"GPU as a service" shows up in vendor decks as if it's one product. It isn't. Renting a bare-metal GPU box by the hour, calling a serverless inference endpoint billed per token, and reserving a cluster for a year are all "GPUaaS," and they have almost nothing in common operationally or on the invoice. Picking the wrong one is either wasted money or a workload that doesn't fit the box you rented.
TL;DR: GPU as a service (GPUaaS) is renting GPU compute on demand instead of buying and racking hardware. It splits into five delivery models — bare metal, VM, container/serverless, and managed inference endpoints — each billed differently: on-demand hourly (~$2.50-$3.99 per H100-hour at neoclouds as of August 2026, versus $6.88/hr at AWS), reserved/committed (roughly 20-40% cheaper for a term commitment), spot/preemptible (60-90% off but reclaimable with little notice), or per-token for hosted inference. The right model depends on whether your job can be interrupted, how long it runs, and whether you need the box or just the output.
What GPU as a service actually means
GPUaaS is the GPU-specific case of the same idea behind any "as a service" cloud product: instead of buying GPUs, racking them, and running your own datacenter, you rent access to someone else's — a neocloud like Lambda or CoreWeave, a marketplace like Vast.ai, or a hyperscaler like AWS. You pay for what you use, provisioning takes minutes instead of a hardware lead time, and someone else owns the depreciation risk on the silicon.
What actually differs between GPUaaS offerings isn't the marketing language — it's two independent axes: how the GPU is delivered to you (bare metal through fully managed endpoint), and how you're billed for it (hourly, reserved, spot, or per-output-unit). Most comparisons collapse these into one spectrum. They shouldn't, because a workload's tolerance for interruption and its need for infrastructure control are different questions.
The delivery models
Bare metal
You get the physical machine, no hypervisor between you and the hardware. Best raw performance, most operational responsibility — you manage the OS, drivers, and orchestration yourself. This is what a neocloud like Crusoe or CoreWeave sells for large training clusters where every percentage point of throughput matters and you already have the ops team to run it.
Virtual machine
The standard model — a VM with GPU passthrough, an OS image, root access. Slightly more overhead than bare metal, much less operational burden: snapshotting, resizing, and provider-managed networking come with the box. This is what most GPU rental — Lambda, RunPod pods, Hyperstack, Vast.ai, and Aquanode's own marketplace — actually is.
Container
You ship a container image, the platform schedules it onto a GPU, and you don't manage an OS at all. Faster cold-start than a full VM in most implementations, and closer to how CI/CD already works for most engineering teams. RunPod's serverless tier and Modal both work this way.
Serverless / per-second billing
The container model taken further: no persistent instance at all. You define a function or endpoint, the platform scales workers up on request and down to zero when idle, and you're billed per second of actual compute — not per hour of a box sitting there. Modal's pricing is a clean example: an H100 SXM5 runs $0.001097/second, which is roughly $3.95/hour of actual usage but $0 when nothing is running. This is the right model for bursty, unpredictable inference traffic where an always-on box would sit idle most of the day.
Managed inference endpoints
The most abstracted tier — you don't think about GPUs or containers at all, you send a prompt and get a response, billed per input/output token or per generated unit (image, video, audio-minute). Together AI's serverless inference pricing runs $0.03-$3.00 per million input tokens and $0.12-$15.00 per million output tokens depending on model, with image generation from $0.0019/image. This is the model for teams building an application on top of someone else's already-hosted model, not training or running their own.
| Delivery model | You manage | Best for | Example provider |
|---|---|---|---|
| Bare metal | OS, drivers, orchestration | Large training clusters, max throughput | Crusoe, CoreWeave |
| VM | OS, app stack | General training/inference, most workloads | Lambda, RunPod, marketplace |
| Container | App image only | Reproducible jobs, CI-style batch work | RunPod serverless, Modal |
| Serverless (per-second) | Function/endpoint code | Bursty, unpredictable inference traffic | Modal |
| Managed inference endpoint | Nothing infra-side | Building on an existing hosted model | Together AI |
The pricing models
Delivery model and pricing model are independent — you can rent a VM on-demand, reserved, or spot, and the same is increasingly true for containers.
On-demand hourly is pay-as-you-go with no commitment, priced at a premium over reserved rates in exchange for the flexibility to stop anytime. As of August 2026, on-demand H100 pricing runs from about $2.50/hr (Hyperstack) to $3.29/hr (RunPod, SXM variant) at neoclouds, versus $6.88/hr per GPU on AWS's p5.48xlarge ($55.04/hr for 8x H100).
Reserved / committed capacity trades a term commitment (commonly 7 days to 1+ year) for a lower rate — Together AI's reserved GPU pricing runs $3.19-$7.99/hr depending on commitment length against $3.99-$8.19/hr on-demand for the same H100/H200/B200 tiers, roughly a 20% discount at the shorter end. This fits steady, forecastable demand — a training pipeline that runs continuously, not a spiky workload.
Spot / preemptible offers a provider's unused capacity at a steep discount, commonly cited at 60-90% off on-demand, with the catch that the provider can reclaim the instance with little notice — sometimes as little as 30 seconds. This only works for jobs that checkpoint frequently and can resume cleanly; losing a run with no checkpoint is not a discount, it's wasted compute.
Per-token / per-output-unit applies to managed inference rather than raw compute — you're billed for what the model produced, not for how long a GPU ran, and the provider absorbs the utilization risk of keeping GPUs warm between requests. This is usually the cheapest option for low-volume or spiky inference and the most expensive at sustained high volume, where renting your own endpoint and running it near-continuously beats per-token pricing.
Decision table: workload to pricing model
| Your situation | Recommended model |
|---|---|
| Training run, hours to days, needs to complete without interruption | On-demand VM or bare metal |
| Steady production training/inference load, predictable for weeks+ | Reserved capacity |
| Batch job that checkpoints often and can tolerate being killed | Spot/preemptible |
| Inference traffic with sharp, unpredictable spikes | Serverless (per-second billing) |
| Building an app on a model you don't want to host yourself | Managed inference endpoint (per-token) |
| Shopping across providers for the lowest current rate on a specific GPU | Marketplace on-demand — check the GPU Availability Index before committing |
What GPUaaS pricing doesn't tell you
The hourly rate on a pricing page is the easy number. Three things it leaves out, worth naming before you commit to a provider on price alone:
Egress and storage fees. Some providers (Lambda, Crusoe) advertise zero egress fees; others charge for moving data out, which can matter more than the GPU rate if you're moving large model weights or datasets between providers. Check this explicitly — it's rarely in the headline price.
Idle cost. An on-demand VM you forget to stop bills the same whether it's training or sitting empty. This is the single most common way GPUaaS costs run over budget, and it's worth estimating before you provision — our idle cost calculator does the arithmetic for a box left running unattended.
What it costs to move. A lower rate on a different provider only saves money if switching is cheap. Re-downloading model weights, rebuilding a Python environment, and losing whatever state was on the old box can turn a 30% hourly saving into a net loss of a day's work. We've written about that portability gap directly in moving a GPU workload to another cloud provider.
How to estimate your actual cost
Before picking a provider or a pricing model, get a real number for the workload, not just the sticker rate. Our training cost calculator estimates total spend for a training run given GPU type, hours, and count, and the live pricing page and marketplace show current cross-provider rates rather than the snapshot in this post — GPU prices move weekly, and a comparison written today will be stale by the time you read it next quarter.
Frequently asked questions
What does GPU as a service mean?
GPU as a service (GPUaaS) means renting GPU compute from a cloud provider on demand — by the hour, reserved for a term, at a discounted spot rate, or per output token for managed inference — instead of buying and operating your own GPU hardware.
What's the difference between GPUaaS and cloud computing?
GPUaaS is cloud computing scoped specifically to GPU compute rather than general-purpose infrastructure. The billing models (on-demand, reserved, spot) are the same concepts used for CPU cloud compute, applied to GPUs, which are more expensive per unit and more supply-constrained.
Is GPUaaS cheaper than buying GPUs?
For most teams, yes, unless utilization is near-continuous for a long period (typically well over a year) at a scale large enough to justify the capital cost, staffing, and depreciation risk of owning hardware. GPUaaS avoids the upfront cost and lets you match spend to actual usage.
What's the cheapest GPUaaS pricing model?
Spot/preemptible pricing is cheapest per hour — commonly 60-90% below on-demand — but only for jobs that can be interrupted and resumed from a checkpoint. For inference with low or spiky volume, per-token managed endpoints are often cheaper than running your own always-on GPU.
How much does an H100 cost per hour as a service?
As of August 2026, on-demand H100 rental runs roughly $2.50-$3.99 per GPU-hour at neocloud providers like Hyperstack and RunPod, versus $6.88 per GPU-hour on AWS's p5.48xlarge instance. Reserved and spot rates can run lower at either end depending on commitment length and interruption tolerance.
Sources
- Hyperstack GPU pricing page, checked August 2026. On-demand hourly rates, NVIDIA A100 through B300.
- RunPod pricing page, checked August 2026 (page last updated July 27, 2026). On-demand and serverless GPU rates.
- Lambda GPU Cloud pricing page, checked August 2026. Per-minute billing, zero egress fees.
- Together AI pricing page, checked August 2026. On-demand vs reserved GPU cluster rates, serverless per-token inference rates.
- Modal pricing page, checked August 2026. Per-second serverless GPU billing.
- Vast.ai pricing page, checked August 2026. Marketplace on-demand/interruptible/reserved instance types.
- Crusoe Cloud page, checked August 2026. Zero egress fees, bare-metal GPU fleet.
- AWS p5.48xlarge on-demand price ($55.04/hr, 8x H100, $6.88/GPU-hr) via Vantage AWS EC2 instance pricing, checked August 2026.
