The Numbers Are the Story
$920 million a month, over $10 billion a year, is a magnitude that ranks in the entire history of AI infrastructure. 110,000 Nvidia chips is roughly the scale of a leading lab's one-generation flagship cluster. SpaceX choosing to sign this kind of long-term deal right before its IPO amounts to showing its hand to the market: its AI ambitions are serious. Starlink's data, rockets' engineering simulation, plus the bloodline with xAI—the room for imagination is indeed large enough.
Rivals Still Rent Each Other's Data Centers
The web of relationships is telling in its own right: Google has its own Gemini, Musk holds xAI and Grok, and the two sides are rivals in models yet closed the business deal. This kind of "competition is competition, but the data center still gets rented" arrangement has grown increasingly common in AI circles over the past two years—there are precedents between Microsoft and OpenAI, and between Google and Anthropic.
The Other Side
A dash of cold water is due: a company before its listing has an incentive to tell the fullest story, a monthly-payment structure means the deal may not run its full course, and the outside world can't see the specific terms. But one thing is real—another giant-scale buyer entering to sweep up supply, Nvidia's capacity continuing to be carved up, and the days of second-tier companies queuing for cards will only get longer. There are only so many chips, and when big customers lock in supply early, small and mid-size companies have a harder time getting stock, with the cost ultimately spread into API prices and subscription fees. Those managing an AI budget would do well to remember: current unit prices may not hold for long, and quotas and rate limits will shift right along with upstream supply and demand. This order will be a reference point for gauging where compute prices are headed.
via: The Decoder, Engadget