We can’t send customer data to an overseas API.
The model runs on cards reserved for you, in India, inside your own network boundary. Nothing is pooled with other tenants and nothing leaves the country.
LLM hosting
Private LLM hosting in India for open-weight models.
Run Llama, Qwen, Mistral, Gemma or DeepSeek on GPUs dedicated to you, in India, from ₹32,000 per card per month. Your weights, your prompts and your data stay in your environment.
Common problems
What we fix.
The model works on a laptop and falls over with ten users.
We serve it through vLLM or a similar stack with continuous batching and the right quantisation, so one card handles many conversations at once.
Our team knows models, not Linux servers.
We install drivers and the serving stack, expose an OpenAI-compatible endpoint and hand you working API keys. You write prompts, not systemd units.
We’re not sure the model we want fits on one GPU.
An 8B model fits an L4. Models up to about 30B fit an L40S with quantisation. A 70B model fits one H100 at 4-bit, or an H200 for longer contexts. We size it with you.
Ways to start
Choose where to start.
One-month minimum on every GPU card; the price shown is per card per month. Prices exclude applicable taxes.
NVIDIA L40S 48GB
₹55,000per month
For Llama 3 8B, Qwen 14B, Mistral Small and similar models with a healthy context window.
- 48GB memory per card
- Runs quantised 30B-class models
- Good fit for internal assistants and RAG
Private LLM, deployed and managed
₹75,000per month
We deploy the model, secure the endpoint and keep it running. This is the AI Production Care service; the GPU card is rented on top at the prices on either side.
- Model, serving stack and API keys set up
- Access limited to your network or VPN
- Our engineers on call 24/7, P1 response under 15 minutes
NVIDIA H200 141GB
₹2,20,000per month
For Llama 3 70B, Qwen 72B and similar models at higher precision, or long documents with many users.
- 141GB memory per card
- Room for large KV cache and batching
- H100 80GB available for quantised 70B models
Common questions.
Which open-weight models can you host?
Any model whose weights you are licensed to run: Llama 3, Qwen, Mistral, Mixtral, Gemma, DeepSeek, Phi and fine-tuned variants of them. Share the Hugging Face repo and we confirm the GPU and serving stack.
Why does an LLM need a GPU?
Each token the model generates needs billions of multiply-add operations and a full pass over the weights. A GPU keeps the weights in fast memory and runs those operations in parallel, so a 70B model answers in seconds instead of minutes on a CPU.
How much GPU memory does my model need?
Roughly two bytes per parameter at 16-bit precision and about half a byte at 4-bit, plus memory for context. An 8B model needs about 16GB at full precision, so an L4 at ₹32,000 a month works. A 70B model needs an H100 at ₹1,80,000 with 4-bit weights, or an H200 at ₹2,20,000 for more precision and context.
Does hosting the model in India make us DPDP compliant?
Hosting in India keeps the data in the country, which is one requirement among many. Your policies, consent handling, access controls and audits are still yours. We can describe exactly where data flows so your compliance team can document it.
Is our data used to train anything?
No. The cards, the storage and the model are reserved for you. We do not log prompts or responses unless you ask us to for debugging, and we delete those logs when you say so.
Can our apps keep using the OpenAI client library?
Yes. The serving stack exposes an OpenAI-compatible chat completions endpoint. Change the base URL and API key and the rest of your code stays the same.