Nvidia’s Nemotron 3 Super anchors a $26B open-model gamble

Nvidia will spend $26 billion over five years building open‑weight AI models and is launching Nemotron 3 Super, a 120‑billion‑parameter system it claims can match or beat leading Chinese and proprietary rivals on core reasoning tasks. The move effectively turns the world’s dominant AI chip supplier into a fully fledged frontier lab — and signals that Washington’s most important AI hardware company is betting on openness, not secrecy, at the top end of the model race.

According to a recent financial filing detailed by Wired, the $26 billion plan covers training, scaling, and maintaining a family of large open‑weight models intended for use by enterprises, researchers, and US government agencies. That level of spending would put Nvidia in the same league as OpenAI, Google, Anthropic, and China’s DeepSeek in terms of frontier‑scale training budgets, but with a distinct twist: most of the resulting models are meant to be downloadable and customizable rather than locked behind APIs.

What Nemotron 3 Super Actually Is

Nemotron 3 Super is Nvidia’s most ambitious open‑weight model family to date: a 120B‑parameter hybrid Mixture‑of‑Experts (MoE) design that combines Mamba‑style sequence modeling with transformer attention, tuned specifically for long‑context “agentic” workflows and tool‑using systems. A technical report from Nvidia describes Nemotron 3 Super as an “open, efficient mixture‑of‑experts hybrid Mamba‑Transformer model for agentic reasoning,” with context windows of up to 1 million tokens and throughput optimized for Blackwell‑generation GPUs.NVIDIA Nemotron 3 Super Technical Report

Benchmarks compiled by independent evaluations suggest Nemotron 3 Super narrowly leads other frontier‑class open models from China and the US on several reasoning and coding suites.Qodo.ai reports that the model tops Qwen 3.5‑397B and MiniMax‑M2.5‑230B on a composite precision metric, while delivering substantially higher tokens‑per‑second on Nvidia hardware. Early community tests on platforms like Hugging Face and OpenRouter indicate that the model can be fine‑tuned and quantized down to consumer‑grade GPUs, bringing near‑frontier capabilities to startups and hobbyists.Hugging Face collectionMarktechpost

Nvidia is not fully open‑sourcing Nemotron 3 Super under a permissive license; instead, it is using its own Open Model License, which allows broad commercial use but retains some safety and attribution requirements.Hugging Face collection Still, the release includes weights, code, and training “recipes,” giving developers enough detail to reproduce or extend the models on their own infrastructure.

A $26 Billion Bet on Open Frontier Models

The $26 billion commitment disclosed in Nvidia’s filings is explicitly framed as a program to develop “frontier open‑weight” models — systems trained at or near the global state of the art whose weights can be inspected, hosted, and fine‑tuned outside Nvidia’s own cloud.Wired Analysts see the move as an attempt to ensure that, as Chinese labs like DeepSeek, Baidu, and Alibaba ramp up their own powerful open releases, American‑aligned infrastructure and tooling remain the default choice for companies and governments worldwide.Epoch AI

This aligns with a broader US industrial strategy built around Nvidia‑powered AI infrastructure: supercomputing projects with the US Department of Energy and national labs, large‑scale AI factories with partners like CoreWeave, and multi‑gigawatt data center plans under initiatives such as Stargate.Nvidia investor releaseWikipedia Open‑weight models trained on that infrastructure can then be deployed on‑premises by allied industries and agencies, rather than relying on foreign or purely proprietary stacks.

Faster Innovation, Faster Proliferation

By pushing frontier‑class capabilities into open‑weight releases, Nvidia is accelerating the diffusion of powerful AI tools into everything from robotics and code assistants to military decision‑support systems. Researchers will be able to study and stress‑test Nemotron 3 Super directly, rather than inferring behavior from API access, which could improve safety tools and evaluation frameworks. Enterprises can fine‑tune the model on sensitive internal data without sending it to a third‑party cloud, addressing compliance and sovereignty concerns.Qodo.ai

But the same openness also lowers barriers for malicious or reckless actors. Policy groups have warned that frontier‑scale open models could make it easier to automate cyber intrusions, disinformation, or biological threat design once capabilities cross certain thresholds.CNAS Nvidia’s license attempts to constrain some high‑risk uses, yet enforcement at global scale remains uncertain.

As countries race to shape guardrails for frontier AI, Nemotron 3 Super and Nvidia’s $26 billion bet crystallize a key strategic question: will open, American‑aligned models undercut China’s influence in AI — or simply speed up a worldwide proliferation of capabilities that no one fully controls?