Infrastructure Is the New Bottleneck

The shortage in artificial intelligence is no longer just chips. On April 8, 2026, TrendForce said Nvidia’s Blackwell line is now expected to make up more than 70% of Nvidia’s high-end graphics processing unit shipments in 2026, while Rubin slipped because memory, networking, power, and liquid-cooling pieces are harder to line up on time. (trendforce.com) That sounds backward until you look at what a modern rack has become. Nvidia’s GB200 NVL72 system packs 72 Blackwell graphics processing units and 36 Grace central processing units into one liquid-cooled rack, so buying the processors is only one part of the job; you also need pipes, pumps, heat exchangers, and a building that can feed it. (nvidia.com) Electricity is the first wall. The United States Department of Energy said in December 2024 that United States data-center load growth had tripled over the prior decade and could double or triple again by 2028, which turns every new artificial-intelligence campus into a fight for substations, transmission lines, and utility approvals. (energy.gov) Utilities are saying the same thing from the other side of the meter. The Electric Power Research Institute’s 2024 utility survey said data centers are creating growing challenges both because there are more of them and because each site is drawing more power for cloud and artificial-intelligence workloads. (epri.com) Once that power arrives, the next problem is heat. Blackwell systems are being sold around direct liquid cooling because air cooling struggles when so much compute is concentrated in one rack, which means older buildings with raised floors and big fans are often the wrong shape for the new hardware. (nvidia.com) (supermicro.com) The network is becoming its own construction project. On March 18, 2025, Nvidia introduced Spectrum-X Photonics and Quantum-X Photonics switches and said they are meant to connect artificial-intelligence factories at the scale of millions of graphics processing units while cutting network energy use with silicon photonics and co-packaged optics. (investor.nvidia.com) Co-packaged optics is a simple idea with a very expensive purpose. Instead of pushing electrical signals across longer copper traces that waste power as heat, companies move the light engines closer to the switch chip, and Broadcom’s Tomahawk 6 launch in June 2025 was built around that exact pitch for larger artificial-intelligence networks. (broadcom.com) This is why land has become strategic again. A useful artificial-intelligence site now needs cheap power, water or other heat-rejection equipment, fiber routes, zoning approval, and enough physical space to expand in phases, so the winning regions look less like software hubs and more like old industrial corridors. (belfercenter.org) (epri.com) Permitting delays can waste more time than chip delays. Harvard’s Belfer Center noted that a July 2024 voltage event in northern Virginia caused 60 data centers to disconnect at once, a reminder that grid operators are now treating these clusters as system-level infrastructure rather than ordinary commercial buildings. (belfercenter.org) The companies know this is turning into an energy-management business. In August 2025, Google signed demand-response agreements with Indiana Michigan Power and the Tennessee Valley Authority so some artificial-intelligence data-center loads can be reduced during peak grid demand instead of running flat out at all hours. (reuters.com) So the race is shifting down the stack. In 2026, the scarce thing is often not the graphics processing unit itself but the fully prepared slot where that unit can actually run: a powered site, a cooled rack, an optical network, and a permit stamped before the hardware arrives. (trendforce.com) (nvidia.com) (energy.gov)