NVIDIA’s Vera Rubin leap
Jensen Huang unveiled the Vera Rubin GPU, promising roughly 3–4x Blackwell AI compute density and a new CG‑HBM memory architecture — NVIDIA says Blackwell+Rubin orders top $1 trillion through 2027 and Q1 revenue hit $78B, up 77% year‑over‑year. The GTC framing shifts NVIDIA from pure silicon to full‑stack AI systems and platform lock‑in, a message investors are already pricing in. (x.com) (youtube.com)
NVIDIA says the Vera Rubin platform now includes seven distinct chips that are in full production to scale so‑called agentic AI factories. (nvidianews.nvidia.com) NVIDIA’s NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs in a liquid‑cooled enclosure and the company and third‑party coverage cite roughly 3.6 exaflops of FP4 inference compute for the rack configuration. (tech.breakingcube.com) Micron confirmed HBM4 designed for Vera Rubin — 36GB, 12‑high stacks — is in high‑volume production with bandwidth above 2.8 TB/s and roughly a 20% power improvement versus prior generations. (markets.businessinsider.com) NVIDIA presented Rubin as a co‑designed systems stack incorporating NVLink‑6, ConnectX‑9 SuperNIC, BlueField‑4 DPU and Spectrum‑6 switches, and the company claims this stack can reduce inference token cost up to 10x and cut GPUs needed to train MoE models by about 4x. (investor.nvidia.com) The Groq 3 LPX language‑processing unit debuted inside NVIDIA’s ecosystem after a December 24, 2025 transaction that transferred Groq assets and key personnel in a deal valued at about $20 billion. (techrepublic.com) Senators including Elizabeth Warren and Richard Blumenthal have publicly questioned whether NVIDIA’s $20 billion Groq arrangement should face antitrust scrutiny. (msn.com) NVIDIA showcased the Rubin rollout at GTC 2026 before an audience reported at more than 30,000 attendees from over 190 countries, a scale the company used to illustrate enterprise demand. (storagereview.com)
