Cheapest GPU to run a 70B model.
Prices checked
A100 PCIe · Community Cloud at RunPod does this job for $1.19 per hour served, the cheapest of 107 machines that can, checked .
Cheapest 10
all 162 rentals- 1
A100 PCIe · Community Cloud
Marketplace
RunPod · 1x A100 80GB PCIe · 80 GB$1.19per hour served - 2
A100 SXM · Community Cloud
MarketplaceRunPod · 1x A100 80GB SXM · 80 GB$1.39per hour served - 3
RTX Pro 6000 · Community Cloud
MarketplaceRunPod · 1x RTX PRO 6000 · 96 GB$1.69per hour served - 4
A100 PCIe · Secure Cloud
Own machinesRunPod · 1x A100 80GB PCIe · 80 GB$1.79per hour served - 5
A100 SXM · Secure Cloud
Own machinesRunPod · 1x A100 80GB SXM · 80 GB$1.79per hour served - 6
H100 PCIe · Community Cloud
MarketplaceRunPod · 1x H100 PCIe · 80 GB$1.99per hour served - 7
vbm-72c-480gb-gh200-gpu
Own machines
Vultr · 1x GH200 · 96 GB$1.99per hour served - 8
NVIDIA A100 PCIe
Own machines
Crusoe · 1x A100 80GB PCIe · 80 GB$2per hour served - 9
1A100.22V · 1x A100 SXM4 80GB
Own machines
Verda (DataCrunch) · $1.12 on spot$2.25per hour served - 10
1x NVIDIA GH200
Own machines
Lambda · 96 GB$2.29per hour served
Other ways to pick
What matters
- Memory before anything else
- A model that does not fit does not run slowly, it does not run. Memory is the first filter and speed is the second, which is why this page counts gigabytes before it counts anything else.
- 80 GB holds it in 8-bit
- One H100, H200, A100 80GB or MI300X takes the whole model at 8-bit weights, with room for a short context. It is the cheapest shape that serves at all.
- 160 GB holds it in 16-bit
- Two 80 GB boards, or one 192 GB board, run the model at full precision. Where the GPUs are in one instance they talk over the node's own links rather than over the network.
- What the context costs
- Every conversation in flight takes memory of its own on top of the weights, so a board that fits the model exactly serves one request at a time.
How we count
A 70-billion-parameter model has to fit in GPU memory before it answers anything, which rules out most of this market and decides the price of the rest. We count only machines with 80 GB on one board, or GPUs adding up to 160 GB, and 107 of the 162 rentals we read qualify. The cheapest of them is $1.19 per hour served at A100 PCIe · Community Cloud, which rents for $1.19 per GPU-hour. They are ranked by what the work costs rather than by the hourly rate, because a faster card at twice the price can still be the cheaper way to buy the job.
A 70-billion-parameter model is about 70 GB of 8-bit weights and about 140 GB at 16-bit, before anything is left over for the context window. We count a single board of 80 GB or more, which holds the 8-bit weights, and any instance whose GPUs add up to 160 GB or more, which holds the 16-bit ones. The price is what the whole instance bills for an hour, because that is what you rent. Each price is that work at the machine's on-demand rate: A100 PCIe · Community Cloud rents for $1.19 an hour, and $1.19 per hour served is what the work costs on it.
- Marketplace
- The provider matches you with someone else's machine.
- Own machines
- The provider owns or leases the machines.
Frequently asked
- What is the cheapest GPU to run a 70b model?
- A100 PCIe · Community Cloud at RunPod, $1.19 per hour served. That is 1x A100 80GB PCIe with 80 GB of memory each, renting for $1.19 per GPU-hour. Read on 1 Nov 2026.
- How is the cost worked out?
- A 70-billion-parameter model is about 70 GB of 8-bit weights and about 140 GB at 16-bit, before anything is left over for the context window. We count a single board of 80 GB or more, which holds the 8-bit weights, and any instance whose GPUs add up to 160 GB or more, which holds the 16-bit ones. The price is what the whole instance bills for an hour, because that is what you rent.
- Why do only 107 machines qualify?
- The job needs 80 GB on one board, or GPUs adding up to 160 GB. The other 55 rentals we read are still in the index; they just cannot do this job as described. Serverless rates are left out entirely, because they buy a container that runs when it is called rather than a machine you hold.
- Can I pay less on spot?
- 57 of the 107 qualifying machines publish an interruptible rate, and each of those rows shows what the same work would cost on it if the run were never preempted. It can be, so treat the spot figure as a floor rather than a quote.
- How current are these prices?
- Prices are read every 3 days from 10 providers' own public pages and keyless APIs, and this ranking recomputes with them. Last refresh: 1 Nov 2026.