The balance of power in open-weight AI has been tilted toward Chinese labs — DeepSeek, Alibaba, Moonshot AI, Z.ai, and MiniMax all release frontier models freely, while the leading American offerings from OpenAI, Anthropic, and Google remain locked behind cloud interfaces. Nvidia is now moving to change that calculus directly.
According to a 2025 financial filing, the company plans to spend $26 billion over the next five years developing open-weight AI models — a commitment that would move it from chip manufacturer and software platform provider into the territory of a frontier model lab. Executives confirmed the investment in interviews, describing it as a deliberate strategic push rather than a side project.
What Nvidia Is Building
The company released Nemotron 3 Super alongside the disclosure — its most capable open-weight model to date, carrying 128 billion parameters. Nvidia claims the model scored 37 on the Artificial Intelligence Index, which aggregates performance across 10 benchmarks. OpenAI‘s GPT-OSS, which carries a comparable parameter count, scored 33 on the same index, though several Chinese models scored higher. The company also says Nemotron 3 Super ranks first on PinchBench, a new benchmark measuring a model’s ability to control OpenClaw.
Beyond benchmark performance, Nvidia disclosed a set of architectural and training techniques applied to Nemotron 3 Super that improve reasoning, long-context handling, and responsiveness to reinforcement learning. The company makes these technical details public alongside the weights, allowing startups and researchers to modify and build on them directly — a deliberate contrast to the approach of proprietary labs.
Bryan Catanzaro, VP of applied deep learning research at Nvidia, stated plainly: “Nvidia is taking open model development much more seriously. And we are making a lot of progress.” He also confirmed that the company recently completed pretraining a 550-billion-parameter model, though that has not yet been released publicly.
The Strategic Logic
The investment serves multiple functions simultaneously. Kari Briski, VP of generative AI software for enterprise, says the models are used internally to stress-test not just chips but also the storage, networking, and datacenter-scale systems Nvidia builds and sells. “We build it to stretch our systems and test not just the compute but also the storage and networking, and to kind of build out our hardware architecture roadmap,” she said.
The open-weight models are also tuned to run on Nvidia‘s own hardware, which deepens the company’s grip on an ecosystem already dependent on its silicon. Since Meta released Llama in 2023 — the first major open model from a large American AI company — much of the global startup and research community has gravitated toward open Chinese models as American alternatives stagnated or moved toward restricted access. Meta CEO Mark Zuckerberg recently signaled the company may not make future models fully open.
That gap created a risk for Nvidia: if the open ecosystem consolidates around models optimized for competing hardware, the company’s position as the default training infrastructure could erode over time. Spending $26 billion to anchor the open-weight ecosystem to its own architecture is, according to the announcement, how the company intends to prevent that outcome. As Catanzaro put it: “It’s in our interest to help the ecosystem develop.”
Photo by Brett Sayles on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
