Nvidia Vera Rubin platform unveiled at GTC Taipei, targets $1 trillion orders by 2027
In brief
- Vera Rubin delivers 10x lower cost per token than Blackwell, Nvidia claims
- Production shipments begin autumn 2026
- Vera Rubin supply chain is twice Blackwell's size
- Platform targets agentic AI workloads
- Nvidia partners with 350+ factories across 30 countries
Architecture and Specifications
Vera Rubin production shipments are scheduled for autumn 2026. The platform is explicitly designed for agentic AI workloads—AI systems that don't just respond to prompts but execute multi-step tasks autonomously.
The NVL72 systems bundle together Rubin GPUs, Vera CPUs, Groq 3 LPX inference trays, Spectrum-6 networking, and BlueField-4 storage. This integrated approach reflects what Nvidia calls "extreme co-design," optimizing every component to communicate with neighbors using minimal latency and overhead.
The cost reduction—up to 10x per token—represents a material shift in AI economics. Inference, not training, dominates operational expenses at scale. Cheaper inference multiplies the addressable market for deployed AI agents.
Supply Chain and Manufacturing
The supply chain for Vera Rubin is twice the size of what Blackwell required. Huang emphasized Taiwan's critical role, noting that Nvidia collaborates with over 350 factories across 30 countries, with 150 located in Taiwan. This geographic concentration and scale highlight both the opportunity and the geopolitical fragility embedded in AI infrastructure buildout.
The doubling of supply chain complexity underscores how constrained semiconductor manufacturing remains, even as demand accelerates.
