Every node sits on a public interface for pulling data and serving traffic, and on a private fabric used for collective operations between GPUs.
New instances accept inbound connections on port 22 only. Everything else is dropped at the edge until you open it, either in the console or through the API. Outbound traffic is unrestricted.
This catches the most common mistake in GPU rental: exposing a Jupyter notebook on port 8888 with no token to the open internet. If you need it reachable, open the port deliberately and put authentication in front of it.
Instances within a region and account can talk over a private /24 with no configuration. Traffic on that network never leaves the facility and is not metered.
Cross-region private links and VPC peering back to your own cloud account are available on enterprise agreements. Ask us about it.
The number that usually matters is not peak throughput but how long it takes to get a dataset onto the box. At 25 Gbps, a 500 GB dataset lands in roughly three minutes if the source can keep up — in practice the bottleneck is almost always the other end, not us.
For repeated runs against the same data, put it on a persistent volume once rather than pulling it each time. The volume attaches in seconds and costs less than the egress you would spend re-downloading.