Sources: xAI ends talks with Oracle to expand an existing arrangement to rent Nvidia chips; source: Oracle is not part of xAI's supercomputer project in Memphis
Later coverage of Dell's prospective server supply for xAI's Memphis system and xAI's planned GPU expansion suggests the company pursued a more directly assembled infrastructure path rather than broadening the Oracle rental relationship.
First-order effects
xAI loses a potential expansion path for rented Nvidia compute from Oracle, while Oracle is excluded from the Memphis supercomputer project described here.
xAI must source the additional chips, servers, and operating capacity for Memphis through other vendors or arrangements; Oracle retains only the existing arrangement, if any.
Second-order effects
The decision concentrates execution responsibility on xAI and its chosen hardware and infrastructure suppliers, making server delivery and deployment more consequential than an incremental cloud-rental expansion.
Oracle must compete for AI-infrastructure workloads without xAI's Memphis project, despite its demonstrated ability to provide Nvidia-backed capacity to major cloud customers.
Third-order effects
If leading AI developers continue to favor dedicated, self-directed clusters for flagship systems, cloud rentals may remain useful for flexibility but become less central to the largest training deployments.
The split raises compute execution risk: the value of chip access increasingly depends on whether builders can turn allocated GPUs into powered, networked, operational systems on schedule.
The trend: Frontier AI builders are increasingly balancing rented cloud capacity against dedicated infrastructure they can control more directly.
@xDaily xAI contracted for 24k H100s from Oracle and Grok 2 trained on those. Grok 2 is going through finetuning and bug fixes. Probably ready to release next month. xAI is building the 100k H100 system itself for fastest time to completion. Aiming to begin training later this mo…
NEWS: Elon Musk's xAI and Oracle have ended talks on expanding their server rental agreement. • Talks broke down due to disagreements over timeline and power supply concerns. • xAI is now building its own AI data center in Memphis, Tennessee. • The potential deal was [image]
scoop: Larry Ellison may be a friend and backer of Elon Musk, but even he isn't immune from Musk's impatience. xAI and Oracle ended talks to expand their cloud deal for Nvidia chips. It could have been a $10B deal for Oracle over several years. https://www.theinformation.com/ ...
This is a super interesting move. Oracle owns all the old Sun Microsystems networking IP which was started when Sun was originally writing Linux for Cray, this is why Oracle cloud is super competitive for distributed training. I wonder what Elon is really cooking on here 👀
Here's a rundown of how the deal talks between Elon Musk and Oracle fell apart: 1) On May 14, we reported that xAI was talking to Oracle about spending $10B over several years to rent servers w/ Nvidia GPUs https://www.theinformation.com/ ...
Elon's advantage is in control of the physical stack which has now become the bottleneck. The move to Texas was the best possible thing he could have done. So many GPUs and no place to build a data center.
100k H100s is around $3bn, $4b including all connects, racks etc The chips alone draw 100 MW electricity, all in perhaps 50-100% more (couple hundred dollars million a year) Given another likely B100 or X100 (Dojo?) build after best to build own datacenter..