Workload qualification
Define duration, traffic shape, and interruption tolerance.
The capacity form fits the workload from the first run.
GPU Cloud
Match workload shape.
Then match capacity.
Qualify duration, continuity, memory, and scale before selecting a capacity form. Capacity matches how the work runs, and the fit is stated before the term starts.
Workload signalPoints toWhy
Short, interruptible bursts
Serverless
Capacity follows demand and scales to zero between runs, so idle time is not held.
Steady inference, flexible timing
On-demand
Start when the work is ready and release the instance when it finishes.
Long training runs, no interruption
Reserved
A fixed term holds the accelerators for the duration of the run.
Memory-bound inference
H200 or B200 tiers
141 GB and 192 GB per accelerator, with NVLink and RDMA between them.
Mixed batch and interactive
Reserved plus on-demand
A reserved base carries the steady load; on-demand absorbs the peaks.
Who it is for
Teams that know the workload but need the right capacity form.
Training, inference, and batch work are qualified through runtime shape instead of a hardware name alone.
- Training teamsRuns measured in daysA fixed term holds the accelerators, and the interconnect is chosen before the run starts, not after a failure.
- Inference teamsTraffic that movesSteady load sits on reserved capacity; peaks go to on-demand, and idle periods do not hold hardware.
- Batch and evaluationWork that can waitInterruptible capacity is acceptable when the job can restart, and that tolerance is stated up front.
What the account receives
Capacity decisions tied to the way work runs.
Memory fit
Match working-set needs to a realistic accelerator tier.
Memory limits enter the decision before a hardware name does.
Scale path
Review single-node and coordinated work through the interconnect need.
A design is proven against the workload before it scales.
Continuity choice
Select on-demand, reserved, or serverless capacity by operating pattern.
Continuity expectations are written down before the term starts.
Ready to secure capacity for your workload?
Tell us how the workload runs: how long, how tolerant of interruption, how much memory. A member of the team replies within one business day.