SpaceX’s “quiet” pivot is now loud enough to change how investors will read the valuation story. Fortune reports that SpaceX agreed in May to give Anthropic access to roughly 325,000 Nvidia GPUs across its Colossus data centers for $1.25 billion per month. Weeks later, it struck a similar agreement with Google, providing about 110,000 GPUs for $920 million per month. Together, those two contracts could generate about $26 billion a year, a figure described as more than SpaceX’s entire revenue last year.

That number matters because it challenges the simplest narrative of SpaceX as only a rocket and satellite company. Fortune frames these compute-rental deals as a way SpaceX can justify its eye-popping $1.8 trillion valuation after its record-breaking IPO last month, especially as Musk’s long-term ambitions like orbital data centers and a Mars colony are likely years away. In other words, SpaceX is monetizing infrastructure it has already built, while still working toward longer-dated bets.

Why does compute rental fit so well right now? In AI, demand for training and operating large models continues to outpace supply. Building a large data center typically takes years of development and substantial investment, which creates a timing gap. Fortune notes that SpaceX’s offering is attractive because customers who need GPUs immediately can pay a premium to access that capacity instead of waiting to build. Sean Cray, a senior analyst covering telecom, media and technology at Moody’s, tells Fortune this works because “demand for AI compute continues to outpace supply as the AI race heats up and companies build increasingly powerful models,” and because companies can’t easily build fast enough.

For a quick sense of where this plugs into SpaceX’s current financial picture, Fortune lays out the mix: SpaceX generated $18.7 billion in revenue last year, including $11.4 billion from connectivity services and about $4.1 billion from launches and other space-related work. Its AI segment contributed another $3.2 billion, but it also recorded an operating loss of roughly $6.4 billion, even as Musk has touted AI as the company’s next money-maker after it acquired xAI in February. Part of the pressure is that Grok, the company’s large language model, requires massive investment in GPUs, electricity, and other infrastructure to train and operate.

Yet the compute-rental business is not tied only to whether Grok wins the race on model quality or revenue today. Fortune includes a key point from Cray: even if Grok doesn’t catch up to models created by Anthropic and OpenAI in terms of revenue and capabilities, the compute renting business shows investors “there’s just different pathways for them to generate revenue in their AI segment,” and “it doesn’t strictly have to come from Grok and their AI enterprise applications.” The second-order implication for decision-makers is straightforward: the market can be offered evidence of profitability even while an AI product’s commercial trajectory is still unfolding.

There is also an operational flexibility story baked into the design. Fortune says SpaceX can rent capacity to outside customers, but it can also use the same hardware to train newer versions of Grok, or redirect it to Starlink’s own internal uses. That dual-use capability is powerful, because it means SpaceX is not necessarily locking itself into one customer or one workload forever. But it also creates a classic contracting tradeoff: flexibility can protect a company from internal surges, yet the same clause can make revenue less durable if customers find cheaper alternatives.

The potential pitfall is explicit in the contracts: Fortune reports that built into both the Google and Anthropic arrangements is a 90-day cancellation provision. That provision gives SpaceX the ability to reclaim its compute if Grok’s needs suddenly increase. At the same time, it could allow customers to walk away if cheaper compute capacity becomes available, raising the risk that revenue from this business could be temporary. Sridhar Tayur, a professor of operations management at Carnegie Mellon University, flags the board-level question Fortune frames as: is this “infrastructure as a service going to become a permanent part,” or is it a one-off that won’t last.

Executives should also notice the competitive signaling. Fortune reports that Meta is reportedly in talks to lease its computing power to Anthropic in a deal that could be worth up to $10 billion over two years, Meta is also in talks with Anthropic to lease its computing power, sources told the New York Times. And SpaceX’s customer development could broaden beyond private AI labs: the Wall Street Journal reported Friday that SpaceX was in talks to provide the Pentagon with data-center capacity potentially worth billions of dollars for running AI models. The deal is not final, but Fortune frames it as strengthening the case that selling compute could become a long-term business rather than a temporary stopgap.

Finally, set aside valuation for a moment and focus on industry mechanics. Fortune notes that SpaceX’s Colossus buildout accelerated after xAI brought its first major Colossus cluster online in just 122 days by converting an existing factory. Since then, SpaceX expanded its Memphis-area complex to roughly 2 million square feet across Colossus and Colossus II, together providing about 1 gigawatt of compute power, with a plan to eventually include 1 million GPUs. When you connect the build speed, the GPU shortage dynamics, and the billion-dollar compute contracts, the strategic stakes get real for any operator in AI infrastructure or capital markets: the winners may not only be the firms that invent models, but the firms that can deliver compute access faster than others can construct it. For boards and finance leaders watching SpaceX, the question becomes whether GPU scarcity is persistent enough that these compute rentals turn from “valuation support” into an enduring line of business.