The category layer

Stock categories

Small equal-weighted groups, each one standing for a single constraint on the buildout.

A stock page tells you what one company did. A category asks the question underneath it: which input is actually binding right now, and is the category that sells it winning? The record shows the constraint moving — power broke out in early 2024, memory not until late 2025, and local opposition to data centres is the only one still setting records in the last complete quarter.

These are not investment products and they carry no track record. Each list was written on the date it was defined, out of names that existed on that date — so everything that already failed is missing by construction, and any earlier stretch of the chart overstates the category. Weighting is equal, because cap-weighting the memory basket would just build a Micron tracker. And equal weight is not diversification: every basket reports how many independent bets it really holds. Several hold barely more than one. The methodology sets out where these rules come from. Looking for one name rather than a category? Every stock we cover is listed in the registry.

Compute — the silicon itself

The silicon everyone else is buying The sellers of the thing the whole buildout is for. Nvidia's own edge mix moved from launch-heavy to competitor-heavy over two years — the category is getting contested, not just bigger. 6 names ≈1.9 real bets Who actually assembles the racks The assembly layer between the chip and the data-centre floor. Thin margins on enormous volume — the part of the buildout that scales without pricing power. 6 names ≈2.4 real bets The other stack Chinese AI listings, cut off from the leading-edge silicon the rest of these baskets take for granted. 3 names ≈1.7 real bets

Memory — the newest binding input

The newest constraint Coverage did not move until 2025Q4 — eighteen months after power. Most of the capacity is listed in Seoul, so this basket is priced in three currencies. 6 names ≈1.9 real bets The machines that stack the memory Dicing, bonding, test and substrates. If HBM is the constraint, these are the tools that decide how fast it loosens — and none of them is a memory maker. 5 names ≈1.9 real bets

Interconnect — the wiring between

The wiring between the GPUs Sells more silicon per rack as racks grow, regardless of whose accelerator wins. 9 names ≈1.8 real bets

Fabrication — where chips get made

Where the chips actually get made The oldest constraint and still the deepest — a step-change in coverage from late 2025. 4 names ≈1.6 real bets

Power — electrons and the gear to move them

The gear the grid is short of Transformers, switchgear, cooling. Thin in the record — 12 articles — but the earnings say otherwise. 14 names ≈2.1 real bets The twenty-year answer to a two-year problem Highest internal correlation of any basket here; treat the equal weight with suspicion. 13 names ≈1.6 real bets Who sells the electrons First constraint to bind: coverage broke out 2024Q1 and has not gone quiet since. 12 names ≈2.6 real bets Filling the gaps in the load Batteries, fuel cells and on-site generation — the answer when the grid connection exists but the supply is not firm. 6 names ≈2.9 real bets

Siting — land, permits and the neighbours

The rigs that became data centres Bitcoin miners repurposing power contracts and shells for AI compute. They already hold the one thing that is scarce: interconnection. 11 names ≈1.5 real bets Where it physically goes The only constraint still setting records in the last complete quarter: local opposition doubled in 2026Q2. 9 names ≈2.2 real bets Who pours the concrete 5 names ≈1.5 real bets

Capital — who pays, and who lends

Who is actually paying for it The demand side. Their capex guidance is every other basket's revenue. 8 names ≈2.2 real bets Renting the compute out again The financing-loop ring — customers, landlords and creditors of each other. 5 names ≈1.6 real bets