China just released a new A.I. model that has led U.S. A.I. companies like Open A.I. as well as chip suppliers like Nvidia to shed trillions of Dollars in value overnight.
So, did China just pop the A.I. bubble?
To answer that question, we need to first ask ourselves some hard questions. Was there actually an A.I. bubble? And, if so, did China pop it? or will A.I. stocks bounce back before we know it?
To answer these questions, let’s first do a quick recap of
What happened?
First, let’s talk about the technology. In 2022 Open AI shocked the world with the launch of it advanced AI chatbot, ChatGPT, built on the back of innovations in so-called large language models , LLMs for short.
Over time, LLMs both got more powerful and more power hungry, leading to an unprecedented boom in the valuation of, mostly, American companies that either built, ran, or provided the infrastructure to power these models.
But then, over the course of the last two months, two new versions of the Chinese Deepseek LLMs were released, alongside various Chinese alternatives. These models rivalled the performance of America latest models. But, even more importantly, they were far cheaper to train and run, leading to a massive sell-off of in American A.I. infrastructure stocks like Nvidia.
However, if we zoom out a bit, we can see that, so far, that sell-off, while big, does not really seem THAT bad. ZOOM, So, was this all just a panic?, soon to be forgotten, or is this the start of the end for America’s A.I. bubble? To answer that question, we first need to answer:
Was America’s A.I. sector in a bubble?
According to ChatGPT, a bubble
refers to a situation where the price of an asset—such as stocks, real estate, or cryptocurrencies—rises significantly above its intrinsic value due to excessive speculation, investor enthusiasm, and irrational expectations.
And yes, I made sure chatGPT was not hallucinating on that definition.
Anyway, one of the most famous bubbles in recent history, was the dotcom bubble in 2001, during which hundreds of companies that made close to zero revenue were valued at millions of Dollars simply because they promised to do ‘something’ with the internet.
Crucially though, some companies that were highly valued in the overall stock market bubble, turned out to be correctly valued because they indeed would become the tech titans of today.
Therefore, in hindsight, we can say that while the 2001 tech hype was a bubble because investors were wrong about most companies that were part of it, the overall bet on the internet as a technology, and some companies in them, proved to be correct.
To make matters even more tricky, we can never be completely sure whether or not we are in a bubble right now because we do not know what the intrinsic value of any company is because it is based on the future profits, which are inherently uncertain.
That being said, economists who use bubble detection algorithms have certainly found strong signs of bubble like behavior in US tech stocks today.
However, before Deepseek arrived on the scene there were good reasons to believe that
Much of the A.I. industry was NOT a bubble
Now, to understand why, we first need to talk about the pre-Deepseek economics of A.I., where 3 types of companies were emerging as A.I. winners.
The first were model providers like openA.I. or Anthropic, which trained huge large language model and then provided light user access to it via a subscription service, while charging power users for each time they use the model via the API. These model providers needed to train their models using huge data servers provided by the cloud providers, which in turn relied on Nvidia graphics cards.
But, much like the tech giants who survived the last tech bubble, what really made these companies valuable is not their technology, but rather the fact that they are protected from competition. Ebay and Amazon were protected by so-called network effects, which meant that their product marketplaces worked better the more people use it. However, so far, it seemed like the A.I. titans were shielded from competition in a different way, by the massive capital required to train and run these models, build server centers, and develop state of the art graphics cards.
This made their business model more like that of Airbus and Boeing than that of Facebook. Airbus and Boeing are less profitable than Facebook, but they are still extremely profitable and very well protected from competition.
So, in that sense, the sky high valuations for big model providers, datacenter providers, and Nvidia, were arguably NOT a bubble at least up to the Deepseek model releases. So,
How did Deepseek disrupt the economics of A.I.
Well, interestingly, Deepseek did, arguably, not disrupt the economics of A.I… first. It was actually open A.I. itself that changed the economics of A.I. when it released it’s 01 model in late 2024.
You see, 01 is a so-called reasoning model, which is different from the LLMs that came before because it reasons more like a human being, meaning that instead of being extremely expensive to train, it is extremely expensive to run.
But, given that the big model makers were mainly protected by a high capital requirements to train these models this could have already endangered them, if it wasn’t for the fact that OPEN!!… A.I. does not OPENLY share its models.
However, this was no threat yet to the data server providers, nor to Nvidia, because their chips and servers would still be needed to run Open A.I.’s new reasoning model. Therefore, ironically, Open A.I.’s new reasoning model initially seemed to vindicate the crazy investments in A.I. infrastructure and research.
That is, until Deepseek came out with it’s own, open-source, cheap to train, and cheap to run, regular AND reasoning models. And, despite the fact that these models were still trained on Nvidia chips, they were trained on much weaker, CHEAPER Nvidia chips. On top of that, the Chinese models can in many cases even be run on consumer computers, rather than on data servers, meaning that now the main line of defense for the 3 sectors of American champions is potentially gone.
So, while this is great news for large language models as a technology, and great news for companies that use or build services on top of these models, it just made the valuations of Nvidia, cloud-service providers, and especially Open Ai. seem irrationally high.
So, while the sector was arguably not completely in bubble territory before, it now all of the sudden LOOKS A LOT MORE like a bubble. A bubble that could very well deflate further now that it has become clear that competition will be fierce.
However, that bubble deflation may look very different for the companies that we have discussed. So far, the arrival of Deepseek looks really bad for model providers like open A.I., as it showed they are not as well protected from competition as previously thought. And while cloud providers and Nvidia, now face more competition because of the lower cost of running these models, they could actually be saved by the fact that cheaper model may mean that MORE people want to start running them, meaning that their valuations may not be so crazy after all.
So, actually, I would argue that China just popped ‘parts’ of the A.I. bubble, not all of it. In fact, much like Amazon valuations were not a bubble within the Dotcom bubble, it could be that, even if stock prices from giants like Nvidia fall further, that they were not technically a bubble and that their stock prices will come back with a vengeance in a couple of years time.
But, then again that is what it seems like to me, right now. This is a new technology. The race to economically dominate this technology is far from over, so I’ll keep you posted on any new developments. But, in the meantime, if you want to dive deeper into the economics of A.I., I highly recommend you check out our video sponsor The Economist. Specifically, for this video I relied a lot on this amazing article about the economics of reasoning models, this deepdive on the market impact of deepseek, and this super interesting article about the other Chinese models that are rapidly being developed.
All of these articles are linked in the top comment under this video, just alongside the link to our exclusive link giving you a 20% discount The Economist, which gives you full access to their daily, in-depth global reporting for an entire year.
The Economist been part of my media diet long before I started this channel because they stand for independent journalism for independent thinkers. Whether you prefer to enjoy the print edition over a cup of coffee or their app’s digital weekly edition, you’ll always stay ahead of the news no matter where you are.
So, don’t miss out – click the link in the description or top comment, or head over to economist.com/moneymacro to claim your exclusive 20% discount today.