NVIDIA
2026-07-21
NVIDIA Vera Rubin Platform Specs Revealed: 10x Tokens per Watt, Monolithic CPU+GPU Design
Summary
Key Takeaways
NVIDIA detailed the Vera Rubin platform in a media briefing, featuring a monolithic design that tightly integrates Vera CPU and Rubin GPU (2 GPUs per 1 CPU). The flagship Vera Rubin NVL72 superchip packs 36 Vera CPUs and 72 Rubin GPUs, using liquid cooling. NVIDIA claims 10x tokens per watt and 3x memory bandwidth over Grace Blackwell. Ian Buck emphasized NVIDIA's continued CPU push, while Hannah Coutand argued that multi-chiplet designs stress memory bandwidth and data movement, whereas monolithic design accelerates data transfer and reduces server installation time from hours to minutes. The Vera CPU will be sold as a standalone product, with shipments to China starting August 2026. First customers include Microsoft, OpenAI, and Oracle, with mass production in H2 2026. Notably, performance data comes from NVIDIA's internal and customer benchmarks (e.g., CoreWeave's 10x token performance) and not from independent third-party tests. Mass production scale, actual deployment volumes, and independent verification of performance claims remain pending.
Why It Matters
Defensive encirclement: By selling Vera CPU standalone and tightly coupling it with Rubin GPU in a monolithic design, NVIDIA aims to encircle AMD and Intel in the AI CPU space, forcing users into full-stack lock-in. Hidden lock-in: Monolithic design prevents separate CPU/GPU upgrades; liquid cooling requirements increase infrastructure dependency. Performance claims of 10x tokens per watt are based on internal benchmarks, likely workload-specific, and lack independent verification. Physical limitations: Large die size risks yield issues; liquid cooling retrofit costs are downplayed; China shipments face export control uncertainties.
PRO Decision
[Vendors] Competitors (AMD, Intel) should exploit the lack of independent verification of Vera Rubin's performance claims by pushing for third-party benchmarks across diverse workloads. AMD should highlight its chiplet architecture (e.g., MI300X) flexibility and ROCm open ecosystem. Intel should promote Xeon+Gaudi and oneAPI, emphasizing open standards to reduce lock-in. Competitors should accelerate development of interconnects rivaling NVLink and secure key customer proofs-of-concept.
[Enterprises] CIOs and architects should perform zero-trust audits: demand independent benchmarks covering multiple AI workloads; assess liquid cooling retrofit costs and cycles; require interoperability guarantees in contracts; adopt multi-vendor strategies and explore OCP open accelerator alternatives to avoid single-platform lock-in.
[Investors] Look beyond PR: focus on actual mass production and customer deployment scale. Performance claims need independent validation; beware of overpromising. NVIDIA's CPU standalone sales expand its addressable market but face CPU market barriers and export control risks. Long-term trend favors integration, but competition may compress margins. Compare roadmaps of AMD and Intel to assess NVIDIA's moat depth and supplier concentration risk.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)