Cloud bottlenecks move beyond chips

The AI infrastructure story used to sound simple. Get more GPUs. Build a bigger cluster. Wait for chip supply to catch up. But that framing is getting stale. The harder part now is often everything around the accelerator — the memory attached to it, the copper and optics tying racks together, and the local power system that has to feed the whole thing. Micron, Nvidia, Schneider Electric, Turner & Townsend, and Vertiv have all put out pieces over the last year that point in the same direction: the bottleneck has spread. ### Why isn’t the GPU the whole story anymore? A modern AI server is not just a processor with a power cord. It is a tightly coupled package of GPU, high-bandwidth memory, networking, cooling, and utility capacity. If one piece is short, the whole build slips. That is why memory has become so strategic. In Micron’s fiscal Q1 2026 remarks, the company said it had completed agreements on price and volume for its entire 2026 HBM supply, including HBM4. That is a clean signal that buyers are reserving memory far in advance, not treating it like a commodity add-on. (investors.micron.com) ### Why does HBM matter so much? HBM is the memory that sits right next to the AI processor and feeds it enormous amounts of data at very high speed. For training and large-scale inference, that is not optional. If compute is the engine, HBM is the fuel line. Micron now pegs the HBM market at about $35 billion in 2025 and around $100 billion by 2028 — two years earlier than it had previously expected. That tells you the value pool is moving toward memory, not just logic. (investors.micron.com) ### What changed in networking? As clusters get bigger, the network stops being background plumbing. It becomes part of the compute system. Nvidia’s latest AI networking gear is built around that idea. The Quantum-X800 InfiniBand switch offers 144 ports of 800 Gb/s, and Nvidia’s Spectrum-X photonics line is aimed at scaling AI factories to millions of GPUs with much higher bandwidth density than traditional Ethernet setups. Basically, if thousands of accelerators cannot stay synchronized, expensive compute sits idle. (investors.micron.com) ### Why are copper and optics suddenly strategic? Because AI clusters are denser, hotter, and more latency-sensitive than older cloud builds. Short-reach copper, high-speed transceivers, and co-packaged optics all matter more when every rack is pushing huge east-west traffic. Nvidia is leaning into silicon photonics partly for scale, but also because power and signal integrity become harder as links multiply. The network is no longer a cheap wrapper around the server — it is one of the things defining how big the cluster can get. (nvidia.com) ### Isn’t power the real chokepoint now? In a lot of U.S. markets, yes. Schneider Electric says 7 of 13 NERC regions are set to operate below capacity safety margins through at least 2030, and it notes that data centers are already shifting to secondary markets as traditional hubs run into energy constraints. That means the gating item for a new site may be substation upgrades, interconnection queues, or tariff structures — not whether the building shell can be finished. (nvidia.com) ### What does that do to project economics? It raises the cost of “AI-ready” capacity. Turner & Townsend says 2025 is an inflection point as high-density, liquid-cooled AI facilities overtake traditional cloud assumptions, and its U.S. analysis points to a premium for those higher-density builds. In plain English — a megawatt of AI data center is getting more expensive because the power train, cooling stack, and delivery risk are all tougher. (se.com) ### Who benefits from that shift? The obvious winners are not just GPU vendors. Memory makers gain leverage when supply is booked out. Networking vendors gain leverage when cluster performance depends on fabric quality. Power and cooling vendors gain leverage when “speed to power” becomes the schedule. Vertiv’s April 2026 deal with CPower is a good tell — the pitch is not backup power, but using batteries and on-site energy assets to accelerate interconnection and improve ROI in grid-constrained markets. (turnerandtownsend.com) ### Bottom line AI infrastructure is moving from a chip shortage story to a systems bottleneck story. The scarce thing is now the whole stack — memory, fabric, cooling, and power access. That does not make GPUs less important. But it does mean the next pricing power may sit one layer out from the chip. (investors.micron.com) (vertiv.com)