Meta Platforms – Jackpot!

Better lucky than good.

• Meta’s failure to produce a leading model is playing to its advantage as it now has the opportunity to sell its compute capacity into the market for an excellent return.
• There is a fierce debate raging on both the internet and in the market about whether Meta selling its capacity into the market is a negative signal or not.
• The negative argument is that this is a sign that Meta has overbuilt capacity based on its needs, which, combined with tokenmaxing giving way to rationing and a buyer of compute becoming a seller, means that demand has peaked.
• The bears are also latching onto comments made at Meta’s AGM about possibly selling the compute if Meta found that it had overbuilt as another sign that the peak of demand is now behind us.
• I do not doubt that at some point, the industry will find itself in a position of oversupply of compute, but I do not think that this is what is going on with Meta right now.
• There is no doubt that Meta is having trouble with its AI, as the API for Muse Spark was supposed to launch weeks ago and has still not materialised, and its leading position in open source has long been ceded to the Chinese.
• However, at the same time, Meta was unable to buy more capacity from Google that it needed to run the AI that sits behind its Family of Apps business, and xAI has very recently been able to command prices for its compute that imply very high returns and cash generation.
• The shortage argument is also supported by the fact that data centres are difficult to build and some like QTS (owned by Blackstone) have abandoned plans due to local opposition and problems with getting planning permission.
• This means that it is going to take longer to bring supply into the market, meaning that the value of supply that is available now will be higher.
• Hence, I think that the situation at Meta is not one of oversupply of compute, but a decision to take advantage of a crazy market while it gets its house in order.
• This is precisely the same decision taken by xAI, which was to park Grok 5 and sell compute to Anthropic and Google at prices that will earn it more than 60% per year.
• In a crazy market, the cure for high prices is always high prices, and this is Meta acting rationally as opposed to panicking and dumping excess capacity into the market.
• The net result is that if more players who have compute decide to park their AI products and sell their capacity, then prices are going to come down, but that has yet to come to pass.
• Consequently, I have had a look at Meta’s core business, which is becoming much more efficient thanks to finally being able to automate its content vetting procedures and concluded that it is worth around $700 per share, even including $135bn capex spend this year and similar going forward.
• This assumes that all of this capex is only used to support internal processes, which don’t do very much other than keep the Meta ecosystem viable and steadily growing, albeit more slowly in the long term.
• If I then assume that Meta sells 250MW in 2026, 1.2GW in 2027, growing to 5GW in 2030 at prices starting at $37.7bn/GW today but falling to $25bn / GW in 2030, I end up with an extra $70 per share.
• This is because the return on investment of selling into the market is higher than keeping it internally, which is why I think Meta will go ahead with this project, but only for 5 years.
• This will be temporary because at some point the cost of compute will fall to a point where the returns of using it boost its in-house products and services will be higher, and it will withdraw from the market.
• Hence, if Meta starts selling its compute, I get $770 per share (33% upside) instead of $700 per share (21% upside) if it does nothing.
• Furthermore, I think the risk of an imminent collapse in the price of compute remains remote, meaning that the risk-reward on Meta is disproportionately skewed to the upside.
• I am considering taking a long position in Meta with a 9-to-12-month time horizon.