The world's first! "NVIDIA's favorite child" CoreWeave launches GB200 servers on a large scale

CoreWeave seizes the opportunity again, taking the lead in deploying NVIDIA's GB200 system, as AI giants rush to enter the market.

NVIDIA announced today on its blog that AI cloud computing provider CoreWeave has become one of the first cloud service providers to deploy the NVIDIA GB200 NVL72 system at scale. Cohere, IBM, and Mistral AI are among the first users.

According to the latest MLPerf benchmark tests, these systems offer 2-3 times the performance improvement over the previous generation H100 chips, significantly accelerating large model training and inference capabilities.

CoreWeave CEO Michael Intrator stated that this achievement showcases the company's engineering strength and execution speed, as well as its focus on the next generation of AI development:

“CoreWeave is designed to act faster — we have proven this time and again, always being among the first to operate the most advanced systems at scale.”

NVIDIA's Vice President of Hyperscale and High-Performance Computing (HPC) Ian Buck stated:

“Businesses and organizations around the world are racing to transform inference models into agent-based AI applications, which will change the way people work and entertain themselves.”

“CoreWeave's rapid deployment of the NVIDIA GB200 system provides the AI infrastructure and software that make AI factories a reality.”

Wallstreetcn previously mentioned that in recent years, NVIDIA and CoreWeave have built a "parent-child relationship" through investments, supply, and reverse leasing, leading to the latter's rapid rise. In 2023, NVIDIA committed to paying CoreWeave $1.3 billion over four years to lease back the chips it produces.

Recently, CoreWeave also announced a partnership with OpenAI, further solidifying its leading position as an AI infrastructure provider.

This series of actions indicates that, in the context of a continuous shortage of AI computing resources, cloud service providers that can offer cutting-edge chipsets will still hold a favorable position in market competition.

Performance Breakthrough: New Benchmark Sets Industry Record

Last year, CoreWeave was one of the first suppliers to provide NVIDIA H100 and H200 GPUs and was also one of the earliest companies to demonstrate the NVIDIA GB200 NVL72 system.

CoreWeave recently set a new industry record in AI inference testing using NVIDIA's GB200 Grace Blackwell superchip, a result published in the latest MLPerf v5.0 results.

Notably, without specific optimization for the Blackwell architecture, the GB200 NVL72 server can help Cohere achieve up to 3 times the performance improvement in training a 100 billion parameter model compared to the previous generation NVIDIA Hopper GPU. According to reports, by further optimizing the use of GB200 NVL72's unified memory, FP4 precision, and the NVLink domain of 72 GPUs (with each GPU connected to work collaboratively), Cohere is achieving significantly higher throughput, reducing the generation time of the first computation tokens and subsequent tokens, and providing higher performance and more cost-effective inference