Is DeepSeek AI’s Dotcom Moment?
In the 1960s the American auto industry was on top of the world. It became complacent. By the 1970s, it was cranking out poor quality product designed for rapid obsolescence. Anyone of my generation remembers all too well the infamous Pinto and Vega. I had a 1975 Mustang that, once when rolling down the window, the crank snapped off in my hand. Every inspection required exhaust or braking system service. I had to keep a case of oil in the trunk to add some for any trip longer than a few miles and had to be careful where I parked because the only way to start it was to push it, jump in, and throw it in gear. And this was all before the car had 60,000 miles on it.
This was the opening for Japan to eat their lunch. It took another decade for the US auto industry to once again make the cars people wanted and regain its competitive position.
Not to stretch the parallels between the 2020s tech industry and the 1960s auto industry too far. The US tech industry doesn’t produce products that break down, so it makes them obsolete. It doesn’t respond to what its customers really want so much as tell them what they will get. It’s main product has been its stock, and it’s been been only a matter of time before its complacency left it open.
The Mag Seven have been flying close to the sun for a couple years. Nvidia in particular seemed invulnerable to any threat.
Enter China’s DeepSeek.
The short version is that DeepSeek has found a way to make AI cheaper. It works with much less powerful chips and much less costly training. Nvidia will still sell its chips, but they won’t command such exorbitant price premia. Its stock and those of its supercap cousins may not either.
Some have called it US AI’s Sputnik moment. If my auto industry precedent has any validity, the outcome may less happy for US Big Tech.
Financial media are asking what is Washington going to do about it.
You can’t get the right answers by asking the wrong questions.
So the Nasdaq is down sharply, over 3% just today. The Dow, meanwhile, is up. Dividend, quality and value stocks are doing just fine. Investors who thought there was no price too high to pay for a stock and that a good story was good enough are getting a valuable lesson in the value of crunching the numbers.
They could ask themselves, given the obvious fact that huge amounts of money attract competition from every corner of the world, whether Nvidia became a half trillion dollars lower in value today or whether the value assigned before wasn’t realistic to begin with. They appear to already be asking such questions about its Mag 7 brethren.
https://x.com/IranObserver0/status/1883942426894828019?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5Etweet
I’ve seen various articles about DeepSeek and they’ve been vague about how it’s “cheaper.” It doesn’t to be less expensive from a perspective of hardware infrastructure as this video implies that DeepSeek has 50,000 of NVIDIA’s highest end processors which cost about $50,000 each. That’s $2.5B right there and that’s not counting the price of the computers to host those processors, the data centers with their air conditioning and back up power systems, and the software technologists who have to write the software and train it with data sets.
About the only things that I can see being truly less expensive are the technologists and perhaps the data sets. Assuming I’m not wrong about the hardware requirements, everything else is opaquely subsidized by the Chinese government.
Beyond all that, though, DeepSeek does not appear to be anything world breaking in the space of artificial intelligence. It’s still a large language model system that is based on some sort of open source product. It is my belief that LLMs are nothing but better expert systems which have been around for decades. They are not any sort of thinking machine at all.
That said, maybe it is believed that DeepSeek will sell whatever product it has but undercutting competitors and thus wiping out the competition. Regardless of whether it is done ethically or not, I don’t think it really matters. It will be, in my opinion, a Pyrrhic victory as I don’t think this iteration of AI is going to generate enough business value to justify the valuations of the companies involved in this latest bubble. DeepSeek could just be the big winner of a smoldering cinder of a new thing.
I’ve spoken to a friend and told him that DeepSeek does not have the feel of Google when they first appeared. Google’s search engine was so much better and it had a tremendous barrier to entry (their source code is not open source) that all of the competition lost tremendous amounts of market share and the coup de grace came when the dot-com bubble crashed and the competition all went bankrupt.
The LLM AIs have some interesting functions such as aggregating data from multiple sources and presenting them in a reasonably well-written summary. However, the correctness of that information is not guaranteed (AI hallucinations) and for serious work it requires that a person not knowledgable in the subject to verify every assertion.
I guess my point is this. I think DeepSix could be a real spoiler in that it is the madman (perhaps backed by the Chinese government that doesn’t care about a financial ROI) that undercuts everybody else, creating an environment where no one can turn a profit. However, it is not checkmate for China in the AI race, not even close.
Here’s the 30-second video clip where someone claims that DeepSeek is using a lot of NVIDIA chips that are not supposed to be sold to China.
https://x.com/KobeissiLetter/status/1883878322071023807
For what it’s worth, I have not attempted any analysis of the technical aspects of DeepSeek versus the prior state of the art. I’m mostly focusing on the market aspect, both in terms of its impact on securities prices and the reporting that it today overtook ChatGPT in app store downloads, and the big picture state of the tech industry. Thanks for adding some color to the merits of the DeepSeek issue.
I guess I went off the rails 🙂 because it is practically impossible to avoid being inundated with the hype around AI these past few years and, while I could be wrong, this iteration of AI looks to be a really huge sub-bubble similar to the networking bubble during the dot-com bubble.
I still think it’s intriguing that the bubble can be disrupted and maybe even popped by an entity that may be dissembling on its costs in creating a competing product.
Despite the one-day carnage, I was surprised to find that the equities I hold in tobacco companies and the international oil companies were up over 1% for the day. The dividend yields in these companies are reasonably good and they have been consistently paid for decades.
It seems that if there is a correction or crash, there will first be a rotation into quality or perceived quality. Apple was up 3% yesterday.
If it wasn’t this it would be something else. Like any market crash, the ultimate cause is having gone too high to begin with. All else is just catalyst.
I don’t what know what the next catalyst will be, but these stocks are still too high, so if this one doesn’t do it, another will.
For sure it was most hyped that took the most losses. The Dow Jones Industrials rose. My model equity portfolios, emphasizing dividends, quality and value, fared well. They did not rise as much as the Mag 7 they underweight, but that left much less downside to be taken back. Most of the funds rose yesterday.
Today it’s back to the hype. Wall Street and Silicon Valley are making way too much money from glitz-addled investors to give up easily.
I for one am glad to see the the US infotech oligopoly getting some serious competition. Japanese competition made the US auto industry more responsive to its customers needs and wants in the 1980s, and Chinese competition could do the same for the US tech industry.
That a positive tech development caused so much weeping and gnashing of teeth on Wall Street and in Silicon Valley is a shot across the bow.
Another skeptical take:
https://www.zerohedge.com/markets/chatccp-deepfake
Sorting through the technical claims and counterclaims isn’t my bailiwick, but it strikes me the truth is likely somewhere between the extremes. DeepSeek has probably achieved some new efficiencies – certainly outdoing ChatGPT at the Apple App Store is no cheap feat – but if there weren’t any exaggeration in its claims it would be unusual for the tech space.
For the record, it’s never been my position that AI isn’t real, valuable or revolutionary, but that it’s not new and the stock valuations it’s being used to justify are way out of proportion. The only scenario in which they could be even remotely appropriate is if a handful of these companies take over the world. And God help all of us if they do.
the claim is it was done with older chips with less memory due to restrictions, which caused them to optimize things whereas other players just through more money and gear at problems. they were able to write lower level code apparently that would share memory across chips, more cleverly focus where training/inference should occur in internal structuring of the model, and also use some techniques where i believe existing LLMs like chatgpt would be interrogated to make deepseek learn faster and be smarter as well. the oft published 5m not being inclusive of *all* training. there is also some speculative claims that they do indeed have the latest gear and got around hardware restrictions, although the techniques they say they used all appear quite valid and a lot of stuff to pay lip service to if they truly didn’t figure out and actually do it. nonetheless there’s no denying it is much cheaper and portable to train/infer on. i think the real takeaways are simply this; 1. generative AI can be done an order of magnitude cheaper and 2. this is a win for chinese ingenuity and setting themselves on the world stage to say they are serious software players in the game. i don’t really expect this to do long term harm to nvidia, and it should be a win for anyone looking for cheap use of these models, which seem to be trending toward a commodity.
Meanwhile, speaking of barriers to entry, Meta (formerly known as Facebook) has four teams of people to pore over DeepSeek to reverse engineer what DeepSeek has done.
https://markets.businessinsider.com/news/stocks/meta-sets-up-war-rooms-to-analyze-deepseek-s-tech-the-information-reports-1034271747
I didn’t state this in my earlier posting but I believe that the real breakthrough in AI is not going to come from any of the work that is getting so much publicity. I suspect it’ll be something quite different from what they’re currently doing, it might come from a company we’ve never heard of, and I even suspect that the hardware platform will be very, very different from what is currently used (processors with thousands of execution units per die).