The cluster’s existing pages mostly answer the buy-or-API question. Cloud GPU rental is the third option the cluster has not addressed: rent a card on someone else’s machine, run the same self-hosted stack on it, and pay per second. For someone who needs a 24 GB or 80 GB card for a few hours a week but cannot justify spending $1,000-$3,000 of their own money, rental is often the cheaper answer than the API and the cheaper answer than buying.
Verified September 2026 against current vendor pricing pages, the Vast.ai billing reference, the Hugging Face Inference Endpoints pricing doc, and the Ollama library for the model sizes the comparison uses.
What Rental Actually Sells You
Each of the four vendors below exposes the same product: a remote machine with one or more NVIDIA GPUs, billed by the wall-clock second (or per minute) the machine is on. The differences are in three places: how much bandwidth and storage cost on top of the GPU bill, how the GPUs are procured (RunPod and Vast.ai run on a mix of provider data centers and independent hosts; Lambda runs its own data centers; Hugging Face Inference Endpoints sits on AWS, GCP, and Azure), and what minimum commitment is required. None of the four charges a setup fee for the GPU itself, and none requires a long contract for on-demand usage.
The catch is that the published pricing page is rarely the whole bill. Storage is extra; bandwidth is extra; the cloud GPU rate may or may not include system RAM and vCPUs. A 4090 at $0.74/hr sounds like $17.76 a day, but the actual day cost includes the persistent volume the model weights live on and the egress if you push results back to your own machine.
RunPod (Pods, Serverless, Clusters)
RunPod publishes full per-hour pricing on three product lines; the Pods table is the one most local-AI readers will use, the Serverless line is for hosted inference endpoints on top of the same GPUs, and the Clusters line is multi-GPU. The pricing page is dated “Updated September 13, 2026” and shows per-second billing on Pods and Serverless (RunPod pricing).
| GPU (VRAM) | Pods $/hr | Serverless $/hr |
|---|---|---|
| RTX 4090 (24 GB) | 0.74 | 1.10 |
| RTX 5090 (32 GB) | 0.99 | 1.58 |
| L40S (48 GB) | 1.09 | 1.75 |
| A100 SXM (80 GB) | 1.59 | 2.72 |
| H100 SXM (80 GB) | 3.49 | 4.79 |
| H200 (141 GB) | 4.59 | 5.93 |
| B200 (180 GB) | 6.79 | 8.64 |
Storage on Pods costs $0.10/GB/month for container disk and $0.20/GB/month idle for volume disk; network storage runs $0.05-$0.07/GB/month over 1 TB or under 1 TB respectively (RunPod pricing). RunPod does not state a minimum commitment or a free credit amount on the public pricing page. Reserved capacity exists for “Up to 50% Off” on 1, 3, or 6 month terms, but the page directs enterprise volumes to sales.
Lambda Labs (owned data centers, no egress)
Lambda publishes per-GPU hourly rates on its own data centers; on most cards the rate is identical across 1x and 8x boxes (A6000 1x/8x both $1.09, A100 40GB 1x/8x both $1.99), and on two cards it actually drops at 8x (H100 SXM $4.29 at 1x falls to $3.99 at 8x; B200 $6.99 at 1x falls to $6.69 at 8x). The page states “Transparent pricing with no egress fees” (Lambda GPU Cloud), which is the differentiator that matters once a job processes real data.
| GPU (per GPU) | 1x $/hr | 8x $/hr/GPU |
|---|---|---|
| Quadro RTX 6000 (24 GB) | 0.69 | - |
| A10 (24 GB) | 1.29 | - |
| A6000 (48 GB) | 1.09 | - |
| A100 SXM (40 GB) | 1.99 | - |
| A100 SXM (80 GB) | - | 2.79 |
| H100 PCIe (80 GB) | 3.29 | - |
| H100 SXM (80 GB) | 4.29 | 3.99 |
| GH200 (96 GB) | 2.29 | - |
| B200 SXM6 (180 GB) | 6.99 | 6.69 |
Lambda’s published per-GPU rate actually drops as you scale up: an 8x H100 SXM box is $3.99/GPU/hr, cheaper per-GPU than the $4.29/GPU/hr on the 1x configuration. The page advertises “Pay by the minute” billing with no published minimum commitment terms, and all rates are per-minute. Lambda does not publish an RTX 4090 SKU at all; its cheapest current-gen card is the Quadro RTX 6000 at $0.69/hr.
Vast.ai (marketplace; rates vary per host)
Vast.ai is a marketplace, not a data-center operator. Individual GPU hosts list capacity and set their own hourly rate, so the rates are a moving target. The billing reference page confirms the three charge buckets - “Active rental (GPU) (in $/hr)”, “Storage costs (in $/GB/hr)”, “Bandwidth costs (in $/TB)” (Vast.ai billing) - and the per-second billing model: “You are charged the base active rental cost for every second your instance is in the active/connected state.”
What is verifiable:
- Per-second billing with no minimum commitment (the docs explicitly answer a “10-minute rental” question with: “You will only be charged for the 10 minutes of usage”)
- No refunds after credits are spent: “After spending credits, there are absolutely no refunds”
- Storage and bandwidth add to the base rate; rates “vary considerably from machine to machine”
What is not verifiable on this page: a specific RTX 4090 or H100 hourly figure. The pricing is per-offer and must be read off the live search page (Vast.ai pricing) before each rental.
Hugging Face Inference Endpoints (managed inference)
Hugging Face Inference Endpoints is a different product from the other three. The page does not sell raw GPU time; it sells a managed deployment with autoscaling, billed while the endpoint is initializing or running, on hosts in AWS, GCP, and Azure (HF Inference Endpoints pricing). The formula is instance hourly rate * ((hours * min replicas) + (scale-up hours * additional replicas)).
| GPU (1x, AWS) | VRAM | $/hr |
|---|---|---|
| nvidia-t4 | 14 GB | 0.50 |
| nvidia-l4 | 24 GB | 0.80 |
| nvidia-a10g | 24 GB | 1.00 |
| nvidia-l40s | 48 GB | 1.80 |
| nvidia-a100 | 80 GB | 2.50 |
| nvidia-h200 | 141 GB | 5.00 |
GCP’s H100 is $10/hr/GPU on a 1x, more than double AWS’s H200. The page does not state storage or bandwidth costs; those come from the underlying cloud provider. Endpoints requires an active Hugging Face subscription with a credit card on file. PRO accounts and Enterprise Hub organizations can request higher quotas.
Renting vs Buying at the Crossover Math
A useful comparison point is the RTX 4090 on RunPod at $0.74/hr, which works out to $532.80 for a 720-hour month of continuous inference. The same card retails for around $1,800-$2,200 in mid-2026 (per the GeForce RTX 4090 launch pricing pages this cluster references in best-used-gpu-for-local-ai). If a workload runs less than about five hours a day, rental is cheaper than buying the same VRAM tier outright.
For 80 GB cards, the comparison changes. There is no consumer card at 80 GB, so the choice was always between an enterprise A100, an H100, or a Mac Studio M5 Ultra with up to 512 GB of unified memory (per the best-gpus-for-local-ai hub). Rental puts the 80 GB tier in reach without the enterprise contract. An H100 PCIe at $3.29/hr/GPU on Lambda is $2,368.80 for a 720-hour month, before electricity. The electricity saving is the largest line item: at 13.63 cents per kWh (EIA 2025 average), a self-hosted 700 W-class card running 24/7 costs about $143 of electricity per month plus the amortised hardware.
The other way rental saves money is when the workload is variable. A household that runs a 32B local model nightly on a workstation-class card has a low duty cycle; a research workflow that bursts to multi-GPU for a week of fine-tuning and then idles for a month pays nothing during the idle month if it rents.
The Catches the Pricing Page Does Not Show
Three structural limitations apply to all four vendors, not just one.
Throughput is bounded by the GPU you rent. A rented RTX 4090 cannot serve more users than an owned RTX 4090; the VRAM-tier guides on this site explain why. Heavy multi-user serving looks like a different problem and a different bill than inference for a few clients.
Cold start time is real. First-request latency on Lambda, RunPod, and HF Endpoints is governed by container pull, model load, and KV-cache warmup. If every request is the first request of the hour, the per-token economics shift.
No-data-leave-the-building requires care. Vast.ai in particular routes data through hosts whose physical location and audit posture are not under your control. Lambda and HF Endpoints pin to the underlying provider’s data-residency rules. The privacy-sensitive case still favours local execution, and the no-GPU entry point explains how to keep inference off-network at the lowest tier.
Bottom Line
For a model size the cluster has already mapped (24-32 GB consumer card, 80 GB prosumer, or 141+ GB workstation) and a workload that runs fewer than a few hours a day, rental is genuinely cheaper than buying and competitive with the API on cost. RunPod is the clearest entry point for a single-card workload because the rates are published down to the second. Lambda’s no-egress policy is the differentiator that matters when a job outputs hundreds of GB. Hugging Face Inference Endpoints is the right pick if the goal is hosted inference rather than raw compute. Vast.ai is the budget pick when the workload tolerates host-level variability. Run your own numbers on your own usage; the vendors publish their rates and this site will not pretend to know your volume better than you do.