Storage for AI training: what 50 TB of datasets and 20 TB of egress actually costs
Four AI-storage searches land on our AI-datasets tool page from positions 18 or worse, all asking the same thing: where should training data live. Every figure below is one workload, stated so you can argue with it: 50 TB stored, 20 TB of egress per month, global region, 2 million class-A and 15 million class-B operations per month. That is the tool's own AI-datasets preset, and the same ranking recomputes live in the tool. This article answers "for AI training" only. The general cheapest-storage question, at a different workload, is answered and ranked in full in our generic object storage article, which this page does not repeat. One honest boundary first: the branded queries our own pages already hold at single-digit positions (vendor S3 pricing pages) are not re-answered here.
The cheapest five for the AI workload
| Provider | Monthly total | Storage | Egress | Storage rate | Egress rate | S3 API |
|---|---|---|---|---|---|---|
| Hetzner Storage Box | $111.94 | $94.00 | $2.00 | $0.0032/GB/mo | $0.0001/GB | no |
| MEGA S4 Object Storage | $161.62 | $138.61 | $0.00 | $0.00263/GB/mo | $0.00263/GB | yes |
| iDrive e2 | $300.00 | $300.00 | $0.00 | $0.006/GB/mo | $0.01/GB | yes |
| Telnyx Cloud Storage | $308.60 | $307.00 | $0.00 | $0.00614/GB/mo | free | yes |
| B2 Cloud Storage | $347.50 | $347.50 | $0.00 | $0.00695/GB/mo | $0.01/GB | yes |
The top row is a backup box, not an object store: Hetzner Storage Box has no S3 API, so no training pipeline can read from it directly. It is the cheapest place to park 50 TB at $111.94 a month, and the worst place to train from. Excluding it, the cheapest S3-compatible archive for the workload is MEGA S4 Object Storage at $161.62 a month.
What the cheapest five misses about training
Training reads the same bytes thousands of times and ships finished weights to whoever wants them, so the egress line matters more than the storage line. Cloudflare R2 ranks 15th on the raw monthly total at $756.15, but its egress line for 20 TB is $0.00: zero egress is what makes it the weight-distribution pick, and the raw ranking cannot show that.
For the archive side, Backblaze B2 sits 5th at $347.50 a month with a $0.00 egress line inside a 3x-stored free allowance, which is why it is the raw-dataset parking pick. Storj lands 10th at $490.00 with a $140.00 egress line, and its decentralized multi-stream reads are the throughput pick for sharded datasets. The ranking above decides the bill; these three picks decide the pipeline, and all three figures come from the same captured prices.
How these numbers stay true
Every figure above is computed from storage-catalog.json, not typed by hand: per-GB rates, base plans, free allowances and API fees all come from the catalogue rows, and the totals are the tool's own cost model at the AI-datasets preset. Totals are normalised to USD at the fixed rate the tool uses; native currencies stay in the catalogue behind each row. A drift test fails the site build when a price moves or the preset changes, so the page is regenerated instead of rotting. The figures on this page describe the catalogue snapshot of 2026-09-01.