On June 5, SpaceX filed with the SEC that Google would pay $920 million per month from October 2026 through June 2029 for access to approximately 110,000 Nvidia GPUs, along with CPUs, memory, and other components. This is SpaceX’s second major compute deal in a single month—Anthropic had already agreed in early May to pay $1.25 billion a month for all of Colossus 1’s 220,000 GPUs, with a contract also running through 2029.
Together, the two deals would bring SpaceX roughly $26 billion in annual compute leasing revenue. For a company about to IPO at a $1.75 trillion valuation, these numbers directly underpin the AI infrastructure narrative. The mainstream interpretation goes in two directions: either SpaceX is transforming from a rocket company into an AI compute provider, or the bottleneck in frontier AI competition has shifted from “who has the best model” to “who can deploy the most hardware.”
Both readings have merit, but they miss a more fundamental question: why can SpaceX provide something that Google cannot? Google operates one of the world’s largest cloud infrastructures, with 2026 capex plans of $175–185 billion. Why would it need to rent GPUs from a rocket company?
The answer lies not in engineering capability, but in institutional constraints.
Let’s be clear: Google’s compute needs are not fabricated.
Sundar Pichai said something unusually candid on the Q1 earnings call: “Obviously, we are compute-constraint in the near term. As an example, our cloud revenue would have been higher if we were able to meet the demand.” Translation: we have customers, contracts, and cash, but not enough servers to run the workloads.
How large is this gap? Google Cloud’s revenue backlog reached $462 billion in Q1, nearly double the prior quarter. The backlog is growing faster than Google can deploy new data centers. Meanwhile, Nvidia GPU lead times stretch to 36–52 weeks, 30% to 50% of large data center projects face delays, and U.S. grid interconnection queues now exceed five years—only 13% of projects in the queue ever reach commercial operation.
For Google, paying a premium to rent immediately available GPUs is cheaper than waiting. In an AI market that reshuffles every quarter, an 18-month deployment delay is far more expensive than paying a few extra dollars per GPU-hour.
Anthropic faces the same bind. Claude’s user base grew 80x, but the company planned for 10x. CEO Dario Amodei publicly acknowledged “difficulties with compute”—paying subscribers hit rate limits during peak hours, which is unacceptable at the $200/month Max tier.
So the demand side checks out. The real question is the supply side: how did SpaceX pull this off?
Colossus 1 sits on a former Electrolux factory in Memphis, Tennessee. It went from groundbreaking to operational in 122 days. By comparison, traditional data centers typically take 2–4 years from planning to commissioning.
The speed advantage isn’t better engineering—Colossus uses the same Supermicro liquid-cooled racks and Nvidia Spectrum-X networking as any other H100 cluster. What SpaceX skipped were the two slowest steps in traditional data center construction: grid interconnection and environmental permitting.
Colossus doesn’t connect to the grid. It runs on 35 methane gas turbines with a combined capacity of about 421 megawatts. xAI claimed these were “temporary-mobile” units requiring no federal permit—because they’re mounted on flatbed trailers, they fall into the same regulatory category as a traveling circus’s generators. Under the rules at the time, mobile turbines could bypass air quality permits as long as they didn’t stay in one location for more than 365 consecutive days.
But the Southern Environmental Law Center (SELC) confirmed through thermal imaging that 33 of the 35 turbines were continuously operating for over a year. University of Tennessee researchers used NASA satellite data to find that peak NO2 concentrations spiked 79% near Colossus. Shelby County, where Colossus is located, already has the highest asthma emergency rates in Tennessee.
The NAACP has filed an intent-to-sue letter alleging Clean Air Act violations. But the current EPA has shuttered its environmental justice office and cut $2 billion in funding, making near-term federal enforcement unlikely.
Google cannot do this. Even as Goldman Sachs projects that U.S. data center power demand will grow 165% by 2030, requiring roughly $720 billion in grid investment, Google must follow standard interconnection procedures, comply with environmental reviews, and navigate local opposition—188 community groups are active across 40 states, and moratorium bills have been proposed in more than 10 states.
In other words, SpaceX’s competitive advantage isn’t technical. It’s that SpaceX is willing and able to do things in the current regulatory vacuum that Google cannot. This is not engineering superiority; it’s institutional arbitrage.
If Colossus were genuinely built to meet Grok’s training needs, its GPU utilization would be near industry norms. The reality is the opposite.
The Information reported in early May that Colossus’s GPU fleet utilization was around 11%. For comparison, Meta’s training clusters run at about 43% and Google’s at about 46%. Out of 550,000 installed GPUs, only about 60,000 are doing actual compute work. The rest sit idle, still consuming power and cooling.
The implication is straightforward: Colossus wasn’t built to meet existing demand. It was built on the assumption that demand would materialize—and the assumption didn’t hold. Grok holds only about 3.4% of global generative AI web traffic share (Similarweb, January 2026), and its growth lost momentum in early 2026 as Claude and Gemini gained ground.
xAI isn’t “monetizing spare capacity.” It’s repackaging an asset facing potentially tens of billions in write-downs as “cloud revenue.” By routing the customer contracts through SpaceX—a separate entity about to go public—xAI’s balance sheet avoids the direct hit of GPU depreciation, while SpaceX gets exactly what it needs pre-IPO: third-party revenue to show public market investors.
The termination clauses make the story clearer.
Both deals include 90-day mutual termination provisions. After December 31, 2026, either party can cancel with 90 days’ notice. Musk himself spelled it out on X on May 28: this is “a 180 day lease with 90 day notice mutual cancellation thereafter”. He added, “The short term was our request, not Anthropic’s”—because “if compute gets super tight I said we might need it back at some point.”
This structure means several things.
First, neither contract is locked-in revenue. The S-1 language saying “through May 2029” is legally accurate but economically misleading. If Google’s own replacement capacity comes online, or GPU rental prices drop significantly, $920 million a month could vanish in 90 days.
Second, Musk’s “take it back” logic reveals Colossus’s split strategic identity—it’s trying to be both a neutral third-party compute platform and a resource it can reclaim on short notice. This kind of “we might need the servers back” clause is unheard of in cloud computing. AWS doesn’t tell Netflix “if we need the servers, you’ll have to move.”
Third, the 90-day clause effectively downgrades both deals from “long-term partnerships” to “each side buying a time option”: Google is buying insurance (“a backup if I can’t build fast enough”); SpaceX is selling a floor (“someone covers operating costs while my AI business finds its footing”).
To understand the political dimension of this deal, follow the timeline.
In February 2026, Musk posted a series of attacks on Anthropic on X. He wrote that Anthropic was “doomed to become the opposite of its name” (misanthropic), called it “the most hypocritical company,” and declared that “Anthropic hates Western Civilization.”
That same month, Trump signed an executive order banning federal agencies from using Anthropic’s models. Defense Secretary Hegseth designated Anthropic a “supply chain risk to national security”—a label previously reserved for foreign adversaries. The Pentagon had been negotiating a contract worth up to $200 million with Anthropic, but talks collapsed when Anthropic insisted on two restrictions: no fully autonomous weapons and no domestic mass surveillance.
On the same day the ban was announced, OpenAI announced its own Pentagon contract—with the exact same restriction clauses Anthropic had demanded, which the Pentagon agreed to in writing. Meanwhile, xAI had already secured Pentagon frontier AI contracts worth up to $200 million in July 2025, joining OpenAI and Google on the DoD’s AI vendor list.
Three months later, SpaceX announced that Anthropic would pay $15 billion a year to lease all of Colossus’s compute. Musk’s tone on X had completely reversed: “I spent a lot of time with senior members of the Anthropic team… Everyone I met was highly competent and cared a great deal about doing the right thing. No one set off my evil detector.”
A competitor’s CEO going from “you hate Western civilization” to “you’re good people” in three months—with a $45 billion revenue contract in between—tells you everything about the nature of this deal.
Musk is simultaneously the person cutting off Anthropic’s government revenue (through his influence in the Trump administration) and the person collecting Anthropic’s commercial revenue (through SpaceX’s Colossus lease). This isn’t “compute neutrality.” It’s a vertically integrated political-commercial loop.
Morningstar analyst Nicolas Owens issued a rare bearish report on June 2, valuing SpaceX at a fair value of $780 billion—less than half the $1.75 trillion IPO target. The core logic is simple: the AI business “poses a material threat of value destruction,” Grok “has not demonstrated significant performance advantages over leading peers,” and orbital data centers “may not be commercially viable.”
Owens assigned probabilities to three scenarios for SpaceX’s AI business: the best case (7% probability) at $1.3 trillion, the base case (50%) at $88 billion, and the worst case (43%)—the business model proves unviable.
The $1 trillion valuation gap between Goldman’s buy-side clients and Morningstar’s sell-side analysis is, at its core, a bet on how long Musk can operate an AI infrastructure empire outside the constraints of environmental regulation and competition law. Goldman’s IPO investors are betting the regulatory vacuum lasts. Morningstar is betting it doesn’t.
Colossus’s 35 gas turbines are the most fragile brick in this empire’s load-bearing wall. If the EPA resumes enforcement, a federal court issues an adverse ruling, or Tennessee tightens its air quality standards—any one of these would fundamentally change Colossus’s cost structure. At that point, the 90-day termination clause stops being a boilerplate contract term and becomes Google and Anthropic’s emergency exit.
This is why the deal resembles submarine cable overbuilding in the late 1990s more than AWS in 2006. Back then, telecom companies laid fiber far beyond actual demand based on forecasts of “internet traffic doubling every 100 days.” Then DWDM technology doubled per-fiber capacity every 18 months, and over 90% of fiber went dark. Global Crossing and 360networks went bankrupt.
Google’s $462 billion cloud backlog plays the same role today as those internet traffic forecasts did then: it’s real, but “real demand growth” and “demand growth that must be met at current price levels” are two different numbers. Algorithmic efficiency improvements—MoE architectures, quantization, more efficient inference—continue to increase output per GPU-hour. That means satisfying the same level of demand will require fewer GPUs in the future.
SpaceX chose to sign short-term, cancellable leases at the peak of the demand curve. It is essentially selling a time window that could close within 18 to 24 months: before Google’s own capacity comes online, before alternative compute suppliers enter the market, before environmental regulation tightens, before the next efficiency leap shifts the GPU demand curve downward. Any one of these four clocks reaching its endpoint could reset that $920 million monthly payment to zero.
The real question isn’t how much SpaceX will earn from these contracts. It’s what Colossus’s 550,000 GPUs will be worth when the contracts end, and whether Grok will have enough demand to fill them by then.