Cloud GPU hosting lets you rent an NVIDIA card by the hour instead of buying hardware, and the honest price range right now runs from under $0.50/hour for entry cards to $5-plus/hour for a B200. Picking well matters more than picking a famous name, because the same A100 80GB can cost anywhere from $0.80 to $2.21/hour depending on the platform. This guide covers how to match a card to your workload, what the major providers actually charge, where Vultr fits in that picture, and the billing details that quietly decide your real monthly bill.
Three buying models cover almost every situation, and Vultr is the sensible default for the middle one.
Predictable, data-center-grade instances across many regions, billed hourly: Vultr's Cloud GPU line, with the H100 from $2.99/hour and $300 in free credits for new accounts.
Absolute lowest hourly rate, tolerance for variable reliability: marketplace-style clouds, where Vast.ai's H100 sits at $1.65 and distributed consumer cards drop even lower.
Serverless inference that scales to zero: RunPod and Modal bill per second and suit spiky API traffic better than a standing instance.
👉 Check Vultr's current GPU rates and start with the free credit
Start with memory. Every other decision, including which platform you rent from, follows from how many gigabytes your models actually need.
Around 24GB (RTX 4090, L4, A10G): image generation with SDXL or FLUX-class models, 7B–13B LLM inference, embeddings and vector work. Rates run from about $0.20/hour on the cheapest marketplaces to around $0.45 at mainstream providers.
Around 48GB (L40S): 30B-class inference, video generation and transcoding, 3D rendering, mid-size fine-tunes. Roughly $1/hour at current rates.
80GB (A100, H100): 70B inference and fine-tuning, multi-day training runs. The A100 80GB spans $0.80–$2.21/hour across current comparison boards; the H100 spans $1.65–$4.76.
141GB and up (H200, B200): very long contexts and frontier-scale training. The H200 starts at $3.15/hour and the B200 at $4.99.
Consumer cards such as the RTX 4090 give you the strongest price-to-performance in the 24GB tier, but the ceiling is real. A 34B model already wants more than 24GB once you add KV cache, and image pipelines with heavy batching fill 24GB faster than people expect. When a workload outgrows the card you picked, the migration cost between platforms is part of the price, because budget marketplaces and premium clouds rarely share instance formats.
Vultr sells its NVIDIA GPUs as hourly cloud instances rather than subscriptions. Each plan bundles the card with a fixed set of vCPUs, RAM and NVMe storage, billing accrues hourly with the monthly rate as a cap, and new accounts get $300 in free credits to test before paying anything. GPU pricing has been moving quarterly across the whole industry, so treat the numbers below as advertised starting rates at the time of writing and confirm the live figure before planning a long run.
NVIDIA A16 — from $0.35/hour
An unusual entry card built around four 16GB GPU dies, aimed at virtual desktops and light per-user workloads
Fits: VDI, dev and test boxes, small-batch embedding jobs
Heads-up: the cheapest entry point on the lineup, but not a training card
👉 Check Vultr's current rate for the A16 plan
NVIDIA L4 — from $0.67/hour
24GB VRAM on an efficiency-focused Ada card with low power draw
Fits: SDXL/FLUX-class image generation, 7B–13B inference, embeddings
Heads-up: the 24GB ceiling pushes larger models up to the L40S
👉 Check Vultr's current rate for the L4 plan
NVIDIA L40S — from $1.19/hour
48GB VRAM, the general-purpose workhorse of the Ada generation
Fits: 30B-class inference, video generation and transcoding, rendering, mid-size fine-tunes
Heads-up: the practical sweet spot when 24GB runs out but 80GB HBM pricing is overkill
👉 Check Vultr's current rate for the L40S plan
NVIDIA A100 80GB — from $1.99/hour
80GB of HBM, the Ampere-generation card built for sustained training
Fits: 70B fine-tuning, multi-day runs, scientific computing
Heads-up: older than the H100, so compare throughput per dollar for inference-heavy work
👉 Check Vultr's current rate for the A100 plan
NVIDIA H100 80GB — from $2.99/hour
80GB of HBM3 with Transformer Engine acceleration for large-model work
Fits: large-model training and fine-tuning, production inference at scale
Heads-up: undercuts hyperscaler list prices, with Google Cloud listing the same card at $3.35 and AWS at $3.50 on recent boards
👉 Check Vultr's current rate for the H100 plan
The catalog rotates as new silicon lands. Some regions carry professional Ada cards, Blackwell-era options appear as supply arrives, and a separate bare-metal line covers whole-machine workloads. If you manage everything as containers, Vultr's managed Kubernetes engine (VKE) handles that side, and the usual platform pieces such as block storage, snapshots, an API and a CLI work the same way they do on Vultr's standard compute.
On raw hourly price, Vultr is rarely the cheapest option, and pretending otherwise would waste your time. What it offers is a middle position: an H100 at $2.99 undercuts hyperscaler list prices, while the platform around the GPU behaves like normal cloud infrastructure rather than a single-purpose box.
The market sorts into three rough groups. At the budget end, marketplace models rent spare or distributed capacity: Vast.ai's H100 sits at $1.65/hour with the A100 80GB at $0.80, TensorDock advertises entry rates from $0.10, and Salad runs distributed consumer cards from $0.02. Those prices are real, but reliability and location vary by host, which matters a lot once a training run spans days.
Specialist GPU clouds cluster in the middle. RunPod pairs per-second billing with Secure and Community tiers plus serverless endpoints, listing an H100 at $2.69. Lambda preinstalls its CUDA and PyTorch stack, charges nothing for egress, and rents an 8×H100 node at $19.92/hour, though it has no spot tier and stays US-focused. Nebius ($2.00) and Hyperstack ($1.95) both push European availability and data residency, Crusoe ($2.45) runs on renewable-powered capacity, Modal targets serverless Python workflows ($3.95 for an H100), and CoreWeave sits at the premium end ($4.76) with InfiniBand clusters and Kubernetes-native scheduling aimed at large teams.
Hyperscalers trade price for breadth. Google Cloud's catalog stretches from a T4 up to GB300 systems, bills per second, allows up to 8 GPUs per instance, and hands new customers $300 in credits. AWS lists its H100 at $3.50 and wraps GPUs in the SageMaker toolchain, while Azure appeals mainly to enterprises already committed to the Microsoft stack.
Against that field, Vultr makes sense when you want one account for the whole project: GPUs for the model work, standard compute and managed Kubernetes for the app around it, block storage for datasets, and 30-plus data center locations to place things near users. It makes less sense if you want the absolute floor price, a managed ML stack out of the box, or mature serverless inference as the core workflow.
Sticker rates mislead because platforms measure time differently, and over a month the difference compounds.
Billing granularity: Google Cloud, RunPod and Modal bill per second. Hostinger rounds up to the next full hour, so a session of 1 hour and 1 second bills as 2 hours, per its own FAQ. Vultr accrues hourly and caps the bill at the listed monthly rate, which protects a long-running instance from stacking up hundreds of separate hourly charges.
Spot and interruptible capacity: current comparison boards put spot discounts at 30–70% below on-demand, and RunPod's own spot tier undercuts its on-demand H100 by roughly half. Interruptible pricing suits checkpointed training and batch jobs, and it's the wrong tool for a latency-sensitive API.
Idle billing: a running instance bills whether or not you're pushing work to it. Scaling to zero with a serverless product, or destroying instances when idle, is where the real savings come from.
Egress and extras: Lambda charges nothing for egress, and most platforms meter bandwidth beyond an included allowance. If your workflow moves large datasets in and out constantly, price that in before comparing headline rates.
Destroying an instance is permanent. Anything not saved to persistent storage disappears with it, so snapshot the disk or move data to block storage before you tear a machine down.
Use someone else's money first. Both Google Cloud and Vultr hand new accounts $300 in free credits, and a couple of hours on the exact card you're considering tells you more than any comparison table.
Create the account and confirm the credit has been applied.
Pick the GPU plan closest to your real workload, in a region near you, after checking availability.
Deploy, SSH in, and run your actual model rather than a benchmark toy. Tokens per second on your own prompt is the number that matters.
Save a snapshot of your setup so redeploying later takes minutes instead of an afternoon.
Destroy the instance when the test ends, then do the math on what a month at your usage pattern would cost.
👉 Create a Vultr account and test its GPUs with the free credit
Entry 24GB-class cards start under $0.50/hour, and the cheapest advertised rates run around $0.02–$0.16 on marketplace and distributed-capacity clouds. Polished mainstream providers charge $0.35–$0.45 for a 24GB card. The 80GB tier spans roughly $1–3/hour on-demand, a B200 starts at $4.99, and spot or community pricing typically shaves another 30–70% off.
On the H100, yes at list price: $2.99/hour against Google Cloud's $3.35 and AWS's $3.50. On the A100 80GB the gap narrows, with hyperscalers listing $1.40–$1.65 and marketplaces running from $0.80. Price alone rarely settles it, though; the $300 trial credit, hourly-capped billing, and having compute, storage and Kubernetes in one account do most of the persuading.
Not in full. A single 80GB card handles 70B inference and light adapter fine-tuning with quantization, while a full fine-tune spreads across multiple GPUs; Lambda's 8×H100 node at $19.92/hour gives a sense of what that tier costs.
It's gone unless it lives on persistent storage. Rebooting restarts the machine and keeps your setup intact; destroying deletes it. Snapshot before teardown, and keep datasets on block storage rather than local NVMe if you plan to redeploy often.
Match the platform to the job and the decision mostly makes itself.
Occasional image generation or small-model inference: grab the cheapest 24GB card you can find. On Vultr that's the L4, and marketplace-style clouds go lower if you accept variable reliability.
Fine-tuning in the 70B class or training for days at a stretch: an 80GB card is the floor, and running a throughput test with Vultr's $300 credit before committing to a long reservation is cheap insurance.
Spiky, per-request inference APIs: serverless GPU products from RunPod or Modal fit that traffic shape better than any standing instance, Vultr included.
A long-running project that needs GPUs plus ordinary cloud plumbing, meaning compute, Kubernetes, storage and load balancers in one account: this is where Vultr's hourly-capped model and 30-plus regions earn their keep.
One last practical check before you deploy anywhere: confirm the card is actually in stock in your chosen region, and re-check the live rate, because GPU prices have moved repeatedly over the past two years. 👉 See Vultr's full GPU lineup and live pricing