Vera Rubin at GTC — Full-Stack AI Factory for Long-Running Agents
Jensen Huang’s GTC 2026 keynote (March 16): Vera Rubin platform — seven co-designed chips (Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6, Groq 3 LPUs) in NVL72 and CPU racks for agentic inference. Claims 10× perf/watt vs prior; $1T visible POs for Blackwell+Vera Rubin through 2027 — NVIDIA demand signal, not booked GAAP.
NVIDIA’s story is no longer “buy more GPUs.” At GTC the proprietary frame is rack-scale AI factories co-designed for persistent agents — prefill on Rubin, decode on Groq LPUs — with a trillion-dollar purchase-order visibility claim attached.
NVIDIA (March 16, Jensen Huang GTC keynote) unveiled Vera Rubin: full-stack of seven specialized chips — Vera CPU (sequential reasoning), Rubin GPU (high-throughput prefill), NVLink 6 Switch (260 TB/s all-to-all scale-up), ConnectX-9, BlueField-4, Spectrum-6 Ethernet, integrated Groq 3 LPUs. Rack offerings: Vera Rubin NVL72 GPU racks and dedicated Vera CPU racks. Company targets 10× better performance per watt in key workloads vs prior generations; architecture for full agentic inference lifecycle. Reference designs: Vera Rubin DSX AI Factory and Omniverse blueprints. Huang: $1 trillion in visible purchase orders for Blackwell and Vera Rubin systems through 2027, doubling prior estimates. Partners (hyperscalers, OEMs, labs) preparing 2H 2026 deployments.
Claims vs checks
Chip list and platform branding are NVIDIA primary (newsroom + keynote); CNBC/Tom’s Hardware corroborate same-day. Perf/watt “10×” and token-rate claims are vendor. “$1T visible POs” is demand-signal language — not the same as recognized revenue.
Limits
- H2 2026 partner readiness ≠ customer GA everywhere.
- Groq LPU integration details and supply are NVIDIA-framed.
- Visible POs can slip, cancel, or double-count across generations.
Sources
- NVIDIA: “NVIDIA Vera Rubin Opens Agentic AI Frontier” (March 16, 2026).
- CNBC: “Nvidia GTC 2026: CEO Jensen Huang keynote” (March 16, 2026).
- Tom’s Hardware live blog and GTC keynote coverage.