이미지: METAL LAB 생성
Summary
- The largest version of NVIDIA's open-weight Nemotron 4, currently in development, is reportedly designed with at least 1 trillion parameters, double the size of Nemotron 3 Ultra
- NVIDIA has reportedly tripled its cloud investment for training its own models to $28 billion through 2031, with release expected as early as this fall
- Even at 1 trillion parameters, the model falls short of Moonshot AI's Kimi K3 (2.8 trillion) and DeepSeek V4 Pro (1.6 trillion), and a gap in intelligence scores has also emerged
- 넴트론4 최대 모델 파라미터
- 최소 1조 개, 넴트론3 울트라의 2배
- 자체 모델 클라우드 투자
- 2031년까지 280억 달러로 3배 증액
- 공개 시점
- 이르면 2026년 가을
- 경쟁 모델 파라미터
- 킴K3 2.8조 · 딥시크 V4 프로 1.6조
- 지능 지수 비교
- 넴트론3 울트라 38점 · 킴K3 약 60점 (Artificial Analysis Intelligence Index)
- 넴트론3 울트라 위상
- 출시 당시(6월) 오픈 미국 모델 중 최상위, 킴K2.6에는 뒤처짐
- 출처
- The Information (디코더 인용)
1 trillion parameters, still not enough for second place
NVIDIA's next-generation open-weight model, Nemotron 4, is being designed with at least 1 trillion parameters in its largest version, according to a report by The Information. That's twice the size of Nemotron 3 Ultra, released last June. But the number doesn't necessarily put it at the top of the industry. Moonshot AI's Kimi K3 reportedly has 2.8 trillion parameters, while DeepSeek's V4 Pro comes in at 1.6 trillion. Even if NVIDIA hits the 1 trillion mark, it would only be catching up to a scale Chinese labs have already surpassed.
A performance gap remains
Parameter count alone doesn't guarantee performance, but the current benchmark gap is clear. Nemotron 3 Ultra was the strongest openly released American model at launch, but it still fell short of Kimi K2.6, which held the top spot at the time. In the latest version of the Artificial Analysis Intelligence Index, Nemotron 3 Ultra reportedly scored 38, while Kimi K3 scored around 60. That 20-plus-point gap is the wall Nemotron 4 will need to clear.
| Model | Parameters | Intelligence Index | Developer |
|---|---|---|---|
| Nemotron 4 (upcoming) | 1T+ | Undisclosed | NVIDIA |
| Kimi K3 | 2.8T | 60 | Moonshot AI |
| DeepSeek V4 Pro | 1.6T | Undisclosed | DeepSeek |
| Nemotron 3 Ultra (current) | Half of Nemotron 4 | 38 | NVIDIA |
Why NVIDIA is pouring in $28 billion
NVIDIA has reportedly tripled its cloud investment for training its own models to $28 billion through 2031. It's an unusual position for a GPU seller to also be building large language models itself — but there's a clear interest at stake. The more companies run open models on their own servers, the more GPU demand grows. As our publication reported on August 8, NVIDIA signed the "Open Weights and American AI Leadership" open letter last July, joined by more than 200 companies and institutions, signaling its support for the spread of open-weight models. The physical AI model Cosmos 3, released around the same time, is part of the same open strategy.
Caught between regulation and customers
At the same time, NVIDIA has also put its name on petitions opposing regulation of open models. This comes as the Trump administration is reportedly considering restrictions targeting specific Chinese models. The catch is that the more successful Nemotron 4 becomes, the more NVIDIA finds itself in direct competition with major customers like OpenAI. Selling hardware while also selling models built on that same hardware could send a subtle but uncomfortable signal to those customers.
So what changes
Nemotron 4 is expected to launch as early as this fall. Once it does, benchmark performance is likely to matter more than parameter count as the real measure of success. This case shows that hitting 1 trillion parameters no longer guarantees a top-tier position. It suggests that the open-weight race between U.S. and Chinese labs is shifting from a scale competition to one focused on closing the actual intelligence gap. For NVIDIA, it remains to be seen how long its two business lines — selling GPUs and developing its own models — can coexist without friction.
