Managed GPU environment · EU region
Compute you don't have to run.
You need somewhere to train and serve models. You don't need a cloud account, quotas, firewall rules, driver installs or an on-call rota. We take that part — you get an environment that's ready, and someone who answers when it isn't.
Talk to us →Most people who need a GPU don't need a cloud platform. They need it to work. Our first customer said it unprompted: far less overhead than dealing with subscriptions and access management. The machine isn't the value — everything you avoid is.
Three levels, by what you're actually doing.
We don't sell machine types. You pick the level of ambition; we pick and run the capacity that covers it — and change it underway if your needs do.
You're trying something out. Smaller models, short runs, lots of starting and stopping. The environment stops itself when you're not using it, so you pay for work — not for idling.
You're training or fine-tuning something that will be used. More memory, longer runs, and monitoring that catches a stall before you notice it yourself.
Something has to stay up and answer. At that point it isn't about GPUs any more — it's uptime, accountability and someone to call. We define response times together.
Two different memories, and they're easy to confuse. GPU memory decides how large a model and how large a batch fit on the card. System memory decides how much data you can hold while loading and preparing it. It's almost always system memory that stops people first — and that isn't the number the level is named after. We set it from what you're actually doing, and expand it along the way if the dataset grows.
Your data doesn't have to die with the machine.
Most people need compute in bursts — three days now, maybe two weeks in the autumn. Without somewhere to leave your data, every pause means uploading everything again. The store is a room that stays put between engagements.
When the environment is deleted, the store remains. Come back in two weeks or four months and your datasets, models and results are exactly where you left them.
Moving from Bench to Workshop, or on to Plant, the store connects straight to the new environment. You copy nothing — and you don't pay for a machine sitting idle while you upload.
The store costs 0.36 NOK per GB per month — around three and a half times less than the disk on the environment itself, because it doesn't have to sit ready for a GPU. Keeping data between engagements is cheap. It's leaving a machine running that's expensive.
Your data is one thing — the hours you spent setting the environment up are another. We can keep an image of the machine as it stood: drivers, libraries, your own installs. The next environment then starts ready rather than empty. The image costs 0.79 NOK per GB per month and counts only the space actually in use — a 200 GB environment with 60 GB in use costs around 47 NOK a month to keep, against around 311 NOK to leave the environment stopped with its disk and static address.
The store sits in the same region as the environment, and within a region AWS doesn't charge for data transfer. We copy everything onto the local NVMe disk while the environment is being set up — 500 GB takes around ten minutes — so you always work against the fastest disk, not across the network. When you stop, the changes sync back.
The price is per month and runs on calendar time, not on usage. You can empty or delete the store at any time, and the cost stops immediately. Network charges only apply if you pull the data out of the cloud to your own machine.
Three ways to leave it be
If you don't need the machine for a while, the only question is how long. Compute costs nothing in all three cases — it's the storage that separates them.
The watchdog stops the environment when nobody is using it. Everything stays untouched.
Same install, same disk, same address. You pay for allocated space, not used.
Around six and a half times cheaper, and safer: an image can be restored onto newer hardware. A stopped machine is tied to the hardware it started on, and that may not be sold in a year.
There is no crossover point in money — the image is cheaper from the first month. What you pay for by leaving the environment stopped is being back in seconds rather than ten minutes. We tell you when an environment has been stopped long enough that the wait has become expensively bought.
Setup is paid for the first environment. Restore from the store and it falls away — the work is already done, and you've paid to keep it. The image buys away the setup fee next time.
Nothing stays behind unless you said yes
- When you end an environment we ask explicitly whether the data and the machine image should stay. We show what it costs per month before you confirm. Choose to delete and it is deleted — we keep nothing «just in case».
- You get a receipt every month with what the store holds and what it costs. No silent renewal to go hunting for.
- If a payment goes unpaid we warn you three times over 60 days before anything is deleted. You have that whole period to download the contents or start an environment and fetch them yourself. We never delete without warning.
The operations are the product.
Everything below actually happened to us, and we've engineered it away. That's why it's here — not because it looks good in a bullet list.
A customer ran the machine out of memory after two hours. The whole server froze, he was locked out and lost unsaved work. The cloud default has no swap — it goes straight from full to dead. Swap and a guard process are now set up from the start.
Cloud hardware has a lifespan. A machine type available today may be gone in two years, and an environment left stopped for a long time may suddenly refuse to start. The watchdog follows the notices and moves you across in good time — a stopped machine can be converted directly, and an image can be restored onto newer hardware. You get told it has been done, not handed a task.
That time, we spotted the freeze ourselves — the customer hadn't. Now the watchdog runs every quarter hour and checks health, whether login responds, and whether the address has changed. It fixes what it can fix itself and alerts us to the rest. The alert goes to us, not to you.
Without a static address the machine changes address on reboot, mid-incident. Static address from day one.
Regions fill up — we've been refused capacity more than once. Finding room in another region is our job, not your wait.
The environment shuts down when idle. For daytime work that halves the bill, and it's a saving you don't get renting a machine by the hour.
Environments have an expiry date. You get warned well before anything is deleted — not an email afterwards.
The operations automation is being built out and is set up per customer today. We don't promise more than we actually do.
What's included, and what costs.
The line runs at the machine. If something is wrong with the environment, that's our job and it costs you nothing — even when it takes all day. If we're going into your own code, that's consulting, and we say so before we start.
Included in operations
- The environment won't answer, hangs or has frozen — we spot it, fix it, and tell you what happened.
- Memory runs out, the disk fills up, swap and memory protection need setting up or adjusting.
- The address, the firewall, keys and access.
- Capacity disappearing in the region, machine types being retired, moving you to newer hardware.
- Round-the-clock monitoring, and stepping in before you notice.
- Working out which side of the line a case falls on. If it's quick, we just tell you the answer.
Charged separately
- Why your own code uses more memory than you think — a review of your data loading, notebook or pipeline.
- Moving over to streaming or memory-mapping, set up together with you.
- Performance work inside your training job or model.
Excl. VAT. The minimum unit is half an hour — 600 NOK — and after that each started quarter hour. A 35-minute case costs 900 NOK, not 1,200.
We never start chargeable work without you saying yes first. If you're unsure what something costs, ask — you'll get an estimate, not an invoice after the fact.
And sometimes the advice is not to. We've talked a customer out of upgrading to more memory, because the problem was in how the data was being loaded and would have come back on a more expensive machine a few days later. We make more selling you hardware than saying that — but you'd have paid for the wrong thing.
What does it cost?
Setup once, operations while the environment is yours, and capacity for the hours the machine is actually on — our cost price plus 25 %. You can find compute cheaper elsewhere. What you're paying for here is that it's run for you.
Four plain questions. Then we suggest a level and work out what it costs. You can override all of it afterwards.
This is an estimate, not a quote. What you actually pay depends on how much you use the environment, how much storage you need, and the exchange rate and capacity pricing in the period. We always send a concrete quote before anything starts.
Answer the first two questions and we'll start calculating.
Get a concrete quote →When not to choose us.
We don't win on price, and we don't pretend to.
Specialist providers sell compute below our cost price. If you just need many hours as cheaply as possible, and have your own people to run it — use them.
With cloud skills in-house you'd be paying for something you already do. Then our role is advice, if anything — not operations.
Several environments and the store under one agreement are part of the product — we keep track of that for you. But if you need databases, queues, your own networks and your own console access to all of it, what you want is a cloud platform. Then the infrastructure work comes back regardless of who invoices you, and you should have your own account. We'll gladly help you set one up instead.
Where does it run?+
In the EU, and you'll know exactly which region before we start. That's the whole difference from the marketplaces, where the machine can sit anywhere with anyone — disqualifying if you handle personal data.
Do you have access to my data?+
The environment is set up with no links to our systems, with its own key pair and firewall group. We hold operational access so we can help you. If you'll process personal data, we sign a data processing agreement first.
What if the machine goes down?+
We notice and fix it. What response time we commit to is agreed concretely — we won't put a number here that isn't covered in your agreement.
Can I change level along the way?+
Yes. Needs change, which is exactly why you're not buying a machine type. We move you, and you pay infrastructure on actual usage.
Why are you more expensive than the cheapest?+
Because you're buying something else. With us you get a guaranteed EU region, a Norwegian company to invoice and hold accountable, and someone else doing setup, monitoring and fixes. If you don't need that, we're the wrong choice — see the section above.
How quickly can we start?+
The first environment was ready about ten minutes after it was ordered. Setup isn't the bottleneck — the clarifications around it are.
Tell us what you need to compute.
We'll come back with something concrete: which level covers the need, what it costs, and what we take responsibility for.