Some thoughts on this morning’s $META news, what it means for AI, and the implications for semis:
I don’t view the news as a shock. Susan insinuated they’d been speculatively buying GPUs in March, Mark alluded to the neocloud possibility last month, and $SPCX has already paved the way and demonstrated that near-term returns on a GPU cloud business are quite high. To the extent this happens, it gives $META some incremental utilization/monetization of its GPU infrastructure assets that it would probably have forgone otherwise.
Demand for Meta AI, and the company’s first Muse Spark model, remains in its earliest stages. It is certainly possible, and likely probable, that Meta AI demand does not yet require an overwhelming amount of infrastructure to support it. It’s a good start for MSL, but it’s not frontier and it also does not have a coding presence, which is where the vast majority of usage and spending on AI is being generated. Point being, when you look at management commentary and the quality/usage of the current models, I don’t find potential near-term oversupply all that surprising.
Now, does this have negative implications for Meta’s own AI ambitions via MSL, and does it mean AI infrastructure investment more broadly will crater (first SpaceX, now Meta, who’s next)? I believe the answer to both is no.
On the former, MSL is clearly on a positive trajectory. They rebuilt the entire team and technology stack, Muse Spark is a step-change in quality off the admittedly low base of the Llama series, and the lab has repeatedly emphasized that training bigger models shows reliable scaling with more compute. Point being, bigger and more intelligent models are on the way, and when they arrive, demand for Meta AI, including its integration with the company’s social platforms, will increase commensurately. At that point, Meta’s AI demand will much better match, and likely exceed, its near-term supply. I believe this strongly, and disagree the news means something is wrong at Meta AI.
On the latter for semis, I do not expect this to have any meaningful impact on the pace of AI capex, by $META or anyone else. These companies have longer investment and time-horizons than the public markets, and they're buying chips for years out. That means both Elon and Zuck are preparing for a world where $SPCX and $META are competing at the frontier and demand for their 1P AI products is insatiable. Whether or not you believe in that world (which I do), both are very unlikely to slow spending anytime soon. Moreover, with everyone at the frontier agreeing compute will be the most strategic and scarce resource going forward, I see it as highly unlikely Elon or Zuck stop buying and let that capacity go to competitors. They’d much rather procure and rent it out if need be. That’s a far more favorable position than not having the capacity at all. And suppliers would also prefer less customer concentration, so I think allocations will remain broad. All said, I don't believe we'll see any change in the pace of AI infrastructure investment at $SPCX, $META, or any of the large players for the foreseeable future.
So bottom line/TLDR: This isn’t a shock and it makes sense for now, but I don’t believe it has negative implications for Meta’s 1P AI ambitions over time, or for AI capex more broadly. Buy the semis dip.
