简体中文
English
Website review
This review may include AI-generated content and is not individually verified. Domain, certificate and other facts come from the observed data on this page.
What is gpuperhour.com?
gpuperhour.com is a live cloud GPU price comparison site. It tracks per-GPU-hour rates from 31 providers and shows the cheapest current secure on-demand offer for each GPU model, alongside a 30-day historical catalog average.
What you'll find there
- A market snapshot table covering 14 popular GPUs, including H100 (SXM5, PCIe, NVL), H200, B300, A100, L40S, RTX 5090, RTX 4090, RTX A6000 and MI300X.
- For each GPU: the lowest current price, the provider offering it, the 30-day average, and how many providers have stock.
- Guidance pages on choosing a GPU for a workload and renting by the hour.
- A downloadable price history / CSV index for tracking trends over time.
What its numbers mean, and what they don't
The "current price" column covers only secure on-demand instances observed within roughly 15 minutes of the last refresh. Spot, reserved and peer-to-peer marketplace listings are tracked but excluded from that column, because mixing them would compare unlike things. The historical figures are medians of daily catalog averages over 30 days and include every pricing type and availability state — so they are a catalog baseline, not a record of what you could actually have paid.
A practical example: the site lists H100 SXM at $2.27/hr as the cheapest current offer, while the 30-day catalog average for all H100 listings sits at $2.18/hr. The gap is not a price rise — it reflects the different populations being averaged. Treat the two columns as separate signals.
How to use it
Start with the snapshot to see which GPU classes have real competition and stock, then click through to the provider to confirm the rate, region and terms before committing. Availability changes quickly, so a listed price is a lead worth verifying, not a quote.
If you want a second reference point for on-demand rates, Lambda and RunPod publish their own pricing directly, and Vast.ai is a marketplace whose host-set prices sit outside the secure on-demand tier this site compares.
How do I find the cheapest H100 or H200 cloud GPU price right now?
Check the live secure on-demand listings for H100 and H200 on GPUPerHour, then confirm the price and availability directly with the provider before you book. The snapshot in your evidence shows H100 SXM5 at $2.27/hr (Vast.ai) and H200 NVL at $3.43/hr (QuantaCloud), but those are point-in-time observations, not standing quotes.
How to read the comparison
GPUPerHour normalises every listing to a price per GPU per hour and separates secure on-demand from spot, reserved and peer-to-peer marketplace offers. That matters because a cheap marketplace host and a secure on-demand instance are different products: one can disappear mid-training, the other carries a contract and a support path.
| What you're comparing | Use the "current price" column | Use the 30-day historical average |
|---|---|---|
| Spot vs on-demand | No — spot is excluded | No — it mixes pricing types |
| Peer-to-peer vs secure | No — peer-to-peer is excluded | No — availability states are mixed |
| Trend over a month | No | Yes, as a rough sense of direction |
The historical averages include unavailable and spot listings, so they are a catalog median, not a record of what you could have paid. Treat them as context, not as a savings estimate.
A practical booking sequence
- Filter to H100 or H200 and to secure on-demand only.
- Note the cheapest price and the provider's stock count.
- Open the provider's own pricing page and re-check the rate, region and GPU variant (SXM5, PCIe, NVL and SXM differ in interconnect and memory).
- Confirm whether the price is per GPU or per instance, and whether storage, egress and minimum rental periods are extra.
- Only then commit, and prefer a short initial rental if you're testing a new provider.
Decision criteria
- Short experiments or fine-tuning: H100 SXM5 is usually the better value; the cheapest listings cluster there.
- Large-model inference with big memory needs: H200's larger memory can reduce the number of GPUs you need, which can beat a lower per-hour rate on a smaller card.
- Price-sensitive batch jobs: if your workload tolerates interruption, spot can be far cheaper, but compare it separately rather than against on-demand numbers.
- Anything long-running: weigh provider reliability and support over a few cents per hour; a failed job costs more than the difference.
For a second opinion on general market ranges, Vast.ai and RunPod publish their own live rates, which is useful for sanity-checking a comparison site's numbers.
Which cloud GPU should I choose for AI training versus inference?
For AI training, prioritise GPUs with large memory and high interconnect bandwidth; for inference, prioritise cost per GPU-hour and availability. On GPUPerHour's snapshot, training-class options like the H100 SXM5 (cheapest listed at $2.27/hr) and H200 (from $3.43/hr) sit well above inference-friendly cards such as the RTX 4090 ($0.74/hr) or RTX 5090 ($0.53/hr). The trade-off is straightforward: training throughput and multi-GPU scaling versus inference cost efficiency.
H3 Quick comparison
| Workload | Typical fit | Example from the page | Main trade-off |
|---|---|---|---|
| AI training (large models, multi-GPU) | H100, H200, B300 | H100 SXM5 $2.27/hr; H200 NVL $3.43/hr; B300 $7.89/hr | Higher hourly cost, but faster time-to-train |
| Fine-tuning (medium models) | A100 80GB, L40S | A100 SXM4 80GB $0.67/hr; L40S $0.97/hr | Memory capacity vs. newer architecture |
| Inference (high volume, latency-sensitive) | RTX 4090, RTX 5090, RTX A6000 | RTX 4090 $0.74/hr; RTX 5090 $0.53/hr; RTX A6000 $0.44/hr | Lower cost, but less memory and scaling headroom |
| Inference (large models, high throughput) | H100, MI300X | H100 PCIe $2.59/hr; MI300X $3.39/hr | Cost per token vs. latency targets |
H3 How to decide
Start from the model size and whether you need multi-GPU communication. If you are training from scratch or doing full fine-tuning of a large model, the H100 SXM5 or H200 class is the practical baseline; the page shows only a handful of providers with stock for each, so availability matters as much as price. For inference, the RTX 4090 and RTX 5090 are the cheapest listed options, but check memory: a 24GB card may not fit a large model without quantisation or sharding.
A concrete scenario: a team fine-tuning a 7B model can often use an A100 80GB at $0.67/hr rather than an H100, saving roughly $1.60/hr per GPU while still fitting the model and optimiser state. The same team serving that model to users might run it on an RTX 4090 at $0.74/hr if latency and throughput targets are modest.
Next step: use the site's per-GPU-hour normalisation to compare like for like, then verify current stock before booking, since the page notes availability can change and its snapshot reflects a specific observation time. For broader market context, GPUPerHour tracks 31 providers, and you can cross-check listed providers directly, such as Vast.ai and RunPod.
How do I rent a cloud GPU by the hour securely?
Rent securely by treating the booking as three separate checks: the provider's security posture, the actual per-GPU-hour price after normalisation, and whether the listing is on-demand rather than spot or peer-to-peer. The last point matters most, because marketplace and interruptible listings are often the cheapest headline number but can be reclaimed mid-run or expose you to another tenant's host.
GPUPerHour tracks live offers from 31 providers and deliberately restricts its current-price column to secure on-demand instances, excluding spot and peer-to-peer hosts. That exclusion is the useful part: it tells you the comparison is like-for-like, so a $2.27/hr H100 SXM5 figure is not being flattered by a preemptible listing that may vanish.
A practical booking sequence
- Pick the GPU by workload, not by price. Large-model training and fine-tuning want H100/H200-class SXM parts with fast interconnect; inference and rendering often run fine on L40S, RTX 4090 or RTX PRO 6000. A cheap card that lacks the memory or interconnect you need is not cheap.
- Filter to secure on-demand. Skip spot, reserved and peer-to-peer marketplace listings if you need a stable session. Expect to pay more per hour for that stability.
- Normalise to per-GPU-hour. Multi-GPU instance prices must be divided by GPU count before comparing. A $9.08/hr eight-GPU instance is $1.14 per GPU-hour.
- Check stock and freshness. Availability moves; confirm the provider still shows live stock at the moment you book.
- Verify isolation and data handling. Prefer providers with clear tenant isolation, private networking and encrypted storage. For sensitive data, ask whether the host is single-tenant and where the machine sits.
- Cap your exposure. Start with a short session, confirm throughput and network performance, then commit to longer runs. Set a spending limit or auto-shutdown so an idle instance does not bill overnight.
What the price table actually tells you
On the snapshot dated 2026-10-01, the spread across 14 popular GPUs was wide. Cheapest current secure on-demand offers included H100 SXM5 at $2.27/hr (Vast.ai), H200 NVL at $3.43/hr (QuantaCloud), A100 SXM4 80GB at $0.67/hr (Vast.ai), RTX 4090 at $0.74/hr (RunPod) and MI300X at $3.39/hr (Hot Aisle). Note that the 30-day historical catalog averages sometimes sit below the current price — for example RTX 4090 at $0.33/hr against a current $0.74/hr — because those averages include unavailable and spot listings. Do not read them as savings you can actually capture.
A concrete scenario
Suppose you need to fine-tune a 7B model over a weekend. You would look at H100 or A100 80GB class cards, filter to secure on-demand, and compare the per-GPU-hour rate plus storage and egress. A single A100 80GB at roughly $0.67–0.68/hr is a reasonable starting point; two of them doubles the hourly cost but may halve wall-clock time, which is usually the better trade if your deadline is fixed.
Next step: check the live index and daily history at GPUPerHour before booking, then confirm the same listing directly with the provider, since availability can change between the snapshot and your session. For a second opinion on provider reputation, Vast.ai and RunPod publish their own instance terms, which are worth reading alongside any comparison.
Why do GPU prices differ between providers for the same GPU model?
Prices differ because the headline rate reflects a bundle of choices the provider made — not just the chip. The GPU model is only one input into what you actually pay per GPU-hour.
The main reasons the same GPU costs different amounts
- Pricing type. On-demand, spot, reserved, and peer-to-peer marketplace listings aren't comparable. A cheap spot rate can vanish mid-run; a reserved rate trades flexibility for a lower hourly figure. GPUPerHour tracks these but excludes non-on-demand, non-secure listings from its current-price column for that reason.
- Security and tenancy tier. A dedicated bare-metal instance, a virtualized secure instance, and a shared marketplace host carry different isolation, compliance, and overhead costs.
- Hardware variant. "H100" covers SXM5, PCIe, and NVL form factors with different interconnect and memory bandwidth. The same name can mean meaningfully different performance.
- Hosting economics. Power cost, region, cooling, and whether the provider owns or resells capacity all move the floor price.
- Stock pressure. Scarcity raises the clearing price. GPUPerHour notes that availability can change before booking, so a low rate with one provider in stock is fragile.
- Instance configuration. Providers normalize to price per GPU-hour by dividing a multi-GPU instance price by GPU count, which can hide differences in CPU, RAM, storage, and network bundled in.
Reading a comparison table correctly
The gap between "current price" and "historical catalog average" on GPUPerHour is a useful signal, but it measures different things. Current price is a live secure on-demand offer; the historical figure is a 30-day median of daily catalog averages including unavailable and spot listings. A current price below the historical average is not automatically a saving — it may just mean the cheap listing type was filtered out of one column and not the other.
| What you see | What it usually means | What to check |
|---|---|---|
| Current price far below historical avg | Different pricing type or a single-stock listing | Whether it's on-demand and secure |
| Multiple providers, similar price | Competitive, commoditized tier | Region, egress, storage costs |
| Only one provider with stock | Scarcity premium likely | Whether you can wait |
| Big spread across variants of one name | SXM vs PCIe vs NVL differences | Interconnect and memory bandwidth |
A practical next step
Pick your GPU model, then filter to secure on-demand only, and note how many providers have stock. If only one or two do, treat that price as volatile and budget for a fallback. Then compare the all-in cost for your actual workload — a slightly higher hourly rate on a provider with faster interconnect can finish a multi-GPU training job sooner and cost less overall.
For a second opinion on live rates, GPUPerHour publishes per-GPU-hour comparisons across providers, and Vast.ai and RunPod list their own current marketplace and on-demand pricing directly.
How are the live GPU prices and historical averages on this site calculated?
Prices on GPUPerHour come from scraping 31 providers' public pricing and inventory endpoints, then normalising each instance to a single comparable number.
Live prices
- Only secure on-demand listings count. Spot, reserved, and peer-to-peer marketplace offers are tracked but excluded from the headline figures because their terms differ too much to compare directly.
- A listing must be marked available by a provider that tracks live stock, and observed within 15 minutes of the page's last refresh.
- Multi-GPU instances are divided by GPU count, giving a price per GPU per hour.
- The summary refreshes at most every 5 minutes, so numbers can move while you read.
Historical averages
- These are medians of daily catalog averages over the last 30 days.
- Unlike the live column, they include every pricing type and availability state — spot, reserved, unavailable listings, everything in the catalog.
- Missing days are dropped rather than filled in.
The practical consequence: a historical average is not a record of what people actually paid, and it is not a savings figure. It is a median of catalog asking prices. That is why several GPUs show a current price above their 30-day average — B300 SXM6 at $7.89/hr against a $5.97 average, RTX 4090 at $0.74 against $0.33 — while others sit below, such as A100 SXM4 80GB at $0.67 against $1.06.
If you are budgeting, use the live figure for what you would pay today and treat the average as a rough sense of the listing landscape. A useful next step is to check the daily history and CSV downloads on the site's price index, which let you see whether a current quote is an outlier or part of a trend. For a longer view of how provider pricing has shifted, Cloud GPU Pricing keeps the underlying series, and you can sanity-check specific machines against RunPod or Vast.ai, both of which appear among the tracked providers.
Cloudflare
Google Workspace
User reviews (0)