How they use Nebius AI Cloud

KissanAI uses Nebius GPUs to train and serve Dhenu, its family of language models for agriculture. The models support assistants that answer farmers' questions in Indian languages. Founder Pratik Desai described GPU-cluster training, cloud inference and data distillation.

KissanAI's GTC account describes short, bursty training jobs and farmer-facing inference costs. It says per-query inference costs across millions of farmer interactions dominate periodic pretraining costs, making inference optimization a priority. It explains that farmers lose confidence when a spoken question takes more than two seconds to answer. Serving must be close to farmers and the retrieval knowledge graph, which constrains region choices. A separate customer video identifies H100 GPUs and discusses agricultural vision-model work as a further training plan.

In the GTC spotlight interview, Desai described using Nebius inference endpoints for DeepSeek R1, synthetic-data queries and distillation to train KissanAI's own models. He reported uninterrupted access to GPU nodes and straightforward cluster access for his team. He also described obtaining a requested cluster that enabled experiments with larger agricultural image models for disease detection. The GeekWire spotlight summary describes training agricultural advisory copilots and credits Nebius with scalable infrastructure, low-cost inference and responsive support.