On May 6, Anthropic announced it will consume the entire compute capacity of SpaceX's Colossus 1 data center — over 220,000 NVIDIA GPUs. The cluster Elon Musk built so xAI could fight OpenAI and Anthropic is now leased to Anthropic, because xAI no longer has a competitive model to run on it.
That's the headline. The real story is what happened to the people.
By late March, all 11 original xAI co-founders had walked. So had 80+ researchers and engineers, including chief engineer Igor Babuschkin. Two years ago Musk had the capital, the GPUs, the brand, and a roster you couldn't have assembled on Stripe's best day. Today he has the GPUs.
The talent didn't disappear. They went to Anthropic, OpenAI, smaller labs, their own startups. The same companies that benefit when xAI has no model to ship are the ones that hired its bench.
Here's the part nobody in the AI infrastructure conversation wants to admit: a 220,000-GPU cluster without the people who can train a frontier model on it is a very expensive landlord business. Capital can buy chips in 18 months. It cannot buy a team that ships GPT-5 or Claude Opus in 18 months. There is no shortcut. There is no acquisition that closes the gap fast enough.
This is the part of the AI story that staffing leaders and CTOs need to internalize, because it's not just an xAI story. Every enterprise that's staring at a $50M, $100M, $500M AI budget right now is making the same mistake xAI made if they assume the budget is the hard part.
The hard part is the team that survives the build.
I've watched companies sign nine-figure AI commitments with vendors before they've hired the third engineer who can actually integrate any of it. They have the budget. They have the platform contract. They have the slide deck. What they don't have is the durable engineering org that can absorb new tooling every six months without losing its center of gravity. When the senior people start walking — and they walk first, because they have the most options — the org rebuilds nothing. The contracts keep paying.
That's xAI right now. Multi-billion-dollar infrastructure. Senior bench gone. Renting the cluster to a competitor.
Three things to pull from this if you're sitting on a serious AI build:
Treat your senior AI/ML engineers like the actual scarce resource. GPUs are a commodity now. People who can ship a frontier-quality model on them are not. If you have any, your retention plan is more important than your procurement plan.
Hire as if you'll lose 30% of the team in 18 months. Because you might. The labor market for this skill set is the most liquid I've seen in 20 years of staffing. Build redundancy. Build the bench. Document everything. Pretend you're going to be raided, because someone is going to try.
Don't confuse a contract with capability. An OpenAI seat license, an Anthropic enterprise contract, a ServiceNow AI Control Tower — none of these become outcomes without the team that operates them. The vendor sells you the platform. The platform doesn't run itself.
Musk built the most expensive AI cluster in the world. The bottleneck wasn't the chips. It was never going to be the chips. It was the 80 people who walked out the door, and the fact that nobody at xAI seemed to register that as the actual emergency until it was too late to fix.
If you're building anything serious in AI right now, the GPUs are not your moat. The team that can use them is. Plan accordingly.