A Shift in the Matrix

By Michael Coolbaugh, Chief Investment Officer

Note: The following comes from our daily research note published to clients on the morning of January 28, 2025. Given the raging debate, we’ve decided to make it public.

Whatever the media is trying to tell you, please do not believe that yesterday (January 27, 2025) was fearful or a sign of panic in any way. Per various trading desks, there was “strong cash buyers” immediately off the open. If we look at the ETF flows, we see the following…

SPDR S&P 500 Index ETF (SPY): +$6.1B; Invesco QQQ Trust (QQQ): +$4.3B (largest one-day flow since 2021); and Direxion Daily Semiconductor Bull 3x ETF (SOXL): +$1.3B.

This is not fear. In fact, it’s the farthest thing from fear. Every single Wall Street note that was published yesterday, except for maybe BMO’s piece on the “AI Power” theme, emphatically stated that the sell-off was overblown and that investors should be aggressively buying the dip. Read through the comments on Twitter (X) and you’ll see the same sentiment. People who have never heard of Jevons Paradox have suddenly become experts on the topic. For every one comment suggesting caution, there were at least ten saying this reaction is absurd and is a generational buying opportunity.

I am not an expert in Jevons Paradox, but here’s the short of it… It says that increased efficiency in resource use can lead to increased consumption of that resource. Basically, what people are saying is that because Deepseek proved to be far more efficient than OpenAI or other US models, it will naturally lead to a boom in demand for more semiconductor chips, AI data infrastructure, power, etc.

That certainly may be true. I mean, look at oil. There’s no question that cars have become much more fuel efficient but that hasn’t led to a collapse in the demand for oil. In fact, we use more oil now than ever. So, I’m certainly not one to argue that the same can’t be said for semiconductors.

What I am saying is that the breakneck speed with which AI infrastructure has been built – and the eye-watering amounts of CAPEX intentions from major Technology companies – might be overdone in the short-run. Remember, one of the promises of AI is that it rendered the cyclicality of semiconductors obsolete.

My point being, even with the ever-increasing demand for oil, that hasn’t canceled out the boom/bust cycle in oil. And the same can certainly be said for semiconductors.

You know, all of this reminds me of my former boss at PointState Capital. Back in 2014, he went on stage at the Ira Sohn Conference in New York City and presented a short thesis on oil.

At the time, oil was all the rage. WTI was trading around $100/bbl and investors were throwing endless amounts of money at E&Ps (exploration & production companies). It didn’t really matter the efficiency of each project or even the margins. All investors wanted to be told was that their money was being used to drill more because oil at $100/bbl meant GOBS OF MONEY for investors.

I actually love the slide that Zach used on stage – it was a clown car, with a bunch of clowns piled in and bursting out of every opening. Now think of that parallel to today’s environment. NVIDIA chips are insanely expensive. But the mad dash for AI supremacy has led to companies spending $60-, $70-, $80 billion per year in trying to reach the top.

To me, what Deepseek has shown is that those gargantuan figures of AI capex may not be necessary, even if it means more demand for chips in the long-run. What it could mean for the short- and intermediate-term is that AI infrastructure has been massively overbuilt because everyone accepted Sam Altman’s claims that all we needed was “more compute”. Remember, Altman is the one who claimed he needed to raise $5-$7 Trillion for increased chip-building (and let’s not forget that the last round of private financing for OpenAI included Nvidia as an investor…).

So, let’s say for just a moment that Deepseek’s claims are fabricated by just a tad. A lot of people are arguing that Deepseek is lying and, in fact, used ~$1-$2 billion worth of Nvidia chips. They also layer on top the few hundred million per year in operating expenses. Even if that is the case, even if it were $5 billion worth of Nvidia chips and $1 billion per year in operating expenses. What does that say for the major US Tech companies that are spending that PER WEEK?!?!

Just like everyone was caught up in the liquid gold rush back in 2014 and asking very few questions about the viability of projects, investors are doing the same today with the AI gold rush. And that’s why I think Semiconductor and these other AI Infrastructure plays are so vulnerable to a bursting bubble. It’s not that there won’t be demand for these things in the future. It’s that investors have extrapolated the never-ending good times for every single year into the future. Like Microsoft’s $80 billion and Meta’s $65-$75 billion in AI capex this year.

The trouble with these massively-inflated expectations is that any deceleration, or perception of deceleration, in demand can be catastrophic for careless investors. With oil, it was the massive drop in oil prices back in 2014 that exposed so many of the inefficient projects. With AI, maybe it’s the cost effectiveness and superior performance of Deepseek (which I’m sure US companies are trying to copy ideas from given it is ‘open source’) that reveals the inefficiencies once again.

At the very least, perhaps it serves as a wake-up call to all the sleepy investors who’ve been blindly throwing money at everything AI and causes them to start asking a few more questions about all those lofty promises.

With hype stories, it’s always about the sentiment. And thus far, there’s been nothing to disrupt the euphoria. But if you don’t believe me that the sentiment has clearly begun to shift, just look at comments from prominent Tech figures, like Marc Andreesen (Prominent Tech VC) or Marc Benioff (CEO of SalesForce).

Per Beniofff on Friday night, “Deepseek is now #1 on the AppStore, surpassing ChatGPT – no NVIDIA supercomputers or $100M needed. The real treasure of AI isn’t the UI or the model – they’ve become commodities…”

Per Marc Andreesen on Friday, as well: “Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen – and as open source, a profound gift to the world.”

And Microsoft’s own Satya Nadella (emphasis my own): “To see the DeepSeek new model, it’s super impressive in terms of both how they have really efficiently done an open-source model that does inference-time compute and is super-compute efficient. We should take the developments out of China very, very seriously.”

Again, I ask; Have the vibes started to shift?