With the market panic (head fake? overreaction?) around DeepSeek last week, we have been getting a lot of questions about the hyperscalers’ capex plans for this year. The pillar of the AI investment thesis in semis right now holds that the hyperscalers are going to spend a few hundred billion dollars combined to build out all their new AI data centers. DeepSeek’s model led some to the conclusion that AI can be done with much less compute and so no one will need all those data centers. We do not think that thesis holds – DeepSeek is not quite that revolutionary and if someone can make AI math more efficiently they can just do a lot more of it. That being said, it does highlight a growing sentiment among investors. How long will the hyperscalers keep spending?
For the moment, we think the spending will continue, but we do think sentiment is changing. For the past year or so, any time a hyperscaler announced a major capex plan their stock went up. More recently logic seems to have come creeping back into the conversation. Heavy capex is dilutive to earnings, and so some investors have begun to ask where is the return on that investment. Most of the Street seems to remain in the hopeful belief camp, but the DeepSeek panic laid bare the fact that the other side of the argument has a growing coteries of adherents.
Looking through the latest comments from the hyperscalers we detect growing awareness of the problem. No one we have spoken with thinks the big capex numbers for 2024 are in jeopardy, but we also detect a new note of hesitancy in these conversations. In the past week, we get the impression that many of the teams building out all those new data centers all suddenly have the sense that someone is watching them.
Our guess is that this year will see a lot more meetings with the finance team on everyone’s schedules. Purchase decisions will have to undergo an extra layer of scrutiny and approval. No one is going to say no, or deny a major initiative, but the CFOs are watching now, and getting bold enough to speak up.
Leave a Reply