SpaceX's rocket-speed data-center build skipped backup power — now a $920M/month Google contract runs out September 30

For months this year, at least two of SpaceX's newest AI data centers — one in Tennessee, one in Mississippi — operated with no backup power or cooling. When a primary system dropped, the machines went down and a Grok training run could die mid-step. Tennessee relied on more than 100 mobile chillers and still ran below the company's own 99.9% uptime target; Mississippi leaned on temporary gas turbines because the local grid never delivered power on time.

That is the "setback" buried in the recent headlines about SpaceX throttling its data center strategy. And the instinct to read it as engineering noise is exactly wrong. This is the load-bearing fact cracking the foundation of a roughly $2 trillion market value, because the company's most valuable business is not launching rockets — it is renting AI compute to Google, Anthropic, and others. That landlord business was built so fast it skipped the redundancy that a power grid demands, and now SpaceX is slowing expansion to fix it, on a deadline. A $920 million-a-month contract with Google comes due September 30.

The speed that broke reliability

SpaceX built its first AI data center in 122 days, a pace its own executives called unheard of in the industry. The same philosophy that made rockets cheap — iterate fast, test in production — was applied to grounded compute. Data centers are a different physics. A launch failure kills one vehicle and you fix the next one cheaper; a power failure inside a data center kills a training run that cost real money, and no amount of iteration makes a missing backup generator arrive sooner.

The market has rewarded the speed. In Q2, SpaceX spent $15.8 billion on AI-related capital expenditure — double Q1 — and $23.6 billion in the first half of the year, as compute capacity scaled from 0.4 gigawatts to 1.4. Q2 AI revenue rose more than 210% to roughly $2.5 billion. But the segment still lost $3.7 billion in the first half. The build, in other words, is converting enormous capital into leased capacity at a loss today, on the bet that selling compute to hyperscalers becomes a giant, recurring landlord business.

That bet now carries a near-term tripwire. SpaceX agreed to supply Google with about 110,000 Nvidia GPUs for $920 million a month. If the chips are not delivered by September 30, Google has the right, after a one-month grace period, to walk away, accept fewer GPUs, or cut the payment in proportion to the shortfall. This is not a vague strategic risk; it is an enforceable cap on the exact revenue stream the valuation is premised on, arriving in a matter of weeks.

The throttle in practice

SpaceX's response is to put the brakes on itself. It is prioritizing power and cooling redundancy over raw build speed, delaying new projects outside its Memphis hub, including proposed sites in Texas. The leadership got reshuffled for it: Jake Palmer, Zach Wells, and Pablo Mendoza left, replaced by executives from the rocket and Starbase operations — Wesley Salandro and Logan McConnell — while Starlink's Michael Nicolls was put in charge of AI data center operations. The company is even building a foundry in Bastrop, Texas to make gas-turbine blades in-house, in hopes of bringing new turbines online up to 18 months faster.

There is a grim logic to these moves, and it is worth separating what they fix from what they cannot. Vertical integration and manufacturing discipline can fix a supply-chain bottleneck — sure, build your own turbines. What they cannot fix quickly is the thing this business actually runs on: a terrestrial power grid that did not show up. Mississippi still ran on 69 permitted temporary turbines, double the number SpaceX had disclosed, while 41 permanent units were delayed and air-permit constraints locked in the temporary arrangement.

The grid it tried to escape

This is where the story decomposes into its real shape. The orbital data center narrative — the million-satellite plan, the FCC filing, the idea of bouncing compute off the grid entirely — was sold to investors as an escape from exactly this problem: the years-long utility queue that is throttling every AI builder. That narrative got the attention. But the valuation is not being tested in orbit; it is being tested in grounded buildings in Tennessee and Mississippi, and those buildings are hostage to the very terrestrial grid the space story was supposed to bypass.

Put the two next to each other and the gap is the point. SpaceX's S-1 priced a future where AI infrastructure — lease to Google, lease to Anthropic, at $1.25 billion a month through May 2029 — is the growth engine that carries a company that did less than $19 billion of revenue and lost money in 2025. The same S-1 the market used to justify a $1.75 trillion target after the largest IPO in history, at $135 a share, with shares trading near $147 now after dipping below the IPO price on a share unlock. That valuation is a bet that renting grounded compute at hyperscale works. The throttling is evidence that the buildout's genuine bottleneck — reliability under grid conditions — was underestimated, and the September 30 Google delivery will show whether it was underestimated badly enough to matter.

The clean lesson for an investor new to this stock is narrower than the headlines. Set aside the space data centers; they are a decade-plus story. Watch the one observable the near-term case actually turns on: whether SpaceX delivers the 110,000 GPUs to Google by month's end, and whether the move from speed to redundancy lets the 2-gigawatt year-end target slip. The company throttled its own growth to fix what speed broke. The open question is whether the customers underwriting that growth wait for the fix.