CoreWeave Reports 10x Token Efficiency on NVIDIA Vera Rubin Platform

NVIDIA has announced that measurements conducted by CoreWeave on its Vera Rubin platform show a tenfold improvement in tokens generated per second per megawatt of power. The announcement, made via NVIDIA's official AI infrastructure account on X on Thursday, August 20, 2026, was tied to a broader transformation in the economics of AI infrastructure. In the post, NVIDIA stated that this performance reflects a significant shift in how the industry builds AI factories, with the focus now moving toward efficiently converting computational power into tokens, and then tokens into revenue. The statement underscored that the metric of tokens per second per megawatt has become a critical new benchmark for AI infrastructure leaders.

Dion Harris, a figure mentioned in NVIDIA's post, linked economics to technology through five critical questions that every AI infrastructure leader should ask. These questions cover computing, networking, storage, software, and security. This suggests that the performance gain does not come from a single component but is the result of extreme co design across the entire technology stack. NVIDIA emphasized that achieving such results requires extreme co design across the full stack, combined with an open ecosystem built for innovation and scale. This approach differs from merely boosting hardware specifications; it integrates every layer, from silicon to software, to achieve maximum efficiency. The Vera Rubin platform itself is the latest generation of NVIDIA's data center architecture, designed for the era of agentic AI reasoning.

According to NVIDIA's official website, the Vera Rubin NVL72 is purpose built to handle AI workloads that require agentic reasoning, a category that is rapidly growing as more complex AI is adopted across industries. The announcement comes amid a global rollout of Vera Rubin. Based on information from NVIDIA's official blog published on July 21, 2026, the platform is supported by 300 global partners and is currently ramping up capacity worldwide. CoreWeave is listed as one of the partners deploying Vera Rubin, alongside Google Cloud, Microsoft Azure, and Mistral. The NVIDIA blog also asserted that Vera Rubin provides benchmark leadership in performance per watt and the lowest token cost for its partners.

This reinforces the efficiency claims measured by CoreWeave, as lower token costs directly impact the operational economics of AI service providers. For CoreWeave, a GPU specialized cloud infrastructure provider, these measurement results represent a significant competitive differentiator. The company can offer computational capacity with better energy efficiency, which translates into lower operational costs and healthier margins. It also signals to the market that competition in AI infrastructure is no longer solely about the number of GPUs, but also about the efficiency of converting power into valuable output. From a broader industry perspective, the metric of tokens per second per megawatt is becoming increasingly relevant as electricity demand from AI data centers soars.

Data center operators and cloud providers now face dual pressures: meeting exponentially growing computing demand while keeping energy consumption in check. The efficiency measured in this announcement directly addresses that challenge. The five critical questions posed by Dion Harris provide a framework for infrastructure leaders to evaluate their readiness. In computing, the questions revolve around processor and accelerator architecture. In networking, the focus is on interconnect capabilities between nodes. Storage is crucial for the data flow required by AI models, while software and security ensure the entire system runs optimally and is protected. The full co design approach emphasized by NVIDIA shows that efficiency cannot be achieved by optimizing a single component alone.

Vertical integration across hardware, networking systems, and the software stack is a prerequisite for achieving metrics like those reported by CoreWeave. This also explains why NVIDIA is building an open ecosystem, as innovation at the application level requires flexibility that only an accessible platform can provide. The adoption of Vera Rubin by major partners such as Google Cloud, Microsoft Azure, and Mistral indicates that the platform has passed initial validation and entered the commercial deployment phase. With 300 global partners supporting it, the ecosystem around Vera Rubin is growing rapidly, creating network effects that accelerate innovation and lower adoption costs for end users. Going forward, efficiency metrics like the one announced are likely to become the new standard for comparison in the industry.

Analysts and market participants will increasingly focus on a platform's ability to generate tokens per watt, rather than just raw compute capacity. This aligns with the long term trend toward more sustainable and economical computing. For companies building or operating AI factories, the questions posed by Dion Harris serve as a practical guide for infrastructure audits. A thorough evaluation of those five areas can reveal hidden inefficiencies and open up optimization opportunities that directly impact profitability. NVIDIA's announcement also reinforces its position as a leader in defining the direction of AI infrastructure technology. By continuously pushing the boundaries of efficiency, NVIDIA is not just selling hardware; it is setting standards that competitors must strive to meet.

This strengthens the company's bargaining position amid increasingly fierce competition in the AI accelerator market. The next phase of development will depend on how quickly Vera Rubin adoption expands among service providers and enterprise companies. With the support of 300 partners and strong measurement results from CoreWeave, this momentum appears likely to continue, driving new standards in the global AI computing economy.