For ML infrastructure leads and training teams

    Build around
    the clock.

    Training-data storage at $4.99/TB flat. No egress on dataset reads. fsspec, PyArrow, and HuggingFace datasets work natively — change the endpoint, keep the code.

    Start for freeTalk to an expert

    No credit card required · No egress fees · Connects in minutes

    Fast storage or affordable storage. Training teams have been told to pick one.

    In-cluster storage (EFS, NFS, proprietary object stores) is fast but expensive. Standard cloud object storage is cheap but slow from the compute side — and charges egress every time a run reads the dataset. Fil One is both flat and S3-compatible.

    The EFS tax

    $307/TB for in-cluster storage.

    AWS EFS costs $0.30/GB — $307/TB — because it is optimised for latency, not cost. A 10 TB training corpus costs $3,072/month. Teams pay that to avoid the alternative: slow, metered S3.

    The egress trap

    Every training run reads the full dataset.

    Using standard S3 instead of EFS saves on storage but adds $0.09/GB egress per read. 20 training runs over a 10 TB dataset costs $18,432 in egress alone that month. The compute bill is not the whole story.

    The iteration limit

    Storage cost constrains how often you can train.

    When each run carries an egress cost, teams gate training iterations. Ablations get skipped. Re-runs get deferred. The research output is shaped by the infrastructure bill.

    10 TB training data. 20 training runs per month.

    Same dataset, three storage options. Monthly cost for storage plus the egress cost of reading the full set 20 times.

    training_data.py
    import fsspec, os
import pyarrow.dataset as ds
from torch.utils.data import DataLoader

# Training data on Fil One — $4.99/TB, $0 egress per run
fs = fsspec.filesystem(
    "s3",
    endpoint_url="https://eu-west-1.s3.fil.one",
    key=os.environ["FIL_ACCESS_KEY"],
    secret=os.environ["FIL_SECRET_KEY"],
)

# PyArrow reads Parquet shards directly — same API as S3
dataset = ds.dataset(
    "s3://training-data/imagenet/shards/",
    filesystem=fs,
    format="parquet",
)

# HuggingFace datasets — same endpoint
from datasets import load_dataset
ds_hf = load_dataset(
    "parquet",
    data_files="s3://training-data/instruct/train.parquet",
    storage_options={
        "endpoint_url": "https://eu-west-1.s3.fil.one",
        "key": os.environ["FIL_ACCESS_KEY"],
        "secret": os.environ["FIL_SECRET_KEY"],
    },
)
    StorageRateEgressTotal/mo
    AWS EFS (gp)$307/TBN/A$3,072
    AWS S3 + 20 runs$23.55/TB$0.09/GB$18,668
    Fil OneYou$4.99/TB$0$50

    AWS EFS gp bursting us-east-1: $0.30/GB. AWS S3 Standard: $0.023/GB storage + $0.09/GB egress per read. Computed: EFS 10,240 GB × $0.30 = $3,072; S3 storage $235.52 + 20 runs × 10,240 GB × $0.09 = $18,432 egress. Fil One: 10 TB × $4.99 = $49.90, egress $0. Q2 2026 public rate cards.

    Training cost that scales with dataset size, not run count.

    62× cheaper than EFS

    AWS EFS costs $0.30/GB ($307/TB). Fil One costs $4.99/TB. At 10 TB of training data, that is $3,072/month vs $50 — before a single training run reads a byte.

    No egress on training reads

    Each training run reads the full dataset. On AWS S3, 20 runs a month over 10 TB costs $18,432 in egress alone. On Fil One, every read is included in flat storage.

    fsspec / PyArrow / HuggingFace native

    PyTorch DataLoader, JAX, HuggingFace datasets, and PyArrow all support S3-compatible storage via fsspec. Change the endpoint — nothing else changes.

    Flat cost at any run frequency

    Run training 5 times or 500 times. The storage bill is the TB you keep times $4.99. Run frequency is an engineering decision, not a cost one.

    One rate. $4.99/TB/month.

    Storage. That is the whole bill. Run training 5 times or 500 times — the storage invoice does not change.

    Start for freeTalk to an expert

    No credit card required · No egress fees · Connects in minutes

    Run as many times as the model needs.

    Free 1 TB evaluation. Point fsspec or PyArrow at the endpoint and run the training loop — the egress line will not be there.

    Start for freeTalk to an expert

    No credit card required · No egress fees · Connects in minutes