PitchBook senior research analyst Harrison Rolfes told Axios that Elon Musk has a recurring habit of overbuilding compute infrastructure: he builds at maximum scale, but his own products can't use it all, leaving competitors to pick up the excess capacity.
The first instance was in 2024. xAI was negotiating a roughly $10 billion server lease with Oracle to train Grok 3. Musk thought Oracle was building the cluster too slowly, walked away from the deal, and built his own data center in Memphis instead. Oracle's freed-up GPU capacity was snapped up by OpenAI. As Rolfes put it: "xAI's Colossus 1 ended up with capacity that Grok's user base could never fill."
Musk didn't just rely on purchasing to stockpile chips for xAI. In 2024, CNBC obtained internal NVIDIA emails showing he asked NVIDIA to prioritize shipping 12,000 H100s originally destined for Tesla to X and xAI, delaying more than $500 million worth of chips for Tesla by months and slowing down autonomous driving and robot training. Those chips ended up underutilized as well.
Two years later, the same pattern repeated. All 220,000+ NVIDIA GPUs in Colossus 1 were rented out entirely to Anthropic. SpaceXAI claims training has moved to the larger Colossus 2, freeing up Colossus 1. But The Information previously reported that xAI's model compute utilization across hundreds of thousands of GPUs was only 11% — Grok's user base simply can't support that scale of compute. It wasn't a proactive upgrade; the capacity was never being fully used.