NEW
Anvil 192 clusters are live in Amsterdam →
Train bigger models.
Wait less.
Ferrite gives your team fabric-connected GPU clusters in about 90 seconds, billed by the second. No reservations, no quotas, no three-week sales cycle.
live
Utilisation
87
%
Jobs running
21
jobs
Just started
train · vision-7b
64 GPU
finetune · support-bot
8 GPU
eval · coder-34b
16 GPU
train · protein-fold
128 GPU
serve · search-rerank
4 GPU
Training on Ferrite
Northwind AI
Lumen Bio
Okapi
Fathom Labs
Quill
Meridian Health
Why Ferrite
The boring parts, done properly
Fast hardware is the easy bit. We spend our time on the networking, storage and scheduling that decide whether a run finishes on Tuesday or next month.
Our own halls, our own racks
We design and run every data hall ourselves, so a cluster you launch today is the same cluster next month.
90s
From click to cluster
Median time to a ready 64-GPU cluster, measured over the last 30 days.
3.2 Tb/s between every GPU
Non-blocking fabric across the whole hall, so 512 GPUs behave like one machine.
Fibre to the rack
Private links to your cloud and on-prem storage.
Storage that keeps up
Parallel file system at 40 GB/s per node, with checkpoints that save in seconds.
Billed by the second
Stop a cluster and the meter stops. No minimum hours, no idle reservations, no surprises.
100% renewable power
Every hall runs on wind and hydro, and the waste heat warms 2,000 homes nearby.
Developer first
Five lines from laptop to 64 GPUs
Use the Python SDK, the CLI or plain HTTP. Bring any container, and your code runs the same as it does on your machine.
Works with PyTorch, JAX and anything in a container
Automatic checkpoints and restarts if a node fails
SSH into any node, any time
Python
CLI
curl
Copy
from ferrite import Cluster
# 64 GPUs, fabric-connected, ready in about 90 seconds
cluster = Cluster.create(gpu=“forge-80”, count=64, region=“ams-1”)
job = cluster.run(
image=“ghcr.io/acme/trainer:latest”,
command=“torchrun train.py --config big.yaml”,
checkpoint=“s3://acme/ckpt/“,
)
job.logs(follow=True)
14,000
GPUs across seven regions
99.98%
cluster uptime last quarter
38%
lower cost than the big three clouds
<4 min
median support response
“We moved a six-week training run to Ferrite and it finished in nine days. The cluster just stayed up, which is the highest praise I can give infrastructure.”
Kenji Mori
Head of ML, Northwind AI
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.