CoreWeave Plans to Build NVIDIA BlueField-4 DPUs Into Its AI Cloud

What BlueField-4 brings, according to CoreWeave

In a blog post, CoreWeave describes NVIDIA BlueField-4 as a generational jump in DPU performance, with 6x the compute power and 800 Gb/s of throughput. A DPU is a processor that takes over infrastructure duties from the main CPU. CoreWeave presents the new chip as a natural next step for its platform, which it calls the Essential Cloud for AI.

The company says the argument for purpose-built infrastructure rests on the growth of AI models and the complexity of their workloads. According to CoreWeave, traditional cloud architectures are hitting their limits as both increase.

How CoreWeave already uses BlueField

CoreWeave says it has been integrating BlueField DPUs into every one of its servers for years. The DPUs offload networking, security and storage management tasks from CPUs, so that GPUs can concentrate on training and inference. According to the company, BlueField-4 extends this approach and enables faster data access, lower latency and stronger tenant isolation in multi-tenant AI clusters.

Working with NVIDIA, CoreWeave says it has used the NVIDIA DOCA software framework to move core networking and security functions directly onto the DPUs. This allows hardware-enforced tenant isolation, using EVPN VXLAN to create separate, secure virtual networks for each customer. The company claims the result is no software bottlenecks and no noisy neighbors, and it contrasts this with general-purpose, legacy hyperscaler clouds.

Three benefits CoreWeave highlights

CoreWeave lists three areas where it expects BlueField-4 to help customers:

• Performance and scale: the 6x performance boost and 800 Gb/s throughput are meant to strengthen CoreWeave's networking and storage fabrics, with faster model access and lower inference latency for demanding AI workloads.
• Security: with the new BlueField Advanced Secure Trusted Resource Architecture, the company says it will extend its zero-trust infrastructure model with finer-grained, hardware-based isolation.
• Efficiency: by shifting more infrastructure tasks to the DPU, CoreWeave says more CPU and GPU cycles go to customers' compute jobs, which it expects to drive higher utilization and better economics.

What it means for customers

The post is a statement of direction rather than a detailed rollout plan. It gives no availability dates, pricing or benchmark results for BlueField-4 on CoreWeave, so the performance and efficiency benefits remain the company's own claims for now.

For developers, researchers and enterprises running large AI workloads, CoreWeave's pitch is infrastructure that "fades into the background", in its own words, letting them move faster. The company also points to its MLPerf results in both training and inference as evidence of its performance standing, and invites prospective customers to talk to its AI experts.