Nusapod

Deploy a model. Get an endpoint.

NUSAPOD is an AI inference platform for your own GPUs. Pick a model — Llama, Qwen, DeepSeek, or your own weights — and minutes later it's live behind an OpenAI-compatible endpoint. GPU scheduling, allocation, model download, runtime, and networking: handled. You just call the API.

minutes to a live endpointOpenAI-compatible APIon GPUs you own
3 clicks
From model pick to live endpoint
OpenAI
Compatible API — swap base_url, keep your code
100%
Your GPUs, your data, your network
1 GPU
Serves many models — sliced, never oversubscribed
Everything handled

You bring the model. NUSAPOD runs it.

Scheduling, GPU allocation, model download, serving runtime, networking, and the API endpoint — one flow. Under the hood, hardware-enforced VRAM slicing packs many models onto each GPU without oversubscribing a single gigabyte.

One-click model catalog, OpenAI-compatible API, per-hour GPU rental, and bring your own model
Model catalog

Curated, ready to deploy

Pick a model — it downloads, deploys, and serves behind an OpenAI-compatible endpoint in minutes. vLLM and llama.cpp runtimes, chosen automatically.

deepseek-r1-671bdeepseek-v3-671bllama-3.1-405b-instructllama-4-maverick-400bqwen3-235b-a22bmixtral-8x22b-instructdbrx-instruct-132bcommand-r-plus-104bqwen2.5-72b-instructllama-3.3-70b-instructqwen2.5-coder-32bgemma-2-27b-itmistral-7b-instructllama-3.1-8b-instructqwen2.5-7b-instructphi-4+ bring your own — safetensors or GGUF
How it works

From GPU to running pod in three steps

From catalog to live endpoint in three steps
GPU virtualization

Slice a GPU — don't waste it

NUSAPOD partitions every physical GPU into isolated VRAM slices. Multiple workloads share one card safely — no oversubscription, no fragmentation — and whatever you don't use stays available.

NUSAPOD slices each GPU into isolated VRAM partitions: workloads share GPUs with isolation while unused capacity stays available — versus whole-GPU allocation that fragments and wastes VRAM
Use cases

One platform, many ways to run AI

From a single appliance to a distributed fleet — the same console deploys, orchestrates, and operates the workloads on top of it.

Private AI, inside your walls

Serve LLMs on infrastructure you own. Prompts, weights, and usage data never leave your network — with scoped API keys and exact per-key token accounting.

GPU-as-a-Service for data centers

Turn a GPU estate into a service: carve right-sized slices per team, meter requests and tokens per key, and keep utilization visible per card.

Mixed-use GPU estates

Coexists with workloads it doesn't manage. Physical-VRAM awareness sees every process on the card and refuses over-commits before they OOM.

AI product backends

Every model ships as an OpenAI-compatible endpoint with custom paths, streaming, capacity alerts, and one-click VRAM resizing as traffic grows.

Research & experimentation

Provision a slice per experiment in minutes — warm model caches make restarts instant, and released VRAM goes straight back to the pool.

Edge & appliance-class hardware

Runs where standard tooling breaks: unified-memory GPUs (NVIDIA GB10 / DGX Spark-class) are first-class, verified on real hardware.

Ready to serve your first model?

From a model pick to an OpenAI-compatible endpoint in minutes — on GPU infrastructure you own and control.