Strike
3.79per GPU hour
Billed per second, 60 second minimum. Nothing to sign.
Best for Experiments, fine tuning, a run that starts today.
Hot GPUs.
Cold country.
Sixteen NVIDIA B200s on the Reykjanes peninsula, cooled by North Atlantic air and powered by geothermal steam. Rent them by the second, prepay an hour bank, or take the whole node.
- Load, mean
- ··
- GPUs at work
- ··
- Jobs running
- ··
- Hottest GPU
- ··
- Draw
- ··
Each bar is a node, each block a GPU, lit the colour its load would heat steel to. Blue is idle. Tap a block to strike it.
- NVIDIA B200 · 180 GB HBM3e
- Billed to the second
- From 1.39 USD per GPU hour
- 100 % geothermal and hydro power
- Data stays in the EEA
- 5 TB egress included every month
- Every hour published in the heat ledger
- 96 hour burn in before any paid job
The site
The cheapest cooling is the weather.
71°C
A B200 under full load
5°C
The air outside the wall, yearly mean
A GPU turns nearly every watt it draws into heat. Most clouds pay twice: once for the power, again to pump the heat away. In Reykjanes the outside air does that job eleven months a year.
The power comes from geothermal and hydro plants on a fixed price contract that runs to 2031. That is where our prices come from, and why they hold still.
- Grid
- Geothermal and hydro, fixed price contract to 2031
- Outside air
- 5 °C yearly mean, cooling runs on it 11 months a year
- PUE
- 1.15 measured, first 60 days
- Uplink
- 2× 100 Gb/s, cables to Ireland and Denmark
Median round trip to F1 Preview
London38 ms
Amsterdam44 ms
New York82 ms
San Francisco141 ms
A job, start to finish
From queue to quench.
Four stages. You pay for one of them, to the second.
Billed so far 21,600 s
Stage 01
Queue
The job waits until every GPU it asked for is free at once. Nothing starts half placed.
FreeStage 02
Heat
The node pulls your image, clears the GPUs and mounts scratch. Usually under 40 seconds.
FreeStage 03
Strike
Your container runs. Every GPU reports load, heat, memory and power once a second.
Billed per secondStage 04
Quench
The container exits. Scratch is wiped, GPU memory cleared, the invoice line written.
FreeThe hardware
Sixteen B200s. No neighbours.
Ask for 8 GPUs and you get one node on one NVLink fabric. Ask for 16 and you get both, joined by 3.2 Tb/s of InfiniBand. GPUs are never sliced or shared between jobs.
Every node runs 96 hours flat out before it takes a paid job. If anything fails, the clock starts again from zero.
BLACKFORGEF1 / NODE A–B
- GPU
- 8× NVIDIA B200, 180 GB HBM3e
- GPU link
- NVLink 5, 1.8 TB/s per GPU
- Node link
- 8× 400 Gb/s InfiniBand NDR
- CPU
- 2× Xeon Platinum 8570, 112 cores
- Memory
- 2 TB DDR5
- Scratch
- 30 TB NVMe, wiped per job
- Cooling
- Direct to chip, outside air
- Power
- Geothermal and hydro grid
S/N F1-0728-26BURN IN 96 H · PASSED
Plans
Four ways to pay. One price each.
Per B200 GPU hour, in USD. CPU, memory, scratch and traffic inside the forge are included, with 2 TB of shared storage and 5 TB of egress every month.
Hour bank
3.09per GPU hour, prepaid
Buy 1,000 GPU hours at a time. Spend them on any job within 120 days.
Best for Teams that know roughly how much they will use this quarter.
Anvil
2.89per GPU hour, as a whole node
One node, all 8 GPUs, yours for a month or more. 16,640 USD per node per month.
Best for Long training runs and endpoints that cannot queue.
Ember
1.39per GPU hour, preemptible
Runs on idle GPUs. 30 seconds of notice before it gives the GPU back.
Best for Evaluation, embeddings, sweeps that checkpoint.
What would a month cost?
- Strike9,096 USDCheapest that never stops · 2,400 GPU h at 3.79
- Hour bank9,270 USD3 blocks of 1,000 h
- Anvil16,640 USD1 node, whole month
- Ember3,336 USDonly if the job can be stopped
Heat ledger
Every GPU hour, on the record.
On the second business day of each month we publish the ledger for F1: the hours the forge had, what each plan used, what sat cold, and every minute it was down and why.
It is built from the same feed that lights the bars at the top of this page.
How the ledger worksF1, September 2026Preview
11,488GPU hours available
| Anvil | 5,760 h | 50.1 % |
|---|---|---|
| Strike | 2,214 h | 19.3 % |
| Hour bank | 1,906 h | 16.6 % |
| Ember | 1,071 h | 9.3 % |
| Cold | 537 h | 4.7 % |
| Uptime | 99.97 % | 13 min down, 1 incident |
| Median Strike wait | 1 min | p95 18 min |
For your scripts
One call to start. One to watch.
The console is built on the same API you get. Start a job from the forge CLI, from curl, or from twelve lines of standard library Python.
$ forge login
key ok, team lumen-lab
$ forge run --gpus 8 --plan strike \
--image ghcr.io/lumen/train:2026.10 \
-- python train.py --epochs 4
job_2a4k1 queue > heat > strike
node a, gpu0 to gpu7
$ forge watch job_2a4k1
gpu0 97 % 71 °C 962 W 152 GB
gpu1 96 % 70 °C 955 W 152 GBForge F2
Québec is next.
F2 is the same build as F1, beside a hydro dam on the St. Lawrence. We list it when it passes burn in, not on a date. Each step moves in the monthly ledger.
- Site lease, Beauharnois, QuébecDone
- Hydro power contractDone
- Two nodes deliveredIn progress
- 96 hour burn inOpen
- Listed in the capacity endpointOpen
Tell us what you want to run.
Access is by invitation while F1 is in preview. Send a work order and a person replies within one business day with start dates and a quote.