Research/SemiconductorsAIPositioning

The AI capex cycle is now a second-derivative trade

Growth is still growth. But in semis, the market has stopped paying for the level of spending and started paying for the rate of change. That distinction decides the next twelve months.

Daniel OkaforHead of Equity Strategy·November 6, 2026·9 min readPremium

For two years the trade was simple: own the compute. Every quarterly capex raise from the hyperscalers was read as a green light, and every accelerator name re-rated in sympathy. That regime is ending, not because spending is falling, but because the market has learned to price the spending in advance.

When a sector trades at 30 times forward earnings, the level of demand is no longer the question. The question is whether the pace of demand growth is accelerating or merely high. Those are different trades with different holders, and the transition between them is where most of the damage happens.

What the order books actually say

Order backlog for advanced accelerators still runs ahead of deliverable supply, and networking attach rates are climbing as clusters get larger. But the incremental growth rate has decelerated for three consecutive quarters across the leading merchant supplier. Nothing about that is bearish on the business. It is simply bearish on a multiple that assumed the second derivative stays positive indefinitely.

The tell is in the guidance cadence rather than the absolute numbers. Companies that can still raise guidance mid-quarter are operating in a demand-constrained market. Companies that guide conservatively and beat are operating in a supply-normalising market. Watch which pattern shows up next quarter.

In a cyclical boom, the multiple peaks before the earnings do. That is not a reason to short — it is a reason to size differently.

The three names that carry the beta

Exposure to AI compute is concentrated in a handful of tickers, and their trends are highly correlated. If you own three of them, you own one position at triple weight. Our trend model currently grades the accelerator complex as a hold-to-accumulate: price above the 50-day, 50-day above the 200-day, but momentum decelerating into the print.

Merchant accelerator exposure is the highest-beta expression and the first to break if guidance disappoints.
Custom silicon programmes are a second lane, but they are customer-concentrated and lumpy by contract.
Software and networking attach is the quiet compounder — lower beta, better multiple durability.

How we are positioned

We are not fading the trend. We are reducing concentration and shifting weight toward the infrastructure layer where revenue is recognised on usage rather than unit shipments. If the capex cycle merely plateaus, the usage-linked names keep compounding while the unit-linked names de-rate. If it re-accelerates, both work.

The invalidation is clean: a break of the 50-day on heavy volume across the complex, followed by a lower high. Until that happens, the trend is still up — it is just no longer early.

Premium research

Continue reading “The AI capex cycle is now a second-derivative trade”

Full-position reasoning, invalidation levels and the model breakdown are available on the Premium plan.

View plans

Demo build: checkout is not connected, so the preview button bypasses the paywall.

Disclosure

This article is analysis and education, not investment advice, and contains no recommendation to transact. The desk holds no positions in the instruments discussed. Figures referenced are generated from a sample dataset in this build. Read the methodology.