◆ Vajra Mesh · Kathmandu, Nepal · --:--
The software layer that turns a GPU you already own into a shared, in-country AI platform.
No GPUs, no racks — just the systems, running on the GPUs you already have.
live demo · updates every second
See how it works →Why Vajra Mesh
Your data is stored and processed in Nepal — it never has to cross a border to reach your own compute.
Pay in rupees through eSewa, Khalti, bank transfer or a formal invoice. No overseas settlement.
A team on your timezone. When a training run matters, the people answering are here.
Vajra Mesh is software. The GPUs stay yours or your partner's — we build and run the systems on top. Nothing to buy, just the tech.
"We're one person, in Kathmandu, building the layer that lets a GPU you already own serve everyone who needs it — in-country, in rupees. No cloud, no markup on silicon, no overseas settlement."
— Samundra, founder
The platform
Four systems, one control plane: the sharing engine, the model gateway, GPU notebooks, and the dashboard that ties them together. Deploy it on the GPUs you or your partner already have — we build the software, not the silicon.
The HAMi + HAIShare engine that puts many tenants on one card — billed by the second, only while a job runs. The card stays yours.
$ vajra share --card L4 --tenants 8 → one card, 8 fair sharesRoutes to open models (vLLM) and frontier models (Anthropic / Claude), priced per token. Change one URL, keep your code.
GPU-backed Jupyter with a fair share per person and idle sessions that shut themselves down.
The dashboard where you operate the whole system: tenancy, quotas, scheduling and the model gateway, across every GPU the platform runs on.
Hardware it runs on
Vajra Mesh doesn't ship hardware. It's designed to run on the GPUs that actually serve the region's workloads — your fleet, or a partner's.
The workhorse card for fast, affordable inference and visual AI — the platform runs on it.
Room for the largest models on one card — big-VRAM serving and fine-tuning.
Dense, cost-efficient inference — the card our multi-tenant PoC already runs on today.
How it works
The model, the dataset size and the GPU hours you expect. In your own words, or on the console.
$ vajra quote --model 70b --hours 24h →A capacity plan and a price in rupees, from our team in Nepal. No sales call, no back-and-forth.
$ plan.pdf · firm NPR price · reserved capacityWe stand up the systems on your GPUs (or a partner's). You start sharing cards, routing models and running notebooks — the meter runs only while your jobs run.
$ vajra deploy --on your-fleet --sharing on --api claudeContact
Bring the workloads. We bring the systems — multi-tenant sharing, one model API, and notebooks — running in-country, billed in rupees. No GPUs, no racks, just the tech.
We reply within one working day.