A GPU is not
one indivisible thing.
Most clouds sell a GPU as "one job, one card", so it sits idle whenever that job waits. Our scheduler treats each card as memory, compute and bandwidth that different workloads stress at different times, and packs compatible work onto it safely.
Most clouds sell you a whole card that sits idle.
We pool it.
Six GPUs running a mixed fleet. Switch to "Reserved GPU" to see the gaps a one-job-per-card cloud leaves empty. Our scheduler fills them with other tenants' work, so the hourly price is up to 47% lower.
$4.25 vs $7.99 /hr80 GB · DynamicDynamic · hourlyWorkloads don't use
all of a GPU at once.
Inference is bound by memory bandwidth, training by compute, a notebook mostly by your typing speed. We schedule on what each one actually uses.
Measured, not reserved
The scheduler tracks how each workload uses memory, compute and bandwidth over time, not just whether a card is occupied.
Isolated in software
GPU calls are mediated per tenant. Memory limits are enforced, and every tenant gets its own container, storage and network.
Rebalanced continuously
As workloads start, burst and stop, placement adjusts so busy tenants aren't stacked on the same card.
Sharing a GPU,
three ways.
Other ways to share a card trade flexibility for isolation, or the reverse.
| NVIDIA MIG | NVIDIA MPS | packet.ai Dynamic | |
|---|---|---|---|
| How the card is split | Fixed hardware slices | Concurrent processes | Scheduled pool, whole card |
| Resize without a reset | No | Yes | Yes |
| Per-tenant memory limits | Yes | No | Yes |
| Uses idle capacity | No, slices sit idle | Yes | Yes |
| Works on any recent NVIDIA card | Data-centre cards only | Yes | Yes |
packet.ai runs on hosted.ai's GPU cloud platform, the software GPU providers use to run, share and bill their own fleets. packet.ai is that platform, run by its own team, sold by the minute.
Your next run
starts at $0.36.
Sign up, top up, launch. You'll be on a GPU in minutes.