None of the AI revolution happens without silicon. The back half of 2027 brings NVIDIA's Rubin Ultra, which packs four reticle-limited GPU chiplets into one socket for around 100 petaflops of FP4 performance and 1TB of HBM4E memory.
At rack scale, the VR300 NVL576 is targeted at roughly 21× the performance of today's GB200 NVL72 — about 15 exaflops of FP4 for inference. Pair that with the new Vera CPU and you have the engine room for the agentic era.
For the full hardware picture, see our deep dive on Rubin Ultra in the PC & Tech hub.