Who amongst us has not looked at the consistency of the transformer model
loss curves and thought, "yeah okay, but how do I make a trade on this?"
I've spent a bit of time trying to get allocation on secondary lab stock,
with some success. Private markets, however, don't allow you to make
leveraged bets.
So the question becomes - how do you best trade on short AGI timelines on
the public markets? Finding a company that's part of the AI supply chain
is necessary but not sufficient.
After some research, I've found these to be the most important factors:
1) AI beta. The single most important factor. Basically: what % of the
company's top line depends on AI.
The reason Nvidia ripped much more than TSMC is that TSMC had a broad and
diversified portfolio of customers, of which only a smaller minority was AI
related. So even one doubling of AI demand isn't enough to make AI a
substantial part of their revenue. Nvidia on the other hand, was heavily
concentrated in AI, and one doubling in demand was basically enough for
Nvidia itself to double.
To figure out where I should invest my savings to be consistent with a
strong belief in short AGI timelines, I've worked with Claude to map out
the AI supply chain; singling out specifically the companies with high AI
beta:
AI Revenue %<15%15-30%30-50%50-75%75%+
Color by
Box width = market share within layer; color = AI revenue mix. Each card shows market share, AI revenue % (rev), and AI operating income % (op). Click a company or layer for detail.
2) Priced-in status. As a former believer in the efficient market
hypothesis, I've tried to figure out the gap between what demand the market
currently forecasts, and what you'd actually expect on short AGI timelines.
One way to quantify that is: what's the delta between GW of buildout
capacity currently planned, and what will actually be required if we get a
"drop-in remote worker" within a year?
Each company's current orders or stated capacity expressed as % of 15GW of additional AI-infrastructure demand expected in the next 12 months. Below 60% suggests room for upside; above 100% is effectively saturated.
<60% (upside)60-100% (moderate)100%+ (saturated)Vertical line = 100% coverage
Each company's current orders or stated capacity expressed as % of 15GW of additional AI-infrastructure demand expected in the next 12 months. Sub-60% suggests upside; above 100% is saturated. Cleveland-Cliffs and GE Vernova read off-scale because they sell into many other end-markets too.
3) Idiosyncratic risk. In any given layer of the supply chain, there
will be several companies supplying something. You want to avoid the
failure mode of betting on the right technology, but on the wrong company.
For instance, in HBM memory used in AI chips, there are only three players
(SK Hynix, Micron, Samsung) that could compete, and so the substitution
risk is lower. Compare that to the neocloud market, where there are many
competitors, and it's hard for a non-expert to decipher which ones to bet
on, even if you're certain there's a compute crunch coming and model
companies ready to be paying meaningful premiums for compute.
siliconphysical infrapowerother
CompanyCoverage / Sub. risk
Coverage = current orders / stated capacity expressed as a share of 15GW of additional AI-infrastructure demand expected in the next 12 months. Bottom-left = low coverage (room to surprise) + low substitution risk (can't be displaced). Dashed line marks 100% coverage.
Data assembled from SEC 10-K/Q filings, earnings calls, and press releases
through Q1 2026. Where exact disclosure was not available, values are marked
with "~" or "est". The research was made for my personal use, and then
shared at the encouragement of a friend. This is certainly not investment
advice for others.