Meta & AI Compute – Money Factory?

Meta becomes interesting.

• Meta is thinking of moving into the business of selling compute, which, given the parlous state of its in-house AI, will be a great way to make money while it gets its house in order.
• However, this will only work as long as it has compute to sell, and here, supply remains extremely constrained.
• The most illuminating fact to come out of the SpaceX IPO is just how constrained current compute demand is and the crazy prices that those without it are willing to pay.
• Immediately before its IPO, xAI (a division of SpaceX) struck a deal with Anthropic to lease Blackwell compute for $37.7bn per GW and another with Google for Vera Rubin (I suspect) for $55.2bn per GW.
• At a price of $37.7bn per GW, RFM Research estimates that a Grace Blackwell data centre will deliver a 5-year annual return of 62% per year while Vera Rubin will deliver a 5-year annual return of 61% at $55.2bn / GW.
• These are dream returns, which explains why every man and his dog now wants to move into the business of selling compute.
• In a normal market, this will mean a very rapid increase in supply, leading to falling prices and a normalisation of returns, which I would think should be in the 10% to 20% range.  
• The problem is that this is not a normal market and the sellers of AI services have far more demand for their products than they can supply, and they are trying to grab as much share as they can before the market matures.
• Hence, demand for compute is growing far more quickly than the market can supply it, leading to skyrocketing prices.
• This is because it takes at least 2 years to build a data centre, meaning that it is data centres started in 2024 (well before demand went parabolic) that are coming online this year.
• Hence, supply is likely to remain extremely constrained for at least another 12 months, meaning that anyone who has capacity to sell now is going to make a lot of money.
• This is precisely the maths that led to xAI parking Grok 5 and renting out the capacity instead, and it looks like Meta may do something similar.
• The problem here is that Meta doesn’t currently have any spare capacity to sell and has been unable to purchase the capacity from Google that it needed.
• This means that it will have to park its development of superintelligence and AI in its apps to free up capacity to sell to everyone else and Anthropic in particular.
• Meta’s 1GW, Blackwell-powered Prometheus data centre is expected to come online imminently, and it also has the first 1GW of its deal with AMD starting to come online in H2 this year.
• This means that if it parks its plans, then it could have slightly more than 1GW of capacity that it could sell into the market this year.
• My biggest issue with Meta’s huge capital expenditure plans in 2026 has been that there was no way to see a return on investment, as the capacity would all be used for internal (possibly value-destructive) purposes.
• If Meta instead sells the capacity to the market, then these investments start to look much more interesting.
• For example, the present value of a 1GW Grace Blackwell data centre about to enter service with a 5-year life, where the capex is a sunk cost, is $137bn at a 10% discount rate and an annual revenue of $37.7bn / GW.
• It would also significantly boost operating cash flow and would have an incremental effect on EPS share, pushing the 2026 PER below the 18.5x where it is currently trading.
• If the pivot is real and my calculations are correct, I should be taking another look at Meta, which has lost 17% of its value to the chip makers so far in 2026.
• Meta is getting interesting once again.