Inception raises $50 million to bring fast, affordable diffusion models to text and code

AI start-up Inception just secured $50 million in fresh capital. The round was led by Menlo Ventures, with backing from Microsoft’s M12, Nvidia, Databricks, and Snowflake.

Inception is betting on diffusion models, or dLLMs, which don’t generate text word by word like autoregressive LLMs. Instead, they refine content step by step—a method that, until now, has mostly powered image generators. Inception wants to bring this approach to text and code.

Inception first introduced its models in February 2025, and Google demoed its own version, Gemini Diffusion, in May 2025. The company’s updated Mercury model claims to generate over 1,000 tokens per second. By comparison, classic autoregressive models like GPT-5 usually top out at 40 to 60 tokens per second.

Mercury is available through partners like OpenRouter and Poe, with pricing at $0.25 per million input tokens and $1 per million output tokens—offering clear speed and cost advantages over standard LLMs.

Correction: An earlier version of this article confused Inception with "Inflection AI," the former AI start-up founded by Mustafa Suleyman.