CoreWeave launches global backbone network for AI workloads

A network built for AI traffic

CoreWeave is expanding its cloud infrastructure by launching a global network backbone specifically designed for AI workloads. While standard cloud networks are built to handle millions of small, predictable flows for web apps and databases, AI requires a different approach. Training and inference pipelines create what the company calls 'elephant flows,' which are high-volume, long-lived data transfers that can last for hours or days.

These flows involve moving petabytes of data across global workflows, such as replicating datasets near available GPUs or distributing model artifacts to different regions. A network optimized for small transactions is often not equipped to sustain this level of continuous, large-scale movement. CoreWeave's new backbone addresses this by focusing on the performance, security, and resilience needed for these specific patterns.

The infrastructure links 51 operational facilities throughout Europe and North America, with eight of those facilities joining the network in 2026. This footprint provides more than 1.5 gigawatts of active power, with approximately 4.2 gigawatts of contracted power. By operating as one global network, the system helps customers optimize data movement and deploy workloads closer to where the demand actually exists.

Performance and security at scale

To ensure high performance, CoreWeave has deployed multiple 400 Gbps connections at every point of presence (PoP) across its unified optical IP network. This Petabit-scale capacity allows customers to stage large datasets and reach available compute without bandwidth becoming a constraint. To put this scale in perspective, the network could simultaneously broadcast a 4k video stream to 66 million viewers.

Security is another core pillar of the design, as AI datasets and model assets represent highly sensitive intellectual property. CoreWeave uses MACsec encryption, which is a hardware-based method for encrypting traffic in-flight at the network interface. By running this in hardware rather than software, the company can protect data without sacrificing the throughput required for massive AI tasks.

The encryption is built into the network from day one, covering backbone links between regions, direct connect traffic from a customer's own network, and traffic inside the data centers. While some other providers offer MACsec on direct links, CoreWeave claims to support it end-to-end over the backbone. This ensures that security does not create a performance bottleneck for organizations moving their most valuable assets.

Redundancy and high availability

The global backbone is engineered to treat infrastructure disruptions as an expected part of daily operations. Each CoreWeave PoP maintains links to at least two other PoPs through four distinct physical routes. This means no two regions depend on a single path between them, which helps prevent a single cable failure from isolating an entire geographic area.

The network includes built-in path intelligence that can automatically redirect traffic if a link, terrestrial fiber route, or subsea cable becomes unavailable. This automated failover happens without requiring customers to intervene or manage their own complex routing configurations. It allows for the confident staging of data near available GPUs, knowing the path is resilient.

This high availability is critical for maintaining the flow of model artifacts and inference traffic. By removing the operational uncertainty of manual data staging and network oversubscription, the network aims to keep accelerators fed. The goal is to eliminate the delays that often occur before AI work can actually begin, ensuring that compute time is used productively.

Connecting the global AI map

While the network is currently operating across North America and Europe, CoreWeave is preparing for significant international expansion. Starting in Q4 2026, new points of presence will come online in the United Arab Emirates, India, Indonesia, Singapore, Australia, South Korea, and Japan. These locations will be connected via a mix of terrestrial and subsea fiber cables.

In Indonesia, the new network point will connect three data center facilities announced in August, bringing local capacity onto the same global network. India will see a significant expansion with 240 MW of data center capacity in Navi Mumbai. These additions will link the Middle East and Asia Pacific regions back to the existing infrastructure in Europe and North America.

The overarching goal is to provide a single AI-native network with global reach, rather than a collection of disconnected regional hubs. By connecting these diverse regions, CoreWeave intends to help customers move data efficiently across the globe. This unified system is designed to convert more GPU capacity into productive work by removing the bottlenecks that typically slow down large-scale AI development.

From backbone to GPU fabric

CoreWeave's design philosophy extends from the global backbone into the internal fabric of the data centers themselves. The network inside the data center is engineered to keep NVIDIA GPUs supplied with data constantly. This architecture aims to deliver 20% higher Model FLOPs Utilization (MFU) and 96% goodput, which is a measure of the actual useful data transferred over a network.

When data reaches a CoreWeave data center, it flows into a non-blocking, multi-plane fabric where every leaf has enough uplink capacity to sustain full line rate. This two-tier architecture links up to 128,000 GPUs using fewer hops, which reduces the congestion and latency that can cause accelerators to wait for data. By minimizing these delays, the system ensures that the hardware can work at its maximum potential.

To further reduce latency, CoreWeave uses bare-metal infrastructure to eliminate hypervisor overhead, which is the software layer that manages virtual machines. They also utilize LOTA in their AI Object Storage to reduce read latency by up to 8x. Together, these components ensure that the journey from data storage to the GPU is as fast and direct as possible, supporting faster results for AI researchers and developers.