CoreWeave's first Vera Rubin NVL72 racks

AI Infra Summit: NVIDIA announces Amazon Annapurna Labs NVHBM collaboration, d-Matrix NVLink Fusion integration, and tokens-per-megawatt efficiency metric

On September 10, Together AI published a technical account of porting its ThunderKittens GPU kernel library to the Vera Rubin NVL72, the Nvidia architecture whose first production racks Dell delivered to CoreWeave on September 4.

The team rebuilt its NVFP4 GEMM implementation for the new hardware, lifting performance from 42% of roofline efficiency to over 22 PFLOPS, which Together AI described as competitive with cuBLAS and CuTe DSL. The post also details changes in the Vera Rubin ISA and how they were leveraged.

CoreWeave published five engineering lessons from building a non-blocking, multi-plane network fabric for AI workloads.

On September 8, CoreWeave recapped the delivery of the first production Nvidia Vera Rubin NVL72 racks by Dell Technologies, which CoreWeave called 'a historic moment,' and said further hardware-related announcements are forthcoming.

Michael Dell had announced the milestone on social media, with CoreWeave's engineering and data center teams bringing up the units.

CoreWeave announced on September 7 that its Dedicated Inference platform now supports Kimi K3, which CoreWeave described as the largest open-source model in the world at 2.8 trillion parameters with a 1 million token context window.

The Dedicated Inference workflow lets users supply their own model weights, with CoreWeave handling deployment, autoscaling, and routing on Nvidia GPUs.

Dell Technologies delivered what Michael Dell described on social media as the world's first production Nvidia Vera Rubin NVL72 racks to cloud provider CoreWeave on September 4.

CoreWeave confirmed receipt, crediting NVIDIA and Dell Technologies as partners, and said its engineering and data center teams are in the process of bringing up the units.