Let's try to contextualize DeepSeek.
Demand for intelligence is infinite. But time is needed to unlock it. The risk is real that capacity is being overbuilt right now.
Disclaimer: The information contained in this article is not and should not be construed as investment advice. This is my investing journey and I simply share what I do and why I do that for educational and entertainment purposes.
This article is entirely free to read.
TLDR Summary
A Chinese open-source app was launched last week and matches OpenAI’s premium offering for free because its technology requires only a fraction of compute for its inference. The details are still unclear with respect to how they did that and what that means more broadly. But it is sending shockwaves through financial markets right now.
The key question is: Do we need all the chips we are currently installing? Investors are nervous about the answer. NVIDIA stock is down 17%, Microsoft is down 3% and the S&P 500 is down 1.8% today as of this writing.
AI bulls are quick to point out that demand for intelligence is infinite and this changes nothing. If it says anything, then that AI will proliferate even faster which should be a boon for most players involved. I am not satisfied with that response.
Of course, demand for intelligence will grow forever without a firm limit. Time is the crucial constraint for growth. AI is a tool for humans to help other humans. And humans need time to develop solutions and scale them commercially. Even the best imaginable AI agent in the world will need time to convince customers and change their habits. There is a constant race between demand for intelligence and supply of intelligence that will determine its price. I would be nervous if I owned stock in a company that produces and/or sells GPUs or has bought a lot of them recently.
We are clearly in a hype cycle that raises the odds that supply is being overbuilt. I have raised that concern in previous articles way before the DeepSeek catalyst. Computing demand is cyclical. It has always been so. It will always be so. Many assets have moved in the last three years as if this is not the case anymore. As if all growth is structural. Investors who believe so are in for a rude awakening. The most important question for everyone else is: How much of the S&P 500’s bull market was the AI picks & shovels play? Your answer on that question will help you to determine how bullish you will be going forward.
What is DeepSeek and why does it matter?
DeepSeek is a Chinese AI company that released a new model on January 20, 2025, called DeepSeek R1. The associated research paper can be accessed here. It instantly made a splash because it seems that it can match OpenAI’s flagship product while using a fraction of the computing resources.
Its app became very popular instantly. But it seems that the true importance of this event just occurred over the weekend. At least that is what the stock market is signaling today.
Community reactions
I am going to refer to two examples that I find helpful to make a few points I deem relevant in the current debate.
Satya Nadella
Last night, Microsoft boss Satya Nadella felt compelled to post a link to a Wikipedia (!) article explaining Jevons paradox. Jevons paradox states that the demand for a resource can increase even when efficiency gains reduce its necessity because falling costs induce an overall demand increase. In information technology this means more efficient calculations reduces the need for them, which in turn makes their output cheaper, which in turn increases the need for those calculations again.
This post is important for three reasons. Firstly, it indicates that DeepSeek is indeed real and relevant and that it likely is considered a threat. Whenever I read about DeepSeek over the weekend, there was always a suspicious undertone. Are the Chinese telling the truth about what they did? And if they did, how much can this actually be generalized to for the entire global AI industry? The fact that the CEO of Microsoft posts about this moment as an important one signals that it is indeed an important one.
Secondly, and more importantly, this post suggests that US Tech leaders were caught by surprise here. Nadella tries to pretend it isn’t so. But he is in my opinion trying too hard with his excitement signaling to be credible. Industry insiders at the very top are clearly surprised. This is remarkable because it doesn’t make them look like they understand their technology very well. And if they don’t, what does that mean about their ambitious capex targets? How firm can these be when DeepSeek is such a surprise to everyone?
Lastly, Nadella seems to be trying to spin DeepSeek positively for Microsoft. He may very well be right eventually. Faster than anticipated AI cost declines make more commercial applications viable from which Microsoft will benefit as the chief computing distributor in the world. But a post like this on a Sunday night feels odd and weak. It seems that he is scared about the narrative forming on Monday. Perhaps he is concerned about his stock price? Perhaps he is concerned how history will judge Microsoft’s involvement with OpenAI? Either way, if he was feeling in control and was comfortable with Microsoft’s place in the industry, would he bother posting this trivial Wikipedia article in that moment? I doubt it.
Nate Jones
The video below is from an AI influencer called Nate Jones which I found to encapsulate quite well what the sentiment in the AI community is this morning. He argues that there is no question that all the compute in the world is needed because demand for intelligence is endless. The more efficient AI computing gets, the faster we will simply achieve our goals.
Tiktok failed to load.Enable 3rd party cookies or use another browser
I find this take entirely unconvincing to be frank. It may be true from a pure product perspective, but certainly not from a financial market perspective. Of course there is a lot of demand for intelligence and there will be even more demand when cost is lower. Nobody would ever dispute that and I don’t think anyone assumed before last week that costs would stop falling.
The question is whether this demand (which depends on our ability to develop and scale use cases) can keep up with supply growth. AI is a tool for humans to help other humans. And humans need time to develop solutions and scale them commercially. Even the best imaginable AI agent in the world will need time to convince customers and change their habits. There is a constant race between demand for intelligence and supply of intelligence that will determine its price. We are clearly in a hype cycle that raises the odds that supply has been overbuilt.
Structural growth and cyclical growth
NVIDIA’s operating performance is an absolute stunner. In 2017, they generated $7bn in revenues and $2bn in EBITDA. This year, they are expected to log $129bn in revenues and $85bn in EBITDA. By 2028, consensus predicts them at $276bn revenues and $175bn in EBITDA. This implies a 64% EBITDA margin, more than 2x of what they generated just two years ago and what would have been a very good year in the 2010s. This consensus predicts that the current computing shortage will continue at least until 2028.
At $3tn market cap, the stock is trading at 45x current earnings and 25x current revenues. It doesn’t sound too out of place among Big Tech. But NVIDIA is a special case in that market cap range.
Virtually all of their revenues are hardware sales. Their revenues are their customers’ capital expenditures. There is certainly some depreciation and technological obsolescence to consider. But in principle, for NVIDIA to post linear growth, their customers’ capex must grow exponentially. And for NVIDIA to post exponential growth, their customers must grow double exponentially.
For that to happen, their customers must extract high RoIs from their investment. If the demand curve for compute ends up showing a dent in the nearterm or perhaps a flatter slope for longer, those RoIs will fall quite quickly. It would not be the first time that a supply shortage precedes a supply glut.
Computing demand is cyclical. It has always been so. It will always be so. A leap in computing innovation makes more commercial applications viable. Demand increases. Supply responds and then overresponds. And demand flattens until the next leap. Absolutely nobody knows for sure where NVIDIA is in the current cycle. This makes a valuation of the company completely impossible. It’s simply musical chairs. Every time.
Beyond NVIDIA specifically, there is a more important question: How much of the S&P 500’s bull market is the AI picks & shovels play? I don’t have a good answer for that question. But everyone’s personal answer to it should be an important input to their overall bullishness right now.
Sincerely,
Rene
I truly appreciate the cautious tone of your recent articles. However, as the saying goes, "The market can stay irrational longer than you can stay solvent."
Should we remain vigilant of the red flags? Absolutely.
Do we know the exact timing or moment when things will unravel? Absolutely not.
For now, I’ll remain cautiously bullish while keeping exit strategies in place for when the market inevitably gets wild to the downside. Please continue posting my friend, cheers!