xAI Signs $5 Billion Mega-Deal with Anthropic—No Matter How You Crunch the Numbers, It’s a Win

Author: Jamin Ball

Translated and edited by TechFlow

TechFlow Intro: xAI’s Colossus 1 data center has just found its first buyer—Anthropic has leased the entire compute capacity, committing to $5 billion annually. This deal instantly elevates xAI into the ranks of new cloud computing giants, while Anthropic gains access to compute resources expected to drive $15 billion in annual revenue—making it a win-win for both parties.

Let’s run some back-of-the-envelope math! (All rough estimates...)

Assume Colossus 1 houses 220,000 GPUs.

Assume these comprise 150,000 H100s, 50,000 H200s, and 20,000 GB200s.

Assumed rental rates:

• H100: $2.30 per hour
• H200: $2.60 per hour
• GB200: $5.00 per hour

The blended rental rate across the entire cluster is approximately $2.60 per hour.

Assume a take-or-pay agreement (i.e., payment for full 24/7/365 utilization).

This translates to $5 billion in annual revenue for xAI. We’re witnessing the birth of a new cloud computing giant!

But there’s more—on Dwarkesh’s recent podcast, Dario ran a rough unit-economics calculation (he explicitly emphasized this uses industry-standard assumptions—not Anthropic-specific data—a critical point, as he disclosed no confidential information about Anthropic).

He used a hypothetical $100 billion compute expenditure (just an illustrative round number). This amount would be split between training and inference. If too much goes to training, revenue suffers; if too much goes to inference, future R&D progress is weakened. He noted that the industry currently allocates compute spending roughly 50/50 between training and inference. He added that, as an industry, the $50 billion spent on inference could generate $150 billion in revenue (he specifically flagged this as a likely unit-economics model for the industry within the next 1–2 years).

Applying this logic to xAI’s deal: under the above assumptions, Anthropic pays $5 billion annually—and could convert that into $15 billion in annual revenue (with 60–70% gross margin).

Win-win!!

Dario’s exact words were:

"Think of it this way. Again, these are simplified facts—the numbers aren’t precise. I’m just building a toy model. Suppose half your compute goes to training and half to inference. Inference carries a certain gross margin."

@GS_CapSF: “We’ve signed an agreement with SpaceX to utilize the full compute capacity of their Colossus 1 data center. This gives us over 300 megawatts of new capacity within one month—more than 220,000 NVIDIA GPUs.”