The fact that Blackwell B200 consumes 1,200W and Rubin VR200 consumes 1,800W is not just a story inside the GPU substrate. The liquid cooling system that extracts that heat out of the substrate, the power grid that draws that power into the data center, and the interconnect that converts communication between tens of thousands of GPUs from copper to light — three physical infrastructures must evolve simultaneously for one GPU to come alive.
The AI bottleneck in 2026 is not the GPU, but power. Dominion Energy Virginia announced that the power connection capacity requested by data center companies as of February 2025 reached 40.2GW. Double compared to just 6 months ago (21.4GW in July 2024). The lead time for large transformers has exceeded 128 weeks (2.5 years), and cases where two completed data centers cannot operate due to lack of transformers have already been reported in Silicon Valley.
In terms of communication, NVIDIA will launch the Spectrum-X Photonics switch in the second half of 2026 and begin converting the wiring of the AI factory from copper to silicon photonics (light). Optical component companies such as Lumentum, MACOM, and AAOI will be the first to feel the impact.