Live Feeds
● TRACKER Updated 13d ago · 38 sources tracked

Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar

Nvidia's Vera Rubin NVL72 platform has demonstrated 67x better performance per dollar for agentic inference according to SemiAnalysis. The system integrates Rubin GPUs with 288 GB HBM4 and Vera CPUs, enabling a single rack to hold 20.7 TB of unified memory. This capacity allows trillion-parameter AI models to run without cross-rack traffic. Production deployment has begun via CoreWeave for customers like Cognition, while software support has expanded through vLLM, SGLang, and Miles to optimize throughput and RL training.

🎙️

Listen to Live Briefing

Real-time synthesized voice briefing · Live Feeds Desk

⏱ ~3 min
Speed:
RSS Source map (41)
⚡ Key Developments & Real-Time Context
Text size:
  • ✓ The Vera Rubin NVL72 provides 20.7 TB of unified HBM4 memory in a single rack.
  • ✓ CoreWeave has deployed the Vera Rubin NVL72 and Vera CPU for production workloads for Cognition.
  • ✓ vLLM provides day-0 support for the Vera Rubin NVL72 via nightly containers.
  • ✓ The Rubin GPU features 288 GB of HBM4 memory.
🛡️ Source Corroboration: 38 independent reporting domains (85% confidence) ⏱ Read time: ~2 min

What changed

SemiAnalysis reported a 67x performance per dollar gain for agentic inference and vLLM announced day-0 support for the platform.

Live updates

  1. Vera Rubin NVL72 achieves 67x performance per dollar in agentic inference

    Nvidia's Vera Rubin NVL72 platform has demonstrated 67x better performance per dollar for agentic inference according to SemiAnalysis. The system integrates Rubin GPUs with 288 GB HBM4 and Vera CPUs, enabling a single rack to hold 20.7 TB of unified memory. This capacity allows trillion-parameter AI models to run without cross-rack traffic. Production deployment has begun via CoreWeave for customers like Cognition, while software support has expanded through vLLM, SGLang, and Miles to optimize throughput and RL training.

    Why it matters

    The NVL72 replaces previous generations like the GB200 to lower the total cost of ownership for AI factories. These facilities require massive capital investment, with single megawatt sites costing approximately $60 million. Efficient rack-scale inference is critical for scaling agentic AI workloads.

    What is confirmed

    • The Vera Rubin NVL72 provides 20.7 TB of unified HBM4 memory in a single rack.
    • CoreWeave has deployed the Vera Rubin NVL72 and Vera CPU for production workloads for Cognition.
    • vLLM provides day-0 support for the Vera Rubin NVL72 via nightly containers.
    • The Rubin GPU features 288 GB of HBM4 memory.

    Still unconfirmed

    • Cognition reports up to 4.8x GB200 token throughput on the new platform.
    • Qwen3.8 Flash Next running on Vera Rubin NVL72 with SGLang reaches 4.23x performance per dollar versus GB300 NVL72.
    • vLLM delivers 7.8x the per-GPU throughput on the platform.

    What to watch next

    • Verification of the 67x performance per dollar claim by independent benchmarks
    • Expansion of production workloads beyond Cognition on CoreWeave
    • Release of official TCO comparisons between GB200 and Vera Rubin NVL72
    Sources used for this update (19)
    1. agenticlibrary.net — Vera Rubin NVL72 Agentic Inference: 67x better Performance ...
    2. vllm.ai — vLLM Support for NVIDIA Vera Rubin NVL72: 7.8x Throughput ...
    3. agenticlibrary.net — Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture ...
    4. www.insidedeeptech.com — NVIDIA Vera Rubin: Full Guide to Rubin GPU, Vera CPU and NVL72
    5. origh.com — Origh Internacional - From Training to Production, NVIDIA and ...
    6. www.metirai.com — CoreWeave Vera Rubin NVL72 Launch: Cognition's 4.8x Claim
    7. github.com — vllm-project.github.io/_posts/2026-10-09-vera-rubin-preview ...
    8. www.lmsys.org — SGLang and Miles on NVIDIA Vera Rubin - lmsys.org
    9. alphasignal.ai — vLLM Runs MiniMax M3 at 7.84x Faster on NVIDIA's Vera Rubin ...
    10. agihunt.info — vLLM ships Vera Rubin NVL72 support with 7.8x… · AGI Hunt
    11. secure.vera.com — Vera | Log In
    12. www.verahealth.ai — Vera Health - Evidence-Based Clinical Answers
    confidence 85%

  2. CoreWeave Deploys Nvidia Vera Rubin NVL72 Rack-Scale System

    CoreWeave has integrated Nvidia's Vera Rubin NVL72 rack-scale AI platform into its cloud services. The deployment starts with production workloads for its customer Cognition. This rollout follows earlier industry debates concerning the system's throughput per total cost of ownership gains, which previous reports placed between 1.4x and 3x at specific token rates, alongside MLPerf Inference v6.1 results showing up to 3.7x performance improvements over the GB300. AI factory operators face high capital costs, with individual megawatt facilities priced at roughly $60 million.

    Why it matters

    Infrastructure economics hinge on system performance, scaling efficiency, and ongoing software optimization because higher token generation directly drives revenue. Capital allocation on a megawatt or gigawatt scale requires strict validation of return on investment across multiple hardware generations. CoreWeave relies on a decade of co-engineering to incorporate Nvidia compute, networking, and software into its cloud infrastructure.

    What is confirmed

    • CoreWeave has made Nvidia's Vera Rubin NVL72 rack-scale AI platform available in its cloud.
    • The Vera Rubin NVL72 deployment begins with production workloads for customer Cognition.

    Still unconfirmed

    • GPT-6 Astra Ultrafast on Nvidia Blackwell GPUs is available now in the OpenAI API and to eligible ChatGPT Work and Codex users.

    What to watch next

    • Broader enterprise adoption metrics for the Vera Rubin NVL72 platform
    • Independent benchmark data verifying the system's throughput and cost efficiency under production loads
    Sources used for this update (4)
    1. blogs.nvidia.com — Inference Archives | NVIDIA Blog
    2. www.storagenewsletter.com — CoreWeave Fully Connected 2026: CoreWeave Delivers Nvidia ...
    3. www.networkworld.com — CoreWeave brings Nvidia Vera Rubin, AI tools to its cloud ...
    4. www.zacks.com — CRWV's Vera Rubin Push: Can AI Infrastructure Fuel Its Growth ...
    confidence 95%

  3. SemiAnalysis disputes Vera Rubin 67x performance claims

    SemiAnalysis reports that the Vera Rubin NVL72 throughput per TCO gain is actually 1.4x to 3x when measured at 60 to 100 tokens per second. This contradicts earlier claims of a 67x improvement over the GB300. CoreWeave has already deployed the system for production workloads for its customer Cognition. While MLPerf Inference v6.1 results showed up to 3.7x performance improvements over the GB300, the discrepancy in TCO metrics suggests significant variance depending on the specific benchmark software and throughput targets used.

    Why it matters

    The Vera Rubin NVL72 is a high-end NVIDIA system designed for agentic AI inference. Accurate throughput per total cost of ownership (TCO) metrics determine the economic viability of scaling AI agents. Discrepancies in these numbers impact how enterprises budget for next-generation hardware.

    What is confirmed

    • CoreWeave has made NVIDIA Vera Rubin NVL72 available for production workloads for customer Cognition.
    • MLPerf Inference v6.1 results show up to 3.7x performance improvements over GB300.

    Still unconfirmed

    • SemiAnalysis produced benchmark results using early pre-release software with NVIDIA assistance.

    What to watch next

    • Confirmation of throughput metrics from additional production users beyond Cognition.
    Sources used for this update (2)
    1. fourweekmba.com — Vera Rubin's 67x Gain Is 1.4x–3x at 60–100 TPS - FourWeekMBA
    2. markets.ft.com — CoreWeave Delivers NVIDIA Vera Rubin NVL72 Performance at ...
    confidence 70%

  4. NVIDIA Vera Rubin NVL72 Boosts Agentic AI Inference Performance

    CoreWeave has made NVIDIA Vera Rubin NVL72 available for production workloads, starting with customer Cognition. The system offers improved performance per dollar for agentic AI inference, with up to 3.7x performance improvements over GB300 in MLPerf Inference v6.1 results. This development extends a nearly decade-long partnership between NVIDIA and CoreWeave.

    Why it matters

    The Vera Rubin NVL72 is part of NVIDIA's effort to enhance AI infrastructure. CoreWeave's deployment of this system in the cloud marks a significant step in making advanced AI capabilities more accessible. The performance improvements are crucial for companies relying on AI for various applications. This collaboration builds on years of productive use of NVIDIA infrastructure.

    What is confirmed

    • CoreWeave has made NVIDIA Vera Rubin NVL72 available for production workloads.
    • The system offers improved performance per dollar for agentic AI inference.
    • Vera Rubin NVL72 demonstrated up to 3.7x performance improvements over GB300 in MLPerf Inference v6.1 results.
    • Cognition is the first customer to run production workloads on the Vera Rubin NVL72.
    • NVIDIA & CoreWeave are now offering Vera Rubin NVL72 & Vera-standalone in the cloud.

    What to watch next

    • Further performance benchmarks for Vera Rubin NVL72
    • Expansion of Vera Rubin NVL72 to other cloud providers
    • Details on Cognition's production workloads and their impact
    Sources used for this update (2)
    1. blogs.nvidia.com — From Training to Production, NVIDIA and CoreWeave Close the ...
    2. wccftech.com — NVIDIA’s Ian Buck Says CoreWeave’s V100 GPUs ... - Wccftech
    confidence 100%

  5. CoreWeave Deploys NVIDIA Vera Rubin NVL72 for AI Workloads

    CoreWeave has made NVIDIA Vera Rubin NVL72 available for production workloads, starting with customer Cognition. The system offers improved performance per dollar for agentic AI inference. Vera Rubin NVL72 demonstrated up to 3.7x performance improvements over GB300 in MLPerf Inference v6.1 results. Cognition is the first customer to run production workloads on the system.

    Why it matters

    The deployment is significant for AI infrastructure scaling, especially for complex workloads. NVIDIA Vera Rubin NVL72 is designed to enhance performance per dollar for agentic AI inference. This follows CoreWeave's plans to scale its next-generation AI infrastructure using multi-rack Rubin clusters. The development aligns with the growing demand for efficient AI computing solutions.

    What is confirmed

    • CoreWeave has made NVIDIA Vera Rubin NVL72 available for production workloads.
    • Cognition is the first customer to run production workloads on Vera Rubin NVL72.
    • Vera Rubin NVL72 demonstrated up to 3.7x performance improvements over GB300 in MLPerf Inference v6.1 results.
    • Vera Rubin NVL72 showed a 67x increase in performance per dollar for agentic inference.
    • Cognition measured up to 4.8× higher SWE-2 inference token throughput per GPU compared with GB200 NVL72.

    Still unconfirmed

    • SemiAnalysis reported a 67x increase in performance per dollar for agentic inference.

    What to watch next

    • Further MLPerf Inference v6.1 results for Vera Rubin NVL72
    • CoreWeave's plans for expanding Vera Rubin NVL72 deployments
    • Comparative performance benchmarks with other AI infrastructure solutions
    Sources used for this update (6)
    1. finance.yahoo.com — CoreWeave Delivers NVIDIA Vera Rubin NVL72 Performance at Production Scale, Starting With Cognition
    2. www.coreweave.com — CoreWeave Delivers NVIDIA Vera Rubin NVL72 at Scale
    3. datacenterrichness.substack.com — CoreWeave: Vera Rubin NVL72 Has Arrived - by Rich Miller
    4. shattered.io — Vera Rubin NVL72 MLPerf Results: 3.7x Over GB300
    5. convergedigest.com — CoreWeave Puts NVIDIA Vera Rubin NVL72 Into Production with ...
    6. www.europesays.com — CoreWeave: Vera Rubin NVL72 Has Arrived - AI Europe
    confidence 90%

  6. Supermicro Ships Vera Rubin NVL72 Racks as CoreWeave Expands AI Infrastructure

    Supermicro has begun shipping NVIDIA Vera Rubin NVL72 racks featuring its Data Center Building Block Solutions and direct liquid cooling stack. This development aligns with CoreWeave scaling its next-generation artificial intelligence infrastructure using multi-rack Rubin clusters and new storage capabilities designed for complex workloads. The deployment follows NVIDIA previewing MLPerf Inference v6.1 results that demonstrated up to 3.7x performance improvements for agentic artificial intelligence workloads, while SemiAnalysis reported a 67x increase in performance per dollar for agentic inference.

    Why it matters

    The Vera Rubin NVL72 system is engineered to optimize tokens per watt for AI factories focusing on agentic workloads. High-density liquid cooling solutions are required to handle the thermal demands of these next-generation multi-rack deployments. Hardware providers and cloud infrastructure partners are rapidly integrating these platforms to support enterprise scaling.

    What is confirmed

    • Super Micro Computer, Inc. is now shipping NVIDIA Vera Rubin NVL72 racks integrated with its Data Center Building Block Solutions and direct liquid cooling stack.
    • CoreWeave is scaling next-generation artificial intelligence infrastructure with multi-rack Rubin clusters and new storage capabilities.

    What to watch next

    • Commercial adoption metrics and benchmark validations for Supermicro's direct liquid cooling racks in production environments.
    • Further announcements from cloud providers regarding deployment timelines for multi-rack Vera Rubin clusters.
    Sources used for this update (2)
    1. finance.yahoo.com — CoreWeave Boosts Agentic AI With Multi-Rack NVIDIA Vera Rubin Clusters
    2. www.techpowerup.com — Supermicro Now Shipping NVIDIA Vera Rubin NVL72 Racks
    confidence 100%

  7. NVIDIA Vera Rubin NVL72 Launches with High Agentic Inference Efficiency

    NVIDIA has introduced the Vera Rubin NVL72 system, targeting agentic AI workloads with significant gains in energy and cost efficiency. On September 16, 2026, NVIDIA published MLPerf Inference v6.1 preview results showing the system delivers up to 3.7x performance improvements. CoreWeave has already deployed multi-rack Vera Rubin NVL72 clusters into its cloud environment to support scale-out agentic AI. The system focuses on optimizing tokens per watt for AI factories, with SemiAnalysis reporting a 67x increase in performance per dollar for agentic inference.

    Why it matters

    The shift toward agentic AI requires chip architectures that optimize for specific inference patterns rather than just raw training power. NVIDIA is integrating these accelerators into its DSX AI Factory platform to maximize energy efficiency. This transition represents a broader transformation in data center design to avoid obsolescence.

    What is confirmed

    • NVIDIA published MLPerf Inference v6.1 submission results for the Vera Rubin NVL72 on September 16, 2026.
    • The Vera Rubin NVL72 system delivered up to 3.7x performance according to MLPerf preview results.
    • CoreWeave deployed multi-rack NVIDIA Vera Rubin NVL72 clusters on its cloud platform on September 16, 2026.
    • The Vera Rubin and DSX platform focus on optimizing tokens per watt for AI factories.

    Still unconfirmed

    • The Vera Rubin AI accelerator delivers 2X profit per GW compared to Blackwell.

    What to watch next

    • Full MLPerf Inference v6.1 final results
    • Additional cloud provider deployments of Rubin NVL72 clusters
    • Official NVIDIA performance per dollar benchmarks for agentic AI
    Sources used for this update (9)
    1. SemiAnalysis — Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar
    2. NVIDIA Blog — AI Infra Summit: NVIDIA Vera Rubin and DSX Platform Advances Showcase Energy Efficiencies of Optimizing Tokens Per Watt for AI Factories
    3. Forbes — The CPU Is Back Thanks To Agents, And What That Means For You
    4. Broadband Breakfast — How Shifts in Computing from AI Are Changing Chip Architectures
    5. Unite.AI — NVIDIA Reports Early Production Results for DSX AI Factory Platform
    6. 36Kr — NVIDIA Leading AI Data Center Major Transformation: Chinese Manufacturers Aim to Avoid Sidelining
    7. Yahoo Finance — NVDA Stock Rises Premarket: Vera Rubin AI Accelerator Delivers 2X Profit Per GW Than Blackwell, Semianalysis Says
    8. www.unite.ai — NVIDIA Vera Rubin NVL72 Posts First MLPerf Inference Preview Results
    9. finance.yahoo.com — CoreWeave Brings Up Multi-Rack NVIDIA Vera Rubin NVL72 Cluster
    confidence 90%
📊

Community Sentiment: How do you assess this situation?

Voice your perspective · Real-time aggregated sentiment from the Live Feeds community