Pricing
Billed by the second, in rupees.
Rates for general availability are being finalised. Enterprise early access on NVIDIA H200 is open now — tell us your workload and we'll send a firm NPR quote within one working day.
NVIDIA H200 available now. RTX PRO 6000 Blackwell and L40S coming soon.
How billing works
You pay for what runs. Nothing else.
Four things are metered. Everything is quoted, invoiced and paid in Nepali rupees.
GPU time
per second
Pods and notebooks are metered from the moment they're running, with a 60-second minimum. Whole cards and slices each have their own rate.
Model endpoints
per token
Input and output tokens are priced separately, and repeated prompt prefixes cost less.
Dedicated servers
per month
Reserved by the month, with lower rates for 6, 12 and 36-month terms.
Storage
per GB-month
Volumes that outlive a pod, billed for the space you keep.
How you pay
In Nepali rupees, on the rails you already use. No dollar card and no approval to pay a foreign provider.
- eSewa
- Khalti
- Bank transfer
- Corporate invoice on net terms
What you never pay for
- Time in the queue
- The meter starts when your container is running, not when you press launch.
- Time after you stop
- Stop a pod and the meter stops with it. Nothing accrues after it ends.
- Hourly rounding
- After the 60-second minimum, you pay for the seconds you use.
- Exchange-rate spread
- Quoted, invoiced and paid in rupees. No USD card, no FX.
Ways to rent a GPU
Pick the slice that fits the job.
Every option is billed by the second. The difference is how much of the card is yours.
A whole card
The entire GPU: all 141 GB on an H200. No neighbours, full speed.
Best for: Large training runs and production serving.
A hardware slice (MIG)
A walled-off part of the card with its own memory and compute. A neighbour's crash can't reach you.
Best for: Fine-tuning and steady work that needs isolation.
A shared slice (HAMi)
A share of a card's memory and compute. The cheapest way in, but you share the card, so it isn't fault-isolated.
Best for: Notebooks, experiments and light inference.
Rent vs buy
Why teams rent rather than buy.
Enterprise GPUs carry capital cost, power, cooling and maintenance. Renting removes all four, and the support comes from Kathmandu.
Upfront cost
Buying hardware: A large capital purchase
CoreValley: None
Power, cooling, maintenance
Buying hardware: Your responsibility
CoreValley: Included, on hydropower
Time to first experiment
Buying hardware: Weeks to months of procurement
CoreValley: Days, not months
Scaling
Buying hardware: Limited to what you own
CoreValley: Up or down as your work needs
Technology refresh
Buying hardware: You carry the obsolescence risk
CoreValley: New GPUs as they land
Billing
Buying hardware: Often USD, plus import duty
CoreValley: NPR, on local payment rails
Support
Buying hardware: Your own staff
CoreValley: Engineers in Kathmandu, on Nepal time
| Aspect | Buying hardware | CoreValley |
|---|---|---|
| Upfront cost | A large capital purchase | None |
| Power, cooling, maintenance | Your responsibility | Included, on hydropower |
| Time to first experiment | Weeks to months of procurement | Days, not months |
| Scaling | Limited to what you own | Up or down as your work needs |
| Technology refresh | You carry the obsolescence risk | New GPUs as they land |
| Billing | Often USD, plus import duty | NPR, on local payment rails |
| Support | Your own staff | Engineers in Kathmandu, on Nepal time |
Always included
In every plan.
A ready AI stack
CUDA, PyTorch, TensorFlow, Jupyter, vLLM and DeepSpeed, installed from first boot.
Help in Nepal time
Engineers in Kathmandu who know the platform and the local context, working in Nepal time.
Data that stays in Nepal
Compute and storage stay in Kathmandu, so regulated and sensitive work stays in the country.
A meter you can check
Usage shows up as it accrues, and the same numbers appear on your invoice.
Questions
Common questions.
When will you publish rates?
Soon. We're finalising rates for general availability. Until then, every early-access customer gets a firm NPR quote within one working day, based on the workload you describe.
Which GPUs can I get today?
NVIDIA H200, through enterprise early access. RTX PRO 6000 Blackwell and L40S are coming soon.
Can I pay in NPR?
Yes. Everything is quoted and invoiced in Nepali rupees. Pay by eSewa, Khalti, bank transfer or a corporate invoice on net terms.
What's the difference between a hardware slice and a shared slice?
A hardware slice (MIG) is a walled-off part of the GPU with its own memory and compute, so a neighbour's crash can't reach you. A shared slice (HAMi) shares the card in software. It packs more work onto each GPU, so it costs less, but it isn't fault-isolated.
How precisely is GPU time billed?
Per second, with a 60-second minimum per pod. The meter starts when your container is running and stops when you stop it. No hourly rounding, and nothing after it ends.
Do you offer academic pricing?
University access is coming soon. Early access is open to enterprises first; universities and labs can join the list now, and we'll be in touch when it opens.
Is there a minimum commitment?
No commitment for pods, notebooks or model endpoints: you pay for what you use. Dedicated servers are reserved for a term, and longer terms cost less.
Get an NPR quote.
Tell us the GPU hours you expect, your model sizes and what you want to run. Within one working day you get a capacity plan and a firm price in rupees.
Early access is open to enterprises on NVIDIA H200. More GPUs, and access for more teams, are coming soon.