The week closed with the AI trade looking less like a contest over models and more like a contest over industrial position. Compute demand is still there, probably stronger than ever, but that stopped being the interesting part. What mattered across the tape was the steady migration of value toward the businesses that can secure electricity, move it, cool it, condition it, and feed the materials chain underneath it.
You could feel that shift first in semiconductors, where TSMC posted record quarterly profit of $18.1 billion, up 58% year over year and signaled another step up in capital spending, while hyperscalers kept widening their silicon options through partnerships and custom-design ecosystems at Google, Broadcom, Marvell, and others. But the week did not really belong to chips alone. The deeper message was that the AI buildout is now constrained by everything surrounding the chip: advanced packaging, network fabric, land, generation, backup systems, and the supply agreements that keep the whole machine from stalling.
That is why power kept showing up not as background scenery but as the governing variable. Constellation Energy’s expansion plans and restart efforts sat in the same frame as Vistra securing 20-year power purchase agreements with Amazon and Meta, Centrus advancing uranium enrichment expansion in Ohio, and a Washington policy climate that keeps making nuclear look less speculative and more strategic. By the end of the week, the important question was no longer whether AI needs more power. It was who can lock up dependable baseload power, fuel-cycle access, and backup resilience before the next surge in demand makes those positions even harder to buy.
That power story broadened quickly into the rest of the electrical stack. Transmission, support gear, and power quality stopped reading like minor line items. HVDC infrastructure markets are still climbing, ABB is pushing high-voltage UPS systems aimed directly at AI data-center conversion losses, and next-generation backup power is becoming table stakes for AI facilities. The market is learning, maybe a little late, that once campuses move into the hundreds of megawatts, the premium shifts toward whoever can deliver stable power quality and uptime, not just raw nameplate capacity.
Cooling and facility systems moved in the same direction. Eaton’s $9.5 billion Boyd acquisition and Vertiv’s record $15 billion backlog were telling because they framed liquid cooling and electrical architecture as strategic control points rather than support services. Equinix’s booking strength and fuel-cell capacity strategy made the same point from a different angle: the winners in this phase are increasingly the operators who can orchestrate power, interconnection, financing, and deployment speed in one package.
Networking and interconnects also came closer to the center. Arista doubling its 2026 AI revenue target to $3.25 billion, Credo’s positioning in optical interconnects, and POET’s surge on optical enthusiasm all pointed at the same pressure point. Once cluster density rises far enough, moving data becomes its own power problem. Copper can still do a lot of the work, but optics start looking less like a nice upgrade and more like a necessity. Even cables and connectors began to read like real bottleneck markets rather than commodity afterthoughts.
Then there was the raw-materials layer, which may be where the week’s quieter signal lived. Uranium kept surfacing, not just as a sector bet but as a strategic AI-adjacent input. Copper kept reappearing for the more obvious reason that every grid upgrade, every campus build, every cable run, and every transformer yard eventually resolves into conductors and metal. The copper wire and cable market’s projected expansion from $182 billion in 2025 to $324.7 billion by 2032 gave that story some scale. So did reminders that helium outages can pinch semiconductor manufacturing and that DRAM tightness may persist for years. Not glamorous. Still decisive.
What changed this week was not the existence of AI demand, but the market’s understanding of where the leverage has moved. The easiest expression of the theme used to be owning the obvious compute winners. Now the sharper read is to look for the companies setting the pace of deployment by controlling scarce physical inputs: foundry capacity, enrichment, turbines, cooling systems, optical links, power-conditioning gear, workforce pipelines, and metals exposure. The bottleneck owners are starting to look like the real AI incumbents.
What to watch from here: whether nuclear policy support turns into actual siting and interconnection speed, whether hyperscalers keep signing long-duration power and supplier agreements that lock smaller players out of capacity, whether optics begin taking a larger share of the interconnect burden inside dense clusters, and whether commodity constraints in uranium, copper, helium, and memory start biting hard enough to slow the buildout. If they do, the next leg of the AI trade may belong less to the model layer than to the firms already standing upstream with the keys.