Pratik on the Nebius Spotlight at GTC 2025

Pratik recorded a Nebius Spotlight conversation at the GeekWire Studios booth during NVIDIA GTC 2025. The format is short, under ten minutes, but it is one of the few public conversations where Pratik talks specifically about the GPU infrastructure behind training the Dhenu family, and what running a domain LLM at the cost discipline we run it at actually requires of a cloud partner.

The part on training economics is worth a listen if you are budgeting one of these builds. A general-purpose foundation model trains on the largest available compute and chases benchmark scores. A domain model trains on a smaller, higher-quality corpus and chases grounded accuracy on a narrow set of tasks, which means a lot of cheap early experiments and very few large-cluster runs. Pratik’s point in the episode is that this calls for a cloud partner who is comfortable with short bursty workloads instead of selling only long reservations.

Serving comes up too, and it is the piece I would flag for anyone specifying an ag deployment. A farmer who speaks a Marathi question into the phone and waits more than two seconds for the response loses confidence in the system. Latency matters before throughput does, and the serving stack has to sit close to both the farmer and the knowledge graph doing the retrieval. That constrains region choices in ways a generic AI-product team would never have to think about.

The stretch that landed hardest in the room was about cost. The economics of an agribusiness AI deployment are dominated by the per-query inference cost across millions of farmer interactions, not by the periodic cost of pre-training a new checkpoint. A cloud relationship set up around training contracts misses where the money actually flows. If a team is comparing infrastructure quotes for a farmer-facing rollout, that per-query line is the one to price first, and it is the reason a partner optimising for inference matters more to us than one selling training capacity.

The full spotlight is at youtube.com/watch?v=89MJYk4i2_g. GeekWire Studios shot a clean set at GTC 2025 with a number of agricultural AI founders, and the Nebius track is a useful walk-through of how the cloud layer is showing up for vertical AI companies right now.