CoreWeave plans petabit-scale backbone with new PoP in Australia

GPU cloud operator CoreWeave is growing its Global AI Network into a backbone of petabit scale, tying together AI compute sites in North America, Europe, the Middle East and Asia Pacific. The plan, announced on 27 October, puts new points of presence in seven further markets, on top of the network it already runs in North America and Europe. Those markets are the United Arab Emirates, India, Indonesia, Singapore, Australia, South Korea and Japan.

Construction starts in the fourth quarter of 2026, and CoreWeave aims to close the intercontinental loop by mid-2027.

At each point of presence the design calls for several 400 Gbps links, MACsec encryption done in hardware, and redundant fibre paths over land and under the sea. Every PoP is meant to reach at least two other sites across four physically separate paths, while each region is to be dual-homed to a pair of PoPs. Converge Digest's analysis said the arrangement is meant to lessen reliance on any one transport route, whether terrestrial or subsea. The IEEE 802.1AE standard is intended to protect traffic in transit without the processing burden of software-based encryption.

CoreWeave said the backbone will move big AI datasets, model checkpoints and inference workloads among GPU clusters in different regions. Inside a facility, its networking architecture supports clusters of up to 128,000 GPUs; the global backbone covers traffic between sites. Converge Digest pointed out that the petabit figure refers to total backbone capacity, not the speed of any single 400 Gbps link, and that CoreWeave has not released enough per-route data to work out overall throughput.

In a separate update, CoreWeave said most of the 22.5 MW of IT capacity planned at atNorth's DEN01 data centre in Copenhagen is now online, running NVIDIA Blackwell systems and Quantum-X800 InfiniBand.