The chips cooled and power kept rising
An AI datacenter does not run on chips alone. You can fabricate the chip and mount it on the substrate, but without electricity to run it and cooling to carry off the heat, the rack never switches on. That constraint is stamped on the 2026 returns. NVIDIA, at the head of the chain, cooled from +239% in 2023 and +171.3% in 2024 to +38.9% and +5.1% in 2025-26. Yet Vertiv, in datacenter power and cooling, still rose +101% in 2026, Caterpillar in generation +85.3%, and Quanta Services, building the grid, +70.3%. Where the chip re-rating passed, the binding constraint moved from silicon to electrons. Once the bottleneck that moved from silicon to the package is filled all the way through the substrate, the last constraint sits outside the chain, in the physical world. It is power.
Power runs on a longer clock than semiconductors
Here the chain leaves the semiconductor, and the clock outside it turns more slowly. A semiconductor bottleneck clears by building fabs, roughly two years. Power runs on the clock of plant permitting and grid construction, so its re-rating lasts longer. A GPU is a matter of ordering; power is a matter of construction. That lag stays in the returns.
Even in 2026, as the chips slowed, the power layer went on rising. Vertiv, in datacenter power and cooling, went +251.6% (2023), +136.9% (2024) and another +101% in 2026 (FY2025 revenue $10.23B, backlog $15B). GE Vernova, on gas turbines, rose +99.2% in 2025 and +66.4% in 2026 (~$45B electrification backlog). Quanta Services, building the grid, was +70.3% in 2026 (backlog $44B); Caterpillar, in datacenter generation, +85.3%. The chip re-rating ran about two years; the power one has run longer.
The power bottleneck runs several layers deep
Building one plant does not clear the constraint at a stroke. Across this wiki's power layer it is spread through generation, the rack and fuel. In generation, gas turbines (GE Vernova) and nuclear (Constellation +92.7% in 2024, Vistra +263% in 2024, Talen +214.8% in 2024) run down to small modular reactors (Oklo +238% in 2025, NuScale +445% in 2024) and fuel cells (Bloom +291.2% in 2025). Backup generation is split between Caterpillar and Cummins (+43.5% in 2026).
Inside the rack there is a bottleneck too. NVIDIA shifts to an 800 VDC architecture that ABB (+29.1% in 2026), Delta Electronics (+183.8% in 2025) and Schneider design alongside it. Cooling adds Delta and Modine (+200.6% in 2023) with 1MW-class CDUs beside Vertiv. AI datacenters run 5-10x the heat per rack, so as the chips slowed these layers filled the gap in turn.
The chain grows above and below generation
Generation sits in the middle of the power chain. It stretched both upward and downward. Above, utilities selling electricity directly to datacenters were added (NextEra, AEP). AEP sees +24GW of load growth by 2030 and ordered 1GW of Bloom fuel cells to bridge that load before transmission is built, on-site generation standing in where the grid cannot keep up.
Below, the fuel surfaces: uranium for nuclear (Cameco +78% in 2025) and natural-gas pipelines for gas turbines (Williams +63% in 2024, $5B+ of datacenter generation). The equipment layer that physically connects it all filled in as well: extra-high-voltage transformers from Siemens Energy (+314.2% in 2024) and HD Hyundai Electric (+300.5% in 2024, $6.73B transformer backlog), and datacenter REITs leasing the land (Equinix +43.4% in 2026, Digital Realty +27.1%). From fuel to generation, grid, transformers and real estate, the power chain closes into a single line. The trickle-down seen in the AI cycle's propagation repeats inside the power layer itself.
Several clocks tick inside the layer
That power is the bottleneck did not lift the whole layer in one direction. Constellation jumped +92.7% in 2024, then rested at -23.7% in 2026; the pre-commercial SMRs, where expectation ran ahead, pulled back with Oklo -28.9% and NuScale -28.9% in 2026. Many clocks tick separately even within power. Still, the earnings layers of generation, cooling and equipment carried their re-rating through 2026. Where expectation ran ahead, the stock rested; where physical capacity became earnings, it kept rising.
Verdict
The read that power is the next bottleneck is borne out in the data. The bottleneck moved from silicon to electrons, and power runs on a longer clock than semiconductors. A fab is built in two years, but plants and grids run on the clock of permitting and construction, so the power bottleneck lasts longer. That is why, even in 2026 as the chips slowed, Vertiv (+101%), Caterpillar (+85.3%) and Quanta Services (+70.3%) kept rising.
Power rose because it is a physical capacity bottleneck. Yet this chain holds a layer equally essential that did not rise. There is no chip without design, and yet EDA is essential and still lagged. If essentiality does not summon a re-rating, under what condition does the propagation actually hold?