Request access
Private Preview

Data and networkStorage Private Preview

Block, file and object storage beside your GPUs.

Volumes for boot and data, shared filesystems that every node in a cluster mounts at once, and S3-compatible buckets for datasets and checkpoints. Storage sits in the same region as your compute and lands on the same bill.

Request access Talk to the solutions desk
fx.storage.filesystems.create(size="20TiB")
Opens with the first cohort. S3 access to shared filesystems is in early access.
TAB 01Which storage for which job Source · Fantasti Reviewed 2026-10-09
Which storage for which job. For each job: the storage to use and why.
Job Use Why
Boot and system disks SSD volume Redundant. Survives hardware failures.
Kubernetes node disks Non-redundant SSD volume Fast, for data you can rebuild.
Databases and storage services Mirrored SSD volume High IOPS with three copies.
Datasets during training Shared filesystem, or a bucket Every node reads the same files.
Checkpoints Shared filesystem, then a bucket Fast to write now, simple to keep.
Scratch and caches Local NVMe Fastest. Erased when the node stops.
Weights for autoscaled inference Shared filesystem New replicas mount the same weights.
Results Bucket Reachable by any S3 tool.
§01 Code Preview API
Preview API · subject to change 
fs = fx.storage.filesystems.create(name="datasets", size="20TiB")

cluster = fx.clusters.create(
    name="pretrain-a", gpu="H200:8", nodes=8,
    mounts={"/data": fs},  # same files on every node
)

bucket = fx.storage.buckets.create(
    name="acme-checkpoints",
    storage_class="HIGH_THROUGHPUT",
)
export AWS_ENDPOINT_URL="$FANTASTI_S3_ENDPOINT"  # shown in the console
aws s3 cp ./ckpt/step-9000 s3://acme-checkpoints/run-12/ --recursive
aws s3api put-bucket-lifecycle-configuration \
  --bucket acme-checkpoints \
  --lifecycle-configuration file://to-standard-after-30d.json
curl -X POST https://api.fantasti.ai/v1/storage/filesystems \
  -H "Authorization: Bearer $FANTASTI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "name": "datasets", "size": "20TiB" }'

Response

{ "id": "fs_3k9d…", "status": "creating" }

Your S3 tools already work.

Buckets speak the S3 API. Point the AWS CLI or any S3 SDK at your Fantasti endpoint and keep your scripts.

  • Endpoint$FANTASTI_S3_ENDPOINT, shown in the console
  • Mountmounts={"/data": fs} puts one filesystem on every node

Read the API

§02 Four kinds Schematic

Four kinds of storage

One node, four places for data. Schematic · not to scale.

One 8-GPU node and the four kinds of storage it uses. Local NVMe, 6 × 3.84 TB, is inside the node, is not encrypted and is erased when the node stops. A volume (SSD; SSD, non-redundant; SSD, mirrored) is attached to one machine at a time. A shared filesystem is mounted by many nodes at once, at the same path. A bucket is reached over the S3 API and has three storage classes: STANDARD, TIERED, HIGH_THROUGHPUT. Connections: 8 GPUs to Volume: attach; 8 GPUs to Shared filesystem: mount; 8 GPUs to Bucket: S3 API.

Node8 GPUs
  • HGX B300 8-GPU baseboard, line drawingHGX B300 · 8 GPUs
  • Local NVMeInside the node
    drives
    6 × 3.84 TB
    on stop
    erased
    encrypted
    no
  • VolumeOne machine at a time
    • SSD
    • SSD, non-redundant
    • SSD, mirrored
  • Shared filesystemMany nodes at once
    mounts={"/data": fs}

    Every node reads the same files.

  • BucketS3-compatible
    • STANDARD
    • TIERED
    • HIGH_THROUGHPUT
Node8 GPUs
  • HGX B300 8-GPU baseboard, line drawingHGX B300 · 8 GPUs
  • Local NVMeInside the node
    drives
    6 × 3.84 TB
    on stop
    erased
    encrypted
    no
  • Volume
    • SSD
    • SSD, non-redundant
    • SSD, mirrored
  • Shared filesystem (highlighted)
    mounts={"/data": fs}
  • Bucket
    • STANDARD
    • TIERED
    • HIGH_THROUGHPUT
Node8 GPUs
  • HGX B300 8-GPU baseboard, line drawingHGX B300 · 8 GPUs
  • Local NVMeInside the node
    drives
    6 × 3.84 TB
    on stop
    erased
    encrypted
    no
  • VolumeAttach

    Three SSD types.One machineat a time.

  • Shared filesystemMount
    mounts={"/data": fs}
  • BucketS3 API

    Three storageclasses.

  • Mounted by many nodes at once
  • Attached to one machine
  • Over the S3 API
§03 Volumes Reviewed 2026-10-09

Volumes.

TAB 02Volume types Source · Fantasti Reviewed 2026-10-09
Volume types
Question SSD SSD, non-redundant SSD, mirrored
Use for Boot disks, system data Scratch and rebuildable node disks Databases, storage services
Size 1 GiB to 8 TiB 93 GiB to 262,074 GiB (about 256 TiB), in 93 GiB steps 93 GiB to 262,074 GiB (about 256 TiB), in 93 GiB steps
Throughput, up to 450 MiB/s 2 GiB/s 2 GiB/s
Read IOPS, up to 20,000 75,000 75,000
Write IOPS, up to 20,000 75,000 75,000
Per step 32 GiB: 15 MiB/s, 1,000 IOPS 93 GiB: 110 MiB/s read, 82 MiB/s write, 28,000 read IOPS, 5,600 write IOPS Same as non-redundant
Full performance from 960 GiB 2,325 GiB 2,325 GiB
Redundancy Erasure coding; tolerates two concurrent hardware failures None Three mirrored copies
Encryption at rest Yes Optional Optional
Price, per GiB-month $0.0888 $0.0663 $0.1475
Snapshots and how storage is billed
Sheet
01 / 04
Title
Volumes
Table
TAB 02
Reviewed
2026-10-09
MAX Shared filesystem Reviewed 2026-10-09

100 PiB maximum size of one shared filesystem.

One filesystem · mounted by every node
EQ 01 Source · Fantasti storage limits Reviewed 2026-10-09
§04 Shared filesystems Schematic

One filesystem, every node.

A shared filesystem is file storage that many machines mount at the same time. Every node in a cluster sees the same datasets, code and checkpoints, with no copy step.

It is backed by SSDs, protected by erasure coding that tolerates two concurrent hardware failures, and encrypted at rest.

Throughput grows with size: a larger filesystem serves more data per second in aggregate.

On Kubernetes, mount it as a persistent volume shared between pods.

One filesystem, many nodes. Schematic · not to scale.

4 8-GPU nodes in a row, each joined by one link to the single shared filesystem below them. Every node mounts it at the same path, /data. Connections: Node 1 to Shared filesystem; Node 2 to Shared filesystem; Node 3 to Shared filesystem; Node 4 to Shared filesystem.

  • Node 1
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Node 2
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Node 3
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Node 4
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Shared filesystemMounted by every node
    mounts={"/data": fs}

    Every node reads the same files.

  • Node 1
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Node 2
    HGX H200 8-GPU baseboard, line drawing
    mount
    /data
  • Shared filesystem (highlighted)
    mounts={"/data": fs}

    Every node reads thesame files.

  • One namespace

One shared filesystem

Reviewed 2026-10-09
Size
1 GiB to 100 PiB
Largest file
2 TiB at the default 4 KiB block size
Redundancy
Erasure coding; tolerates two concurrent hardware failures
Encryption at rest
Yes
Price
$0.1000 per GiB-month
§05 Object storage Reviewed 2026-10-09

Object Storage.

TAB 03Object storage classes Source · Fantasti Reviewed 2026-10-09
Object storage classes
Question Standard Tiered High throughput
API value STANDARD TIERED HIGH_THROUGHPUT
Use for General storage, large static datasets Data whose access you cannot predict Streaming datasets and weights to GPUs, checkpoints
Transfer speed Shared limit per account and region Shared limit per account and region Not throttled
Tiering None Warm to Cold after 30 days without access; back to Warm on access None
Availability Selected regions Selected regions All regions
Price, per GiB-month $0.0184 Warm $0.0375 Cold $0.0125 $0.1375
Request and egress charges
TAB 04Object storage limits and S3 compatibility Source · Fantasti Reviewed 2026-10-09

Limits

  • 1 billion objects per bucket
  • 100 million versions per versioned bucket
  • Keys up to 1,024 bytes
  • Bucket names 3 to 63 characters
  • Multipart parts up to 5 GiB
  • Encrypted at rest
  • Encrypted in transit

Supported

  • Object put, get, head, copy, delete and batch delete
  • Multipart upload
  • ListObjectsV2
  • Versioning
  • Lifecycle rules
  • CORS
  • Bucket policies
  • Object tagging
  • Conditional requests
  • Checksum headers

Not supported

  • Object ACLs
  • Object Lock and Legal Hold
  • Replication rules
  • Event notifications
  • S3 Select
  • Static website hosting
  • SSE-KMS and SSE-S3 request headers
Sheet
02 / 04
Title
Object Storage
Tables
TAB 03 · TAB 04
Reviewed
2026-10-09
§06 Rules that move data Schematic

Rules that move data for you.

  1. 01

    Lifecycle rules

    Expire objects, old versions and abandoned multipart uploads, or move objects to another class by date, by age or by days since last access. Rules are evaluated once a day.

    A lifecycle rule moves objects to another class. Schematic · not to scale.

    Objects stored in the class HIGH_THROUGHPUT are moved by a lifecycle rule to the class STANDARD. Connections: Objects to Objects: rule.

    • Objects
      HIGH_THROUGHPUT
    • Objects (highlighted)
      STANDARD
    • Objects
      HIGH_THROUGHPUT
    • Objects (highlighted)
      STANDARD
  2. 02

    Data transfers

    Copy or keep in sync between buckets, between regions, or from another S3-compatible store. Run once or on an interval.

    A data transfer copies and keeps in sync. Schematic · not to scale.

    A data transfer copies from another S3-compatible store into a bucket in region A, and keeps that bucket and a bucket in region B in sync, in both directions. Connections: Any S3 store to Bucket: copy; Bucket to Bucket: sync; Bucket to Bucket.

    • Any S3 store

      S3-compatible

    • Bucket
      region
      A
    • Bucket
      region
      B
    • Any S3 store

      S3-compatible

    • Bucket
      region
      A
    • Bucket
      region
      B
  3. 03

    Mount a bucket

    Mount a bucket as a local filesystem for sequential, high-throughput reads. For locking, renames and shared editing, use a shared filesystem.

    A bucket mounted on a node. Schematic · not to scale.

    A bucket is mounted on a node as a local filesystem, for sequential reads. Connections: Bucket to Node: mount.

    • Bucket
    • NodeLocal filesystem
      HGX H200 8-GPU baseboard, line drawing

      Sequential reads.

    • Bucket
    • NodeLocal filesystem

      Sequential reads.

  4. 04

    S3 on a filesystem

    Early access

    Attach a bucket to a shared filesystem and reach the same files through the S3 API.

    One set of files, two ways in. Schematic · not to scale.

    A bucket attached to a shared filesystem: the same files are reached at the path /data and through the S3 API. Connections: Shared filesystem to Bucket: same files; Bucket to Shared filesystem.

    • Shared filesystem
      /data
    • Bucket (highlighted)
      S3 API
    • Shared filesystem
      /data
    • Bucket (highlighted)
      S3 API
§07 What survives what Reviewed 2026-10-09

What survives what.

TAB 05What survives what Source · Fantasti Reviewed 2026-10-09
What survives what
Question Local NVMeVolumeShared filesystemBucket
Survives a machine stop No Yes Yes Yes
Survives a spot stop No Yes Yes Yes
Shared between machines No No, one at a time Yes Yes
Size Fixed by the node You choose You choose Grows with use
Billed Included in the node price By provisioned size By provisioned size By stored size and use
Sheet
03 / 04
Title
What survives what
Table
TAB 05
Reviewed
2026-10-09
§08 Protection Encryption and access

Protection

At rest
Encrypted
Durable storage is encrypted at rest.
In transit
Encrypted
S3 requests are encrypted in transit.
Access
Bucket policies
Per-object ACLs are not used.
Local NVMe
Not encrypted
Wiped before reuse. Keep sensitive data on durable storage.
Trust and security
§09 Billing Reviewed 2026-10-09
TAB 06Storage prices Source · Fantasti price list Reviewed 2026-10-09
Storage prices. Preview list prices in USD, with what is metered and the unit of each.
Item Billed by Price Unit
Volumes
Volume, SSD Provisioned size Provisioned size $0.0888 per GiB-month per GiB-month
Volume, SSD, non-redundant Provisioned size Provisioned size $0.0663 per GiB-month per GiB-month
Volume, SSD, mirrored Provisioned size Provisioned size $0.1475 per GiB-month per GiB-month
Snapshot Provisioned size Provisioned size $0.0888 per GiB-month per GiB-month
Shared filesystems
Shared filesystem Provisioned size Provisioned size $0.1000 per GiB-month per GiB-month
Object storage
Bucket, Standard Stored size Stored size $0.0184 per GiB-month per GiB-month
Bucket, Tiered, warm tier Stored size Stored size $0.0375 per GiB-month per GiB-month
Bucket, Tiered, cold tier Stored size Stored size $0.0125 per GiB-month per GiB-month
Bucket, High throughput Stored size Stored size $0.1375 per GiB-month per GiB-month
Requests and egress
Standard class requests: PUT, POST, DELETE Requests Requests $4.25 per million requests per million requests
Standard class requests: GET, HEAD Requests Requests $0.35 per million requests per million requests
Standard class egress, to the internet or another region Data transferred Data transferred $0.0188 per GiB per GiB
Tiered class egress, above three times the stored volume Data transferred Data transferred $0.0625 per GiB per GiB
TAB 07Storage lines on a statement Illustrative

Statement · Illustrative

Account
acme-research
Period
One month
Unit
GiB-month · metered by the second
Storage lines on a statement, illustrative. Total for one month $3,040.06.
Cost center Line Resource Billed by Size Rate / GiB-month Amount
cc-4102research Volume, SSD boot-pretrain-aus-east Provisioned size 512 GiB $0.0888 per GiB-month $45.47
cc-4102research Snapshot boot-pretrain-aus-east Provisioned size 512 GiB $0.0888 per GiB-month $45.47
cc-4102research Shared filesystem datasetsus-east Provisioned size 20 TiB $0.1000 per GiB-month $2,048.00
cc-4102research Bucket, High throughput acme-checkpointsus-east Stored size 6.4 TiB $0.1375 per GiB-month $901.12
Storage, one month $3,040.06

Sizes are made up. Rates are the prices in TAB 06. Storage lands on the same bill as your compute.

Sheet
04 / 04
Title
Storage billing
Tables
TAB 06 · TAB 07
Reviewed
2026-10-09
§10 Q&A

Questions about Storage.

Can two machines attach the same volume?

No. A volume attaches to one running machine at a time. Use a shared filesystem when several machines need the same files.

What happens to my data when a spot node stops?

Volumes, filesystems and buckets are kept. Local NVMe is erased.

Which S3 features are missing?

Object ACLs, Object Lock, replication rules, event notifications, S3 Select and static website hosting. Use bucket policies for access and data transfers for replication.

How do I bring data in?

Use any S3 tool against your bucket, or start a data transfer from another S3-compatible store.

Can Kubernetes pods share a filesystem?

Yes. Mount it as a persistent volume shared between pods.
§11 Request Private Preview

Tell us how much data, how fast and for how long.

We will size it with you.

Request access

Next