Case studies › AI · Concentration
The AI screen has quietly become a semiconductor screen
Run any screen for AI exposure and the survivors collapse into four buckets: accelerators, memory, packaging and power. The theme is broad and the investable surface is four supply chains long.
That is the arithmetic of a basket. If four of five names are levered to the same capital expenditure cycle, the basket is one position with four tickers, and the diversification is nominal.
The concentration is already in the index
It is not a hypothesis about what could go wrong. It is in the earnings base of the index now. In the third quarter, with results in from companies representing 68.8% of the sector's market capitalisation, tech earnings rose 33.1% on 28.8% higher revenue.
Excluding three names, the sector's growth rate is 27.7% rather than 44.2%. Two of those three are a memory maker and a fabless accelerator designer. That is a 16.5 point swing from a three-name exclusion, which is the cleanest available measure of how much of the sector's growth sits in a handful of tickers.
| Measure | Value |
|---|---|
| Market cap reported, Q3 | 68.8% |
| Tech earnings growth | +33.1% |
| Tech revenue growth | +28.8% |
| Growth excluding three names | +27.7% |
| Growth with all names | +44.2% |
What a real diversification test looks like
Not sector labels. Revenue concentration: how much of each name's top line depends on one buyer, one product or one cycle. It is disclosed in the ten-K customer concentration note and in the segment table, and it is rarely read.
Three questions sort a list quickly. Who is the largest customer as a percentage of revenue? What share of revenue is one product family? How many quarters of backlog cover exist against the current run rate?
A name with one customer above 20% and a single product family is a derivative of somebody else's capex committee. A name with a diversified base and multi-year cover is a business. Those two should not carry the same multiple, and in a screen they usually do.
The power leg is different, and that is the point
Three of the four buckets are levered to the same order book. The fourth — power — is levered to a permitting calendar, which is a different clock and a different failure mode.
That makes power the only genuine diversifier in the set and simultaneously the only leg with political risk attached. Buying it as AI exposure and treating it as an AI stock is the common error: it hedges the capex cycle and imports a regulatory cycle in exchange.
The correlation evidence
On 5 November the Nasdaq fell 5.5% and erased US$2.5 trillion. A basket built on four correlated buckets produced roughly one drawdown on that day. That is the assumption being tested, in public, with the whole book.
If the drawdown continues, correlation rises before it falls, because forced selling is indiscriminate. Diversification bought for a 5% day does not work on that day. It works over the quarters that follow, and only if the underlying revenue bases differ.
Where the case breaks
The concentration argument breaks if the four buckets diverge in their order books. Memory pricing is a commodity cycle, accelerator allocation is a capacity cycle, packaging is a fab cycle and power is a regulatory cycle. They are correlated through the buyer, not identical in themselves, and a screen should separate them instead of stacking them.
The risk framing breaks if end demand broadens beyond the few hyperscalers committing capital today. Enterprise adoption at 10% to 15% understanding is the indicator, since enterprise budgets are set independently of each other and of the hyperscalers.
What the arithmetic says
A 16.5 point growth differential from excluding three names, four correlated buckets, and one regulatory leg. The theme is real and the index is already carrying the concentration. The screen is not wrong, it is just measuring a supply chain and calling it a portfolio.
Exclude three tickers and sector growth falls from 44.2% to 27.7%. That is the whole diversification argument, in one arithmetic step.
Sources
Sector growth, reported market capitalisation and the with-and-without comparisons are third-party earnings-season statistics. The five-session index move is market data. The concentration questions and the bucket correlation argument are derived and can be checked against filed customer concentration disclosures.
Published for information only. Nothing here is investment advice or a recommendation. This note is not prepared in accordance with the legal requirements designed to promote the independence of investment research. No position is held in any security named.