Rent an H100 on Google Cloud and the cheapest on-demand rate is $5.38 per GPU-hour, sold as an 8-GPU machine. Rent the same card from Lium, a GPU marketplace, and it is $1.80. Same chip, same hour, three times the price.
That gap is not one odd card. We compared every GPU that both a hyperscaler (AWS, Google Cloud, Azure, Oracle Cloud) and at least one other cloud rent on demand, from the live prices FastGPU collects across the market. On October 29, 2026 at 16:20 UTC there were 18 such cards, and the cheapest big-cloud rate for a card was a median of 3.2x the cheapest rate at which it could be rented anywhere.
Card by card
The cheapest hyperscaler rate for each card against the market floor (the cheapest on-demand or community price at which the card could be rented that morning, in stock wherever any offer was), per GPU-hour:
| GPU | Cheapest hyperscaler | Market floor | Hyperscaler / floor |
|---|---|---|---|
| V100 | $2.49, Google Cloud (8-GPU instance) | $0.07, Clore.ai | 35.6x |
| L4 | $0.71, Google Cloud | $0.09, Vast.ai | 7.9x |
| RTX PRO 4500 | $2.52, AWS | $0.35, Clore.ai | 7.2x |
| A100 40GB | $2.16, Google Cloud (16-GPU instance) | $0.35, Vast.ai | 6.2x |
| L40S | $1.86, AWS | $0.34, Vast.ai | 5.5x |
| A100 80GB | $3.22, Google Cloud (8-GPU instance) | $0.60, Vast.ai | 5.4x |
| A10 | $1.01, AWS | $0.24, Vast.ai | 4.2x |
| MI300X | $6.00, Azure (8-GPU instance) | $1.71, TensorWave | 3.5x |
| RTX PRO 6000 | $3.36, AWS | $1.00, Vast.ai | 3.4x |
| MI355X | $8.60, Oracle Cloud (8-GPU instance) | $2.85, TensorWave | 3.0x |
| T4 | $0.36, Google Cloud (4-GPU instance) | $0.12, Vast.ai | 3.0x |
| H100 SXM | $5.38, Google Cloud (8-GPU instance) | $1.80, Lium | 3.0x |
| B200 | $14.00, Oracle Cloud (8-GPU instance) | $4.89, DeepInfra | 2.9x |
| H100 NVL | $6.98, Azure | $2.59, RunPod | 2.7x |
| H200 | $7.91, AWS (8-GPU instance) | $3.00, Vast.ai | 2.6x |
| B300 | $15.00, Oracle Cloud (8-GPU instance) | $5.89, DeepInfra | 2.5x |
| GB200 | $16.00, Oracle Cloud (4-GPU instance) | $8.00, GMI Cloud | 2.0x |
| GB300 | $18.00, Oracle Cloud (4-GPU instance) | $12.93, Verda | 1.4x |
Each ratio is the first price divided by the second. The live table recomputes all of it on every visit and links every rate to the page that proves it.
Two patterns stand out. The gap is widest on older and mid-range cards (V100, L4, A100, L40S), where marketplaces full of hosts with paid-off hardware set the floor. And it narrows on the newest silicon: a GB300 costs only 1.4x more at Oracle than at the cheapest GPU cloud, because few providers rent one at all and none of them is cheap.
Cloud by cloud
Each big cloud's own on-demand rate against the floor, across every card it lists, gives a median per cloud (on the same snapshot):
- AWS: 4.3x across 12 cards
- Google Cloud: 5.4x across 7 cards
- Azure: 5.5x across 11 cards
- Oracle Cloud: 4.0x across 14 cards
A cloud with few cards is a small sample, so read these as a shape rather than a ranking. Each link goes to that cloud's full pricing guide, where every rate is a live figure.
What the premium buys, and what the floor leaves out
The comparison is honest only if you know what each side includes.
The floor is often a marketplace. Vast.ai, Clore.ai and similar platforms rent machines that independent hosts own. Network quality, data security and uptime vary by host, and none of them publishes an uptime SLA for a single machine. Across the last 25 daily fixings, Vast.ai alone held the cheapest price on 41% of GPU-days, and the cheapest provider for a card changed about one day in six.
Many hyperscaler rates come only as whole machines. The per-GPU rate on an 8-GPU instance is real, but the smallest bill is eight times it.
The big clouds charge for data leaving them. Moving 1,000 GB out costs $90.00 at AWS and $87.00 at Azure, (each after the first 100 GB a month), and $120.00 per 1,000 GiB at Google Cloud. Many GPU clouds charge nothing for egress. For a job that downloads a fine-tuned model or serves files, that can outweigh the GPU bill.
Meters differ. AWS and Google Cloud bill by the second, Azure by the minute; some GPU clouds bill by the hour, and several keep billing a stopped machine's disks. The hidden costs page lists each provider's egress, meter and uptime terms, each linked to the provider's own page.
What the premium does buy: region coverage, existing enterprise contracts and compliance, the surrounding platform (storage, networking, managed services), and published SLAs. For a team already committed to one of them, the premium is often the price of not moving data and not adding a vendor. For a portable training or inference job, it is the first number worth questioning.
How the numbers are made
FastGPU collects posted on-demand prices from each provider's own API, rate card or price list, every day and again while pages are being read, and normalises them to a per-GPU-hour rate. Spot capacity is excluded from both sides. A machine its provider no longer rents to new customers is left out, even when a price list still shows it. The methodology has the details, and the recorded prices are an open dataset under CC BY 4.0.
These figures are a snapshot from October 29, 2026. The live comparison shows today's.
Top comments (1)
A 3.2x spread looks like clean arbitrage on a spreadsheet, yet the practical ceiling is interruptibility. When an unmanaged host resets mid-run, you restart from the latest checkpoint. If a distributed job loses an hour of progress and pulls weights again over commodity peering, that hardware discount vanishes into duplicated compute and engineering triage. Marketplace pricing works best for stateless batch workloads and local evals saving state after every shard, whereas multi-node jobs end up treating the hyperscaler mark-up as an insurance policy against node churn.