SpaceX (SPCX): Google's $30 Billion Compute Contract Is Cancellable on 90 Days' Notice
The number that carried SpaceX into the largest initial public offering in history was not a launch cadence figure. It was roughly $75 billion of contracted artificial intelligence revenue, assembled in the six weeks before the June 12 listing and presented to investors as the evidence that the company had become something more durable than a rocket manufacturer. Anthropic agreed in late May to pay $1.25 billion a month through May 2029 for exclusive access to the Colossus 1 facility outside Memphis. Google followed on June 5 with $920 million a month from October 2026 through June 2029 for approximately 110,000 Nvidia GPUs, plus the CPUs, memory and supporting infrastructure around them. Two customers, three-year terms, a combined backlog larger than the annual revenue of most S&P 500 constituents. The roadshow wrote itself.
The contract language tells a narrower story. If SpaceX fails to deliver the committed GPU capacity to Google by September 30, 2026, Google receives a one-month grace period and may then terminate outright or accept the delivered capacity at a proportionally reduced fee. More consequentially, after December 31, 2026, either party may exit the agreement on ninety days’ notice. The headline figure is $30 billion over thirty-two months. The contractually enforceable floor is a fraction of that: payments begin in October, the mutual exit right opens ten weeks later, and everything past the first quarter of 2027 is held together by the customer’s continued preference rather than by obligation. A backlog with a customer-side termination right at will is not a backlog. It is an option the customer owns and the seller has already booked into its valuation.
That distinction matters more than it normally would because of who the customer is. Google is not a compute-constrained startup with no alternative. Alphabet raised its 2026 capital expenditure guidance to $195 billion to $205 billion this week, up from $180 billion to $190 billion in April. The SpaceX payment represents under six percent of that budget. Every dollar of it is bridge capacity, procured while Google’s own data center construction catches up, and the termination clause is precisely what a buyer negotiates when it expects the need to be temporary. The instructive detail is not that Google is renting. It is that a company spending two hundred billion dollars a year on infrastructure still needed to rent at all, which locates the binding constraint in the industry: not silicon, not capital, but energized shells and interconnection. xAI built Memphis fast, on gas turbines, and speed-to-power is the entire reason that asset commanded a premium tenant. It is also the reason the advantage is contestable. Turbines can be ordered by anyone. Permits can be won by anyone with the same appetite for local opposition.
This is where the competitive analysis breaks along segment lines, and where the market appears to be reading the company incorrectly. In launch, SpaceX owns a genuine and compounding moat: reusable boosters, a cadence no competitor approaches, and a cost structure that has taken more than a decade to build and would take a decade to replicate. In satellite broadband, Starlink owns spectrum, orbital slots and a terminal manufacturing base that constitute a real barrier. In compute, the company owns none of that. It buys the same Nvidia hardware available to every hyperscaler, houses it in facilities whose principal virtue is that they were finished early, and leases it to tenants who can leave. The inputs are commodities, the differentiation is a construction timing advantage that decays as the industry’s shell pipeline fills, and the revenue is cancellable while the depreciation schedule underneath it is not. A leasing business with multi-year asset lives against sub-annual enforceable revenue duration carries a duration mismatch that shows up violently when demand softens. Investors paying a launch-company multiple for a segment with landlord economics are making a category error, and the segment they are mispricing is the one that justified the listing.
The circularity deserves naming as well. Alphabet disclosed this week that its investment portfolio holds $94.1 billion of SpaceX stock as of June 30, roughly $80 billion of it under short-term sale restrictions and $14.1 billion locked through the third quarter of 2027, an effective stake somewhere near five percent after dilution. The position traces to January 2015, when Google put approximately $900 million into a $1 billion round alongside Fidelity for a combined holding just under ten percent. Google is now simultaneously the company’s second-largest disclosed AI tenant and one of its largest outside shareholders. Its $11 billion of annual payments feed the contracted revenue base that supported a $1.75 trillion offering valuation, and that valuation is the mark against which Alphabet carries the stake on its own balance sheet. The loop is tighter and better documented than the vendor-financing arrangements elsewhere in the AI complex that have drawn far more scrutiny, and unlike most of them it sits entirely in public filings.
The mark is also already stale, which is the second thing the disclosure inadvertently reveals. Back the share count out of the $94.1 billion value and Alphabet holds somewhere in the range of six hundred to six hundred forty million shares. SPCX closed at $115.26 on July 23 after touching an all-time low of $110.85 the same session. The position is worth roughly seventy billion dollars today. Alphabet booked a $99.0 billion gain on equity securities in the second quarter that lifted net income 298 percent; it is on track to book something on the order of a twenty to twenty-five billion dollar unrealized reversal in the third, on an asset it is contractually barred from selling. Investors treated the second-quarter windfall accordingly, sending the shares down about seven percent on a quarter that delivered $119.8 billion of revenue and 24 percent growth, with free cash flow swinging to negative $5.9 billion and the company having raised $49.6 billion of fresh equity in June while sitting on a locked position it cannot monetize.
For SPCX itself the price action has already anticipated some of this without naming the mechanism. The stock priced at $135, closed its first session at $160.95, peaked at $225.64 on June 16, and has since given back roughly half, trading below the offering price and near ninety times trailing revenue for a business losing billions annually. Only four to five percent of the roughly thirteen billion shares outstanding currently trade. The unlock schedule is staged rather than cliffed: approximately twenty percent of locked shares release after the second-quarter earnings report, seven percent tranches follow around August 21 and September 10, an additional ten percent frees if the stock reclaims $175.50, and the balance of the 180-day pool clears December 8. Insiders could become eligible to sell as much as forty-four percent of the company by early September, expanding the float roughly ninefold. Musk’s 6.4 billion shares remain restricted until June 12, 2027.
Two dates now govern the equity, and both fall inside that window. September 30 is the GPU delivery condition, and a shortfall converts Google from a committed customer into one holding a termination right or a fee reduction. December 31 is the mutual exit date, after which a third of the contracted AI backlog becomes discretionary for the buyer. Either date landing badly resets the revenue narrative that supported the valuation at the precise moment the tradeable supply of stock multiplies. That combination, rather than the lockup calendar alone, is the risk the market has not finished pricing.