AI & ComputeLOT 261120 · 3 min read

Nvidia Nears $500 Million-Plus Deal for Diffusion LLM Startup Inception

Nvidia is nearing a $500 million-plus acquisition of Inception, whose diffusion LLMs claim 10x faster, cheaper inference. Neither company has confirmed the single-outlet report.

By Nathan Brooks571 words

Key takeaways

  1. Nvidia is near a deal to buy Inception for more than $500 million, per The Information, Nov. 19; neither company has confirmed.
  2. Inception's Mercury diffusion LLM, launched February 2025, claims up to 10x faster and 10x cheaper inference, benchmarked above 1,000 tokens per second.
  3. Inception raised a $50M seed round in November 2025 led by Menlo Ventures, with Nvidia's NVentures among investors.
  4. The reported price is more than ten times the seed round's size.
  5. It would be one of Nvidia's first reported deals since Jensen Huang resigned as CEO.
Nvidia nears a deal to buy diffusion AI developer Inception for more than $500 million, The Information reports
Nvidia nears a deal to buy diffusion AI developer Inception for more than $500 million, The Information reports — AI-generated

Nvidia is nearing a deal to acquire Inception, a Palo Alto startup building diffusion-based language models, for more than $500 million, The Information reported on Thursday, Nov. 19. Neither company has commented publicly. No second outlet has independently confirmed the talks, and Newsquawk noted that a single media report leaves room for terms to change or the deal to collapse.

The report, by Stephanie Palazzolo, Phoebe Liu and Amir Efrati, is paywalled. Efrati wrote on X that the deal covers Inception "and the 3 ML professors behind it." Secondary summaries of the report said the technology would feed Nvidia's open Nemotron models.

What does Inception actually build?

Most large language models, including those behind ChatGPT and Claude, are autoregressive. They write a response one word fragment at a time, in sequence. Diffusion models, the method behind image and video generators such as Midjourney and Sora, work differently: they start from a rough draft of the whole output and refine it over several passes. That structure lets them process many tokens in parallel rather than one by one. Inception applies the approach to text and code.

The company launched its Mercury model in February 2025 and called it the first commercial-scale diffusion large language model. It claimed "up to 10x faster inference speeds and 10x lower inference costs" than conventional models. Founder and chief executive Stefano Ermon, a Stanford professor, told TechCrunch the models have been benchmarked at more than 1,000 tokens per second.

The claim is a commercial one, not just a technical one. Inference cost is becoming the dominant line item for AI deployment, and a tenfold cut, if it holds in production, would change the economics of serving large models. The parallel-refinement method trades sequential generation for whole-draft refinement, and benchmarks so far come from the company itself.

Who funded it, and at what price?

In November 2025, Inception raised a $50 million seed round led by Menlo Ventures. Participants included Mayfield, Innovation Endeavors, Microsoft's M12, Snowflake Ventures, Databricks Investment and Nvidia's own venture arm, NVentures. Angels included Andrew Ng and Andrej Karpathy.

A price above $500 million would value the startup at more than ten times that seed round, struck within weeks of the reported talks. Nvidia, already an investor through NVentures, would be buying out its co-investors at a steep premium. The company was founded by professors from Stanford, UCLA and Cornell.

Where does this fit in Nvidia's deal pattern?

The talks would be one of Nvidia's first reported deals since Jensen Huang resigned as chief executive. At its first all-hands meeting since his exit, the company promised staff raises and more equity. Nvidia has made a run of smaller talent-and-technology acquisitions, and a purchase of Inception would continue that pattern with one difference: it adds a model team rather than hardware.

For a company that sells the compute on which autoregressive models run, owning a method that claims to cut inference cost tenfold is a hedge. If diffusion LLMs prove out at commercial scale, Nvidia would hold the team that proved it. If they do not, the price is modest against the company's balance sheet.

Nothing is signed, and no company has spoken. Until a confirmation arrives, the $500 million figure stands on one paywalled report — and the diffusion question it raises about parallel inference will outlast whatever this deal becomes.

mlq.ai (Source)

Filed under

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,142 articles

Read also

« Previous article