Build
$0
plus usage
For a side project or a first prototype.
- $10 of credit to start
- Every model, serverless
- Pay per token or per second
- Community support
005
Pricing
— Pay for what you call, not what you keep.
Start with $10 of credit and no card. Every plan runs every model, tokens are priced per million and GPUs by the second.
ScaleMost teams
$499
per month, plus usage
For a team with a product in production.
- Reserved GPUs at 30% off
- Fine-tuning and private models
- Guardrails and audit logs
- 99.95% uptime, in writing
- Support in one business hour
Enterprise
Custom
billed yearly
For a company with its own rules.
- Dedicated clusters, any region
- Your cloud or ours (VPC)
- SSO, SCIM and data residency
- 99.99% uptime and a named engineer
Prices in US dollars, before tax. Usage is billed monthly on every plan.
GPUs, by the second
005.2
arc—light
Priced by the hour, billed by the second. Reserve for a month or more and the hour costs 30% less.
GPU
Memory
On demand
Reserved
Good for
NVIDIA H200
141 GB
$3.49 / h
$2.45 / h
Training and the largest models
NVIDIA H100
80 GB
$2.49 / h
$1.74 / h
Fine-tuning and fast inference
NVIDIA L40S
48 GB
$1.19 / h
$0.83 / h
Mid-size models, vision, speech
NVIDIA A10G
24 GB
$0.69 / h
$0.48 / h
Small models and embeddings
What a month would cost
005.3
arc—light
Move the numbers to your workload. The estimate is the plan, the GPUs you keep and the tokens you call.
Estimated monthly bill
$1,208
- Plan
- $0
- GPUs, 30 days
- $1,195.20
- Tokens, 3 in to 1 out
- $13.12
An estimate, before tax. Reserved GPUs need the Scale plan.
Talk it through with usCompare the plans
005.4
arc—light
What each plan holds, line by line.
Build
Scale
Enterprise
Models and GPUs
Every model, serverless
✓
✓
✓
Reserved GPUs
On demand only
30% off on-demand
Dedicated clusters
Fine-tuning
—
✓
✓
Private models and LoRA adapters
—
✓
✓
Regions
Virginia
All four
All four, or your VPC
Safety and control
Guardrails
Default policies
Your own policies
Your own policies
Audit logs
7 days
1 year
Unlimited, exported
SSO and SCIM
—
—
✓
Data residency
—
✓
✓
Support
Uptime, in writing
—
99.95%
99.99%
Support
Community
One business hour
15 minutes, all hours
A named engineer
—
—
✓
Questions
005.5
arc—light
What teams ask before they pay.
What counts as usage?
How does yearly billing work?
Can I set a limit on what I spend?
What happens to a training run that fails?
Do you train on my data?
Can I move from another provider?