SpaceX (xAI) – Position of Strength

Anthropic was good, Google is better.

• The deals that xAI has struck with Google and Anthropic look very good for xAI, as compute scarcity appears to have allowed xAI to charge prices for compute that are, as of yet, unheard of.
• This will go some way towards reducing the valuation of SpaceX (of which xAI is the main part) from nosebleed to merely outrageous.
• Both of these deals were clearly flagged ahead of last week’s IPO, but looking at them in a little more detail, combined with RFM Research’s conclusions, leads me to believe that they should be very profitable for xAI but much less so for Anthropic and Google.
  • First, Anthropic: which will pay $1.25bn per month to rent 325,000 GPUs from xAI.
  • This is essentially all of the capacity of Colossus and around 125,000 GPUs from Colossus II.
  • Colossus is thought to contain 180,000 H100/H200 and 30,000 GB200 which by my reckoning will consume around 260MW of power.
  • Colossus II is partly operational, where Anthropic will rent a further 110,000 GPUs that I think will be GB300s.
  • This represents a further 140MW of capacity, making up 400MW all-together.
  • This makes 400MW in total for which Anthropic will pay $15bn per year or $37.7bn per GW.
  • Second, Google, which signed a deal recently for 110,000 GPUs beginning in Q3 2026, which must be at Colossus II, which I suspect will be Vera Rubin GPUs.
  • This makes up 200MW of capacity for which Google will pay $920m per month, $11bn / year or $55.2bn per GW.
  • At GTC, Jensen promised that Vera Rubin should be able to generate revenues of $50bn, which seemed outlandish at the time but the Google deal implies he has more than fulfilled his promise.
• RFM research has concluded that to make a decent return from Blackwell, the provider needs to generate $15bn – $20bn per GW in revenue, while this number is closer to $24bn for Vera Rubin.
• If my estimates are correct, these contracts are going to be hugely profitable for xAI and clearly demonstrate that demand for compute continues to substantially outstrip supply.
• These contracts certainly carry a premium price partly because they can be cancelled with 90-day notice, but I don’t think that this fully explains the huge jump in price.
• However, they are much less profitable for Anthropic and for Google and begin to explain their current focus on limiting usage, increasing prices and maximising revenue generated per GW.
• For the last 3 years, OpenAI has only been able to generate $10bn per GW in revenues, while recent statements from Anthropic have led me to think that its revenue is somewhere between $20bn and $24bn per GW.
• Given that Anthropic is paying $37.7bn per GW to rent compute from xAI, it needs to quickly increase its prices even further to avoid selling its services at a negative gross margin.
• Google’s situation is even more acute, and I suspect that this is the main reason why Google announced last month that it would put rate limits on all of its Gemini service plans.
• This is an indication of a substantial shift in market power from the model makers to the providers of compute, given that the commodity that they supply is in very short supply.
• Consequently, if the neoclouds such as CoreWeave, Nebius, NScale and so on can replicate these sorts of prices, then they should be in a very good financial position.
• It does also raise some serious questions about Oracle, which by my reckoning is sitting somewhere between $11bn and $15bn per GW, implying that it struck its deal with OpenAI at a price that is much too low and now it needs to renegotiate.
• It is those with compute to sell now that are going to fare the best, and Oracle should be in this camp, but it needs to be able to raise its prices and raise them very quickly.
• I don’t think this environment is going to last, as the cure for high prices is always high prices, and capacity will be built to meet the demand and the price of compute will come down to reasonable levels.
• However, a data centre takes around 2 years to build, and so the shortage is unlikely to be relieved in 2026.
• In the short-term, it looks like the neoclouds could see a large improvement in financial performance, meaning that the compute sector could have much further to rally.