Nvidia Plans to Acquire Diffusion Language Model Startup Inception for More Than $500 Million
According to The Information, citing people familiar with the matter, Nvidia is in advanced talks to acquire Inception for more than $500 million. Upon completion of the deal, its three co-founders will join Nvidia full-time. Inception primarily develops the Mercury family of diffusion language models, and its technology and talent are expected to bolster Nvidia’s Nemotron model team. Neither Nvidia nor Inception had responded as of publication.
Why read this
AI score 61
The deal details reveal Nvidia’s approach to strengthening its Nemotron team and reflect the growing concentration of diffusion language model talent and technology at major tech companies.
Topics
Body · AI translation
English text is being prepared.
According to the sources, Nvidia’s latest acquisition is intended to strengthen the team behind its Nemotron family of models. Inception’s three co-founders—renowned machine-learning professors Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov—will join Nvidia full-time after the acquisition is completed. It has not been disclosed whether Inception will continue to operate independently after the acquisition or what will happen to its other employees.
In addition, Nvidia will grant earnout bonuses to the Inception employees joining the company based on the achievement of business targets.
Founded in 2024 and headquartered in the San Francisco Bay Area, Inception has more than 30 employees. The company primarily develops diffusion language models, with its flagship products comprising the Mercury series. Its early product, Mercury Coder, was used mainly for code generation, followed by the reasoning models Mercury 2 and Mercury 2.5. Inception’s latest model, Mercury 3, was released on October 19 this year and supports a context window of up to 1 million tokens.
The Mercury models increase text-output speed through parallel generation. While mainstream autoregressive language models generate tokens one at a time from left to right, Mercury can generate multiple tokens in parallel and then repeatedly refine the content.
Taking Mercury 3 as an example, comparative data published by Inception shows that the model outputs 1,068 tokens per second on Nvidia GPUs, while GPT-6 Luna and Claude Haiku 5.5 output approximately 130 and 240 tokens per second, respectively. With all three models using the Medium setting, Mercury 3’s output speed is approximately 8.2 times and 4.5 times that of the other two models, respectively.
According to evaluations published by Inception, with all three models using the Medium setting, Mercury 3 not only generated output faster but also outperformed GPT-6 Luna and Claude Haiku 5.5 on eight of 12 capability tests. On GPQA Diamond, which assesses graduate-level scientific questions, Mercury 3 scored 89.23, higher than the two comparison models’ scores of 86.36 and 84.51. On the instruction-following test IFBench, Mercury 3 scored 76.84, above the comparison models’ 61.82 and 54.65; on the agentic search test DeepSearchQA, it scored 71.33, higher than their respective scores of 60.00 and 65.22.
▲ Test results and model comparison published by Inception (Image source: Inception)
Previously, U.S. AI coding-tool company Augment Code used Mercury 2 to compress context from long coding sessions, retaining decisions already made, files involved, and unresolved issues so that agents could continue working. According to inference-service platform Baseten, citing people familiar with the matter, latency for this step fell by 82%, from approximately 150 seconds to 27 seconds, while costs declined by 90%.
The Information reported that if the transaction is completed, Inception co-founders Ermon, Grover, and Kuleshov will join Nvidia full-time.
According to their official Stanford profiles, the three had a teacher-student relationship during their time at Stanford and later all became prominent professors in machine learning. Ermon was Grover’s doctoral adviser and also supervised Kuleshov’s postdoctoral research. Before starting the company, the three had already collaborated for nearly a decade.
Inception co-founder and CEO Ermon, 42, graduated from Cornell University in 2015 and is now an associate professor of computer science at Stanford University, where he researches machine learning and generative AI.
In 2021, he was the third author of the paper “Denoising Diffusion Implicit Models (DDIM),” published at ICLR, a leading deep-learning conference. The research enabled diffusion models to generate images with fewer denoising steps without retraining. Experiments in the paper showed that DDIM could generate high-quality images 10 to 50 times faster than the previous DDPM (Denoising Diffusion Probabilistic Models) method.
Co-founder and CTO Grover, 32, earned his doctorate from Stanford University in 2020 and is now an assistant professor of computer science at the University of California, Los Angeles (UCLA). He previously conducted postdoctoral research at the University of California, Berkeley, and served as a research scientist at Meta FAIR.
In 2016, as first author, he published the paper “node2vec: Scalable Feature Learning for Networks” at KDD, a leading data-mining conference. The paper proposed a method for learning node features from network connections that can be applied to scenarios such as movie recommendations and social-connection recommendations.
Co-founder and Chief Scientist Kuleshov earned his doctorate from Stanford University in 2017 and is now an assistant professor of computer science at Cornell University and Cornell Tech.
In 2024, he and his collaborators published the paper “Simple and Effective Masked Diffusion Language Models” at NeurIPS, a leading machine-learning conference. The research improved the training method for diffusion language models, narrowing their performance gap with autoregressive models on language-modeling benchmarks.
▲ Inception’s three co-founders, from left: Grover, Kuleshov, and Ermon (Image source: Inception)
The acquisition reportedly began before former Nvidia CEO Jensen Huang stepped down. Although Nvidia is facing turmoil over its leadership change, its deals involving AI startups and its efforts to develop proprietary large models to compete with Anthropic are continuing.
Nvidia and Microsoft had previously invested in Inception through their respective investment arms. On November 6, 2025, Inception disclosed that it had raised $50 million (approximately RMB 336 million) in seed funding, led by Menlo Ventures, with participation from Nvidia’s NVentures, Microsoft’s M12, and others. The Information, citing people familiar with the matter, said Microsoft had also expressed interest in acquiring Inception before Nvidia began acquisition talks. However, as of the publication of the report today, neither Nvidia nor Inception had responded.
As noted above, once the acquisition is completed, Inception’s diffusion language-model technology and research talent will support research and development for Nvidia’s open-source model project Nemotron. The project aims to help Nvidia build systems for creating and deploying future AI and to contribute to the open-source AI community.
On May 19 this year, Nvidia open-sourced Nemotron-Labs-Diffusion, a “tri-modal” diffusion language model that switches attention mechanisms to combine autoregressive decoding, diffusion-based parallel decoding, and self-speculative decoding within a single model, thereby increasing decoding throughput.
Nvidia has reportedly invested tens of billions of dollars in the Nemotron project, including several billion dollars to bring in technology and AI researchers through acquisitions or similar transactions. The Nemotron R&D team currently has several hundred employees.
Major technology companies are supplementing the talent and technology required for AI research and development by acquiring companies, bringing in core teams, and securing technology licenses. In March 2024, Microsoft brought in two co-founders and some team members from U.S. large-model company Inflection AI to establish Microsoft AI. In June 2024, Amazon brought in a co-founder and some employees from U.S. AI-agent company Adept and obtained licenses for its model and agent technology. Through such transactions, major companies can bring in teams with which they already have an established basis for collaboration and apply their accumulated research to the development of their own products.
Major companies gain experienced teams, researchers obtain the funding and computing resources needed to continue development, and some model-development efforts that were previously pursued independently are brought into the major companies’ organizations. If the acquisition of Inception is completed, Nvidia will further strengthen its in-house model team, and the transaction will also accelerate the concentration of research talent and technology in diffusion language models within major technology companies.
Sources: The Information, Inception, Nvidia
AI Trends