H200 available nowenterprise early access
Nepal’s own GPU cloud: NVIDIA H200s, rented by the second from a hydro-powered Kathmandu datacenter.
Pay in rupees, get help in Nepal time, and keep every byte of your data in the country.
We reply within one working day with a capacity plan and a rupee quote.
memory per H200
room for big models on one card
latency in Kathmandu
served in-country, not overseas
of your data stays in Nepal
stored and processed in Kathmandu
billed in rupees
eSewa · Khalti · bank · invoice
Why CoreValley
Your data, your money and your support all stay in Nepal. Not because a contract says so, but because the GPUs, the billing and the people are all in Kathmandu.
Data stays here. Your datasets, models and every request are stored and processed in our Kathmandu region (np-ktm-1), and nothing is copied abroad. Regulated projects block all outbound traffic by default.
Pay in rupees. No dollar card and no exchange-rate surprises. Pay by eSewa, Khalti, bank transfer or corporate invoice.
Help in Nepal time. The engineers who answer you work in Kathmandu, in Nepal time. You talk to the people who run the GPUs, not a ticket queue abroad.
Hydro-powered. Our datacenter runs on Nepal's hydropower: low-carbon compute, at a power price that isn't tied to gas markets on another continent.
Close to your users. Network latency in Kathmandu is under 5 ms, so your apps answer people in Nepal without a round trip abroad.
one h200 · everything it touches stays in kathmandu
Early access
More GPUs, and access for more teams, are coming soon.
We reply within one working day.
The platform
Start on a slice of an H200, move to whole cards, then reserve dedicated servers. One account, one rupee invoice and one country the whole way.
h200 · slice → whole cards → dedicated
one h200 · 2 of 7 slices
4 × h200 · whole cards
dedicated · reserved
01 / slice
Rent a hardware slice of an H200 — 35 GB of its memory — billed by the second. Enough to fine-tune a 7B model tonight and shut it down by morning.
02 / card
Same files, same code, same API key — on a full H200 with all 141 GB, then on four or eight of them when the job outgrows one.
03 / rack
Dedicated H200 servers that nobody else touches, reserved by the month — with the same usage meter and audit log you already use.
Hardware
Available now: the NVIDIA H200 — 141 GB of memory on one card, built for training and serving the largest models. Coming soon: the RTX PRO 6000 Blackwell for fast, affordable inference and visual AI.
Hopper · H200
available nowTrain and serve the largest models — 141 GB on one card.
Memory
HBM3e · 6 HBM3e stacks
+47% vs RTX PRO 6000
Memory bandwidth
Platform
up to 7 hardware-isolated instances · 18 GB each
Blackwell · GB202
coming soonFast, affordable inference, fine-tuning and visual AI.
Memory
GDDR7 ECC · 512-bit bus
single PCIe card
Memory bandwidth
Platform
up to 4 hardware-isolated instances · 24 GB each
* H200 tensor figures with sparsity (SXM5 datasheet). RTX PRO 6000 Blackwell Server Edition figures as NVIDIA publishes them, without a dense/sparse label. FP32 and FP64 are non-tensor.
Also coming soon
48 GB GDDR6 · Ada Lovelace
Tell us the model, the dataset size and the GPU hours you expect. Within one working day you get a capacity plan and a firm rupee price — not a sales call.
01 · about two minutes
The model, the data size and the GPU hours you expect. Set them on the console below, or say it in your own words.
02 · within one working day
A capacity plan and a firm price in rupees, from our team in Kathmandu. No sales call needed.
03 · billed by the second
Start on a slice or on whole cards. The meter starts when your job runs and stops the moment you stop it.
your plan · np-ktm-1
1× H200 · full cardfor 24 hours
