SpaceX Discloses a $1.11 Billion Monthly AI Deal, Won't Name the Buyer
Thursday, September 109 Min Read
SpaceX's finance chief told a Goldman Sachs conference on September 10 that it had closed a new compute hosting deal: $1.11 billion a month, customer unnamed. The same week, the company put executives from the Falcon and Dragon rocket programs in charge of data centers that had gone months without backup power and cooling.
$1.11B
Monthly Fee on the New Deal
$3.4B
Expected Monthly Hosting Revenue in December
110,000
Chips Due to Google, September 30
99.9%
The Company's Own Internal Uptime Target
In This Piece5 Sections
On Thursday a finance chief announced a new customer.
That same week, two executives from rocket manufacturing were put in charge of the data centers that customer will run on.
Both stories belong to SpaceX.
The Deal: $1.11 Billion a Month
CFO Bret Johnsen spoke on September 10 at the Goldman Sachs Communacopia and Technology Conference.
He said the company had signed a new customer renting AI compute capacity. The monthly fee is $1.11 billion, with revenue recognition starting December 1. Annualized, that is $13.3 billion.
The customer was not named.
In the same session Johnsen explained why the company is cautious about locking capacity into long contracts:
"You don't want to get in a situation where you constrain your own products in the years to come."
Most of what SpaceX collects in monthly hosting revenue this December rests on three contracts.
The Anthropic contract was signed in late May and covers the entire capacity of the Colossus 1 site near Memphis. The Google deal was announced on June 5. Who holds the third one is not known.
The Same Week, Rocket Executives Moved Into the Data Centers
The Information reported on September 1 that the company had rebuilt its AI infrastructure team.
Wesley Salandro, vice president of production for the Falcon and Dragon rockets, and Logan McConnell, who ran product support at Starbase, moved onto the team. An engineer was brought over from Neuralink for civil works.
The reason was not speed itself but what speed left behind. According to the report, some sites in Memphis and Southaven ran for months "without backup cooling and power."
The Tennessee facility ended up leaning on more than 100 mobile chillers. Uptime fell below the company's own 99.9% target, and model training runs were interrupted repeatedly.
There were departures as well. Jake Palmer, who led infrastructure, left in late July, followed by Zach Wells and Pablo Mendoza. Brent Mayo went to OpenAI and Liz Balke to Anthropic.
Some of the people who built the data centers left for the companies renting them.
SpaceX AI Capacity
What SpaceX sells is not chips.
It is uninterrupted hours.
An Outage Costs More Than the Outage
An ordinary cloud customer renting a server is paying for access. Something drops, the user reconnects, and what is lost is those few minutes.
Training does not work that way. A large training job is spread across thousands of chips that all advance through the same step together. If one rack loses power or cooling, the job halts as a whole and restarts from the last saved checkpoint.
The time lost is not the length of the outage. It is the outage plus everything computed since the last checkpoint.
In a 30-day month, 99.9% permits 43 minutes of downtime. Backup power and backup cooling exist to protect exactly those 43 minutes.
Procuring chips is a purchasing problem. Building a redundant cooling loop is a construction problem, and it does not compress at the same rate.
Which Document Actually Binds
The week's two stories do not carry equal weight.
The $1.11 billion monthly deal is a statement. It was disclosed verbally at a conference, the counterparty is unnamed, and revenue recognition starts on December 1.
The Google contract is written and public, sitting in SpaceX's offering documents. It runs from October 2026 to June 2029 and covers access to roughly 110,000 Nvidia chips.
Its term reads like this: if access to the committed chips is not delivered by September 30, 2026, Google may end the agreement. After a one-month grace period there is also the option of accepting fewer chips at a reduced fee. After December 31, either side can exit on 90 days' notice.
The calendar is set not by the deal announced this week but by a clause signed in June.
The Spending Side, and What the Market Did
Capacity went from 0.4 gigawatts to 1.4 gigawatts in a year, with a year-end target of 2 gigawatts.
The price of that was $15.8 billion of capital spending on AI infrastructure in the second quarter. First-half losses on the AI side came to $3.7 billion.
Revenue beat expectations in the company's first public quarter; the stock fell 13.6% the next day. The line the market reacted to was not revenue but spending. More: What Is Cash Flow?
Johnsen spoke on Thursday. On Friday SpaceX shares closed at $150.28, up 1.42%.
The Dow, the S&P 500 and the Nasdaq each rose about 1% the same day. An announcement of $13.3 billion in annualized revenue did not price the stock apart from the index.
The chart offers a frame; it does not show which of the week's two stories the price answered.
Timeline
- Late MayAnthropic signs a contract worth $1.25 billion a month.
- June 5The Google deal is announced: $920 million a month, 110,000 chips, a September 30 condition.
- Late JulyJake Palmer, who led infrastructure, leaves the company.
- September 1The Information reports the reorganization and the sites running without backup.
- September 10Johnsen discloses the new $1.11 billion monthly deal.
- September 30The chip delivery date in the Google contract falls due.
Nvidia chips are the common input across this chain, and the 110,000 units due to Google come off the same production.
This piece draws on accounts of Bret Johnsen's September 10 presentation at Goldman Sachs Communacopia, on the terms of the Google contract as disclosed in SpaceX's offering documents and relayed by TechCrunch and CNBC, and on The Information's September 1 report on the data centers. The details on backup power, cooling and uptime come from a single outlet citing unnamed sources and are not company statements. The outage-and-checkpoint calculation is a worked example built to show the mechanism, not measured SpaceX data.
Correction (September 13, 2026): prose tightened for rhythm; figures and sources unchanged.
Companies in This Piece
Since The Event