According to the news app Zhitong Finance, Musk's AI startup xAI and Oracle (ORCL.US) have concluded a potential server trade worth $10 billion. Oracle's stock once fell 4.8%, and as of Tuesday's close, Oracle fell 3% to $140.68.

Musk posted on his social network X on Tuesday that his company xAI has decided to establish an internal system for training AI models because "our basic competitive advantage depends on being faster than any other AI company."

According to several negotiators involved in the talks, xAI and Oracle have ended discussions on expanding their existing agreement due to disagreements over timetable and electrical utilities. Under the existing agreement, xAI has been leasing Nvidia's AI chips from Oracle. xAI has signed a partnership to train AI models on Oracle's Gen2 Cloud platform. Now, xAI is purchasing chips to build a datacenter in Memphis, Tennessee.

xAI was originally negotiating a multi-year agreement to lease Nvidia processors from Oracle for a planned supercomputer, but talks were stalled due to some issues, such as Musk's requirement to build a faster supercomputer that made Oracle feel difficult to achieve. Furthermore, Oracle was concerned about xAI's preferred location not having enough electrical utilities. The specific processing power discussed by Oracle and xAI has been outsourced to another customer. xAI has leased about 0.016 million Nvidia chips from Oracle. Musk plans to build a supercomputer using 0.1 million Nvidia GPUs to train the Grok 3.0 big model. However, Oracle has signed an agreement with Microsoft, to provide servers powered by Nvidia chips to OpenAI, involving the upcoming GB200 chips.

Musk responded on X that xAI is building a system using Nvidia's H100 graphics processor to "finish in the shortest possible time." "xAI has ordered 0.024 million H100 processors from Oracle for training Grok 2. Grok 2 is in the process of fine-tuning and fixing bugs and is expected to be released next month. Meanwhile, xAI is building a system using 0.1 million H100 processors on its own to be able to train the Grok 3.0 model as quickly as possible, with the goal of starting training later this month. This will be the world's most powerful training cluster, with significantly superior performance."

Analyst Anurag Rana wrote that Musk's decision to establish internal AI training infrastructure highlights the challenges facing cloud providers, despite their funding, in expanding. "We believe these issues are not limited to Oracle and may also bedevil Microsoft and AWS, not only due to a shortage of dedicated chips, but also due to a shortage of electrical utilities."

Anurag Rana, an analyst, wrote that Musk's decision to establish internal AI training infrastructure highlights the challenges facing cloud providers, despite their funding, in expanding. "We believe these issues are not limited to Oracle and may also bedevil Microsoft and AWS, not only due to a shortage of dedicated chips, but also due to a shortage of electrical utilities."