NVIDIA's self-developed large model Nemotron 3 Ultra debuts: cost-performance ratio maximized, inference costs drop 90% compared to closed-source models

The Zhitong Finance APP noted that NVIDIA (NVDA.US) stated on Wednesday that its open-source Nemotron 3 Ultra AI model has surpassed closed-source models in terms of cost.

NVIDIA stated in a statement that the open-source framework Langchain claims that its Deep Agents suite, specifically designed for Nemotron 3 Ultra, can achieve inference costs that are 10 times lower per run than some "leading" closed-source models, completing more tasks and achieving higher throughput.

NVIDIA added that when compared to LangChain's Deep Agents benchmark tests, Nemotron 3 Ultra's performance on business tasks is also on par with the highest-scoring models.

Harrison Chase, co-founder and CEO of LangChain, stated at the press conference: "The way to build better agents is to continuously improve the systems around the models. When teams can put them together for fine-tuning, memory, tool usage, evaluation, and model behavior create a compounding effect. Our collaboration with NVIDIA shows that enterprises can achieve strong performance from the open-source tech stack while maintaining control over the agent systems they are building."

NVIDIA also pointed out that Abridge, Amdocs, and Box are embedding dedicated agents directly into their platforms, and Ernst & Young (EY) is expanding the application of NVIDIA's capabilities around its NemoClaw blueprint.

The "Ultimate Moat" of Integrated Software and Hardware

NVIDIA's core purpose in developing large models has never been to compete directly with OpenAI or Anthropic by selling large model APIs, but rather to consolidate and expand its absolute dominance in the AI industry chain.

NVIDIA's biggest moat is not its chips, but its CUDA software ecosystem. Today, by launching its own large models (such as Nemotron 3 Ultra) and packaging them into the NVIDIA NIM (microservices framework), it enables enterprises to deploy and fine-tune on NVIDIA hardware with one click. NVIDIA's development of large models aims to make its software stack (AI Enterprise) more attractive, thereby firmly binding customers within NVIDIA's hardware and software ecosystem, a concept known as "full-stack monetization."

Closed-source large models like OpenAI's must be accessed through public clouds, and due to data privacy and security restrictions, many government agencies, military, financial giants, and sensitive industries cannot use them. NVIDIA has launched the "weight open-source, allowing private deployment" Nemotron model, deeply collaborating with enterprise software giants like Palantir and DataRobot. This large model, which operates in a completely isolated and secure local environment, can perfectly penetrate the high-profit markets of government and enterprise that public cloud giants cannot access.

Just as Microsoft created Surface computers to showcase the Windows ecosystem, NVIDIA trains large models to demonstrate to global enterprises "how optimized the performance and cost of large models can be on Hopper and Blackwell chips." Taking the latest Nemotron 3 Ultra as an example, NVIDIA, in collaboration with LangChain, has demonstrated that its inference costs are nearly 10 times cheaper than those of closed-source large models on the market.

In addition to general language models, NVIDIA has also made deep investments in Cosmos (world model), GR00T (robot large model), and others. These types of models can simulate the real physical world and mechanical responses, making them indispensable brains for autonomous driving, humanoid robots, and digital twins in smart factories (Omniverse). In this industrial-grade hardcore AI track, traditional internet large model companies are not adept, which is precisely where NVIDIA's chips will find their greatest future demand