Nvidia is reportedly in the final stages of buying Inception, a two‑year‑old AI startup that uses diffusion – a technique first popular for image generation – to build language models. The purchase price is at least $500 million, according to three people familiar with the talks. In addition to the cash, the deal includes performance bonuses for Inception’s 30‑plus staff.
Inception’s approach differs from the token‑by‑token method used by models such as ChatGPT or Claude. Instead of generating each word one at a time, its diffusion models produce an entire draft and then refine it step by step. The most recent version, Mercury 3, demonstrates this iterative refinement on text tasks.
Nvidia says the acquisition will strengthen its open‑source Nemotron lineup and its physical‑robot models. Nemotron is already used by Palantir and several U.S. government agencies as an alternative to models from Chinese vendors. By adding diffusion‑based language capabilities, Nvidia hopes to better compete with rivals like OpenAI and Anthropic, both of which are also developing custom inference chips.
Why the deal matters now
The talks began before Nvidia CEO Jensen Huang announced his resignation and after Microsoft showed interest in the company’s AI efforts. Bringing Inception’s founders – Stefano Ermon, Aditya Grover and Volodymyr Kuleshov – onto Nvidia’s payroll could accelerate research that blends diffusion and large‑scale language modeling. That may translate into faster, more flexible AI tools for developers who rely on Nemotron’s open‑source code.
Why it matters
If the deal closes, developers who use Nvidia’s open‑source Nemotron models could gain access to diffusion‑based language technology, potentially improving the quality of generated text. It also signals Nvidia’s intent to compete more directly with OpenAI and Anthropic in both software and hardware. The exact timeline for integrating Inception’s models into Nemotron remains unclear.