The H100 price per hour depends on who you ask. On 2 November 2026 Silicon Data's neocloud H100 index read $2.94 per GPU-hour. Published on-demand rates read on 3 November ran from $1.99 to $9.80. Both numbers are real. They measure different things.
What providers publish
A rate card is one provider's price for one configuration, on demand, with no commitment. These are the H100 lines on the providers' own pricing pages.
Pages were read on 3 November 2026 unless a row says otherwise. Prices move. Nebius showed $3.85 on 8 October. It showed $4.50 on 3 November. Lambda lists $3.29 for a single PCIe GPU and $4.29 for a single SXM GPU, so even one provider has no single H100 price.
The spread is wide. The top of the table is almost five times the bottom. Part of that is the product: a shared PCIe card is not a dedicated, networked SXM node. Part of it is the buyer the provider has in mind.
What a live index reads
An index tries to turn many rate cards and rentals into one number. Silicon Data publishes one for H100, once per business day. It read $2.94 on 2 November 2026. Its daily readings over the last month:
Silicon Data also publishes a separate hyperscaler reading for H100. Its page showed $7.12 when read on 3 November. The company says neocloud and marketplace capacity typically prices below hyperscaler on-demand rates.
Here is how Silicon Data describes its own method. It converts each rental record to a per-hour, per-GPU rate. It filters out reporting errors. It adjusts each record to a benchmark contract, for rental type, geography, CPU platform and GPU variant. It then takes a weighted average across providers. The company says it covers 95% of the neocloud providers it tracks and about 80% of the global rental market. These are the company's own claims, not tested here.
Ornn publishes a second index. It says its Compute Price Index is built from settled transactions and updated hourly. That is its own description, from a page dated 2 November 2026. No current Ornn figure was readable, so none is quoted here.
An index sits below most rate cards because it blends many sources, including marketplaces. That is why $2.94 and $3.99 can both be honest.
What is and is not in the hourly rate
Two quotes at the same hourly rate can buy different things. Check these before you compare.
- Host resources. Lambda's 8-GPU line works out to 26 vCPUs, 225 GiB of memory and 2.75 TiB of disk per GPU, from its listed totals. DigitalOcean lists 20 vCPUs and 240 GiB of memory per Droplet. Other pages do not say.
- Storage. Nebius charges $0.08 per GiB-month for its shared filesystem. Together charges $0.16. Hyperstack charges for storage volumes and for a public IP address. RunPod charges for disks, and more for idle volume disk.
- Data transfer. Nebius and Hyperstack state no charge for network traffic. DigitalOcean includes 15,000 GiB of transfer per Droplet. Other pages are silent.
- Place. Fireworks says region-restricted deployments cost 1.5 times the standard rate. If your data must stay in one country, ask for that price.
- Tax. Lambda and Nebius state their prices exclude tax.
- Billing. DigitalOcean bills per second with a five-minute minimum, and still bills a powered-off Droplet.
Why the quote is not the contract
A rate card is for today, for one GPU, with no promise. A large buyer changes all three. Providers publish part of that change. Together lists $3.99 on demand, then $3.69 for 3 to 30 days, $3.45 for 31 to 90 days and $3.19 for 91 to 180 days. For longer terms it says to contact sales. DigitalOcean lists $3.26 for a 12-month H100 commitment against $4.41 on demand. Hyperstack lists reserved H100 SXM from $2.72 against $3.20 on demand.
These are still published rates. They are not what any buyer signed. Volume, prepayment, start date and place are settled in private contracts, and nobody publishes them. Casitel holds no executed prices to show you here, and this post does not guess at them. The list price is not the price explains how those deals close.
So read the index as the direction of the market. Read a rate card as the opening position. Read neither as your price.
How to read an H100 quote today
- Convert the quote to dollars per GPU-hour, all in.
- Match the chip: SXM and PCIe are not the same product.
- Match the scope: network, storage, transfer, region and tax.
- Match the term and the start date.
- Compare it with the index and with rate cards for the same chip, then ask what is behind any gap.
Casitel OS reads these rate cards every day. On 3 November 2026 the median H100 list price across nine GPU clouds was $3.90 an hour.
Sources
- Silicon Data, H100 Silicon Index page, read 3 November 2026
- Silicon Data, Silicon Index overview with H100 hyperscaler reading and method notes, read 3 November 2026
- Ornn, home page describing the Ornn Compute Price Index, 2 November 2026
- Lambda, pricing page, read 3 November 2026
- Nebius, prices page, read 3 November 2026
- DigitalOcean, GPU Droplets pricing, read 3 November 2026
- Hyperstack, GPU pricing page, read 3 November 2026
- RunPod, pricing page, read 3 November 2026
- Together AI, pricing page, read 3 November 2026
- GMI Cloud, pricing page, read 3 November 2026
- Fireworks AI, pricing page, read 3 November 2026
- Google Cloud, Vertex AI pricing, read 8 October 2026
- Casitel OS, H100 list prices across nine GPU clouds, 3 November 2026