SpaceX Stock Falls 10% as Wall Street Questions Its Compute-as-a-Service Bet

---
headline: "SpaceX Stock Falls 10% as Wall Street Questions Its Compute-as-a-Service Bet"
slug: spcx-stock-falls-reflection-deal
category: business
story_number: "02"
date: 2026-06-23
---

SpaceX Stock Falls 10% as Wall Street Questions Its Compute-as-a-Service Bet

The deal was supposed to be good news. On Monday, SpaceX announced it had signed open-source AI startup Reflection to a compute agreement worth up to \$6.3 billion, the latest customer to rent capacity inside its Colossus data-center complex. Instead of rallying, shares of SpaceX (NASDAQ: SPCX) fell roughly 10% on June 22-23, the stock's worst single-day session since its June 11 Nasdaq debut at \$135 a share. The selloff dragged SPCX back toward \$170, down sharply from a first-week peak near \$225.

For a company that has spent two weeks being narrated to investors as the next great infrastructure monopoly, the reaction was a jolt. And it crystallized a question that has been building under the surface since the IPO: is SpaceX's pivot into selling compute-as-a-service the cash-printing machine the bulls describe, or a lower-margin, cancellable side hustle dressed up as recurring revenue?

The deal that didn't help

Under the Reflection terms, the startup last valued at \$25 billion will pay SpaceX \$150 million per month beginning July 1, 2026, for access to Nvidia GB300 chips at Colossus, the Memphis-area supercomputer originally built to train xAI's Grok models. The contract runs through 2029 if it goes the distance.

That last clause is where the optimism frays. Like SpaceX's larger deals with Anthropic (roughly \$45 billion, at \$1.25 billion a month) and Google (more than \$30 billion, at \$920 million a month), the Reflection agreement carries a 90-day exit clause: either party can walk away with three months' notice after an initial period. SpaceX has marketed more than \$80 billion in "committed" compute revenue through 2029. But strip out everything a customer can cancel on a quarter's notice, and skeptics argue the genuinely locked-in figure is closer to \$1.5 billion, with the rest a series of rolling quarterly decisions made by customers who themselves are racing to build or buy their own capacity.

"The 90-day clause cuts both ways," said one infrastructure analyst quoted in coverage of the filings. "This is opportunistic monetization of idle capacity, reallocatable the moment xAI needs it back. Whether that idle compute stays idle long enough to justify \$26 billion in annual projections is exactly the question the clause leaves open."

A margin problem hiding inside a growth story

The second worry is structural. Selling raw GPU time is a commodity business. Renting out chips at scale generates revenue, but it does so at margins that look nothing like the software-style economics investors paid up for when they bid SPCX to \$225. AI products, the models and applications sitting on top of the silicon, can carry gross margins north of 70%. Leasing the silicon itself competes with hyperscalers, neoclouds and a glut of new data-center capacity, all of which compress pricing over time.

"Compute-as-a-service is being valued as if it were a software annuity," wrote analysts at TradingKey, which flagged S&P projections of negative free cash flow at SpaceX through 2029. "It isn't. It's a capital-intensive leasing business with cancellation rights baked in, and the market is starting to price that distinction."

The Cursor question

Layered on top is the \$60 billion all-stock acquisition of Anysphere, maker of the Cursor coding tool, a deal that was supposed to give SpaceX a high-margin AI product to pair with its compute layer. So far, the synergy thesis remains a thesis. Cursor's market share has been eroding against rivals, and SpaceX has yet to show revenue lift from owning both the chips and the developer tool that runs on them. A pending \$20 billion bond sale, disclosed the same week, exists largely to refinance the bridge loan that funded the Cursor purchase, a reminder that the acquisition was paid for with borrowed time as much as borrowed money.

Why it matters

SpaceX's IPO was one of the largest in history, and its valuation rests on a bet that it can become the AWS of the AI era, owning the picks and shovels of the compute gold rush. The Reflection deal genuinely advances that story: it proves Colossus can attract third-party customers beyond xAI, and a steady drip of \$150-million-a-month contracts is real money. Moody's cited roughly \$75 billion in third-party compute agreements when it assigned SpaceX an investment-grade Baa1 rating.

But Monday's selloff suggests the market is no longer willing to take "committed revenue" at face value. The gap between \$80 billion in headline backlog and what is actually contractually irreversible is the entire argument. If customers treat Colossus as flexible overflow capacity to be dialed up and down with each quarter's GPU shortage, SpaceX is running a cyclical rental business. If they treat it as mission-critical infrastructure they cannot leave, it is the annuity the bulls imagine. The 90-day clause means nobody yet knows which.

What to watch

Three signals will tell the story. First, renewals: the earliest contracts hit their first exit windows around December 31, 2026, and whether Anthropic, Google and others extend or trim will be the first hard data on stickiness. Second, the August 6 earnings report, SpaceX's first as a public company, which should reveal actual compute gross margins rather than projected ones. Third, the December 2026 unlock of roughly 95% of SPCX shares, a supply wave that will test conviction once insiders can finally sell.

For now, SpaceX has a customer list any infrastructure provider would envy and a valuation that assumes those customers can never leave. Monday was the day Wall Street started asking what happens if they can.