Same GPU.
Less idle on the bill.
A Dynamic GPU runs your work on a pooled NVIDIA card shared with other isolated tenants. The quiet moments your job leaves are filled by someone else, so you pay up to 47% less than for the whole card.
Your GPU is idle more
than you think.
Jobs use a GPU in bursts: loading data, waiting for requests, saving checkpoints. Switch the view to see what a dedicated card bills you for.
Each tenant runs in its own container with its own storage. Memory limits are enforced per tenant by the scheduler.
Dynamic and Dedicated run on the same NVIDIA cards, with root access and the same CUDA stack. Only the bill differs.
Prototype on Dynamic. Move a benchmark or long training run to Dedicated on the same account and wallet.
When to pick Dynamic.
And when not to.
Dynamic and Dedicated live on the same account and wallet, so you can switch per job.
Dynamic is the right call for
- Inference APIs and chatbots with uneven traffic
- Fine-tuning and evaluation runs
- Notebooks, dev boxes and prototyping
- Agents that idle between tool calls
- Anything you'd otherwise leave running
Choose Dedicated instead for
- Benchmarks that need identical runs
- Multi-day training that saturates the card
- Latency SLOs with no room for variance
- Workloads your compliance team wants single-tenant
Every GPU, both modes.
Pick per job.
Dynamic is available on B200, RTX PRO 6000 and A100 today.
| GPU | Memory | Dynamic | Dedicated | Monthly | |
|---|---|---|---|---|---|
B200New NVIDIA · Blackwell | 180 GB HBM3e | $4.25/hr−47% | $7.99/hr | $11,665.40/mo for 2× | Details |
| NVIDIA · Blackwell | 96 GB GDDR7 | $0.76/hr | — | $299/mo | Details |
| NVIDIA · Ampere | 80 GB HBM2e | $0.89/hr | — | — | Details |
| NVIDIA · Ada Lovelace | 24 GB GDDR6X | — | $0.39/hr | — | Details |
Is my data visible to other tenants?
No. Each tenant runs in its own container with its own storage and network. GPU memory limits are enforced per tenant by the scheduler, so one tenant cannot read or exhaust another's memory.
Will my job run slower?
It runs on the same card with the same CUDA stack. When another tenant is busy at the same moment, your kernels may wait their turn, so throughput varies more than on a Dedicated card. Most inference, fine-tuning and dev work doesn't notice.
How much memory do I get?
The GPU memory listed for the product is yours. On a Dynamic B200 that's 180 GB.
Is this the same as NVIDIA MIG?
No. MIG cuts a card into fixed slices that need a reset to resize. Dynamic schedules work onto the whole card in software and enforces each tenant's memory limit, so nothing has to be repartitioned.
Your next run
starts at $0.39.
Sign up, top up, launch. You'll be on a GPU in minutes.