Last Updated on May 10, 2026 by Taya Ziv
Two weeks ago, Anthropic walked away from classified Pentagon AI contracts. The “responsible AI” company chose principles over one of the biggest government deals in tech history.
Then, on May 6, the same company signed a deal to rent every single GPU in Elon Musk’s Colossus 1 data center in Memphis. All 220,000 of them. Over 300 megawatts of raw AI compute power.
And Musk, who has publicly clashed with Anthropic’s founders, sued OpenAI (the company they left to start Anthropic), and built his own competing AI lab, said he spent a week with the Anthropic team and was “impressed.” His exact quote: “No one set off my evil detector.”
If you’re a founder trying to make sense of the AI landscape right now, this single deal tells you more about where the industry is headed than any funding round or product launch this year.
The Signal
Here’s what actually happened. Anthropic’s usage of Claude grew 80x in Q1 2026. That’s not a typo. Eighty times. The demand for Claude Pro, Claude Max, Claude Code, and the API outstripped every compute deal they’d already signed.
So they went shopping. And the biggest GPU rack available, sitting idle since xAI merged into SpaceX earlier this year, was Colossus 1. A Memphis data center packed with 220,000 Nvidia GPUs, including H100s, H200s, and next-gen GB200 accelerators.
The deal gives Anthropic immediate access to over 300 MW of compute within weeks. For context, that’s roughly the electricity consumption of a small city. And here’s the part that should make you sit up: the agreement also includes “exploratory talks” about developing multiple gigawatts of compute capacity in orbit. As in space. Actual orbital data centers.
Tom Brown, Anthropic’s co-founder and chief compute officer, put it plainly: “We’re going to need to move a lot of atoms to keep up with AI demand, and there’s nobody better at quickly moving atoms, on or off planet Earth.”
Why It Matters
This deal reveals three things about the AI industry that most founders are missing.
First, compute is the actual bottleneck. Not models. Not talent. Not funding. The AI companies at the top of the industry aren’t differentiated by their algorithms anymore. They’re differentiated by how many GPUs they can access, and how fast. Anthropic didn’t partner with SpaceX because they admire Elon Musk. They partnered because even when AI gets cheaper to run, the total demand for compute is growing faster than anyone can build it.
Second, ideology collapses under physics. Two weeks between “we won’t build classified weapons AI” and “here’s our deal with Musk’s company.” That’s not hypocrisy. It’s economics. Walking away from the Pentagon was a values call about what they build. Renting Musk’s GPUs is an infrastructure call about how they build it. But the optics matter less than the signal: when you need 300 MW of GPUs, you don’t get to be picky about the landlord.
Third, the AI supply chain is consolidating into a handful of infrastructure gatekeepers. Nvidia makes the chips. A small group of hyperscalers and data center operators control the physical compute. If you’re a startup building on Claude, your product now depends on a chain that runs through Anthropic, through SpaceX, through Nvidia. That’s three companies between your app and your user’s experience. And you have zero leverage over any of them.
The Take
I think this deal is the clearest signal yet that the AI industry has moved from the “model war” phase to the “infrastructure war” phase.
For the past two years, the headlines were about who had the smartest model. GPT-4 vs Claude vs Gemini vs DeepSeek. Model benchmarks. Context windows. Reasoning capabilities. That competition isn’t over. But it’s becoming table stakes.
The new game is about who can secure enough compute to actually serve the demand. Anthropic’s 80x usage growth means nothing if they can’t run the inference. OpenAI’s $122 billion raise means nothing if there aren’t enough data centers to spend it in. The talent exodus that built a dozen AI startups means nothing if those startups can’t access the physical hardware.
This is the part that should worry founders. Because when the infrastructure layer consolidates, everyone building on top of it becomes a tenant. And tenants don’t set the terms.
Remember what happened when AWS became the default cloud layer? Startups built on it, scaled on it, became completely dependent on it. And then Amazon started competing with its own customers. We’re watching the same pattern play out in AI, just faster and with higher stakes.
What To Do About It
If you’re a founder building on AI infrastructure, and in 2026, that’s basically everyone, here are three things this deal should change about how you think.
Understand your compute supply chain. Know exactly which AI provider you depend on, who provides their compute, and what happens if that relationship changes. Anthropic’s deal with SpaceX is good for Anthropic today. But what happens when SpaceX decides to prioritize its own AI workloads? What happens when the lease terms shift? Your startup’s uptime is downstream of decisions made in Memphis and, apparently, soon in orbit.
Build for multi-provider resilience. The founders who win in 2027 won’t be the ones who picked the “best” AI model. They’ll be the ones who can switch between providers without rebuilding their stack. If your product breaks when Claude has a bad day because SpaceX reallocated GPUs, you don’t have a product. You have a hope.
Stop treating infrastructure as someone else’s problem. The promise of cloud computing was that you didn’t have to think about servers. But AI infrastructure isn’t like renting a virtual machine. It’s like renting a brain. And when the brain’s landlord is a rocket company run by the most unpredictable billionaire in tech, maybe think about your backup plan.
The AI industry just showed you that physics beats ideology every time. The companies that win won’t be the ones with the best principles or the best models. They’ll be the ones who locked down enough compute to actually serve the demand.
For founders building on AI: the model wars are becoming the infrastructure wars. And in an infrastructure war, the landlords always win. Make sure you’re not building your entire company on someone else’s lease.
