January 27, 2025 at 07:00PM
Like everyone today (and all weekend), I’ve been trying to understand what DeepSeek means. Some thoughts:
1) I don’t care how they trained it, the cost to run is most important
Whether DeepSeek had a hidden stash of H100 chips doesn’t matter nearly as much as the changes in opex. The model is free, open source, can be run locally and uses about 3% as much compute as OpenAI’s o1.
That has vast implications for chip investment and power build out. Shares of high-flying Vistra are down 27% on that theme while chipmakers are being crushed across the board.
2) Lower costs could broaden applications but I think there’s a limit.
I don’t need AI in my toaster or –at the very least — I won’t pay much for it. The market is sorting through the winners in this. META opened lower but is now higher because they’re not a datacenter company, they’re an AI user (to make better ads) and lower opex directly benefits them and would mean that they don’t need to spend $65 billion a year to stay in the AI race.
3) All this gets complicated on the macro side
Broader, cheaper AI is disinflationary but that’s not the end of the story. By making it cheaper and broader, it also should increase the speed of adoption and make it more-disruptive. That will be destabilizing (and scary). There are second-order effects as well, like how much governments will spend as instability increases, adding to deficits and devaluation.
4) It all makes a good argument for hard assets
I can see that, but in a world of falling prices and lower employment, you can have temporary surpluses in these things that are tough to predict. Prices aren’t exactly rock bottom now.
5) So what’s the trade?
OK, so now we’ve gone through 10 steps of complicating this and aside from shorting things — which is fraught with danger. We’ve stumbled into a bunch of conflicting signals, especially if you’re a longer-term investor. So let’s wash it all away and try to simplify, because that is what usually works.
China scored a huge victory with DeepSeek in every sense. This is a big boost for national pride right at the time of the biggest holiday of the year. China also has political structures and supply chains that are poised to benefit from AI. The main drawback was that the US was cutting off chips, but this shows that might not matter. Or if you think China cheated to build the model, well, it at least shows they’re good at cheating.
So maybe that’s the whole story and the right trade for now. China scored a win. Buy China.
It also helps that Chinese stocks are very cheap and the market will soon fall in love with the idea of stimulus at the March 7 National People’s Congress (there is still time to front-run the front-run. Markets also crave underdog stories, and China fits perfectly—cheap, hated, and now “solving” AI’s cost problem.
This article was written by Adam Button at www.forexlive.com.