AI Dataset Streaming

Train on multi-terabyte datasets without staging a single byte.

Store raw datasets, model weights, and checkpoints as encrypted disk images on Usenet. Mount them on a training node and stream samples straight into your data loader — no local copy, no cloud egress, no per-GB retrieval fees.

Workflow

From dataset to data loader in four steps.

01
Package

Pack your dataset into a disk image (.img/.iso) or a folder archive. Small files auto-pack; large shards stay contiguous for sequential reads.

02
Encrypt & upload

AES-256-GCM on your machine, then uploaded to Usenet with SHA-256 integrity and redundant article copies.

03
Mount on the training node

Point your GPU box at the archive URI. It appears as a browsable network share or a mounted block device.

04
Stream into the loader

Your data loader reads samples on demand. Prefetch + read-ahead cache keeps throughput saturated across epochs.

Performance

Built for the throughput training needs.

Line-rate
Saturates a 1–10 Gbps NIC on a well-provisioned Usenet provider.
<100 ms
First-byte latency to a mounted sample with read-ahead warmed.
Zero egress
No per-GB retrieval fees. Your provider bill is a flat monthly rate.
Any size
Single archives to 100+ TB. Stream a shard, an epoch, or the entire set.

Throughput and latency depend on your Usenet provider, network path, and archive layout. Numbers reflect typical results on a 10 Gbps link with a tier-1 provider.

Example use cases

Mounting and extracting disk images for models.

Train LLMs on raw text corpora

Mount a 20 TB tokenized corpus and stream shards into your DataLoader. Swap corpora by pointing at a different archive URI.

Fine-tune vision models on image archives

Ship a disk image full of WebDataset .tar shards. The training node reads them sequentially without unpacking to local disk.

Restore model weights & checkpoints

Stream-write a checkpoint image straight to a GPU node's local NVMe when spinning up a new training run.

Distribute datasets to multiple nodes

One archive URI, N mount points. Each training node pulls what it needs — no central data server to run out of bandwidth.

Archive completed training runs

Snapshot the dataset + weights + logs into a single encrypted archive for reproducibility. Restore months later on demand.

Move datasets between clouds

Upload once from on-prem or a cheap egress region. Mount from any cloud or lab without a second round-trip through object storage.

Put your training data on Usenet.

Encrypted, mountable, and streamable. License from $29.99.