Your own GPU server, from 20 to 96 GB of card memory.
The whole machine answers to you, with full root access, a driver stack you choose and pin, and every hour of the card spent on your own work.
FromUS$330a month
Get a GPU ServerYou choose the machine and the price you see, and we build it to your order.
- Every core and every card, yours alone
- Full root access and your own driver stack
- Traffic in and out of the machine is included
- One monthly price, locked for as long as you keep it
The machines we build
- Spark
- Forge
- Foundry
- Reactor
Built for the work you are actually doing.
Every machine here holds a whole model in card memory, from 20 GB on the entry machine to 96 GB on the largest, so your work runs on one card and stays there.
- Fine-tuning your own modelBring an open model and your own data, train it on a card that is yours for the whole run, and keep every checkpoint on a machine only you can reach.
- Serving a model to your usersRun inference around the clock with the whole card to yourself, so the response times you measure on Monday are the response times your users get on Friday.
- Batch work that runs overnightQueue renders, transcodes or embeddings and let the machine run flat out until the job is finished, all for the one monthly price.
Choose the machine that fits your model.
The memory on the card sets how large a model you can load, and the memory in the machine around it gives you room to stage your data and keep the card fed while it works.
Spark
- 20 GB of memory on the card
- The whole machine, yours alone
- Managed twin available
Forge
- 96 GB of memory on the card
- 256 GB of memory in the machine
- The whole machine, yours alone
- Managed twin available
Foundry
- 96 GB of memory on the card
- 512 GB of memory in the machine
- The whole machine, yours alone
- Managed twin available
Reactor
- 96 GB of memory on the card
- 768 GB of memory in the machine
- The whole machine, yours alone
- Managed twin available
Every machine is built to order, so you get exactly the configuration you chose, at the price on its card.
Your price is locked for as long as you keep the machine.
Find the machine your model runs on.
Set the size of your model and how you load it, and we show you the smallest machine that holds it and what it costs.
8 billion parameters at 16-bit works out at about 16 GB of weights.
What each one holds
- Sparkon the card20 GB4 GB to spare
- Forgeon the card96 GBin the machine256 GB80 GB to spare
- Foundryon the card96 GBin the machine512 GB80 GB to spare
- Reactoron the card96 GBin the machine768 GB80 GB to spare
Leave yourself room above the weights and the card keeps your work moving at full speed.
Included with every GPU server.
- The whole machine, yours aloneEvery hour the card works, it works on your job, and the machine answers only to you.
- Your driver stack, pinned where you put itChoose the driver and the toolkit, pin them together, and they stay exactly where you put them from one run to the next.
- A clean Linux whenever you want oneReinstall from the panel in a few clicks and the machine comes back with your own SSH keys already on it.
- Rescue mode, on your commandBoot into a rescue system from the panel at any hour, with your own keys, and take back control in minutes.
What a GPU server is.
A GPU server is a computer with a graphics card in it, and a graphics card is very good at doing the same small sum millions of times at once, which is exactly what training and running a model is made of.
Renting one from us means the whole machine is yours: the card, the processor and the memory around them, all yours for as long as you keep it.
You install whatever you like on it, and we keep it powered, connected and cooled.
The questions people ask.
- How do I know which machine I need?
- Start with the memory on the card, because that is what sets how large a model you can load. Serving sits comfortably close to the size of the weights, and training rewards headroom above them, so give yourself room if you are training.
- How do I get one?
- Choose the machine you want at the price on its card and we build it to order. Ask us where a tier stands on the day and we will tell you.
- Who looks after the drivers?
- You do, on an unmanaged machine, and that is what most people here want. Pick the driver and the toolkit, pin them together, and they stay exactly where you put them. On a managed twin our engineers keep the operating system for you.
- How is it billed?
- One monthly price for the whole machine, yours alone for as long as you keep it, and it renews at the price you took it at.
Tell us what you are running.
We will tell you which machine holds it and what it costs.
Get a GPU Server