Nvidia's Vera Rubin platform could improve the economics of AI inference — the phase in which trained models generate responses for users — by increasing the output that data centers can extract from limited electrical power. A new report from research firm SemiAnalysis estimates that Rubin can produce more revenue and profit from each gigawatt than the company's current-generation Blackwell platform, a factor that could influence investor attention around NVDA stock as cloud providers continue to expand their AI infrastructure capacity.

The findings arrive at a time when electricity availability has become a central constraint on how much AI capacity operators can bring online.

Rubin Outpaces Blackwell on Per-Gigawatt Revenue

SemiAnalysis' latest AgentX benchmark found that Rubin NVL72 — the rack-scale configuration of the new platform — produced about 39% more annual revenue per gigawatt than the strongest GB300 setup, built on the same current-generation Blackwell platform that serves as the comparison baseline. The report also estimated roughly 42% more modeled profit per gigawatt on key agentic AI workloads, in which models carry out long-context, repeated-turn tasks.

Power efficiency has become a central issue for AI data centers. When systems generate more tokens — the units of text that models process and generate — for each megawatt of electricity consumed, cloud companies can serve more users from the same power supply.

The report also measured Rubin's performance against total cost of ownership, a measure that combines hardware, power, and operating expenses over a system's life. In one test running at 170 tokens per second, Rubin delivered up to 67 times the total throughput per TCO of GB300.

SemiAnalysis linked the gains to Nvidia's combined hardware and networking design. The platform integrates the Rubin GPU, Vera CPU, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-6 into a single system that targets long-context and repeated-turn AI workloads.

Power Limits Shape AI Infrastructure Spending

The focus on power comes as data center operators face constraints on available electricity in several markets. Purchasing additional accelerators alone cannot resolve the problem when facilities lack enough power to run new systems. Higher output from each megawatt could therefore allow operators to support more AI activity within fixed power budgets.

For NVDA stock, investors may watch how widely customers adopt Rubin as infrastructure spending priorities shift toward efficiency and operating costs.

Production Ramp Underway at Major Cloud Providers

Nvidia announced Vera Rubin in March and said partners would begin offering Rubin systems in the second half of 2026. The company positioned the platform as the next step after Blackwell for large-scale AI training and inference.

By July, Nvidia said CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius had Rubin racks running as NVL72 production ramped. The company also said its production network spanned more than 350 factory sites across 30 countries.

Those deployments give cloud operators early access to Rubin systems ahead of the second-half 2026 availability window, as Nvidia works to expand production throughout the year through partner manufacturing across major global markets.

Source: Blockonomi