Skip to content
senn-techsenn-tech
Back to blog
AI Hardware
Local AI and GPU hardware2026-11-15· By Franz Senn

One H100, 2.69 to 15.98 USD an hour: what we found in the price lists ourselves

On November 9, 2026 109-gpu-pricing sat on GitHub, discussed with 173 points (as of November 15) on Hacker News. The headline of that thread: the same H100 costs 2.69 or 12.29 dollars an hour depending on where you rent it, 4.57 times the money for one and the same chip. We pulled the two largest positions ourselves, through the Azure pricing API and the RunPod pricing page. In our numbers the distance is bigger, 2.69 to 15.98 dollars. And one row inside the same price list brings it almost back down.

Every number here hangs on one unit, the GPU-hour. It means the usage time of exactly one accelerator over one hour, whether a provider rents a single card or a machine carrying eight of them. A rental price turns comparable at the same form factor, region and booking type. A price per machine-hour is no GPU-hour, a spot rate is no on-demand rate, and an H100 with 80 gigabytes is a different product from an H100 NVL with 94 gigabytes.

NVIDIA logo
NVIDIA makes the H100; the rental prices in this post are set by RunPod and Azure, not NVIDIA. (Quelle: Simple Icons, CC0)

What we measured

Unit: US dollars per GPU-hour. Measurement date: November 15, 2026, from about 15:30 UTC (Azure and AWS last queried around 16:00 UTC). Every price in this post was queried that day. Method: scripted queries from a workstation against the Azure retail prices API (price type Consumption, region list in USD, all result pages), the AWS price file and the RunPod pricing page (displayed price plus the page's price attributes for Community and Secure). Azure lists the price per virtual machine; the GPU-hour arises by dividing through the card count Azure documents, so it is our own arithmetic and shows up as derived in the table. Every row is an on-demand list price without discounts, without reserved capacity and without taxes.

Provider and SKUForm factorPrice per machine-hourGPU-hourNote
RunPod Community Cloud, H100 SXMSXM, 80 GB2.69 USD2.69 USDmachines of third parties
RunPod Secure Cloud, H100 SXMSXM, 80 GB4.39 USD4.39 USDown data centers, per the provider
RunPod cluster rate, H100 SXMSXM, 80 GB4.79 USD4.79 USDmulti-GPU cluster
RunPod Community / Secure, H100 NVLNVL, 94 GB2.59 / 3.19 USD2.59 / 3.19 USDdifferent form factor, not an H100 SXM price
Azure ND96is_H100_v5, West EuropeSXM, 80 GB115.034 USD14.38 USD (derived, divided by 8)8 cards, without InfiniBand
Azure ND96isr_H100_v5, West EuropeSXM, 80 GB127.816 USD15.98 USD (derived, divided by 8)8 cards, 400 Gb InfiniBand per card
Azure ND96isr_H100_v5 Spot, West EuropeSXM, 80 GB23.6204 USD2.95 USD (derived, divided by 8)interruptible
AWS p5.48xlarge, us-east-1SXM, 80 GB55.04 USD6.88 USD (derived, divided by 8)on-demand Linux, in effect from November 1, 2026, price file of November 14, 2026
H100 SXM 80 GB: US dollars per GPU-hour, own measurement of November 15, 2026RunPod Community2.69 · 2.69 USDAzure Spot West Europe2.95 · 2.95 USD, derived, interruptibleRunPod Secure4.39 · 4.39 USDRunPod Cluster4.79 · 4.79 USDAWS p5.48xlarge us-east-16.88 · 6.88 USD, derivedAzure without InfiniBand14.38 · 14.38 USD, derivedAzure with InfiniBand15.98 · 15.98 USD, derived018
(Quelle: Own measurement on November 15, 2026: RunPod pricing page (dated 12 Nov by the page), Azure retail prices API West Europe (Consumption, USD), AWS price file us-east-1 (dated 14 Nov))

The dataset names 2.69 dollars (RunPod Community) against 12.29 dollars (Azure ND96isr_H100_v5, East US) for that same group and calculates the 4.57-fold from it. Our measurement for West Europe gives 5.94 times the cheapest row. The regional difference sits between two data points here: the dataset names 12.29 dollars per GPU for East US (as of November 9), which equals 98.32 dollars per machine-hour (derived, times 8), and we measure 127.816 dollars in West Europe (November 15). Which part of that is price movement and which part is a regional surcharge, we cannot separate from two data points; two regions and two days lie between the figures. Our table carries one row per rate, while the dataset counts RunPod as a single company.

Three settings move the gap

Form factor. The string "H100" stands for at least three cards in these lists. Azure carries the eight-card machine ND H100 v5 with H100 SXM and 80 gigabytes, documented per card with NVLink 4.0 and its own 400 Gbps connection on NVIDIA Quantum-2 CX7 InfiniBand with GPU Direct RDMA. Kept separate in the same list are the NCads-H100-v5 machines with one or two H100 NVL of 94 gigabytes each, without InfiniBand: there we measure 9.08 dollars per card in West Europe and 9.074 in Germany West Central, and RunPod rents that card for 2.59 dollars (Community) and 3.19 dollars (Secure). Whoever compares a number per "H100" compares different products.

Networking. For the same eight-card machine Azure lists four variants at unequal prices in West Europe: ND96isr (the r stands for the RDMA equipment) at 127.816 dollars, and ND96is, ND96is_flex and ND96is_noIB at 115.034 dollars each. The difference is 12.782 dollars per machine-hour, an 11.1 percent surcharge, or 1.60 dollars per GPU-hour, derived from the two list entries. InfiniBand is therefore a line item of its own in the offer, and you can ask for it with or without a price.

Region. In the price list of Germany West Central we pulled every result page of the query (eight pages, 7,948 virtual-machine entries, price type Consumption) and filtered by script for "H100" in the SKU name: only NC40ads and NC80adis machines remain, no ND-H100-v5 machine. The small NC machine with one H100 NVL stands there at 9.074 dollars (West Europe: 9.08), and the large machine with eight cards and InfiniBand is absent. Anyone who wants to reserve inside Germany rents in another region and takes data movement and contracts along with it.

What the dataset leaves out and what the debate says

The dataset states its own limits: on-demand list prices only, one region per provider, GPU counts from the product documentation instead of the price tag, and the remark that availability carries no price. It left Lambda out, because the rates listed there only start to apply on November 16, 2026. We pulled the AWS row ourselves from the official AWS pricing file for us-east-1 (481 megabytes, file dated November 14, 2026, retrieved November 15, 2026): the on-demand rate for Linux p5.48xlarge stands there at 55.04 dollars per machine-hour, in effect from November 1, 2026, so the figure is confirmed. The attributes of that entry name 8 GPUs and 640 GB HBM3, so 80 gigabytes per card, with no model name in them; the GPU-hour is the division by 8.

A row that appears in no headline closes part of the gap: Azure prices the same eight-card machine in West Europe as Spot at 23.6204 dollars per machine-hour (Linux), which lands at 2.95 dollars per GPU-hour, 26 cents above the cheapest RunPod row. The gap then depends less on the provider than on the booking type. Spot means Azure takes the capacity back at any time. For a training run that can be interrupted and resumed afterwards that is a good deal. For a service that promises response times at eight in the morning it is the wrong product.

The comments under the thread, community voices without a measurement log, aim at two places. One participant calls any load worth mentioning on RunPod Community or Vast negligent, because the host machines could look into the virtual machines. That stays a suspicion without a measurement log, and it concerns the Community Cloud with its machines of third parties; the comment leaves the Secure rate per the provider out. Another reports that on his last attempt he got no card at all above 20 gigabytes of video memory, and by then the price was secondary. Price lists do not show availability.

What this means for the next inquiry

We run our own cards: four RTX 5090 in the production inference host, one RTX 5090 in the standby host and one RTX 4060 Ti in a KVM machine, three GPU hosts per our host inventory of November 10, 2026 (taken from the inventory file, not recounted). RunPod lists an RTX 5090 in the Secure Cloud at 1.39 dollars an hour (pricing page, November 15, 2026), so one rented hour there sits below every H100 row above; whether the bill works out for our own load is a different question, and the comparison of token prices against GPU rentals from October works it out. We read the H100 prices above from the price lists only; we have not measured the performance of these cards.

Six questions for any provider: which form factor and which memory, InfiniBand yes or no and what the surcharge costs, which region, which booking type (on-demand, reserved, spot), which promise on availability, and how the price per GPU-hour arises from the price per machine. Whoever gets those answers can compare. For the choice between a cloud hour and a card of our own, one figure is missing on top: our own utilisation, meaning how many hours a day gets computed, what power and cooling cost, and who looks after the machine. The overview of local AI and GPU hardware carries our posts on this, and the arithmetic for a single case belongs into an offer, for instance into the advisory on own AI models. Open stays how long the large regions deliver their eight-card machines at these list prices; that day we measured price lists, and a list price holds no delivery promise.

Open points: The Community price of the RunPod RTX 5090 could not be assigned unambiguously on the page, so we quote the Secure price only. We did not query Lambda or other providers. Availability, taxes, discounts and Windows prices are not covered. We did not re-measure the dataset's own figures (East US, as of November 9), and the Hacker News comments are opinions without a measurement log. Price lists change; every number is valid for November 15, 2026.

Read next: NVIDIA snapshot: restore versus cold start · AMD EPYC 9006 Venice with 256 cores · Inkling 2 Small against our VRAM limit

Further Reading

Labels: read and measured ourselves on 15 Nov 2026 unless noted.

Questions?
What does an H100 really cost per hour?+

On November 15, 2026 we measured 2.69 to 15.98 US dollars per GPU-hour in the price lists for one and the same card type: RunPod Community Cloud at 2.69, RunPod Secure Cloud at 4.39, Azure West Europe at 14.38 without InfiniBand and 15.98 with it, each derived from the Azure price for a machine holding eight cards. The Azure spot list sits at 2.95 dollars, and that capacity comes back at any time. Compare one card type, one region and one booking type.

Why is the eight-card machine missing from the German Azure region?+

The price list of Germany West Central carries no ND-H100-v5 machine at all, measured on November 15. Listed there are only the smaller NC machines with one or two H100 NVL at 9.074 dollars per card. The large machine with InfiniBand shows up in the West Europe list.

Does the price tell me whether I will get the card?+

No. The cheapest row in the list carries the weakest promise: RunPod Community Cloud runs on machines of third parties with no data center commitment, and Azure Spot can end at any moment. The publisher of the dataset writes himself that availability carries no price. Anyone who has to hold a date needs a commitment, and that one lives in the contract.