Nvidia Is Close to Buying Inception for $500 Million to Get Diffusion LLMs
Nvidia is close to acquiring the startup Inception for at least $500 million, The Information reports. Inception trains language models to write an answer all at once instead of one token at a time, as ChatGPT does. The model drafts the whole page first, then refines it over several passes, much like an image generator. If the approach works, Nvidia could strengthen its open Nemotron models, and developers would get faster LLMs for agents and code autocompletion.
AI-processed from The Decoder; edited by Hamidun News
Conventional language models write an answer the way a person dictates a letter: word by word, with each new word depending on the ones before it. That makes long answers slow, and in an agent chain of dozens of steps the delays add up. According to The Information, Nvidia has decided to buy a team that is trying a different route: diffusion large language models (LLMs).
These models produce a draft of the entire answer at once and refine it several times, like an editor revising a whole page rather than adding to it from left to right. Three sources told the publication that Nvidia is close to acquiring the startup Inception for at least $500 million. This is not yet a confirmed acquisition: neither Nvidia nor Inception initially commented on the news.
The practical point for developers is simple: if this approach works at scale, code autocompletion, voice assistants and agents could respond noticeably faster.
What Inception does and how it differs
Inception has existed for about two years, is based in the San Francisco Bay Area and has more than 30 employees. The startup builds language models on diffusion, an approach familiar from image generators. There, an image starts out looking like noise and then emerges over several passes.
Inception applies the same idea to text. A conventional model, like those behind ChatGPT and Claude, generates an answer one token at a time (a token is a piece of a word). A diffusion model creates a draft of the whole answer and improves it step by step. In theory this brings a speed gain, because many tokens are processed in parallel.
The startup's latest release is Mercury 3. According to The Decoder, Inception claimed in October that the model outperforms GPT-6 Luna and Claude Haiku 5.5 with significantly lower latency. This is the company's own claim, and the source cites no independent measurements.
Why Nvidia wants it
According to The Information, the acquisition is meant to strengthen Nvidia's open Nemotron line and its physical models for robots. Palantir and US government agencies already use Nemotron as an alternative to Chinese open models. This is an important market for Nvidia: if the company has strong open models, it is easier for customers to stay on its hardware and software.
There is also a competitive backdrop. As the article notes, OpenAI and Anthropic are building their own chips for inference, meaning running finished models. Nvidia earns money because the market runs models on its GPUs, and competitors' in-house chips weaken that model. An architecture that is fast and cheap to run helps show that Nvidia's platform remains the best place for inference.
The deal includes performance bonuses for Inception employees. The co-founders, professors Stefano Ermon, Aditya Grover and Volodymyr Kuleshov, are expected to join Nvidia full time. The publication says the talks began before Nvidia CEO Jensen Huang's resignation and after Microsoft showed interest in the startup. The Decoder reported separately on the resignation: it came on November 12, after a spy scandal that cut the company's market value by $1.4 trillion.
What this means for developers
First, diffusion LLMs are no longer an exotic niche. When the largest accelerator maker pays upward of half a billion dollars for a team of thirty-odd people, it is betting on the technology, not on current revenue.
Second, if you build products where latency matters (code autocompletion, voice assistants, agents with dozens of steps), it is worth looking at these models now. Response time in agent chains adds up at every step, so the gain may matter more than an improvement in quality.
There are also limitations. The deal has not closed yet, the price is stated as a minimum, and the comparisons of Mercury 3 with competitors come from the startup itself. It is unknown how the diffusion approach will behave on very hard reasoning tasks and with long contexts. It is also unclear whether Inception's open weights will appear inside Nemotron. For now, the practical takeaway is this: watch for independent Mercury 3 benchmarks, and test diffusion models on your own tasks where response time is critical.
- Nov 20, 2026Four Hours on an RTX 5090 Cost 348 ₽, or From 1 730 ₽ With a One-Day Minimum
- Nov 20, 2026Nvidia quits selling H200 in China, and new US rules threaten $44–62 billion in lost revenue
- Nov 20, 2026Nvidia's stakes in startups grew from $22 billion to $65.5 billion, but without Huang the deals are in question
- Nov 20, 2026Epoch AI: AI Chip Capacity for Agents Grew 4.6x in a Year
Key Questions Answered
What Inception does and how it differs?
Inception has existed for about two years, is based in the San Francisco Bay Area and has more than 30 employees. The startup builds language models on diffusion, an approach familiar from image generators. There, an image starts out looking like noise and then emerges over several passes. Inception applies the same idea to text. A conventional model, like those behind ChatGPT and Claude, generates an answer one token at a time (a token is a piece of a word).
Why Nvidia wants it?
According to The Information, the acquisition is meant to strengthen Nvidia's open Nemotron line and its physical models for robots. Palantir and US government agencies already use Nemotron as an alternative to Chinese open models. This is an important market for Nvidia: if the company has strong open models, it is easier for customers to stay on its hardware and software. There is also a competitive backdrop.
What this means for developers?
First, diffusion LLMs are no longer an exotic niche. When the largest accelerator maker pays upward of half a billion dollars for a team of thirty-odd people, it is betting on the technology, not on current revenue. Second, if you build products where latency matters (code autocompletion, voice assistants, agents with dozens of steps), it is worth looking at these models now. Response time in agent chains adds up at every step, so the gain may matter more than an improvement in quality. There are also limitations.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.