GPUs by the hour
Draw capacity by the hour.
Accelerator time from deployments Nawa develops and operates. Draw it on demand or hold a reserved block; the meter is the hour.
Positioning
Between a laptop and a long-term contract.
Owning accelerators means capital and an operations bench, and most workloads justify neither. The hourly offer covers the ground in between: accelerator time on Nawa-operated infrastructure, billed for the hours a job consumes.
Access takes two forms. Work that arrives unannounced draws from the pool on demand, and a job with a known window can hold a reserved block ahead of it.
Fit
Who this fits.
-
Outgrowing local hardware
The model no longer fits the machine on the desk, and a term reservation is more commitment than the roadmap supports. Hourly capacity covers the stretch in between.
-
Burst workloads
A pipeline that idles most of the month and works hard for a day has no use for standing capacity. The draw rises with the burst and falls with it.
-
Evaluation runs
Model comparisons and benchmark sweeps need real accelerator hours on hardware the team has no reason to keep. The hours end when the sweep does.
-
Fine-tuning experiments
A training pass that occupies a weekend needs a block of capacity with a defined end. A reserved window fits it; the capacity releases when the job completes.
Structure
What the hour covers.
Structure is public. Pricing, hardware, region, and minimum duration stay in the agreement.
The consumption model
- Metered by the hour
- Billing follows the capacity drawn, hour by hour. A job that stops also stops costing.
- On-demand and reserved
- Work can draw from the pool as it arrives, or hold a block against a scheduled window.
- Elastic within the site
- Capacity scales up and down with the workload, inside what the deployment holds.
- No hardware ownership
- The hardware and the agreements behind it stay with Nawa.
Where it comes from
- Nawa-operated
- Every hour is served from modular deployments Nawa develops and operates on available power.
- Reference configurations
- Each deployment is assembled from the drawn configurations, from a single module up to a hall.
- Placement
- Sites are selected for power first. Where a workload cares about distance, the reserved inference offer places capacity near the user.
- Growth path
- A team that settles into steady consumption can move the workload onto a reservation, on the same infrastructure.
Next
Say what you need to run.
Describe the workload and the hours it wants. Note any constraint on where it can run.