From the source
Together AI is launching Together Storage Accelerator, a distributed caching layer for Together GPU Clusters that caches data on cluster NVMe so repeated reads from object storage and
Hugging Face Hub are served locally.
- The company says the problem is that every cold start pays a WAN tax, with cross-cloud reads slow, rate-limited and expensive, and egress from AWS running up to $0.09/GB.
- Measured on an H100 node, Together reports warm reads about 20x faster than a cross-cloud origin at up to 3.4 GB/s per node, with a 2 GiB object that took 13-15 s from
Cloudflare R2 read from cache in 0.6-0.7 s.
- Cold reads cost nothing extra, the first read through TSA running within about 5% of a direct download, and a 70 GB checkpoint pulled by 200 nodes fetches from the source once instead of 200 times.
- TSA needs no application changes: existing boto3 code, transformers calls and CLI workflows work as-is, with credentials encrypted in a secrets vault.
- It is available now on Together GPU Clusters at no additional cost, enabled from the cluster's Settings page, and works with any S3-compatible store including Together's T4 object storage, which is in limited preview.






