NVIDIA Vera Rubin Scales to 350 Global Factories
Key insights
- NVIDIA Vera Rubin NVL72 delivers 10x agent throughput over Grace Blackwell, pairing 36 Vera CPUs with 72 Rubin GPUs.
- Spectrum-X Ethernet Photonics enters co-packaged optics production for the first time, enabling million-GPU AI factory deployments.
- AWS, Google Cloud, Azure, and Oracle Cloud are confirmed H2 2026 first deployers alongside CoreWeave, Lambda, Nebius, and Nscale.
Why this matters
The simultaneous H2 2026 deployment commitments from all four hyperscale clouds signals that Vera Rubin will define the infrastructure baseline for next-generation agentic AI workloads. The 10x throughput jump over Grace Blackwell compresses the upgrade cycle for AI practitioners, meaning teams that built capacity plans around Blackwell-era cost and latency assumptions need to revise those models before H2 2026 procurement cycles close. At 350+ factories already in active production, NVIDIA has decoupled hardware availability risk from the announcement cycle, shifting the real strategic bottleneck to deployment engineering and software optimization rather than supply.
Summary
NVIDIA's Vera Rubin platform has reached full production across 350+ factories in 30 countries, with 150 Taiwan partners already building at scale.
The NVL72 pairs 36 Vera CPUs with 72 Rubin GPUs for 10x agent throughput over Grace Blackwell. New Spectrum-X Ethernet Photonics in co-packaged optics supports million-GPU configurations for the first time in production.
Essentially: (AWS, Google Cloud, Azure, Oracle Cloud) are confirmed H2 2026 first deployers, with CoreWeave, Lambda, Nebius, and Nscale following in the initial wave.
- 10x agent throughput over Grace Blackwell reshapes how hyperscalers budget agentic AI capacity.
- Co-packaged Spectrum-X Ethernet Photonics enters production, making million-GPU interconnect viable at data center scale.
- 350+ factories in 30 countries are manufacturing now, shifting supply chain risk from ramp to allocation.
All four major cloud providers landing in the same hardware generation simultaneously means competitive differentiation shifts to software and deployment speed, not hardware access.
Potential risks and opportunities
Risks
- CoreWeave, Lambda, Nebius, and Nscale face margin pressure in Q3/Q4 2026 if AWS, Google, and Azure use higher-throughput Vera Rubin capacity to undercut GPU-as-a-service spot pricing.
- A manufacturing base concentrated in Taiwan across 150 partners creates a single-geography disruption point: any supply shock there hits the full H2 2026 deployment schedule for all named cloud deployers simultaneously.
- Enterprise customers who locked CapEx allocations around Grace Blackwell systems in 2025 face stranded-capacity risk if Vera Rubin deployments drive inference cost-per-token down faster than those plans modeled.
Opportunities
- Photonics and networking component vendors supplying Spectrum-X co-packaged optics gain leverage as a constrained production input for every million-GPU NVIDIA deployment going forward.
- Cloud cost optimization platforms (ProsperOps, Zesty, Spot.io) have a lead-time window to build Vera Rubin-specific pricing arbitrage tooling before H2 2026 capacity hits the spot market and creates new volatility patterns.
- AI inference companies (Together AI, Fireworks AI, Groq) have a narrow window before hyperscale Vera Rubin capacity arrives to lock in enterprise inference contracts at current GPU pricing before the cost floor drops.
What we don't know yet
- Pricing and allocation terms for Vera Rubin NVL72 units have not been disclosed, including whether hyperscalers locked rates before the production ramp announcement.
- NVIDIA has not specified the performance breakdown between Vera CPUs and Rubin GPUs for mixed agentic versus training workloads, which matters for H2 2026 capacity planning.
- Timeline for when Spectrum-X Ethernet Photonics co-packaged optics will be available to non-hyperscale buyers beyond the eight named initial deployers remains unaddressed.
Originally reported by nvidianews.nvidia.com
Read the original article →Original headline: NVIDIA Vera Rubin Enters Full Production Across 350+ Factories in 30 Countries — AWS, Google Cloud, Azure Among First H2 2026 Deployers