NVIDIA stops selling chips and starts selling racks — Vera and the move beyond the GPU

A GPU company building its own CPU is following the bottleneck downstream. The constraint at frontier scale isn't the accelerator — it's everything around it, and NVIDIA intends to own all of it.

NVIDIA detailed its Vera data-center CPU — 250 to 450 watts, up to 1.5 terabytes of memory — pushing into the server-CPU market held by AMD and Intel. Building its own CPU is NVIDIA following the real constraint: at frontier scale the bottleneck is moving data between chips, so co-designing the CPU, GPU, and interconnect as one machine is how you keep the performance curve bending.

The 1.5-terabyte tell

That memory figure names the target workload: inference and long-context serving, where feeding the accelerators matters as much as the accelerators. NVIDIA is building for the workload the market actually runs, not the training benchmark that sells headlines — and owning the whole node lets it optimise for exactly that.

The moat is the grid, and the grid is the ceiling

But the rack NVIDIA is perfecting runs into a wall it can't engineer around. Global data-center electricity demand is doubling past 1,000 TWh, with sites drawing up to 750 megawatts each. NVIDIA can own the CPU, GPU, and interconnect; it cannot conjure the megawatts, and that is the one part of the system that moves in years, not product cycles.

Vera extends NVIDIA's lock from the accelerator to the entire compute node. The only thing it doesn't extend to is the power feeding it — which is why the competitive story of the frontier keeps ending at the substation.

CNBC — Nvidia details its next-generation Vera CPU for AI → · TechPlusTrends — AI data center power requirements 2026 →