Last week, Anthropic announced they’d secured compute capacity from SpaceX’s Colossus data center through xAI. On the surface, it sounds like a straightforward infrastructure deal. Dig deeper though, and you hit something messier: a collision between AI companies’ desperate need for compute and the uncomfortable realities of where that compute actually comes from.
The deal itself isn’t particularly surprising. Anthropic is severely compute-constrained, and getting access to an entire data center is genuinely significant for their ability to scale. But the venue matters. A lot.
The Colossus Problem
Colossus isn’t just another data center. It’s the one with the particularly messy environmental and regulatory history. The gas turbines powering it initially operated without Clean Air Act permits or proper pollution control equipment. xAI got away with this partly by classifying these turbines as “temporary,” which anyone paying attention can see for what it is: regulatory gymnastics.
The health impact isn’t theoretical. Credible reports link the facility to increased hospital admissions related to air quality. We’re talking actual people experiencing actual respiratory problems because of how this infrastructure was built and operated.
Andy Masley, one of the most thoughtful voices pushing back against greenwashing rhetoric around data centers, put it simply: he wouldn’t run computing out of this specific data center. That’s a pretty strong signal.
Why This Matters for Developers
Here’s what bothers me about this deal, and why I think you should care if you work with or build on top of AI systems: it exposes a fundamental supply chain vulnerability.
Anthropic is a company explicitly founded on the premise of building AI safely and responsibly. That’s not just marketing; the team actually seems to believe it. But they’re also compute-constrained in a way that forces tradeoffs. They need capacity to train Claude and serve users. The options available to them are limited. So they do a deal with the data center that has environmental and regulatory red flags.
This isn’t unique to Anthropic. Every AI company faces similar pressures. Every AI company has to source compute from somewhere, and those choices get constrained by what’s available, what’s affordable, and what’s politically feasible at any given moment.
If you’re building products that depend on these models, you’re now downstream from those choices. Your infrastructure decisions carry the implicit endorsement of however those models got trained and served.
The Elon Factor
Then there’s the governance layer, which adds another dimension of risk. Elon Musk’s framing of this deal comes with an implicit threat: Anthropic gets to use Colossus as long as they’re building AI that’s “good for humanity,” and SpaceX/xAI “reserve the right to reclaim the compute if their AI engages in actions that harm humanity.”
The criteria for what counts as “harm to humanity” are, conveniently enough, decided by Elon himself.
This is a new form of supply chain risk that didn’t really exist before. It’s not just about whether your infrastructure provider is reliable or whether they’ll go bankrupt. It’s about whether a single individual with other priorities and ego investment in various projects might decide your model doesn’t meet their subjective threshold for social benefit.
For Anthropic developers and API users, this matters. You’re now operating under a business model where your access to compute capacity depends on passing someone else’s judgment call about what constitutes responsible AI.
The Precedent
What gets me is how this deal gets justified. Musk frames it like SpaceX providing compute to competitors (which they legitimately do for satellites) with fair pricing and terms. The parallel breaks down pretty quickly though. Satellite customers don’t get their service revoked because the company that owns the infrastructure thinks they’re being irresponsible. They might if Elon changes his mind about something, but that’s a different problem.
The xAI shutdown of Grok 4.1 Fast, which happened right before this announcement, is a reminder that compute providers can be cavalier about their customer relationships. Developers who’d migrated to that model got two weeks notice before it vanished, with no migration path to alternatives. That’s brutal for anyone who’d built production systems on it.
I don’t think this is a sign that Claude is going away anytime soon. But it does mean that the people running those models are operating under different constraints than we might have assumed.
What Actually Happened
For clarity: there was initially confusion about whether xAI giving all their compute to Anthropic meant Grok was being shut down. That’s not what happened. Anthropic got Colossus 1. xAI kept Colossus 2 for their own work. It’s a partition, not a surrender.
That said, the timing of xAI deprecating models right before announcing they’re leasing out capacity does raise questions about their own priorities and resource allocation.
The Broader Pattern
This whole situation is a window into how AI infrastructure actually works, outside the cheerful press releases and conference keynotes. Companies that want to build responsibly still have to cut deals with less-than-ideal partners because the alternatives are worse or don’t exist. They have to accept governance structures that feel uncomfortable because the compute constraints are real.
The environmental costs aren’t some abstract problem that happens far away and doesn’t affect your daily work. They’re embedded in every model call, every training run, every request that flows through these systems.
For developers building on top of these models, the question becomes: how much do you want to know about where your compute is coming from, and what are you willing to accept in terms of environmental and governance tradeoffs? Because ignoring it doesn’t make it go away. It just means you’re not paying attention to the chain of decisions that made your tools possible.