Nvidia has confirmed plans to invest $26 billion over five years in building open-weight AI models, a disclosure buried in a 2025 financial filing and confirmed by company executives to WIRED. The GPU maker that built its fortune selling chips to AI labs is now trying to become one — putting it in direct competition with OpenAI, Anthropic, Google, and China's open-model contingent. The timing is pointed: the best American models remain locked behind cloud APIs, while top Chinese offerings from DeepSeek, Alibaba, Moonshot AI, Z.ai, and MiniMax are freely downloadable and increasingly the foundation of choice for startups and researchers outside the US.
To mark the announcement, Nvidia released Nemotron 3 Super, a 128-billion-parameter open-weight model that the company claims outperforms OpenAI's GPT-OSS on the Artificial Intelligence Index — scoring 37 versus GPT-OSS's 33, though several Chinese models ranked higher still. The company also disclosed it has completed pretraining on a forthcoming 550-billion-parameter model, suggesting ambitions well beyond today's release. Nemotron 3 Super incorporates architectural techniques targeting improved reasoning, long-context handling, and reinforcement learning responsiveness, and topped a new robotics benchmark called PinchBench.
The investment logic runs in two directions at once. Releasing models tuned to Nvidia hardware deepens ecosystem lock-in: startups and researchers who build on Nemotron are more likely to reach for Nvidia GPUs. VP Bryan Catanzaro also cast it in geopolitical terms, positioning the open model push as a US-made answer to the Chinese open-weight ecosystem that has quietly become the backbone of AI development outside closed American labs. VP Kari Briski added that building the models doubles as an internal stress test, pushing Nvidia's own compute, storage, and networking infrastructure to supercomputer-scale limits.
Nathan Lambert of the Allen Institute for AI's ATOM (American Truly Open Models) Project called himself "a huge Nemotron fan" while arguing that private investment needs government supplementation. Andy Konwinski of the Laude Institute described the commitment as "an unprecedented signal of their belief in openness," noting that few companies sit at the nexus of so many competing AI efforts simultaneously. If the model line delivers, it could shift the foundation on which the next generation of AI agents is built — offering a well-resourced, openly available US alternative to the Chinese models that currently dominate that space.