Musk just admitted his ChatGPT competitor is a ghost town — by renting out an entire supercomputer to his rival.
The Summary
- Anthropic signed a deal to use all computing capacity at SpaceX's Colossus 1 data center — 300+ megawatts, 220,000+ NVIDIA GPUs — to serve Claude Pro and Claude Max subscribers
- xAI moved training workloads to Colossus 2, freeing up the entire first facility for a competitor
- When everyone else is fighting for compute, you're only selling yours if demand isn't there
The Signal
The compute market just told us something Musk won't say out loud: Grok usage is embarrassingly low. While OpenAI throttles free tier access and Anthropic caps Claude usage for paid subscribers, xAI can lease an entire data center to Anthropic without breaking a sweat.
This isn't strategic asset optimization. It's what happens when you build a supercomputer for a product nobody uses. Grok launched with the X platform advantage — 500+ million monthly users, instant distribution, conversational UI baked into the feed. It should be printing queries. Instead, it's so quiet that SpaceX can hand 220,000 GPUs to Claude.
"When everyone else in AI is scrambling for chips, selling compute means you built capacity for demand that never showed up."
The official story: xAI moved training runs to Colossus 2, so why not monetize idle capacity? Training is batch work. Inference is real-time user demand. If Grok inference was anywhere near ChatGPT or Claude levels, Colossus 1 would still be running hot. The fact that xAI can cleanly separate training and inference infrastructure, then lease out the inference side, tells you inference demand is negligible.
Musk's "evil detector" endorsement of Anthropic reads like damage control. He spent last week with Anthropic leadership to "understand what they do to ensure Claude is good for humanity." Translation: he needed a narrative that frames this as strategic partnership, not distress sale of unused capacity. The safety angle is cover. The real story is utilization rates.
What went wrong with Grok:
- X users don't treat the platform as a knowledge destination
- Grok's personality (edgy, "maximum truth-seeking") doesn't match what most people want from AI assistance
- No ecosystem of integrations, no API traction, no enterprise sales motion
Anthropic, meanwhile, is capacity-constrained. Claude Pro and Max subscribers hit usage limits regularly. This deal lets them serve existing demand without waiting on new data center builds. For Anthropic, it's a clean win. For xAI, it's revenue from assets that weren't generating it anyway.
The Implication
Watch what happens when Colossus 2 finishes training xAI's next model. If usage doesn't spike, this becomes the template: build the infrastructure, lease it to someone who can actually fill it, rebrand as "strategic partnerships."
The agent economy runs on inference at scale. If your model isn't generating sustained query volume, you don't have product-market fit. You have expensive hardware and a story problem. Anthropic just became xAI's biggest customer, which tells you everything about where the actual user demand lives.