The phrase "AI capex" has been used so frequently in the past two years that it has lost shape. To analyze it properly, we need to separate three things: who is spending, what they are spending on, and what is actually constraining further spending.

Who is spending

The vast majority of AI capital expenditure is concentrated in a handful of hyperscalers — Microsoft, Google, Meta, and Amazon — plus a small but growing number of sovereign cloud projects and pure-play AI labs backed by private capital. Together, hyperscaler capex grew from roughly $150B in 2023 to over $400B in 2025, and current guidance puts 2026 in the $500-600B range. Add sovereign and private spending and the total infrastructure capex line crosses $1T per year.

This is not a marketing budget. It is permanent industrial infrastructure spending at a scale that rivals the early years of the railroad expansion or the Eisenhower interstate system.

What is being spent on

Three buckets: chips (currently the most-discussed), networking and racks (less discussed but rising fast), and physical infrastructure — buildings, cooling, and power (rapidly becoming the binding constraint).

Chips are still the headline. Nvidia's data-center revenue trajectory has been the most visible expression of this cycle, and the supply-demand imbalance in advanced packaging at TSMC remains the gating factor for training-cluster build-outs. But within 12-18 months, multiple sources of supply — TSMC capacity expansion, Samsung's HBM ramp, Broadcom and AMD competitive offerings, and custom silicon from the hyperscalers themselves — will close that gap.

When the chip constraint relaxes, the next constraint becomes visible. And that constraint is electrical power.

The power bottleneck

The data center industry has historically consumed about 1.5-2% of global electricity. Current trajectories put it at 6-9% by 2030, almost entirely driven by AI training and inference workloads.

In the United States, the picture is more acute. Multiple grid operators — ERCOT, PJM, MISO — have published projections showing data center load growing 8-12% annually through 2030, after two decades of essentially zero load growth. The grid infrastructure required to deliver that power simply does not exist today and cannot be built on the same timeline as the data centers themselves. Transmission projects take 7-10 years. Generation projects take 4-6. AI clusters can be brought online in 12-18 months.

This timeline mismatch is the most important industrial macro story unfolding right now, and it sits at the intersection of utilities, commodities, and industrials.

Implications by sector

Utilities

After decades of stagnant rate base growth, US regulated utilities are entering a multi-decade capex supercycle. Companies with the right geographic footprint — Virginia, Arizona, Texas, the Carolinas — are signing long-dated power agreements with hyperscalers at rates that materially expand returns. We favor names with high data center exposure relative to overall service territory: Dominion, Constellation, Vistra, and NextEra are the cleanest expressions.

Natural gas

Renewables alone cannot fill the load growth, particularly because AI workloads require firm capacity 24/7. The bridge fuel for the next decade is natural gas. US natural gas demand from power generation is projected to grow 15-25% over the next five years, even before accounting for additional LNG export capacity. We are overweight the upstream gas-weighted producers and the midstream infrastructure that serves them.

Grid equipment and electrical components

Behind the power story is an industrial-scale build-out of transformers, switchgear, cabling, and HVDC equipment. Lead times for large power transformers are now 3-4 years and pricing has doubled. The companies positioned in this supply chain — Eaton, Hubbell, ABB, Schneider Electric, and Quanta Services on the construction side — are the most direct industrial beneficiaries.

Software with strong AI integration

Not all software wins from AI capex. The companies that benefit most are those whose products embed AI in ways that materially improve the user experience and create defensible distribution. The losers are commodity SaaS players whose value proposition was workflow automation that AI can now do directly. We are long quality enterprise software with proprietary data assets; short or underweight commoditized SaaS exposed to AI substitution.

The risks

Three things would derail this thesis. First, a cooling of model-training spending if scaling laws hit a hard plateau — possible but currently no evidence. Second, a regulatory backlash that constrains data center siting or imposes power-use ceilings — happening at the margin but unlikely to bite in the US. Third, a recession that compresses hyperscaler capex budgets — plausible cyclically but unlikely to derail the structural trajectory.

The base case remains: AI capex is the largest industrial spending cycle in a generation, the binding constraint has migrated from chips to power, and the equity opportunities follow the constraint. Position for the constraint, not for the headlines.