Anthropic x SpaceX: 220,000 GPUs and an Odd Couple

Dario Amodei and Elon Musk agree on almost nothing. Musk has called Anthropic "misanthropic," questioned Dario's entire approach to AI safety, and generally treated the company as a punchline. So when the announcement dropped that Anthropic had just secured the entire Colossus 1 data center from SpaceX—every last one of its 220,000 Nvidia GPUs and all 300+ megawatts of power capacity—the reaction was basically: wait, what?

The electricity bill doesn't pause for ideology. And apparently, neither does a deal sheet.

Matthew Berman broke this down in a live stream reaction video, and his read is worth sitting with: "This relieves the compute crunch that they were in and once again makes them extremely competitive and maybe buying some goodwill back. It's a weird partnership between Elon and Dario. They are the exact opposite ends of the spectrum politically."

Weird, yes. But also pretty legible once you follow the money.

The bet Dario made—and what it cost

A few years ago, Anthropic made a deliberate choice: don't go crazy on GPU acquisition. Dario Amodei's reasoning, which he's talked about publicly, was essentially that if AI demand didn't scale at exactly the right rate, over-investing in compute could sink the company. It was a risk-management call. Conservative capex, preserve runway, don't bet the whole company on a demand curve nobody could actually predict.

OpenAI made the opposite bet. Raise everything, acquire every GPU possible, leverage up. It looked reckless from a certain angle.

Then AI demand went vertical. OpenAI's bet paid off. Anthropic's caution became a constraint.

Over the past several months, that constraint showed up in ways that genuinely frustrated Claude's user base. Quota limits got quietly lowered. Peak-hour restrictions appeared without explanation. Third-party tools like OpenClaw got cut off with minimal communication. Berman, who's been a vocal Claude user, put it plainly: "I'm paying you. Let me use the tokens. Don't make it difficult for me to use the thing I'm paying for."

That's not a niche complaint. That's the kind of friction that sends developers to competitors.

What the SpaceX deal actually unlocks

Effective immediately, Anthropic announced three changes tied to the new compute capacity:

Claude Code's 5-hour rate limits are doubling for Pro, Max, Team, and seat-based Enterprise plans
The peak-hours limit reduction on Claude Code is gone for Pro and Max accounts
Opus API rate limits are increasing substantially—Tier 1 goes from 30,000 to 500,000 max input tokens per minute; Tier 4 goes from 2 million to 10 million

That last set of numbers is significant if you're building on the API. The practical effect for subscription users is murkier, because Anthropic still hasn't published what the baseline quota actually is. Doubling an undisclosed number is better than nothing, but it's still a black box. Berman flagged this too: "It's still not clear what you get as part of your quota... it's just a complete black box how they operate."

Transparency remains an open question. The compute crunch is, at least temporarily, not.

Why xAI needed this deal too

Here's the part that doesn't get enough attention: this wasn't just Anthropic being rescued. xAI needed to move those GPUs.

Colossus 1 has been operational in Memphis, Tennessee. Idle GPUs aren't just useless—they're expensive. The depreciation clock runs regardless of whether those chips are processing tokens or collecting dust. Berman noted that "they are losing money by the second that those GPUs are sitting there idle."

SpaceX has made significant capital investments into xAI—this is publicly documented context covered by the Wall Street Journal and Bloomberg in their reporting on Musk's cross-entity capital flows (WSJ, Bloomberg)—which means the pressure to generate returns on that hardware is real. Selling compute capacity to Anthropic generates revenue. The fact that Musk has publicly criticized Anthropic's leadership doesn't make the revenue less useful.

Worth noting: viewers in Berman's live chat raised the question of where Cursor fits in all this, given xAI's separate relationship with the coding tool. The suggestion in chat was that Colossus 2—a separate facility—may be earmarked for Cursor. I can't independently verify xAI's specific arrangements with Cursor, and Berman couldn't either in real time, so treat that as an open thread rather than established fact.

Anthropic's bigger compute shopping spree

The SpaceX deal is live now, but it's not the only thing Anthropic has been building. Their announcement blog listed several other infrastructure agreements in motion:

Up to 5 gigawatts from Amazon AWS, with nearly 1 gigawatt of new capacity targeted by end of 2026—this figure comes from Anthropic's own announcement materials, so treat it as their projection, not an independently verified number
A 5-gigawatt agreement with Google and Broadcom, coming online in 2027
$30 billion in Azure capacity via a Microsoft/Nvidia partnership
$50 billion in American AI infrastructure through Fluidstack

Berman interviewed Google Cloud CEO Thomas Kurian earlier this year, and Kurian suggested Google had planned sufficient TPU capacity to serve multiple markets simultaneously—internal Gemini workloads, external inference for partners like Anthropic, and direct chip sales. Sundar Pichai, separately, said Google would have more revenue right now if they had more compute. Those two statements don't fully reconcile, which is its own interesting tension.

The aggregate picture: Anthropic has gone from compute-constrained to compute-hungry-but-shopping, very quickly. Their partnerships team has been busy.

The Nvidia problem nobody's solving

Zoom out from any single deal and the structural reality is the same everywhere: Nvidia can't make chips fast enough. Every major AI company is constrained by the same bottleneck, just in different ways. Berman's framing: "AI demand is going to outstrip supply for probably at least another decade, if not more. There is infinite demand for intelligence in the world."

That might be optimistic on the timeline, but the direction is right. The demand curve for inference capacity isn't flattening—it's accelerating as more applications get built on top of these models.

Berman made a point here that I keep turning over: when he asked Greg Brockman whether OpenAI could move workloads between Nvidia GPUs, Google TPUs, and AWS Trainium, Brockman said it was actually pretty easy. If that's true—if the hardware is becoming interchangeable from a model-serving perspective—then the competitive moat isn't in which chips you use. It's in having enough of whatever's available. The chip manufacturer wins regardless. The model companies are just fighting over allocation.

Which raises a question I don't see answered anywhere: at what point does having all this third-party compute create dependency risk? Anthropic is now distributing its infrastructure across AWS, Google, Microsoft/Azure, Fluidstack, and SpaceX/xAI simultaneously. That's diversification by one reading—and by another, it's five different relationships to manage, five different pricing negotiations coming due, five potential single points of failure. The specific concern isn't abstract: if any one of those partners changes terms, gets acquired, or decides Anthropic is a competitor rather than a customer, what's Anthropic's next move? They still don't own the silicon.

The orbital footnote

Buried in the announcement: SpaceX and Anthropic "have also expressed interest in partnering to develop multiple gigawatts of orbital AI compute capacity." No timeline, no specs, no details beyond that sentence.

Orbital compute is genuinely being discussed at the highest levels of the industry. Jensen Huang has said it's worth exploring. Musk is obviously bullish. (Berman recalled hearing skepticism from a prominent industry voice—I'm not going to paraphrase secondhand; if you know which interview he's referring to, drop the link in the comments.) Whether space-based inference ever becomes real infrastructure or stays in the "interesting press release" category is a genuinely open question.

For now, the 220,000 GPUs in Memphis, Tennessee are the story. Anthropic users who've been hitting rate limits are getting real relief today. The company that made a conservative bet on compute and paid for it with months of user frustration has just made several very large, very expensive bets in the opposite direction.

Dario and Elon still don't agree on much. But they're splitting the electricity bill in Memphis.

By Yuki Okonkwo, AI & Machine Learning Correspondent, Buzzrag. Source material: Matthew Berman's live stream, published approximately 11 hours ago.