Nvidia's Nemotron 4 Aims for One Trillion Parameters — Chinese Labs Already Passed That
Agentic AINvidia is developing a one-trillion-parameter open-weight model. The challenge: two Chinese labs already have models at that scale or above — and Nvidia is selling the hardware both sides are using to get there.
What Happened
Nemotron 4 is Nvidia’s next open-weight model, targeting at least one trillion parameters — twice the size of Nemotron 3 Ultra. A potential fall 2026 release has been reported. Nvidia has tripled cloud spending for model training to $28 billion through 2031.
The competitive context complicates the framing. On the Artificial Analysis Intelligence Index, Nemotron 3 Ultra scores 38 points. Moonshot AI’s Kimi K3 — at 2.8 trillion parameters — scores approximately 60. DeepSeek V4 Pro at 1.6 trillion parameters has also already passed the trillion-parameter threshold that Nemotron 4 is targeting. Nvidia is racing to reach a scale its competitors have already surpassed.
Why It Matters
The strategic tension is specific to Nvidia’s position. Open-weight models are good for GPU sales — companies that run their own models self-host, which means buying and operating clusters. Nemotron 4, if successful, creates an Nvidia-branded option for enterprises that want an open-weight frontier model with implicit hardware optimization for Nvidia infrastructure.
But Nvidia is also competing against its major customers. OpenAI and Google are both large buyers of Nvidia hardware and direct competitors in the model market. Nemotron 4 is a product aimed at enterprises that would otherwise choose OpenAI, Anthropic, or Google for frontier capability — or choose a Chinese open-weight model.
The regulatory angle adds another layer: Nvidia signed a petition opposing regulation of open-weight models while the Trump administration considers targeted restrictions on Chinese AI models. A Chinese model ban would benefit Nvidia’s Nemotron program directly, and Nvidia’s policy position aligns conveniently with its commercial interests regardless of how the regulation resolves.
The So What
The trillion-parameter number is not the interesting part of this story. The interesting part is that Kimi K3 and DeepSeek have demonstrated that open-weight frontier capability at this scale is achievable without Nvidia’s model program — using Nvidia’s hardware, trained at costs that have fallen dramatically. For enterprise teams evaluating open-weight models: the choice is no longer “open-weight vs. closed API” but “which open-weight model, from which origin, on which hardware.” Nemotron 4 adds a US-origin option at scale. It doesn’t determine the competitive outcome.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →