Author: Jamin Ball

Translated by: Deep Tide TechFlow

Deep Tide Introduction: xAI's Colossus 1 data center just found a buyer—Anthropic has acquired all computing power, paying 5 billion dollars annually. This deal not only instantly made xAI a rising giant in cloud computing, but Anthropic can also leverage this computing power to generate 15 billion dollars in revenue, making both sides extremely profitable.

Let's do some rough calculations! (These are draft estimates...)

Assume Colossus 1 has 220,000 GPUs

Assume among them there are 150,000 H100, 50,000 H200, and 20,000 GB200

Rental assumptions:

• H100 costs 2.30 dollars per hour
• H200 costs 2.60 dollars per hour
• GB200 costs 5 dollars per hour

The mixed rental rate for the entire cluster is approximately 2.60 dollars per hour

Assuming this is under take-or-pay agreements (you pay for full capacity 24x365)

This amounts to xAI generating 5 billion dollars in revenue annually. We are witnessing the birth of a new cloud computing giant!

Moreover—during a recent podcast by Dwarkesh, Dario roughly calculated a unit economics figure (he specifically emphasized that this is general industry math and not specific to Anthropic—this is important, he did not disclose any information about Anthropic).

He mentioned that, taking 100 billion dollars in computing expenditure as an example (he just randomly chose an integer). This expenditure would be distributed between training and inference. If the training proportion is too high, your revenue will be insufficient; if the inference proportion is too high, you will undermine future R&D progress. He believes that currently the industry's computing expenditure for training/inference is 50/50. He stated that as an industry, that 50 billion dollars in inference spending could be transformed into 150 billion dollars in revenue (he specifically pointed out that this is likely the industry's unit economics model 1-2 years from now)

Applying this logic back to xAI's deal. According to the assumptions above, Anthropic pays 5 billion dollars annually. Assuming they convert this into 15 billion dollars in annual revenue (with a gross margin of 60-70%)

Win-win!!

Dario said:

"Think of it this way. Again, these are simplified facts. The numbers are not precise. I’m just doing a toy model. Assume you use half your computing power for training and half for inference. Inference has some gross margin"

@GS_CapSF "We have signed an agreement with SpaceX to use all the computing power of their Colossus 1 data center. This allows us to gain over 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within a month."