The headline of 2027's hardware year arrives in the second half: NVIDIA Rubin Ultra. It places four reticle-limited GPU chiplets into a single socket, delivering roughly 100 petaflops of FP4 compute and 1TB of HBM4E stacked memory.
At rack scale, the VR300 NVL576 targets about 21× the performance of today's GB200 NVL72 — around 15 exaflops of FP4 for AI inference and 5 exaflops for training. It follows the standard Rubin parts launching in 2026 and sets up Feynman for 2028.
Why it matters: Rubin Ultra is the engine room of the agentic-AI era — see how it ties into the AI compute race.