Use cases
Work that belongs in Nepal.
The common thread is not the model. It is that the data cannot leave the country, the invoice has to be in rupees, or support has to answer in Nepal time.
H200 available now · enterprise early access
Nepali and Maithili language models
Teach models Nepali properly. Continue pre-training and instruction-tune on Devanagari text — the data that closes the gap is exactly the data that should not leave the country.
- Why here
- Text corpora often carry personal data from local sources. Training in-country keeps where the data came from defensible.
- Typical setup
- typical setup · H200 pods for training, then a model endpoint to serve the result.
Regulated financial workloads
Read KYC documents, flag suspicious transactions and score credit — for banks and finance companies under NRB supervision.
- Why here
- Customer data cannot cross the border. Dedicated servers and private networking, where nothing leaves unless you allow it, make the security review answerable.
- Typical setup
- typical setup · Dedicated H200 servers in your own private cluster, with outbound traffic blocked by default.
Clinical imaging and records
Triage radiology scans, screen retinal images and pull facts out of clinical notes — for hospitals and diagnostic chains, without exporting patient data.
- Why here
- Patient data is the least portable data there is. Where the compute sits is the whole argument.
- Typical setup
- typical setup · GPU pods with persistent storage; dedicated capacity for production.
Government and civic AI
Automate citizen services, digitise land records and build Nepali-language public information systems — on infrastructure inside the jurisdiction.
- Why here
- Sovereignty is a procurement requirement, not a preference. The infrastructure is inside Nepal.
- Typical setup
- typical setup · Dedicated servers, plus model endpoints for public-facing assistants.
Universities and labs
Real GPU access for students and faculty — no procurement cycle, no shared server to look after, and no foreign cloud account nobody can pay for.
- Why here
- JupyterHub shuts idle notebooks down on its own, so a department can give forty students a GPU without forty invoices.
- Typical setup
- typical setup · JupyterHub with a hardware slice of an H200 (MIG) for each student.
Product teams shipping AI
Fine-tune a small model, serve it behind an endpoint, and grow with your traffic — with costs in the currency your runway is in.
- Why here
- Per-second billing and per-token endpoints mean the bill follows your traction instead of running ahead of it.
- Typical setup
- typical setup · GPU pods for fine-tuning, per-token model endpoints in production.
Workloads
Built for the work, not the demo.
Six kinds of workload, each with a product that runs it. All metered in rupees, all kept in Kathmandu.
From lab to production
One platform across the lifecycle.
The same project grows from a shared notebook to a dedicated server — without switching clouds, currencies or countries.
- 01
Explore on a slice
Start in JupyterHub on a slice of an H200 — a hardware slice (MIG) or a cheaper shared one (HAMi). Cheap enough to leave running while you find out whether the idea holds.
- 02
Train on whole cards
Move to whole H200 cards, or a full eight-GPU server joined by NVLink for distributed runs.
- 03
Serve behind an endpoint
Deploy to a model endpoint with an OpenAI-compatible API, or keep a dedicated pod running. Billed per token or per second.
- 04
Reserve capacity
When load is steady, move to a reserved dedicated server and take the discount that comes with a longer term.
Questions
Before you ask.
Who can get access today?
Enterprises, on NVIDIA H200, through early access. More GPUs — starting with the RTX PRO 6000 Blackwell — and access for more teams, including universities and startups, are coming soon. You can join the list now.
Does my data ever leave Nepal?
No. Compute and storage are in Kathmandu, nothing is replicated to a foreign region, and you can block all outbound traffic for a project.
Can a university give a whole class GPU access?
Yes — that is what JupyterHub is for: per-user limits, automatic idle shutdown and shared GPU slices, on one invoice. Access for universities is coming soon; join the list and we will tell you first.
What does a regulated workload look like on CoreValley?
Dedicated servers in your own private Kubernetes cluster, private networking with outbound traffic blocked by default, an append-only audit log and invoices in NPR — the pieces a bank or public body needs for a security review.
How do I move from experiment to production?
Inside one project: from a notebook slice, to whole cards, to a model endpoint or a reserved server. Same account, same rupee invoice, same country.
Tell us what you’re building.
A model, a dataset size and a deadline are enough for us to recommend a setup. We reply within one working day with a capacity plan and a firm rupee quote.
Early access is open to enterprises on NVIDIA H200. More GPUs, and access for more teams, are coming soon.