NVIDIA plans to invest $26 billion in open-weight AI models over the next five years, according to The Decoder’s reporting on a financial filing with the US Securities and Exchange Commission. Company executives confirmed the plans in interviews with WIRED, as cited in The Decoder’s coverage. This is a financial disclosure, not a press release. That makes it harder to discount.
Alongside the investment announcement, NVIDIA released Nemotron 3 Super, a 128-billion-parameter open-weight model the company describes as its most capable to date. On the Artificial Analysis Index benchmark suite, Nemotron 3 Super narrowly outperforms OpenAI’s GPT-OSS and lands roughly on par with Anthropic’s Claude 4.5 Haiku, while falling short of top competitors. NVIDIA also states the model delivers significantly higher throughput for agentic AI use cases, though that specific claim hasn’t been independently verified in available sources.
The context matters here. The Decoder frames this explicitly as NVIDIA stepping into an open-source gap that OpenAI, Meta, and Anthropic have each pulled back from in different ways. A chip company committing $26 billion to open-weight software isn’t a pivot. It’s a platform strategy. Developers who build on NVIDIA’s open models build within NVIDIA’s hardware ecosystem. The open-source investment and the hardware business aren’t separate bets.
For developers evaluating open-weight model stacks, Nemotron 3 Super enters the field as a credible mid-tier option, competitive with Claude 4.5 Haiku on general benchmarks, available without licensing cost, and positioned explicitly for agentic workloads. Whether the throughput claims hold at scale is a question practitioners will answer in production.