NVIDIA GTC 2025: Jensen Huang Unveils Blackwell Ultra and the Vera Rubin Roadmap

NVIDIA CEO Jensen Huang unveiled Blackwell Ultra in his GTC 2025 keynote on March 18, 2025 — the next evolution of the Blackwell AI factory platform, carrying 8-stack HBM3e memory, shipping in systems in the second half of 2025 and optimized for AI inference and test-time scaling.

At the same event NVIDIA laid out the next-gen Vera Rubin architecture roadmap, Spectrum co-packaged optics networking, and the NVAQC quantum research center — Huang's full roadmap toward the 'AI factory' era.

From Chips to 'AI Factories': A Narrative Upgrade

Huang's GTC narrative is no longer about selling chips but selling 'AI factories' — packaging GPUs, networking, racks and software into whole-plant solutions that deliver compute. That narrative directly underwrites hundred-billion-dollar compute deals: Stargate's $500B and the first gigawatt of the Nvidia-OpenAI $100B LOI both build on the Vera Rubin platform (see our coverage). The roadmap itself is a sales tool: publishing code names and timelines early lets customers lock in procurement budgets ahead.

Our Take

With AI compute demand still exploding, Blackwell Ultra is seen as key infrastructure behind next-gen frontier training and inference — and a core pillar of NVIDIA's later climb past $5 trillion. But the real thing to watch is the capital loop of 'the chipmaker investing in customers who turn around and buy its chips' (see our Nvidia-OpenAI LOI coverage): as the data-center bottleneck shifts from chip supply to power supply, whether the hardware roadmap delivers on schedule also depends on energy and capital markets aligning.