Execution Memo No. 13
New position - The value investor's AI bottleneck play
It has been a while since I disclosed any new positions.
At the start of the war I sold out of most of my biotech book, rotated into cash, and took a handful of smaller tail positions: $LYB, $WEAT, $BAK and a few others, mostly expressed through options. Those worked out far better than I had any right to expect. Most of them ran up over 100%, and with the clock ticking on the options, I closed them out and started hunting for the next thing.
I think I have found two very compelling names. Here is the origin story.
The current market is dominated by AI capex. Photonics stocks have risen wildly. Part of me wants to partake in all of these exciting technologies, because I have managed to spot AI trends before every investor poured into them in the past. Companies like Google and Nebius Group are key examples. But unlike photonics today, those stocks were extremely cheap when I bought them. Photonics is not.
So I recently started searching for AI-related bottlenecks, trying to work out where all of this capex is actually going. I asked Grok to help me uncover emerging AI capex spend. It came back with several names, most of them clustered around the same theme. The first one I came across was Accelsius, which you get exposure to via $INV. Accelsius is essentially a bet on AI data center cooling demand.
As data centers become more power-hungry on the back of next-gen chips, the nature of cooling demand is shifting. Today, if you want to cool a data center, you are most likely using air cooling, essentially mimicking how a modern building is cooled, just at far larger scale. There are several problems with this.
First, cooling a data center with air means you need fresh water to cool that air, and we all know local communities are increasingly worried about their fresh water supply. Second, and more fundamental, air cooling can only get you so far before it hits a hard physical ceiling on how much heat can be pulled off a chip. That ceiling sits around 280 watts of thermal design power, and modern AI chips have blown well past 700.
The solution is liquid cooling, and it can be done two ways: either by heating the fluid up, or by letting it boil. Most of the market is currently focused on heating over boiling, because the technology is more mature. But looking at the physics and the trajectory of power demand per chip, two-phase cooling, where the fluid is boiled to extract the heat, becomes inevitable. And the thing being boiled is not water but engineered dielectric fluids and refrigerants that vaporize around 50C, which is exactly where the chemistry comes in.
Yesterday I wrote all my thoughts on $INV in a longer piece that I will link below. In short, the company is building the two-phase cooling solution, but getting comfortable with the management and the execution is another thing entirely. Luckily, there is a far safer way to bet on this market than $INV. So that is what I bought instead.
