STOCK TITAN

CoreWeave Becomes First Cloud Provider to Deliver Multi-Rack NVIDIA Vera Rubin Results in MLPerf® Training v6.1

(Moderate)

Sentiment and the balance of points

Rhea-AI Sentiment reads the wording of the document, how positive or negative its language is on a 1 to 5 scale. The balance of points shown with the takes weighs what the document actually discloses, so the two can disagree, for example when a trial that missed its main goal is described in upbeat language.

Tags
AI
See more from StockTitan in Google Search and AI answers. Adds StockTitan as a preferred source · opens Google
Add on Google

The only cloud provider to submit multi-rack NVIDIA Vera Rubin NVL72 results, CoreWeave trained DeepSeek-V3 671B 2.7x faster than its submission on NVIDIA GB200 NVL72 at the same GPU count

LIVINGSTON, N.J.--(BUSINESS WIRE)-- CoreWeave, Inc. (Nasdaq: CRWV), The Essential Cloud for AI®, today announced its results in the MLPerf® Training v6.1 benchmark suite, including the first multi-rack results on NVIDIA Vera Rubin NVL72 from any cloud provider. CoreWeave competed in the Closed division’s Available category, which compares systems apples-to-apples on the same reference model, using only hardware that customers can buy or rent today. It submitted results on NVIDIA Vera Rubin NVL72 and NVIDIA GB200 NVL72 systems across four models, from a single 72-GPU rack to 256 GPUs across multiple racks.

As frontier models grow larger and more complex, AI teams face a practical planning question: how much faster will a training run finish on the next generation of compute, and will that gain hold once the job spans multiple racks? A new GPU generation only pays off when the networking, orchestration, and operations around it keep every GPU productive at scale. MLPerf Training measures how long a system takes to reach a specified quality target. CoreWeave’s Training v6.1 submissions show how quickly Vera Rubin performance carries into a production multi-rack environment.

“New silicon is only as fast as the cloud it runs on,” said Chen Goldberg, executive vice president of product and engineering at CoreWeave. “Our MLPerf Training v6.1 results show what NVIDIA Vera Rubin can do when the full stack is built for it, from a single rack to multi-rack scale. We ran these benchmarks on the same production clusters our customers use to train frontier models, with CoreWeave Mission Control keeping the fleet healthy throughout.”

CoreWeave’s Training v6.1 submissions covered four model families. Key results include:

  • Multi-rack Vera Rubin delivered leading results with DeepSeek-V3 671B: CoreWeave trained the 671-billion-parameter mixture-of-experts model to target quality in 17.65 minutes on 256 NVIDIA Rubin GPUs and NVIDIA Spectrum-X Ethernet networking across multiple racks. CoreWeave was the only cloud provider to submit results on this benchmark with NVIDIA Rubin.
  • 2.7x faster than CoreWeave’s GB200 NVL72 submission: At the same 256-GPU count, CoreWeave’s Vera Rubin result trained DeepSeek-V3 2.7x faster than its own GB200 NVL72 submission, a generation-over-generation improvement. ¹
  • Fastest 72-GPU submission on Llama 3.1 8B: On a single NVIDIA Vera Rubin NVL72 rack, CoreWeave reached target quality in 7.16 minutes, the fastest 72-GPU submission. That is 1.9x faster than its 72-GPU GB200 NVL72 submission (13.70 minutes). ¹
  • Strong results across model sizes on GB200 NVL72: CoreWeave trained GPT-OSS 20B in 17.18 minutes on a single 72-GPU rack and completed Qwen3.5 397B reinforcement learning in 22.78 minutes on 256 GPUs across multiple racks using NVIDIA NeMo-RL and NVIDIA TensorRT-LLM.

CoreWeave ran the benchmarks in the same production environment customers use. CoreWeave Mission Control provided fleet health monitoring and node and fleet lifecycle management, and a RoCE-based NVIDIA Spectrum-X Ethernet AI fabric with NVIDIA ConnectX-9 SuperNICs connected the racks.

The MLPerf® training v6.1 results build on CoreWeave’s track record in independent evaluations, including its leading MLPerf® Inference v6.1 results earlier this year and its position as the only AI cloud to earn the top Platinum ranking in SemiAnalysis ClusterMAX™ three times in a row.

Full results are available at www.mlcommons.org.

¹ Speedups are derived by CoreWeave from published MLCommons results, comparing CoreWeave’s Vera Rubin submissions with CoreWeave’s GB200 NVL72 submissions at the same GPU count. Result not verified by MLCommons Association. The MLPerf® name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited.

About CoreWeave
CoreWeave is The Essential Cloud for AI®. Built for pioneers by pioneers, CoreWeave delivers a fully integrated platform of technology and teams that enables innovators to move at the pace of innovation, building and scaling AI with confidence. Trusted by 9 of the 10 leading foundation model providers, CoreWeave serves as a force multiplier by combining superior infrastructure performance with deep technical expertise to accelerate human discovery. Established in 2017, CoreWeave completed its public listing on Nasdaq (CRWV) in March 2025.

press@coreweave.com

Source: CoreWeave, Inc.

Key Terms

mixture-of-experts technical
A mixture-of-experts is a computer model design that uses a group of smaller specialist models, with a controller that picks which specialist(s) handle each task—like sending a question to the right expert on a team. For investors, this matters because it can deliver faster, cheaper, or more accurate AI features without building one giant model, affecting a company’s product performance, costs, competitive edge, and regulatory or safety risks tied to how the system is managed.
reinforcement learning technical
A type of artificial intelligence that learns by trial and error, receiving feedback from its actions to favor choices that lead to better outcomes. Think of it like a salesperson learning which pitches close deals by trying different approaches and keeping the ones that work. For investors, reinforcement learning matters because it can power smarter trading systems, optimize business operations, or improve products—potentially boosting efficiency and profits while also introducing model and execution risks.
roce technical
ROCE (return on capital employed) measures how much profit a company earns for each dollar of capital put into the business, comparing operating profit to the total funds used to run the company. Investors use it like a production efficiency score — similar to measuring how much fruit a tree yields compared with the cost to plant and care for it — to judge how well management turns invested money into returns and to compare companies of different sizes.

Keep reading