model. That story hasn't gone away, but it's getting complicated by a more commercially straightforward one.

When your competitor pays you $1.25 billion a month to run their models on your hardware, the line between "AI lab" and "AI infrastructure provider" gets very blurry very fast.

Colossus I, the 300MW facility that anchors xAI's current compute portfolio, is now effectively Anthropic's data center on lease. Cursor, the AI code editing startup that announced it would use xAI data center capacity, adds another reference customer — though that relationship is evolving given SpaceX's plan to acquire Cursor within 30 days of its IPO.

The business logic here is straightforward: xAI built infrastructure at a scale that exceeds what Grok alone requires. Rather than let that capacity sit idle or wind down capital deployment, SpaceX is monetizing it externally. That's not a failure of the original AI mission. It's actually the smarter play, especially ahead of an IPO.

The Scale of What xAI Has Actually Built

Colossus I at 300MW is already massive by any conventional measure. For context, a typical hyperscale data center campus runs between 100MW and 500MW — and those campuses often represent years of phased development. xAI stood up 300MW of GPU-dense compute in a single facility.

Colossus II is where the numbers start getting harder to parse. Musk has claimed 1GW of capacity at launch — a figure repeated prominently in SpaceX's IPO filings. The documents describe this as "nameplate compute draw," calculated by multiplying installed GPUs by their rated power consumption. What the documents also acknowledge, buried in the fine print, is that this figure "does not represent actual power consumption or utilization."

Satellite imagery from January told a more grounded story: Colossus II had cooling infrastructure in place capable of managing approximately 350MW — not 1GW. The gap between nameplate capacity and operational reality is one of the most important numbers in the entire SpaceX data center narrative, and it's being smoothed over in the IPO marketing.

According to the IPO chart, nameplate compute draw hit 1GW in March 2026, up from 300MW a year prior. That's an impressive growth curve on paper. Whether the actual operational capacity tracks that curve is a different question.

The Financial Bet Is Already Enormous

xAI spent $12.7 billion on AI infrastructure in 2025. In the first quarter of 2026 alone, it deployed another $7.7 billion. That's a $20 billion-plus infrastructure build in roughly 15 months.

The Anthropic deal — at $1.25 billion per month, or $15 billion annually — suggests a path toward recouping that capital, assuming the lease holds for its full three-year term. But the 90-day termination clause on either side introduces meaningful uncertainty. Anthropic could walk if better options emerge or if its own financial situation shifts. SpaceX could walk if it decides to redirect that compute toward internal workloads or a higher-paying customer.

At $15 billion in annual revenue from a single tenant, xAI's compute business would already rank among the largest infrastructure-as-a-service operations in existence — if it sustains.

The company is targeting an IPO valuation north of $75 billion. The Anthropic deal gives investors a concrete revenue anchor. The questions are about repeatability and durability.

The Realities That Don't Make the Pitch Deck

Two threads running through the current xAI story deserve more scrutiny than they're getting.

First, Grok's actual performance in the market. Usage has been declining, and the platform has seen significant executive attrition — including the departure of all non-Musk co-founders. That's not a minor footnote. The AI model that xAI built this entire infrastructure to support is losing ground while the company leases that infrastructure to a direct competitor. Anthropic's Claude competes head-to-head with Grok. SpaceX is now, effectively, subsidizing that competition.

Second, the workforce stability question. Data centers of this scale require serious operational talent to run reliably — power engineers, cooling specialists, network architects, GPU cluster managers. A company in the middle of a high-profile IPO push, with visible leadership churn, faces real risk in maintaining operational quality at the pace it's promising customers.

Neither of these issues is fatal. But they're the kinds of things that separate durable infrastructure businesses from ones that look impressive in a single earnings cycle.

Where the Orbital Layer Changes Everything

The most audacious element of SpaceX's compute strategy isn't Colossus I or II. It's the stated ambition to launch up to one million space-based data center satellites.

Orbital data centers aren't a near-term product. The technical challenges — thermal management in vacuum, radiation hardening of GPU clusters, latency characteristics, power generation at scale — are genuinely hard. But SpaceX has something no other company attempting this has: an operational launch infrastructure that can put hardware into orbit at dramatically lower cost than anyone else.

If SpaceX solves even a fraction of the orbital compute problem, it changes the geographic constraints of AI infrastructure entirely — data processing closer to the point of collection, reduced dependence on terrestrial power grids, and a moat that no land-based competitor can replicate.

For the energy and infrastructure sectors watching this space, the near-term implication is simpler: xAI's terrestrial data centers are power-hungry at a scale that will reshape local grid dynamics wherever they're sited. A 300MW facility running at capacity is roughly equivalent to powering 250,000 homes. A 1GW campus — if it ever reaches true operational capacity — would rival the demand of a mid-sized city.

Utilities, land developers, and clean energy project sponsors in the markets where SpaceX is building should be paying close attention. The compute demand is real. The lease revenue is real. And the company acquiring Cursor, leasing to Anthropic, and talking to additional enterprise customers is signaling that it intends to grow its external compute business, not shrink it.

The IPO will tell us how the market values all of this. The more interesting question is which AI company signs the next nine-figure compute lease — and whether SpaceX's infrastructure ambitions eventually extend far enough to make its data center business larger than its rocket business.

[INTERNAL LINK: SpaceX's AI Developments]

[INTERNAL LINK: The Future of AI Infrastructure]

[INTERNAL LINK: Understanding Cloud Computing Trends]

For more insights on the evolving landscape of AI and infrastructure, visit InfraSale Marketplace.