AI teams often store datasets in remote object storage or pull models from Hugging Face Hub, according to Together AI's announcement. The company says every cold start pays a wide-area network tax because cross-cloud and cross-region reads are slow and subject to rate limits. Together AI notes that egress from Amazon Web Services can cost up to $0.09 per gigabyte. When hundreds of nodes pull the same model checkpoint, the company says those repeated transfers waste compute time, raise bills, and risk failed jobs if the origin throttles traffic.1
Together Storage Accelerator runs on every GPU node in a cluster and exposes drop-in endpoints for existing storage workflows, Together AI reports. When any node reads an object for the first time, the service caches it on local cluster NVMe storage. Subsequent reads from any node in the cluster are then served locally instead of crossing the network. According to the post, the system is powered by ASTRA, Together AI's internal caching technology, and works with existing boto3, transformers, and command-line tools while storing credentials in an encrypted secrets vault.1
Together AI reports that a 200-node cluster pulling a 70 GB checkpoint makes only one origin fetch with the cache enabled, rather than 200 separate downloads. The company also says cached data continues serving during upstream outages or rate limits. In tests conducted on October 28, 2026 on a single H100 node, Together AI measured a 2 GiB object download from Cloudflare R2 taking 12.7 to 15.0 seconds without the cache. With a warm cache, the same download took 0.64 to 0.72 seconds. Small 5 KB and 2.4 MB files from the same bucket fell from 320 ms and 680 ms to 5 ms and 8 ms, according to the post.1
The service is available now through a cluster's Settings page, and Together AI says teams can set it up in under ten minutes without migrating data from existing S3 or R2 buckets. It also pairs with T4, Together AI's managed in-data-center object storage, which remains in limited preview through sales. Together AI cautions that its measurements came from a preview deployment running TSA chart 0.6.0 using single-stream reads. Actual performance depends on object sizes, network paths, and regional origin speed, the company notes, so gains are largest when the upstream origin is slowest.1