xAI – Printing Cash

Another deal at very high margins.

• ReflectionAI’s deal with xAI raises a lot of questions, but while ReflectionAI has cash to spend, it is clear that xAI will be making a lot of money.
• ReflectionAI and xAI have announced that Reflection will rent GB300’s in Colossus II for which it will pay $150m per month starting July 1^st and running for 3½ years.
• Each party can terminate the contract with 90 days’ notice.
• Unlike the Anthropic and Google deals, xAI has declined to disclose how many GPUs ReflectionAI is renting in what appears to be an attempt to prevent gadflies like me from estimating the economics of the deal.
• However, we can make some educated guesses.
• I have estimated that Anthropic is paying $37.7bn per GW to rent all the capacity in Colossus I and some in Colossus II, while Google is paying $55.2bn per GW in Colossus II starting in Q3 2026.
• From this, I have concluded that Anthropic is renting GB200s and GB300s while Google is renting Vera Rubin.
• The only way to explain the price disparity between the two different rates is that they are renting different generations, and given that Rubin should produce 3x the number of tokens per GW compared to GB300, the numbers begin to make some sense.
• This means that the going rate for the Blackwell at the moment is $37.7bn per GW, which is an astronomical price, especially when one considers that much of the industry earns just $10bn per GW and even Anthropic is only making $20bn – $25bn per GW.
• This points to a substantial shift in market power away from the AI companies towards those that have compute to sell.
• Using Anthropic as a guide, this would imply that ReflectionAI is renting 40,000 GPUs.
• The most surprising aspect of this deal is that xAI had the capacity lying around and ready to go as all that anyone hears these days is that everyone is sold out many months ahead.
• I suspect that xAI looked at the going rates for compute and decided that they are currently so high and so lucrative that it is better to rent out the compute and come back to Grok5 another day.
• This makes some sense, as at $50bn per GW the 5-year return on Vera Rubin is 55% while at $37bn, the 5-year return on Blackwell is more than 40%.
• These are fantastic returns, which is why everyone is suddenly becoming a data centre provider and is trying to stand up compute as quickly as they can.
• However, the cure for high prices is always high prices and at some point, there will be so much supply that the rates for compute will start to come down again, but this may not be for a while.
• This is because there is not enough silicon or electricity to meet demand, and so for now the current mania looks likely to persist.
• This means to me that prices for AI services have to rise and rise significantly, as even after the most recent price rise Anthropic is only earning $20bn – $24bn per GW, and to make a 37% gross margin on $37.7bn, it will have to charge $60bn.  
• This is more than double what it makes now, and so it looks certain that further substantial price rises are coming, raising the question of what this will do to demand.
• The real winner here is xAI and anyone else who has capacity to rent out at these prices, meaning that CoreWeave is potentially not in so much financial trouble after all.
• The other dark horse here is OpenAI, which struck its capacity deals more than a year ago, meaning that it could be sourcing compute at a rate of less than half of Anthropic.
• This could give an edge, as this would give it the ability to compete far more aggressively than many of its rivals.
• However, this is far from clear at the moment, and so the place to be is still in the picks and shovels who are the recipients of a once-in-a-generation spending spree.
• I am still holding Qualcomm, Samsung and nuclear power to benefit from the ongoing AI boom.