.png)
Legacy hyperscalers have a major problem with networking services.
Today’s GenAI clusters require performant, non-blocking interconnections between GPUs to enable effective communication during training. However, due to their general-purpose nature, traditional ethernet-based designs do not provide the network throughput and latency needed to get the most out of their GPU clusters.
Choosing the right networking services isn’t a matter of luxury—it’s a matter of necessity. Getting higher performance out of GPU clusters is a critical key to enabling effective training time, which in turn impacts cost-to-train. AI enterprises should work with a hyperscaler with purpose-built network architecture to not only support AI model training with high throughput and low latency, but also help improve price-to-performance ratios while doing it.
At CoreWeave, we’ve built out our network services to meet the needs of AI from the start. Here’s how we’ve created a better way to network for generative AI.
Better optimized GPU cluster networking
NVIDIA developed the Quantum-2 InfiniBand Platform as a powerful yet complex solution—and we knew we couldn’t build our infrastructure without it. That’s why we designed an expansive InfiniBand network connecting our GPU clusters with ultra-low latency. Our InfiniBand network offers highly performant and reliable interconnect between GPUs, specialized for generative AI model training.
With InfiniBand, we support the ability to interconnect and create GPU superclusters. One-to-one non-blocking architecture helps AI enterprises get faster data exchanges between GPUs, reducing training time (and, therefore, cost to train). Now, teams can leverage the scale of GPU megaclusters while also maintaining enhanced performance, helping them accomplish more in less time.
This design also helps to enable better GPU cluster networking and faster east-west connections. Our SHARP-enabled network can halve the number of traversals, reduce latency, and double throughput for collective operations. Additionally, SHIELD, which is baked into InfiniBand, helps to enable up to 5000x faster link failure recovery.
Running on InfiniBand cables also helps enable cutting-edge GPU performance and high-speed connectivity. With bidirectional throughput, your teams can reach up to 3.2Tbps on 8NDR.
In a market that moves as quickly as AI, that 15% can make a significant difference between the next groundbreaking model and one that gets left in the dust.
Direct connection for multi-cloud needs
GenAI models need a lot of data to train, experiment, and deploy. When you’re working at an AI enterprise, that data might live in many places. GenAI teams need to trust that data—whether it’s at another cloud or on-prem—moves via a highly performant, highly secure networking service.
We offer direct connections to our data centers so your organization can get the private and performant link it needs to transfer the massive amounts of data needed for training. With no data transfer or egress fees, your teams can train their foundational models with cost-effective connections to other clouds or environments where data might live.
At CoreWeave, we’re aggressively expanding our number of physical locations and working to make them more accessible to our customers. That’s why we offer network POPs throughout a wide geographical range—increasing access and leveraging existing network providers (like Equinix and Megaport). Whether via a dedicated connection or another carrier, our network offers the flexibility AI enterprises need to work with their existing infrastructure.
Plus, our backbone and edge team has established a number of Private Network Interconnects (PNIs) with the hyperscalers. Internet traffic between our platform and major hyperscalers will enjoy heightened performance—enhancing our customers’ multi-cloud experience.
Modern networking infrastructure for cloud resource management
While Ethernet cannot meet the demands for cluster performance required by major AI enterprises, it can support AI projects in other use cases. At CoreWeave, we’re giving Ethernet networking a major upgrade to help ensure clients have predictable and consistent performance between GPU nodes and storage clusters—and enable a higher-performing and better-managed overall cloud infrastructure.
NVIDIA Spectrum switches, purpose-built for generative AI use cases, assist in unlocking ultra-fast port speeds and providing reliable connectivity. Plus, NVIDIA BlueField DPUs help offload networking so more compute capacity can focus on AI workloads.
We’re working with some of our closest partners and clients to deliver a modern Ethernet design that supports enhanced performance. With a new level of specialization enabling east-west traffic, AI enterprises can experience the benefits of predictable and consistent performance.
At CoreWeave, we run InfiniBand with millions of links at a massive scale. Our teams are primed with experience and technical expertise to resolve issues quickly—so much so that 99.9975 of our InfiniBand links at any given point in time are ready to rock to 95%+ goodput.
That means accelerated access to data, leading to faster training and faster time to market.
Better networking services for GenAI
Generative AI demands specialized networking architecture that traditional hyperscalers simply aren’t equipped to deliver.
At CoreWeave, we had a vision for a better networking infrastructure—and then we built it.
This new way to network delivers the low-latency, non-blocking interconnections that AI enterprises need to link GPUs and unlock ultra-fast, ultra-productive, and ultra-reliable model training.
With better performance and scalability, our architecture helps organizations get their models to market ultra-fast, helping reduce costs to boot.
The right networking architecture helps build some of the biggest clusters in the world. Read more about how AI clusters are changing in 2025 and beyond.
