Nvidia is nearing an agreement to acquire Inception for at least $500 million, according to three people cited by The Information and relayed by The Decoder. Performance payments for the startup’s more than 30 employees are also said to be part of the proposed transaction.

Inception builds language models using diffusion, the iterative refinement process best known from image generation. Instead of committing to a response one token at a time, its Mercury models revise a broader output over several steps. The approach is intended to reduce waiting time, although real performance depends on task, hardware and serving software.

The reported plan would fold Inception’s technology into Nvidia’s open Nemotron model family and physical-AI work for robots. Co-founders Stefano Ermon, Aditya Grover and Volodymyr Kuleshov are expected to join Nvidia full time if the agreement closes. Palantir and US government agencies already use Nemotron models, according to the report.

This remains a report about negotiations, not a completed acquisition. Neither Nvidia nor Inception initially commented, and price, retention terms or regulatory review could still change. The strategic logic is clear: Nvidia is expanding beyond chips and inference software into models, while major model developers are designing their own hardware. Whether diffusion language models provide a durable advantage will require broader independent testing after any integration.