Pricing

Simple, per-second billing

Everything is billed per second. Prices change roughly twice a year when power costs move. If you run more than $5k a month, email us and we will find a volume number.

GPU instances

Billed while running, per second, rounded up to the minute. Storage is separate. You stop paying the moment you delete the instance.

GPU	VRAM	RAM	Price / hr	Best for
RTX 4090	24 GB	64 GB	$0.19	Fine-tuning small models, community containers
RTX 6000 Ada	48 GB	96 GB	$0.45	Mid-size training, larger VRAM on a budget
A10G	24 GB	48 GB	$0.35	Inference, batch jobs
L40S	48 GB	96 GB	$0.80	Inference and fine-tuning on one box
A100 40GB	40 GB	128 GB	$0.95	Mid-size training runs
A100 80GB	80 GB	128 GB	$1.45	30B to 70B inference and training
H100 80GB	80 GB	128 GB	$2.10	Large training runs, heavy batch inference

Persistent storage: $0.05/GB/month. Snapshots: $0.03/GB/month.
Networking unmetered at 1 Gbps, 10 Gbps on H100.
Public IPv4 included. IPv6 on request.
NVLink pairs on A100 and H100 available for training.

Model hosting

A flat monthly base per model that covers one replica, plus per-token usage. We add replicas when traffic needs them and you only pay for the extra compute while those replicas actually run.

Model size	Base / mo	Input / M tokens	Output / M tokens	Includes
1B to 7B	$29	$0.05	$0.15	1 replica on a 4090
8B to 13B	$49	$0.10	$0.30	1 replica on an A10G
14B to 40B	$99	$0.20	$0.60	1 replica on an L40S
41B to 70B	$199	$0.40	$1.10	1 replica on an A100

Every deployment has a monthly spending cap you set. We stop serving rather than let it blow past the cap.
Streaming is supported and output tokens are billed the same either way.
Fine-tuning hosted on the same model id is billed as compute hours, not tokens.
If you bring your own checkpoint, there is no one-time setup fee for models up to 70B.

Add-ons

Thing	Price	Note
Object storage	$0.02/GB/month	For datasets and checkpoints between runs
Dedicated IP	$2/mo	If you need a stable egress IP
Slack support	$40/mo	We read it, we answer it
Custom image builds	$15 each	We bake your deps into an image

Common questions

What happens if I hit my cap mid-request?

Current requests finish, new ones get a 429. Your spend never goes over the cap.

Do you do annual billing?

No, everything is per-second. If a longer commitment actually saves you money, email us and we will quote it.

Can I use my own model weights?

Yes. Upload them or point us at a Hugging Face repo. Most checkpoints under 70B deploy in a few minutes.

Is there a free tier?

New accounts get $10 of credit. After that you pay. We do not do a permanent free tier because we cannot afford to.