Skip to content
Access
Request access

KATHMANDU VALLEYNVIDIA H200 NVLCOMING SOON

The valley is coming online.

Nepal's GPU cloud: NVIDIA H200 NVL GPUs in a datacentre in the Kathmandu Valley.

Request accessSee how it works
  • H200 NVLNVIDIA GPUs
  • Per secondbilled in rupees
  • Safe with usyour data, protected

Built on

NVIDIA H200 NVL GPUs, in a datacentre in the Kathmandu Valley

The five-layer cake

Every layer, named. From energy to your app.

NVIDIA's Jensen Huang describes AI as a five-layer cake: energy, chips, infrastructure, models and applications. Here is each layer at CoreValley.

  1. 01ApplicationsFraud scoring · voice in Nepali · assistants · forecasting
  2. 02ModelsOpen models · your fine-tunes · your own
  3. 03InfrastructureDatacentre · cooling · networking · storage · scheduling
  4. 04ChipsNVIDIA H200 NVL · 141 GB HBM3e per GPU
  5. 05EnergyNepal's hydropower · 99.99 % facility uptime

The five layers are NVIDIA's framing1. The facility's uptime is its own commitment for power, cooling and the building2.

The GPU

141 GB on every GPU. HBM3e at 4.8 TB/s.

Room for large models and long contexts on a single card, Tensor Cores for FP8, BF16 and TF32, and NVIDIA's AI software on top.

GPU memory
141 GB
Memory bandwidth
4.8 TB/s
FP8 Tensor Core
3,341 TFLOPS3
Max power (configurable)
Up to 600 W

One GPU, seven ways

141 GB on one GPU. Most jobs need a slice of it.

NVIDIA Multi-Instance GPU splits one H200 NVL into as many as seven GPUs, each isolated in hardware. Pay for the slice, not the card.

An H200 NVL reads its whole memory in about 29 milliseconds — faster than a blink.4

How the GPU is divided

One GPU with 141 GB of memory.

Use cases

Built for the work. Banks, products, classrooms.

Fraud scoring, voice in Nepali, private language models, training on local data — the work that should run close to home.

All use cases

Startups

Startups and IT exporters

Build AI into your product and serve customers at home and abroad, billed in rupees.

Runs on On-demand GPUs · Inferencing

Banking

Fraud, AML and KYC for banks and wallets

Check every payment as it happens and send the suspicious ones for review.

Runs on Reserved GPUs · Inferencing

Education

Universities and AI skills

One GPU, shared fairly by a whole class: a private slice for every student.

Runs on MIG slicing · On-demand GPUs

Voice AI

Voice AI in Nepali

Turn Nepali speech into text and answer callers in their own language.

Runs on Inferencing · MIG slicing

LLM inference

Private language models on your documents

Ask questions of your own documents and get answers that show where they came from.

Runs on Inferencing · Reserved GPUs

Model training

Training and fine-tuning

Teach open models your own data and language, and keep what they learn.

Runs on Reserved GPUs · On-demand GPUs

Energy

Hydropower and grid forecasting

See river flow and power demand coming, hours and days ahead.

Runs on Reserved GPUs · Inferencing

Language

Language AI for 124 mother tongues

Models, speech and translation for Nepali, Maithili, Bhojpuri, Tamang, Nepal Bhasa and more.

Runs on Reserved GPUs · Inferencing

Government

Government services and records

Turn handwritten records into searchable text and answer citizens in plain Nepali.

Runs on Reserved GPUs · Inferencing

Health

Healthcare triage and decision support

Put the chest X-rays that most need a doctor at the top of the list.

Runs on MIG slicing · On-demand GPUs

Story 1 of 10: Startups and IT exporters

Data protection

Your data is safe with us. Protected at every step.

Five commitments for your datasets, models, prompts and answers, in plain words.

Billing

Pay for the seconds you use. Not the hour you started in.

Billing
Per secondGPU time and MIG slices are metered by the second they run.
Minimum
60 s60 seconds per run. Nothing is rounded up to the hour.
Currency
NPRInvoiced in Nepalese rupees, so there is no exchange-rate margin.
Ways to pay
eSewa · KhaltiOr bank transfer and invoice. VAT at the statutory rate.

Rates are being finalised. Every request gets a firm quote in rupees within one working day.

How billing worksGet a quote

Your workspace

The tools you already use. Ready on day one.

Notebooks, pipelines, training, serving and data tools in one workspace, ready when you sign in.

Notebooks and pipelines

  • JupyterLab
  • Kubeflow
  • Kale
  • Airflow

Training at scale

  • Ray
  • Spark
  • Katib
  • TensorFlow

Serving and tracking

  • KServe
  • MLflow
  • NVIDIA NIM

Data and analytics

  • Presto
  • Superset

Products

Six ways to run. One cloud in Nepal.

Compare all six

Start the climb

What are you here for?

Pick one. It takes about a minute.