Five days before its IPO, SpaceX disclosed two back-to-back compute leasing deals worth more than $25 billion annually. That sequencing did not happen by accident.

The assumption in most coverage of the SpaceX-Alphabet compute deal is that Google is the story. Google is capacity-constrained. Google needed Nvidia GPUs it could not build fast enough. Google paid $920 million per month to get them. That read is accurate, but it is looking at the transaction from the wrong end.

SpaceX, through its xAI subsidiary, built the Colossus 1 data center in Memphis, Tennessee to train Grok, its flagship AI assistant. Grok's adoption did not keep pace with that infrastructure investment. What SpaceX was left with was a massive, partially idle GPU cluster drawing enormous power, generating costs against a model that was not scaling revenue fast enough to justify the build. The occupancy problem needed solving before an IPO roadshow could credibly proceed.

The solution was to become a landlord.

Two Deals, One Thesis, Five Days Before Public Markets Open

In late May, Anthropic agreed to pay SpaceX $1.25 billion per month through May 2029 for access to the full compute output of Colossus 1, more than 300 megawatts of capacity and over 220,000 Nvidia processors. SpaceX's own IPO filing described the arrangement plainly: "This allows us to monetize unused compute capacity in our infrastructure." Then, on June 5, five days before the planned June 12 Nasdaq debut, a second deal surfaced. Google agreed to pay $920 million per month, from October 2026 through June 2029, for approximately 110,000 Nvidia GPUs, central processors, memory, and related hardware.

The combined annual committed revenue from these two agreements exceeds $25 billion. SpaceX's total 2025 revenue, from Starlink, launch services, and all other operations, was under $20 billion. The compute leasing contracts, if they run to their full terms, would represent more than $70 billion in aggregate contract value.

That arithmetic reframes the IPO story entirely. SpaceX was not going public primarily as a rocket company or a satellite internet provider. It was going public with a third revenue category, one characterized by multi-year committed contracts with creditworthy counterparties, which is exactly the revenue quality that public market investors price at a premium.

The Termination Clauses Are the Actual Risk Architecture

Both deals include 90-day termination rights for either party. That language is framed in most reporting as flexibility, a reasonable exit ramp given the dollar amounts involved. The operational reality cuts the other way. A 90-day notice period on a contract supplying a third of Google's Gemini Enterprise compute capacity, or all of Anthropic's inference infrastructure, is not a comfortable cushion. It is a concentrated dependency with a narrow window.

Google's deal adds a harder constraint. If SpaceX cannot deliver the committed 110,000 GPUs by September 30, 2026, plus a one-month grace period, Google can terminate immediately or accept a pro-rata reduction in capacity with a corresponding fee reduction. That GPU delivery milestone is now a material public company obligation for SpaceX. Missing it, in the quarter after an IPO, would be the kind of operational miss that moves stock prices.

Enterprise technology buyers looking at either Google Cloud or Anthropic's Claude platform are now one upstream contract failure away from a capacity event. That is a new category of vendor risk that did not exist in the same form six months ago.

What Google's Decision Reveals About Its Own Infrastructure Position

Google has the resources and engineering talent to build its own data centers. It does build them, aggressively, and has for two decades. The decision to pay $920 million per month to rent capacity from SpaceX, rather than wait for its own buildout, tells you something specific about the demand curve for its Gemini Enterprise agent platform. Demand arrived faster than even Google's construction timeline could accommodate.

This is the same dynamic driving the entire hyperscaler buildout cycle, but SpaceX's deal surfaces it in a way that most cloud capacity announcements do not. When a company with Google's capital commits $30 billion to a third-party landlord on a multi-year fixed contract, it is admitting that speed of deployment matters more than long-run unit economics, at least for this capacity window. That is an important signal for any enterprise technology buyer trying to understand whether the AI platforms they are evaluating have stable infrastructure beneath them, or whether those platforms are themselves racing to secure capacity on borrowed time.

SpaceX's Structural Play Is Smarter Than It Looks

The standard critique of this arrangement is that SpaceX is selling capacity that xAI originally built for itself, which implies a strategy that failed. That reading underweights how the resulting position looks from a capital allocation standpoint. xAI built a massive compute cluster. Grok did not scale to fill it. Rather than carry the cost or write down the investment, SpaceX converted the asset into a contracted revenue stream with blue-chip counterparties, then disclosed those contracts in an IPO filing.

The IPO filing itself stated that SpaceX "expects to enter into additional similar services contracts." That sentence is the business model declaration. The Anthropic and Google deals are the proof-of-concept, not the ceiling. If SpaceX can establish itself as a compute landlord to the AI industry's top-tier buyers, the infrastructure it continues to build for orbital data centers, Starlink, and future xAI operations carries a different economic rationale. It becomes capacity that can be rented before it is needed internally, then reclaimed as internal demand grows.

That is a capital efficiency flywheel. Whether it survives contact with public market quarterly reporting cycles is a different question entirely.