Data centers & compute

The data center has become the factory of the intelligence economy. Every model trained and every query answered runs through physical racks of accelerators, networking, and cooling, and hyperscalers are committing hundreds of billions of dollars a year to build more of them. Capital expenditure at this scale has few precedents outside wartime industrial mobilization, and it is being funded from some of the strongest cash flow machines ever assembled.

Some history grounds the scale: the cloud buildout of the 2010s was the largest infrastructure program of its era, and current AI capex already runs at a multiple of it. Each prior compute wave: mainframe, PC, mobile, cloud. Ended up bigger than forecasts, because falling unit costs kept unlocking new demand.

Demand continues to outrun delivery. Training clusters grow with each model generation, but the more durable story is inference, the everyday serving of AI to billions of users and, increasingly, to autonomous agents that run continuously. Inference converts AI from a research expense into an operating utility, and utilities need constant capacity.

The technology stack is deeper than the chip. High bandwidth memory, advanced packaging, optical interconnects, liquid cooling, power distribution units, switchgear, and backup generation all sit between an accelerator and a working data center. Several of these layers are supplied by a handful of companies running at full capacity with multi year lead times.

The economics show up in backlogs before they show up in earnings. Electrical equipment makers, cooling specialists, and data center construction firms report order books stretching years forward; capacity is frequently pre leased before ground breaks. When a customer pays today for delivery in two years, pricing power follows.

The obvious beneficiaries: chip designers, hyperscalers. Are well covered and priced accordingly. The more interesting layer is the enablers: transformer and switchgear manufacturers, thermal management suppliers, specialty contractors, modular builders, and the distributors who route scarce components to site.

Follow one dollar of hyperscaler capex and it fans out across a long chain: land and shell construction, generators and switchgear: cooling plant, racks and cabling, networking, and finally silicon. Each layer has its own leaders, lead times, and pricing dynamics, and each is investable separately.

Many of these enablers are small and mid caps classified under industrial or construction codes, invisible to investors screening for 'AI.' Their revenue mix shifts toward data centers quarter by quarter, and the market tends to reprice them only after several beats, the window where under covered research earns its return.

Competitively, the layer matters more than the logo: chip design concentrates profits in a few names, while construction and electrical layers remain fragmented enough for regional specialists to win outsized share. Neocloud providers renting GPU capacity add a new customer class bidding for the same equipment.

The risks are real. A pause in hyperscaler capex would echo down the chain fast; power constraints can delay projects independent of demand; and any efficiency breakthrough that slashes compute per query would be deflationary for capacity, though history suggests cheaper compute expands usage rather than shrinking spend.

What to watch: hyperscaler capex guidance each earnings season, lead times on transformers and turbines, data center vacancy and pre leasing rates, and backlog growth at the enabler layer. The research tracks these signals across the coverage universe and flags where the buildout is landing next.