Analysis: Musk is deploying too much computing power, and competitors are rushing to seize the opportunity at a bargain.

According to Beating Monitoring, senior research analyst Harrison Rolfes from PitchBook said in an Axios interview that Musk has repeatedly overexerted himself on computing infrastructure: building at the maximum scale but the product side can’t keep up, and the excess capacity is ultimately taken by competitors.

The earliest instance was in 2024. xAI negotiated a roughly $10 billion server leasing deal with Oracle to train Grok 3. Musk found Oracle’s cluster building too slow, the deal fell through, and he turned to self-build the Memphis data center. The GPU capacity freed up by Oracle was signed by OpenAI. Rolfes’s exact words: “xAI’s Colossus 1 ultimately ended up with capacity that Grok users could never fill.”

Musk’s chip stockpiling for xAI isn’t just through procurement. CNBC obtained internal Nvidia emails in 2024, showing he instructed Nvidia to prioritize reallocating the 12k H100s originally allocated to Tesla to X and xAI. Tesla’s over $500 million worth of chips were delayed by months, slowing down autonomous driving and robot training. The chips taken were ultimately not fully utilized.

Two years later, the same thing happened again. All 220k+ Nvidia GPUs of Colossus 1 were leased out to Anthropic. SpaceXAI claims that training has moved to the larger Colossus 2, so Colossus 1 was freed up. But The Information previously reported that xAI’s model compute power utilization with tens of thousands of GPUs was only 11%, and Grok’s user base simply couldn’t support this scale of computing power. The capacity wasn’t actively freed up after an upgrade; it was already underutilized.