GPU servers · A100 and H100
A100 H100 dedicated server with the full GPU, never a slice
An A100 H100 dedicated server, as we rent it, is one whole machine with one NVIDIA GPU passed straight to your own operating system: the A100 with 40 or 80 GB, or the H100 with 80 GB. You pay by the month, nobody else runs on it, and the port is unmetered, so moving datasets and checkpoints costs nothing extra.
- H100 80 GB$1,682.67Hopper · HBM2e · PCIe Gen 5 · FP8
- A100 80 GB$972.47Ampere · HBM2e · 1,935 GB/s
- A100 40 GB$772.47Ampere · HBM2 · 1,555 GB/s
No video encoder on any of the three. ECC memory on all three.
Own an A100 or H100 that sits idle?
List the machine on the Primcast marketplace and rent it out to people who need the compute, month by month.
00Marketplace
Or rent an A100 or H100 from someone who owns one
These are machines their owners list on our marketplace. A listing marked as held by our sister company, Server Room, is rented on serverroom.net instead of here.
01What they are for
An H100 AI server for training, an A100 for serving models
These GPUs do one job: hold a model in fast memory and do the maths on it. If you are training or fine-tuning, or answering requests from a model that is already trained, this is the shelf to look at.
01Memory speed
They read the model quickly
Once a model sits in GPU memory, how fast it answers mostly comes down to how fast the GPU can read its weights. The A100 80 GB reads 1,935 GB/s, more than twice the 864 GB/s of the L40S.
02Training
The H100 adds FP8
Hopper brings the Transformer Engine and FP8, which is why people pick the H100 to train or fine-tune large models. If your run needs to finish sooner, not just finish, that is the reason to pay for it.
03Long runs
Built for weeks of load
Error-correcting memory throughout and data-centre hardware made to run flat out. For a training run that lasts weeks and costs real money, a flipped bit you never notice is the risk you are removing.
04One tenant
The full GPU is yours
No MIG slice, no vGPU profile, no time-sharing, and no other customer on the machine. Some cloud A100 offers are a fraction of one GPU, so check that before you compare prices.
02Before you order
Two things to know about the NVIDIA H100 and A100
They cannot encode video
NVIDIA’s own encode and decode support matrix lists zero NVENC encoders on the A100 and on the H100. They can decode (five engines on the A100, seven on the H100), but every frame that leaves them has to be encoded by the processor. They also have no display outputs, so remote desktops, workstations and game servers do not belong here.
If you need video out, the NVIDIA L40S with three AV1 encoders is the GPU to rent, and our transcoding page shows how many live channels one carries.
The A100 is still a good buy
It is the previous generation, and for a model that fits, it is often the better purchase. The 40 GB A100 holds a 13-billion-parameter model at 8-bit with room to spare, and the 80 GB one takes a 30-billion model at 8-bit.
Rent the H100 when your code uses FP8 and the newer tensor cores, or when time to result is what you are paying for. Rent an A100 when the question is “does it fit, and how fast does it read”.
03Specifications
NVIDIA H100 and A100 specifications
Figures from NVIDIA’s datasheets and its published specification tables. Where we could not match a figure to the exact board we fit, the sheet leaves it out.
H100 80GB
Hopper
80 GB HBM2e
350 W
PCIe Gen 5 x16
None. Decode only (7 engines)
Yes
A100 80GB
Ampere
80 GB HBM2e
6,912
1,935 GB/s
300 W
PCIe Gen 4 x16
None. Decode only (5 engines)
Yes
A100 40GB
Ampere
40 GB HBM2
6,912
1,555 GB/s
250 W
PCIe Gen 4 x16
None. Decode only (5 engines)
Yes
Why the H100 shows no core count or bandwidth: our store record for it mixes two boards, the SXM module that bolts onto a baseboard and the PCIe GPU that goes into our servers. The 80 GB HBM2e memory is the PCIe GPU’s, and that is the one you get. Ask on chat and we will tell you exactly which board goes into your machine.
04Choosing
A100 vs H100: which one to rent
Each GPU is a monthly line on top of the server it goes in, so the real question is whether the step up is worth it for your job. The table puts the memory and the monthly cost side by side.
| Card | Card memory | The card | Cheapest complete machine |
|---|---|---|---|
| A100 40GB | 40 GB HBM2 | $699 | $772.47 |
| A100 80GB | 80 GB HBM2e | $899 | $972.47 |
| RTX PRO 6000 Blackwell 96GB | 96 GB GDDR7 ECC | $1,299 | $1,382.67 |
| H100 80GB | 80 GB HBM2e | $1,599 | $1,682.67 |
The RTX PRO 6000 Blackwell is in the table on purpose: it holds 96 GB, more than any GPU on this page. You give up NVLink and some memory speed, and you gain four video encoders. If your only question is whether the model fits, read the RTX PRO 6000 page before you decide.
Choosing between the two A100s is about capacity first. If your model and its context fit in 40 GB with a fifth left over, take the smaller one. If you are close to the limit, the 80 GB A100 gives you the room and reads faster too.
05The machine
Which A100 server or H100 machine, and the PCIe link it gets
The A100 is a PCIe Gen 4 GPU and the H100 is Gen 5, while our server platforms are Gen 3 and Gen 4. So the H100 never gets the full bus it was designed for here, and we would rather you knew that now.
Intel Xeon Silver / Gold
AMD EPYC
Intel Xeon E5-2600 v3/v4
Intel Xeon Silver / Gold
AMD EPYC
Whether that matters depends on your job. An inference server with the weights already loaded hardly touches the slot. A training loop that pulls a fresh batch from host memory every step feels it on every step. Not sure which you are? Tell us what you will run and we will point you to the machine.
06Sizing
What fits in 80 GB of GPU memory
Weights only, for the 80 GB A100 and the H100. Leave room for the context and the runtime on top.
| Model | 16-bit | 8-bit | 4-bit |
|---|---|---|---|
| 7-8 billion parameters | 16 GB — fits | 8 GB — fits | 5 GB — fits |
| 13-14 billion parameters | 28 GB — fits | 14 GB — fits | 8 GB — fits |
| 30-34 billion parameters | 68 GB — does not fit | 34 GB — fits | 19 GB — fits |
| 70 billion parameters | 140 GB — does not fit | 70 GB — does not fit | 38 GB — fits |
On the 40 GB A100, a 30 to 34-billion model fits with room only at 4-bit. A 70-billion model at 16-bit fits on no single GPU we sell; that takes more than one GPU, which is where NVLink comes in. To work it out the other way round, from your model to the GPU, use the LLM memory sizing page.
07Prices
A100 GPU hosting and H100 prices, with the machine
The GPU is one line on your invoice and the server is another. Both come from the same order catalogue the checkout uses, so each price below belongs to the exact machine beside it.
| Card | The machine it goes in | The card | Complete, per month |
|---|---|---|---|
| A100 40GB | Intel Xeon E5-2600 v3/v4 | $699 | $772.47 Configure this build → |
| A100 40GB | Intel Xeon Silver / Gold | $699 | $782.67 Configure this build → |
| A100 40GB | AMD EPYC | $699 | $916.59 Configure this build → |
| A100 80GB | Intel Xeon E5-2600 v3/v4 | $899 | $972.47 Configure this build → |
| A100 80GB | Intel Xeon Silver / Gold | $899 | $982.67 Configure this build → |
| A100 80GB | AMD EPYC | $899 | $1,116.59 Configure this build → |
| H100 80GB | Intel Xeon Silver / Gold | $1,599 | $1,682.67 Configure this build → |
| H100 80GB | AMD EPYC | $1,599 | $1,816.59 Configure this build → |
The port is unmetered in both directions with no egress charge, and on this kind of work that is often the biggest bill elsewhere. An IPv4 address comes with the server, there is no contract, and you pay one month at a time. For every other GPU we fit, priced the same way, see all GPU servers with prices.
08Delivery and help
How soon you get it, and who answers
These GPUs are the parts most likely to be bought in for you. You are told which of these four applies before you pay.
- About 30 minutes
Already racked
A machine with the GPU fitted is standing in a rack. It is yours about 30 minutes after the order is approved.
- 4 to 24 hours
We hold the GPU
We fit a GPU we have on the shelf into the server you picked.
- 5 to 10 working days
We buy it in
The GPU is ordered for you from a distributor, then fitted to your server.
- 10 to 15 working days
On allocation
Current accelerators are sometimes rationed by the distributor, and this is the longest wait we quote.
Questions before or after you order go to our chat. An AI assistant answers in seconds and can look up your account and your server, and a person from our team steps in whenever you ask. Machines that are racked and ready right now are on instant servers.
09Questions
Questions about renting an A100 or H100
01Can I train AI models on a single H100 server?
Yes. One H100 with 80 GB trains and fine-tunes models whose weights, gradients and optimizer state fit in that memory, and FP8 through the Transformer Engine stretches how far it goes. Full training of a large model usually needs more than one GPU. Several GPUs in one machine, or several machines together, is a build we quote: see multi-GPU and multi-node servers.
02Is this the same as an NVIDIA DGX H100?
No. A DGX H100 is NVIDIA’s own system with eight H100 SXM modules on one baseboard. What we rent is one H100 PCIe GPU in a standard server, which suits single-GPU training, fine-tuning and inference. If you need several GPUs linked together, see multi-GPU servers.
03Can I transcode video on an A100 or an H100?
Not in hardware. Neither GPU has a video encoder, so encoding falls to the processor and runs at processor speed. They do decode, so a job that reads video and outputs something else works fine. For video out, rent the L40S, which has three AV1 encoders.
04Is the GPU split into MIG instances or shared?
No. You get the full GPU passed through to your own operating system, on a machine nobody else uses. Both GPUs support MIG, so you can partition yours yourself if that suits your work. That is your choice, made on your own machine.
05Should I rent the A100 40 GB or 80 GB?
Start with capacity. If the model and its context fit in 40 GB with about a fifth left free, the smaller GPU is the better buy. The 80 GB A100 also reads faster, 1,935 GB/s against 1,555 GB/s, so take it if you are close to the limit on memory or speed.
06Which NVIDIA driver and CUDA version will I get?
The ones you install. The server arrives with a clean operating system and root access, and you pin the driver, CUDA and framework versions your code was tested on. Nothing upgrades behind your back. If you want the driver installed before hand-over, say so on the order and name the version.
07Can I put two A100s or H100s in one server?
Yes, as a quoted build rather than a checkout option. Both GPUs support NVLink, so two of them can share work across a fast bridge. Which chassis, how many GPUs and the lead time depend on the GPU, so tell us what you plan to run and we come back with the build.
08How soon can I have an A100 or H100 server?
It depends on where the GPU is. About 30 minutes after approval if a racked machine already carries it, 4 to 24 hours if we hold the GPU, 5 to 10 working days if we buy it in, and 10 to 15 working days for a GPU on allocation. You are told which applies before you pay. Racked machines are listed on instant servers.
Pick the GPU, then the machine around it
Every A100 and H100 build is priced above with its server. Choose one, configure it, and the order shows the full monthly total before you pay.