More Confidence in Self-Built Clusters
In May, The Information reported that Musk planned to build a supercomputer cluster with 100,000 H100s by fall 2025, collaborating with Oracle. xAI was said to invest $10 billion to lease Oracle's servers.
Recently, Musk addressed the news about terminating the supercomputer cluster collaboration with Oracle. He stated that xAI obtained resources for 24,000 H100s from Oracle to train Grok-2, confirming ongoing server leasing cooperation between xAI and Oracle.
However, for the 100,000 H100 GPU cluster, they chose to build it themselves and pushed forward at the fastest speed, reportedly completing the installation of 100,000 cards in just 19 days.
Dell and Supermicro became Musk's new partners. The CEOs of both companies recently expressed on Twitter that they are collaborating, accompanied by data center photos.
Musk personally visited the site during the cluster construction process. He also revealed on Twitter that Grok is training in Memphis, and Grok-2 will be released in August.
Notably, Oracle previously raised concerns about power supply for the cluster location. Estimates suggest that 100,000 H100s require 150 megawatts of power allocated from the grid, but Musk seems to have resolved this issue.
Latest news indicates that the cluster currently has 8 megawatts. After signing an agreement on August 1st, they will have 50 megawatts. Now, 32,000 cards are online, with 100% expected to be online in the fourth quarter - sufficient to support training and running GPT-5 scale models.
In conclusion, AI giants believe that having computing power in their own hands is more reliable, worth burning through cash for. Estimated costs put each H100 at about $30,000-$40,000. Musk's supercomputer cluster is valued at $4 billion (over 29 billion RMB).
Earlier reports suggested Microsoft and OpenAI were planning a $100 billion data center project called "Stargate". Insiders revealed that Oracle and Microsoft are reaching a deal involving 100,000 B200 GPUs, with the cluster possibly ready by next summer.
Additionally, Meta has been rumored to have a luxurious supercomputer cluster, while cloud providers like AWS are investing even more in data centers.