MarketCatalyst
← Back to Resources
AI Infrastructure

CoreWeave's Vera Rubin Racks Train AI Up to 2.7x Faster in First Cloud MLPerf Test

CoreWeave became the first cloud provider to post multi-rack Nvidia Vera Rubin results in MLPerf Training v6.1, training the 671-billion-parameter DeepSeek-V3 model in 17.65 minutes on 256 Rubin GPUs. That was 2.7x faster than its GB200 systems at the same GPU count, and Llama 3.1 8B ran 1.9x faster.

Published Nov. 16, 2026 · 12:15 p.m. ET · 3 min read·MarketCatalyst

CoreWeave (CRWV) posted the first multi-rack results from a cloud provider on Nvidia's (NVDA) new Vera Rubin platform in the MLPerf Training v6.1 benchmarks, released Monday. Its Vera Rubin systems trained AI models up to 2.7 times faster than its previous-generation GB200 systems at the same GPU count, an early look at how much Nvidia's next chips could speed up real-world AI work.

2.7x
Faster DeepSeek-V3 training than CoreWeave's GB200 result at 256 GPUs
17.65 min
Time to train DeepSeek-V3 671B on 256 Rubin GPUs
1.9x
Faster Llama 3.1 8B training versus GB200 on 72 GPUs

CoreWeave's results

BenchmarkSystemGPUsTimeVs. GB200
DeepSeek-V3 671BVera Rubin NVL72 (4 racks)25617.65 min▲ 2.7x faster
Llama 3.1 8BVera Rubin NVL72727.16 min▲ 1.9x faster
GPT-OSS 20BGB200 NVL727217.18 min–
Qwen3.5 397B (reinforcement learning)GB200 NVL7225622.78 min–

Why the results matter

  • Multi-rack is the hard part. DeepSeek-V3 is a 671-billion-parameter mixture-of-experts model that constantly shuffles data between GPUs. Spreading the job across four racks tests networking as much as raw chip speed, which makes the 2.7x gain more meaningful.
  • Fastest at its size. CoreWeave's Llama 3.1 8B result was the fastest submission at 72 GPUs.
  • Real production clusters. The tests ran on CoreWeave's production clusters using Nvidia's Spectrum-X Ethernet networking and CoreWeave's Mission Control software, across single- and multi-rack setups.

Nvidia's own numbers

Nvidia said Vera Rubin swept every benchmark in the round. In its own submission, Vera Rubin NVL72 trained DeepSeek-V3 in 4.14 minutes on 1,024 GPUs, about twice as fast as its GB300 systems at the same scale. MLCommons said this round drew 24 submitting organizations and included the first results for Nvidia's Rubin and Google's TPUv7 chips.

Why it matters for CoreWeave. CoreWeave rents out Nvidia GPUs, so showing it can run the newest chips at scale on its own infrastructure is central to winning AI customers. Nvidia owns more than 10% of the company, and CoreWeave shares have been hit in recent sessions by worries about AI spending and Nvidia's leadership turmoil, including a sharp drop early Monday.

What to watch

  • Vera Rubin ramp. How quickly CoreWeave can bring Vera Rubin capacity online for paying customers.
  • Nvidia earnings. Tuesday's report and any update on Vera Rubin production and demand.
  • Rival clouds. How Microsoft Azure, Oracle, Nebius and others compare as they post their own Rubin results.

This article is informational only and is not investment advice.

Readers also read

MarketCatalyst is not a registered investment advisor and does not manage client assets. Content on this platform is provided for informational and educational purposes only. It is not investment advice, and MarketCatalyst is not a stock-picking or trade-alert service. Trading stocks and options involves risk, including the possible loss of principal. Consider your own goals, time horizon, and risk tolerance, and consult a qualified financial advisor before making any investment decision.