Nvidia is buying a different generation loop, not merely another model team

Nvidia is in advanced talks to acquire Inception for at least $500mn plus employee earn-outs, The Information reports. The two-year-old company has more than 30 employees. The agreement is not final, and a proposed acquisition price does not independently validate the models.

Inception’s Mercury models use diffusion rather than the autoregressive loop behind most chatbots. An autoregressive model commits to one token and then generates the next; a diffusion language model predicts many tokens in parallel, begins with an answer that resembles noise and repeatedly refines the whole sequence. That can reduce latency on long responses and makes the architecture attractive for code completion, voice and other interactive workloads.

The company’s technical report says its smaller coding models reached 737 and 1,109 tokens a second on Nvidia H100s and were up to ten times faster than speed-optimised frontier alternatives at comparable quality. Those are developer-reported results from 2025, not proof that diffusion now dominates stronger general models. Inception says newer Mercury versions are in production, but its public comparisons still come from the vendor.

The Information says Nvidia plans to fold the researchers into Nemotron and adapt their work to physical models for robotics. The economics are slightly paradoxical: a model that produces the same answer with fewer GPU-seconds can reduce hardware per request. Nvidia’s wager is that much lower latency creates more real-time inference—and that owning the architecture helps it shape the open-model software stack rather than merely sell the chips underneath it.