Load balancing
Put inference, render workers or a custom service behind one address. Health checks, failover between nodes, and optional overflow from reserved onto on-demand.
How this actually runs
Bytes should sit next to the GPUs that read them. In-region transfer - instance to instance, instance to object storage, farm to bucket - is included with compute. Uploading a dataset or downloading frames inside a single site does not generate a surprise egress line.
Object storage is S3-compatible in either site at $0.021 per GB-month. Persistent NVMe attaches to a node, survives stop/start, and is sold in 100 GB steps at $0.09 per GB-month. Static IPv4 is $4/month. These are the only storage add-ons on the public card.
We do not move workload data between sites without instruction. Disaster recovery and active/active across sites are products you turn on, not silent replication. Encryption is in transit and at rest. Deletion and export follow the privacy policy: you can ask, we will do it.
Load balancing is L4/L7 in front of inference replicas, custom render workers or any service you run on reserved or bare metal. Basic health-checked distribution is included on those terms. Cross-site failover and custom routing are quoted. TLS on our edge or yours. Prometheus scrape if you already operate that way.
One front door for GPU work
Several model replicas behind one URL. Multi-model endpoints can switch without a new VIP if you containerise them.
The farm already distributes frames; this is for custom engines or interactive preview services you run yourself.
Take a node out for a driver update without taking the service down. Cross-site failover is opt-in and respects residency.