Home/Solutions/Research and university
Two things are not negotiable in a university lab: the department budget and where the data is processed. Here a hard budget cap is not a warning but a management tool, and region pinning is a line in your ethics committee file.
You pay per second and stop whenever you want. No commitment.
Your work looks like this
If you are a team queuing on an in-house cluster, sharing GPUs between projects and writing up the spend at the end of term, the problems will be familiar.
In an institutional budget an overspend is not corrected next month; it is a signature problem. That is why a cloud bill that depends on guesswork is unacceptable.
Ethics committee approval, obligations under KVKK (Turkish personal data protection law) and GDPR, and sometimes a clause from the data provider you signed with: the country where the data is processed has to be known in writing.
A two-day queue on a shared cluster wrecks a thesis timetable. You need short-lived, predictable outside capacity — without opening a procurement process.
Budget control
The budget card in the console projects every 10 seconds: current spend plus burn rate. Notices go out at 80% and 95%, and at 100% pods are paused — not deleted, their disks preserved. That is why the department budget is not exceeded.
The cap is defined at organisation level; a separate per-pod cap can be added on top. One PhD student's experiment cannot eat the department's monthly allocation.
On a critical training run you can pick "notify only"; the default behaviour is to pause. Lowering the cap below current spend means an immediate pause, and the console asks you to confirm that separately.
Compute, storage and network traffic are separate rows. Each row expands per pod and exports as CSV — you do not fill in a spreadsheet by hand for the end-of-term report.
The billing role sees the spend but cannot start pods; the member role starts pods but cannot edit the budget. The person managing the funding and the person running the jobs do not have to be the same.
When the cap fills, checkpoints and volumes stay where they are. Once the limit is updated the job continues from where it stopped; you do not start over.
Made-up scenario
Not a real institution or customer: an illustrative example constructed to show how the product is used. The numbers are examples.
Data sovereignty
You pin your pod to a country; a placement request that violates the region pin is rejected by the API. The text on the region card in the console pod detail is not a fixed string but comes from the pod's actual placement record — an unpinned pod shows different, honest wording.
SOC 2 Type I has not been obtained yet; in our technical plan it sits as a process within Phase 3. We do not write up a certificate we do not hold as if we did. For your institutional procurement file we can share the current state in writing.
Choosing a card
Most lab work is short experiments; heavy training comes a few times a term. Per-second billing fits exactly that distribution: experiment on a cheap card, step up a tier when you need to.
| GPU | VRAM | Typical academic workload (example) | Example starting price |
|---|---|---|---|
| RTX 3090 | 24 GB | Teaching and workshops, student assignments, small experiments | $0.21/hr |
| RTX 4090 | 24 GB | Thesis experiments, QLoRA, image generation | $0.42/hr |
| RTX A6000 | 48 GB | Long context, multi-model comparison studies | $0.88/hr |
| A100 PCIe | 80 GB | Publication-scale training, large datasets | $1.32/hr |
| H100 SXM | 80 GB | Heavy training, short intensive allocations | $2.19/hr |
Example prices, illustrative only; they move with supply and demand on the marketplace. Storage and traffic are separate line items — all of them are on the Pricing page.
Region, scorecard and image selection: GPU Pods.
A university cluster that is idle outside working hours can join the pool: Become a Host.
Discounts
Our academic pricing policy is not written down yet; rather than describing a programme that does not exist, we suggest sitting down together. Tell us your department's workload, its term funding and its data requirements, and we will say plainly what we can do. If we cannot, we will say that too.
The address is a placeholder; a proper company mailbox is coming: merhaba@kaldera.ai
Frequently asked
The cap is an enforcement mechanism: the budget worker projects every 10 seconds and pauses pods at 100%. Still, let us be honest — while a paused pod keeps its disk, the storage line keeps running. That row is visible on the billing screen and is included in the cap calculation; we do not promise indefinite free retention.
A single invoice is produced at the monthly period close and downloaded as a PDF; the line items separate into compute, storage and network traffic. Institutional payment methods and contract texts are being settled together with the company incorporation — today we do not publish a finalised institutional tariff on this page.
The pod's placement information is a field in the API and is visible in the console; a request that violates the region pin is rejected. You can attach those records to your file. The sovereignty package beyond that — an EU legal entity, an audited Verified tier — is in Phase 3 and does not exist today.
Yes: there are roles under a single organisation. The member role starts pods, the billing role sees the spend, the admin sets the cap. Every pod is recorded together with the user who started it.
Not today. Multi-GPU training on a single machine works; cross-node cluster training is on the roadmap and sits in the nav with a "soon" badge. We do not make promises without a date.
Get started
You don't pay for seconds that don't work, a hard budget cap, per-second billing. If your department's situation is different, tell us; we do not sell a programme that does not exist.
The console runs on the local development address: localhost:3000