Deepseek implications
This post has been crossposted at “The Economics are Just Too Good”
It’s really hard to know when the market is going to start caring about something.
Deepseek was a story in December. I read about it over the holidays. I wrote an ultimately unpublished piece about the length of this one, about how it was yet another sign that AI economics simply don’t make sense. Paired with stories coming out of an AI conference in mid-December, that current AI approaches were hitting limits, and comments from the CEO of Google, that the low hanging AI fruit had been picked, I thought a rethink on AI was obvious. But I’ve thought an AI rethink was obvious for a year and a half based on the prospective economics, and at the end of the year AI stocks found new strength even as there were signs the economics were teetering. Maybe I was missing something.
A couple weeks later, in early January, Hard Fork, the popular and entertaining technology podcast by the New York Times and Platformer also had a story about Deepseek. They wrapped it up in a bow, the cheapness to build, the extended capabilities, the open source-ness of it all. But for whatever reason, Deepseek’s advances didn’t enter the public or at least the market’s consciousness.
But this weekend, the market started to care about Deepseek. A Chinese company that uses several different algorithms to dramatically lessen the computational intensity needed to achieve similar results to the American flagships for a small fraction of the cost. This is tough for a host of reasons.
It highlights a lack of diversity in the approach taken by the American flagship AIs. Importantly, most of America’s AI scientists come out of Google’s coaching tree. At a time and in a field where there should be a multitude of approaches, as Deepseek reveals, the American champions share a general approach with new techniques differentiating themselves at the edges, but with a clear common ancestry and chain of thought. Deepseek has already moved to version 2.0 while the Americans are doing point releases on a 1.0 product.
Today, we have a hundred different technologies that run on our computers and phones in the background. They determine how we see the colors on our monitors, how data gets encoded and transmitted over the internet, how sound is digitized and played back, and none of them are the first or even second version of that technology. The internet, as we use it today, is as much about new and different compression algorithms to get more data through the pipe as it is about improvements in the throughput of those pipes. The Chinese seem to be innovating on the former, while US champions have been focused on the latter.
It reveals an American dependence on advanced hardware. In a lot of ways, we’ve seen this play before and it’s the reason why Honda and Toyota prosper today and Ford, GM, and Chrysler are much diminished from their height. The Japanese made small, fuel efficient cars, with reliable engines, and were initially laughed at (still are in pickup truck world), but quickly became preferred to the large, gas guzzlers American car companies were making and in many ways still do. All the while the American champions bragged about how many horse equivalents their engines mustered. It was the wrong approach then, in cars, and sets us back if our AI companies continue that path now.
It reveals that there are unknown unknowns. ChatGPT was considered the vanguard, Google, a fast follower, Anthropic and Perplexity as dangerous 3rd parties, and Meta as the leader of the open source community. All running on state of the art Nvidia chips. It turns out Deepseek is both a leader with ChatGPT and Google, the leader of the open source community, and can run on less than the most performant chips. We don’t know who else in China has similar if not better technology and when they might emerge. It’s clear that coverage is being shaped or the media has been ill equipped to handle AI developments in other geographies and has led us to believe things that simply aren’t true.
It reveals the US government is behind. The announcement came within a week of the announcement by the US, Oracle, OpenAI, and Softbank of a $500bn investment in data center infrastructure. For economic conservatives that would like the Government to keep out of commerce, it was a painful announcement that reminded us of government largess and Soviet style waste in unneeded infrastructure. Our whole thought process is behind and potentially captured by our AI vanguard that wants the US taxpayer to subsidize their incomes. And this comes on the back of the military AI failure in Israel where to my knowledge, Palantir, as the AI partner of choice, misidentified a multitude of targets early in the war, before widespread devastation became the strategy. Now that we know that the Chinese have commodity AI that runs on much cheaper hardware, the Chinese have a real shot at militarizing competent AI before we do.
So there are a number of implications that need to be confronted, and it's unclear when investors will recognize all the implications. It is clear though that efficient market theory is wrong - even when pertinent information is revealed to the most interested part of the market, the market is slow to pick up on the implications. It took a month to start with the Deepseek news. But in my view, it’s also clear that the reckoning isn’t over. There are still those that believe that the Deepseek data must be wrong or its history misconstrued and just don’t believe the news. There are the issues above. Then there’s that this is a new paradigm in technology investing dynamics. Investors are used to 4-7 companies dominating their markets and using techniques to lock customers in. They’ve forgotten the 60 years before that when technology was a tough place to invest because innovations were constantly leapfrogged and customer lock-in was much reduced. We’ll see how it turns out, but I think we might be heading for at least one sphere where it’s more of the latter than the former and investment multiples with the inability to forecast dominance years out, decline.