For ML engineers and research teams

    Stop deleting checkpoints
    you'll want back.

    S3-compatible storage at $4.99/TB flat. Keep every checkpoint, eval set, and training artifact without rationing by cost.

    Start for freeTalk to an expert

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

    The bill decides which runs survive.

    A checkpoint is 2–10 GB. A training run has dozens of them. At hyperscaler pricing, keeping the full history of an experiment — plus reading it back for evaluation — costs more than the training compute.

    The deletion decision

    Storage pressure picks which runs survive.

    A checkpoint file runs 2–10 GB. Keep 100 runs with 10 checkpoints each and you are holding 2–10 TB. On AWS S3 that is $46–$230/month in storage alone — before a single eval read.

    The eval bill

    Every eval run is an egress event.

    Loading checkpoints to run a benchmark reads them from storage back to compute. At $0.09/GB egress, reading 5 TB of checkpoints for a single eval pass costs $450. Teams learn to run fewer evals.

    The lost run

    The one you deleted was the one you needed.

    Six weeks later an ablation shows that the deleted run outperformed the saved checkpoint on the metric you did not track. The run is gone. The experiment restarts from zero.

    Same boto3. Checkpoints that cost what they weigh.

    Point your existing checkpointing code at the Fil One endpoint. Eval reads are included in flat storage — no egress line.

    checkpoint.py
    import boto3, os

s3 = boto3.client(
    "s3",
    endpoint_url="https://eu-west-1.s3.fil.one",
    aws_access_key_id=os.environ["FIL_ACCESS_KEY"],
    aws_secret_access_key=os.environ["FIL_SECRET_KEY"],
    region_name="eu-west-1",
)

# Save after each epoch — same call as AWS
s3.upload_file(
    f"checkpoint_epoch_{epoch}.pt",
    "ml-artifacts",
    f"runs/{run_id}/epoch_{epoch}.pt",
)

# Load for evaluation — no egress charge
obj = s3.get_object(
    Bucket="ml-artifacts",
    Key=f"runs/{run_id}/epoch_{epoch}.pt",
)
state = torch.load(obj["Body"])
    AWS S3 Standard$230 storage + $450 egress$680/mo
    Fil One10 TB × $4.99 — egress $0$49.90/mo

    AWS S3 Standard us-east-1 Q2 2026: $0.023/GB storage, $0.09/GB egress. Computed from stated inputs — 10,240 GB × $0.023 = $235.52 storage; 5,120 GB × $0.09 = $460.80 egress. Fil One: 10 TB × $4.99 = $49.90, egress $0.

    Storage that scales with experiment count, not bill.

    The only change is the endpoint. The decision of which checkpoints to keep stops being a cost decision.

    No egress on eval runs

    Load any checkpoint, any number of times. Evaluation loops do not add to the bill. Run as many evals as your team needs.

    Keep every run

    10 TB of checkpoints costs $49.90/month. 100 TB costs $499. The rate stays flat. Deleting early runs to save money stops being a decision.

    Predictable cost

    One flat rate per TB. No request fees, no retrieval tiers. The storage line is storage volume times $4.99.

    S3-compatible

    boto3, HuggingFace Hub, PyTorch Lightning checkpointing — any S3-compatible tool connects with an endpoint swap and no SDK changes.

    One rate. $4.99/TB/month.

    Storage. That is the whole bill. No egress fees, no request charges, no retrieval tier. Keep every checkpoint, load it as often as you need to.

    Start for freeTalk to an expert

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

    Keep every checkpoint.

    Free 1 TB evaluation. Point your existing checkpoint code at the endpoint and stop rationing runs.

    Start for freeTalk to an expert

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