Invest1 publisher3 min readPublished
SpaceX shifts AI data-center strategy toward power and cooling redundancy as it races to meet $920m-a-month Google deadline
Cryptopolitan, citing The Information, says sites in Tennessee and Mississippi ran for months without backup power, and Google can trim payments proportionally if about 110,000 GPUs miss a September 30 handover.
The Investor · Invest desk
What happened
- SpaceX is reordering how it builds AI data centers to put power and cooling redundancy ahead of build speed, a change The Information reports may slow its expansion.
- The Google contract gives the buyer access to about 110,000 Nvidia GPUs for $920 million a month, with the chips due to be delivered by September 30.
- Several sites in Tennessee and Mississippi ran for months with no backup power or cooling, and when the primary systems dropped the machines went down, disrupting Grok training.
- The AI division lost $3.7 billion in the first half while compute capacity went from 0.4 GW to 1.4 GW and second-quarter AI revenue rose more than 210 per cent to top $2.5 billion.
- The NAACP sued in April and residents filed a separate class action over air quality and turbine noise, with the Justice Department arguing that cutting power to Colossus 2 would risk national security.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- exposure Google's remedy is graduated, not binary: after a one-month grace period it can walk, take fewer GPUs, or cut payments in proportion, so lateness shows up as a smaller invoice each month while the retrofit continues.
- constraint Vacancies left by departing data center executives are being filled with people pulled off the rocket and Starlink programs, so those programs are down staff while a Starlink executive runs AI infrastructure.
- decision Building a foundry in Bastrop to make turbine blades and vanes in-house commits capital to metallurgy at a company whose stated bottleneck is electricity; Musk said it could pull new gas-turbine deployment forward by as much as 18 months.
- contradiction The publisher flags its own account as unverified and says power and thermal design are industry-wide problems, which leaves the reader with a deal term to price and no confirmed operating record behind it.
Divide the monthly payment by the hardware and the deal prices out at about $8,364 per GPU per month, or roughly $11.46 an hour across a 730-hour month [1]. The account does not say what else the $920m buys beyond access to about 110,000 Nvidia GPUs [3].
The per-megawatt view is the one the failures speak to. Cleanview puts AI data center revenue at $10m to $12m per megawatt per year [6], and the Google contract annualises to $11.04bn [2]. At $12m per MW that implies 920 MW of fully ramped capacity; at $10m, 1,104 MW [3]. Against a fleet that reached 1.4 GW, one customer is the revenue-equivalent of roughly two-thirds to four-fifths of everything built [4], and the contract annualised is larger than the AI business's entire second-quarter revenue annualised, about $10bn [10].
That last comparison is also a test of the Cleanview number. Annualise the quarter and spread it over 1,400 MW and the fleet is earning about $7.1m per MW per year, roughly 71 per cent of the low end of the range [5]. Ramp accounts for part of the gap. Machines going down when the primary power and cooling dropped accounts for some of the rest [5].
Deferring the redundancy work was never free. Second-quarter AI capital expenditure ran about $15.8bn, double the first quarter, taking the half to $23.6bn (the two quarters add to $23.7bn, so somebody rounded) [7][6]. Spread across 181 days that is about $130m a day, which makes one full $920m payment worth roughly seven days of construction spending [7].
The reliability problem started at the meter. The Mississippi site feeding Colossus 2 could not get grid power on time, so the company ran portable gas turbines, legal under state law for a year without air permits [9]. Regulators counted 69 of them at Southaven by July, twice the number disclosed, at a site of about 1,775 MW [10]. Forty-one permanent units were late, and the temporary ones stayed [12].
My read, and September 30 will settle it either way: the binding constraint is turbine and permit lead time, not chips, which puts the real pressure on the 10 GW target for the end of 2027 [15]. That is 8.6 GW to add, against 1 GW added in the first half, or about 4.3 years at the same pace [8]. Two other readings work. Google's remedy is graduated, so a 5 per cent shortfall on 110,000 GPUs trims about $46m a month [9], and a customer that wants the capacity takes the trim and waits. Or the retrofit is small relative to $15.8bn a quarter, gets absorbed inside the grace month, and the only casualty is the 2027 slope.
One caution on all of it. Cryptopolitan says the reporting is yet to be verified, and adds that power and thermal design are industry-wide problems [16]. The same account attaches a NASDAQ ticker to the builder and calls Grok SpaceXAI's chatbot [17]. Anyone pricing counterparty risk on a $920m monthly invoice has to know which entity signed it, and this account does not say [17].
What to watch
- Whether the roughly 110,000 GPUs are handed over by September 30, and whether Google uses the grace month or trims the invoice.
- The air-permit status of the 69 temporary turbines at Southaven and the arrival of the 41 delayed permanent units.
- Third-quarter AI capex against the $15.8bn run rate, which would show whether redundancy work is additive or substitutive.