Nvidia’s Blackwell Push
Jensen Huang framed Nvidia’s Blackwell strategy at GTC as a platform play—projecting roughly $1 trillion in Blackwell/Rubin orders by 2027 and positioning Nvidia as the full‑stack AI orchestrator, not just a chip vendor. The shift emphasizes low‑latency inference, agentic AI workloads that raise CPU and orchestration demand, and broader data‑center power/efficiency changes. (fool.com; forbes.com)
Nvidia unveiled the Vera Rubin platform during its GTC keynote on March 16, 2026, and showcased GB300 NVL72 systems aimed at low‑latency, long‑context inference workloads. (cnbc.com) The company introduced the Groq 3 Language Processing Unit — the first LPU product stemming from Nvidia’s December 2025 asset purchase of Groq — with a target shipping window in Q3 2026. (cnbctv18.com) Nvidia announced Intel Xeon 6 will serve as the host CPU in DGX Rubin NVL8 configurations, a concrete hardware tie that underscores increased CPU responsibilities for Rubin-class inference systems. (digitimes.com) Cloud providers Microsoft, CoreWeave and Oracle Cloud Infrastructure were cited as early deployers of GB300 NVL72 racks for low‑latency agentic use cases such as coding assistants and long‑context inference. (blogs.nvidia.com) Nvidia and third‑party SemiAnalysis benchmarks released at GTC claimed up to 50× performance improvements and roughly 35× lower costs for agentic‑AI workloads on Blackwell Ultra compared with older inference setups. (blogs.nvidia.com) Analysts noted Nvidia had previously estimated roughly a $500 billion opportunity from Blackwell and Rubin through 2026, and said that specific order‑backlog and delivery schedule details are expected to surface on Nvidia’s next earnings call. (cnbc.com) Nvidia also released NemoClaw, an open‑source agent orchestration stack at GTC intended to standardize multi‑agent workflows and create recurring demand for inference hardware and orchestration software. (forbes.com)
