Platform
One platform from first run to production
Train, fine-tune and serve on the same hardware, the same storage and the same bill. Here’s how a run moves through Ferrite.
Forge 80 · 64 GPUs
01
Pick your hardware
From a single 24 GB card for a notebook to 2,048 GPUs on one fabric. Change your mind later without moving data.
torchrun train.py
02
Bring your code
Point us at a container or a Git repo. Your dependencies, your drivers, your training script: unchanged.
8 → 512 GPUs
03
Scale without rewiring
Grow from 8 to 512 GPUs in place. The scheduler keeps jobs on neighbouring racks so the fabric stays fast.
endpoint · 120 req/s
04
Serve what you trained
Turn a checkpoint into an autoscaling endpoint with one command, on the same platform, with the same bill.
Python
CLI
curl
Copy
$ ferrite cluster create --gpu forge-80 --count 64 --region ams-1
✓ cluster “bright-otter” ready in 84s
$ ferrite run bright-otter -- torchrun train.py --config big.yaml
→ streaming logs (ctrl-c to detach)
$ ferrite cluster scale bright-otter --count 128
✓ scaled in place, fabric intact
Quickstart
Your first cluster
Install the SDK with pip install ferrite, set your API key, and launch. That’s the whole setup.
Clusters come up with drivers, NCCL and your container ready
Logs, metrics and GPU health in the console and the API
Spot clusters for sweeps, at up to 60% off
Hardware
Three families, one API
All prices are per GPU-hour, on demand. Commit for a month or more and they drop by up to 35%.
Inference and fine-tuning
Spark 48
Memory
48 GB
Interconnect
PCIe
Cluster size
1–16
Relative training speed
$1.19
/ GPU-hour
Most training runs
Forge 80
Memory
80 GB
Interconnect
3.2 Tb/s fabric
Cluster size
8–512
Relative training speed
$2.29
/ GPU-hour
Frontier-scale models
Anvil 192
Memory
192 GB
Interconnect
6.4 Tb/s fabric
Cluster size
8–2,048
Relative training speed
$4.49
/ GPU-hour
Private by default
Single-tenant nodes, encrypted disks and your own network. SOC 2 Type II and ISO 27001.
Bring your data
Mount S3, GCS or Azure buckets directly, or keep hot data on our parallel file system.
Private links
Dedicated 100 Gb/s links to your VPC or office, with no egress fees between regions.
Ferrite Labs
What comes after GPUs
Our research team is testing quantum and photonic accelerators next to the clusters you use today.
Edge
Inference close to your users
Small endpoints in 30 edge sites, so answers arrive in tens of milliseconds.
Start in minutes
Your first 100 GPU-hours are on us
Sign up, add a card and launch a cluster before your coffee’s cold. No sales call, no minimum spend.