KubeOn GPU pricing
Editions for how far the optimizer goes
Start with Insights to measure every GPU-hour and see what the optimizer would do. Move to Optimize to let it act in Approve or Automatic mode, with burst to other regions of the same cloud, and to Enterprise for burst across clouds and into your data centers.
Insights
Every GPU-hour measured and owned, with proposals and no write access.
Get a quote
Quoted on monitored GPUs, annually
- GPU cost by team, idle GPU time and DCGM utilization
- Sharing, sizing and scale-to-zero proposals
- GPU type fit recommendations
- Budgets and alerts on GPU spend
- GPU policies in Recommend mode
- Managed hosting
- Standard support
Optimize
For platform and ML teams that want the optimizer to act.
Get a quote
Quoted on monitored GPUs, annually
- Everything in Insights
- GPU optimizer in Approve and Automatic modes
- MIG and time-slicing applied in your change windows
- GPU autoscaling with spot and on-demand fallback
- Packing and scale to zero
- GPU burst to other regions of the same cloud
- Audit log with one-click revert
- SSO with SAML or OIDC
- Priority support
Enterprise
For multi-cloud GPU fleets, owned data centers and regulated programs.
Get a quote
Quoted on monitored GPUs, annually
- Everything in Optimize
- GPU burst across clouds and into your data centers
- Owned GPU rate cards and hardware amortization
- Custom roles and audit export
- Security review and custom terms
- Named technical account manager
- 24/7 support with an SLA
Quotes include the rollout plan. GPU compute stays on your own cloud bills and in your own data centers.
Compare
What each KubeOn GPU edition includes
| Feature | Insights | Optimize | Enterprise |
|---|---|---|---|
| Measure | |||
| GPU node cost split by GPU requests | Included | Included | Included |
| Idle GPU time, priced | Included | Included | Included |
| DCGM utilization and GPU memory per pod | Included | Included | Included |
| MIG slices allocated by share | Included | Included | Included |
| Owned GPU rate cards | Not included | Rate cards | Rate cards and hardware |
| Optimize | |||
| Proposals with patches (Recommend mode) | Included | Included | Included |
| Approve and Automatic modes | Not included | Included | Included |
| MIG partitions and time-slicing | Proposals | Included | Included |
| GPU autoscaling, spot first | Proposals | Included | Included |
| Packing and node removal | Proposals | Included | Included |
| Scale to zero | Proposals | Included | Included |
| GPU burst to other regions of the same cloud | Not included | Included | Included |
| GPU burst across clouds and into your data centers | Not included | Not included | Included |
| Control | |||
| Policy limits, windows and scopes | Included | Included | Included |
| Audit log | Included | Included | Included |
| One-click revert | Not included | Included | Included |
| Custom roles and audit export | Not included | Not included | Included |
| Hosting and support | |||
| Fully managed by UnitVane | Included | Included | Included |
| SSO with SAML or OIDC | Not included | Included | Included |
| Support hours | Business hours | Extended | 24/7 |
| Technical account manager | Not included | Not included | Included |
How we count
The terms in a KubeOn GPU quote
Plain definitions, so the quote and the invoice say the same thing.
Monitored GPU
A physical GPU seen by an agent during the billing period, measured as the average number present at once. A GPU split into MIG slices or time-sliced still counts once.
Burst GPU
A GPU attached by GPU burst counts as a monitored GPU only for the hours it is attached. The cloud charges for it stay on your own bill.
Optimized cluster
A cluster running the GPU optimizer. There is no separate charge per cluster: GPUs are counted the same way with or without it.
User
A person who signs in to UnitVane. People who only receive reports, alerts or approval requests by email are not counted.
Your quote
What sets the price
We share prices with each client, based on their requirements. Every KubeOn GPU quote is built from the same four inputs.
- Monitored GPUs
- The average number of physical GPUs present at once over the billing period, across every cluster with an agent. CPU nodes are not counted.
- Edition
- How far the optimizer may go: Insights for measurement and proposals, Optimize for Approve and Automatic modes with burst inside one cloud, Enterprise for burst across clouds and into owned data centers.
- Data region
- Choose your region when your workspace is created. European data stays in a European region. The optimizer itself runs in your clusters.
- Term
- Quoted annually, with multi-year terms and volume pricing available. UnitVane, KubeOn, UnitVane AI and the GPU optimizer can share one agreement.
Worked example: monitored GPUs for a 30-day month
GPUs are averaged over time, so autoscaling peaks and burst count only for the hours they run.
| Pool | Peak | Average | |
|---|---|---|---|
| ml-inference8 L40S all month, autoscaling to 12 for 10 hours a day | 8 L40S all month, autoscaling to 12 for 10 hours a day | 12 | 9.7 |
| ml-notebooks2 A10G on weekdays from 07:00 to 20:00, scaled to zero otherwise | 2 A10G on weekdays from 07:00 to 20:00, scaled to zero otherwise | 2 | 0.8 |
| dc1-gpu8 A100 all month, one split into three MIG slices | 8 A100 all month, one split into three MIG slices | 8 | 8.0 |
| burst4 L40S in eu-central-1 for 3 days of the month | 4 L40S in eu-central-1 for 3 days of the month | 4 | 0.4 |
| Total | 26 | 18.8 |
This counts as about 19 monitored GPUs, not its 26-GPU peak.
- Autoscaling and scale to zero count for the hours GPUs run: GPUs the optimizer removes stop counting when they leave.
- MIG slices and time-sliced replicas count as the one physical GPU they share.
- Burst GPUs count only while they are attached to your cluster.
- Owned GPUs in your data center count the same way as cloud GPUs.
- Growth during the year is handled the way your quote sets out, with no surprise charges month to month.
FAQ
KubeOn GPU pricing questions
Billing
Editions
Tell us about your GPU fleet. We will send a quote.
GPU types, how many run at once, which clouds and data centers, and how far you want the optimizer to go.