5 AI Infrastructure Stocks Powering the $2T Buildout

Add up announced hyperscaler capex, sovereign AI programs and enterprise buildouts through 2028 and you get a number around $2 trillion. Capital cycles that large don't reward everything equally — they reward chokepoints: places where supply can't respond quickly to demand. Here are the five we own in the FurlPay Research model portfolio, and why.

NVDA$225.16-0.06%Bullish

Still the systems company, not a chip company — NVLink fabric + CUDA keep switching costs brutal. Rubin cycle ramps into H2.

MU$971.66+2.30%Bullish

HBM sold out into 2027 at software-like margins (see our Q3 review). The cheapest way to own the memory bottleneck.

AVGO$392.99-5.94%Bullish

Every hyperscaler designing custom silicon runs it through Broadcom — XPU + networking gives two ways to win.

TSM$426.35-0.96%Bullish

The only foundry that matters at the leading edge. CoWoS advanced-packaging capacity is the industry's real unit constraint.

ANET$198.82-2.36%Bullish

AI clusters are networks first — 800G Ethernet backlogs and hyperscaler share gains compound quietly.

The power problem is the next trade

Every incremental gigawatt of AI datacenter needs grid interconnects, transformers, switchgear and liquid cooling — markets where lead times run two to four years. Eaton (ETN) and the utilities with nuclear-adjacent capacity (NEE, and the URA uranium complex) are the second-derivative plays most portfolios still underweight. When memory and packaging bottlenecks clear, power becomes the binding constraint.

In a gold rush, sell shovels. In a compute rush, own whatever the shovel factory can't get enough of.

Risk to the thesis: a visible deceleration in hyperscaler capex guidance this earnings season (see our Q2 preview) would compress every multiple in this list simultaneously. Size accordingly — concentration in one theme is a feature until it isn't.