2025 08 15 dwani gpu grant

Open Source Grant Proposal – GPU Resources from Lambda.ai

Request

We request two H200 GPUs from Lambda.ai for six months to support ongoing open-source development and deployment of dwani.ai, our multimodal inference platform.

GPU Allocation Plan

GPU 1
- Purpose: Run Gemma3‑27B‑Instruct with vLLM for large-scale document, image, and text inference.
- Goal: Deliver fast and efficient LLM inference for Indian languages.

GPU 2
- Purpose:
- Automatic Speech Recognition (ASR)
- Text-to-Speech (TTS)
- Translation
- Arm64 library builds and optimization
- Goal: Ensure arm64 compatibility and expand multimodal capabilities.

Technical

Lambda.ai offers discounted GH200 pricing due to arm64 library compatibility challenges.
We have rebuilt most required libraries from source, enabling arm64-compatible multimodal inference on dwani.ai. - https://github.com/dwani-ai/vllm-arm64/releases - https://github.com/dwani-ai/docs/tree/main/docs/misc/v2

Cost Breakdown

Resource	Cost/hour (USD)	Hours	Days	Total (USD)
1× H200	1.49	24	180	6,436.80
2× H200	2.98	24	180	12,873.60

Total GPU Cost (2× H200 / 6 months): $12,873.60
Grant Budget: $15,000
Balance: $2,126.40

Use of Remaining Balance

The remaining funds will be used for:
- Developer workshops
- Conference participation
- Community event engagement

Other Grants & Status

We received $7,500 in GPU credits from Lambda via the NVIDIA Inception Program.
100% used for building dwani.ai and offering open API access.
lack of additional resources will require suspending API access for workshops + student projects.
We received USD 25k Google cloud credits
we use for UI + API router deployment.
GCP GPU costs are 5x compared to nebius/lambda labs, it costs 10K USD per H100 GPU per month.
Its not suitable for inference API for dwani.ai

Impact

This grant will allow us to:
- Continue multimodal inference without interruption
- Keep API access free for workshops and student developers - Contribute arm64-supported open-source tooling to the AI community