Nvidia's $28B Nemotron 4 bet: One trillion parameters to chase Moonshot and DeepSeek

Nvidia is pouring serious money into a problem most people didn’t think it wanted to solve: building frontier-scale models of its own. The company is developing Nemotron 4, an open-weight model family whose largest member will pack at least one trillion parameters — double the size of Nemotron 3 Ultra. Cloud spending on in-house training has tripled to $28 billion through 2031, according to The Information. An earliest release window of this fall is on the table.

Here’s the uncomfortable part: even hitting a trillion parameters only gets Nvidia to where Chinese labs already are. Moonshot AI’s Kimi K3 runs 2.8 trillion parameters. DeepSeek V4 Pro sits at 1.6 trillion. When Nemotron 3 Ultra launched in June, it was the strongest open US model on the Artificial Analysis Intelligence Index — but it still trailed Kimi K2.6. On the current index, Nemotron 3 Ultra scores 38 points while Kimi K3 lands around 60. That’s not a gap; that’s a gulf.

The strategic logic is genuinely interesting. Nvidia signed the petition opposing regulation of open models, and the more companies self-host open models, the more GPUs Nvidia sells. Nemotron 4 fits neatly into that flywheel: build a competitive open model, convince enterprises they can run frontier AI on their own hardware, and watch data center orders pile up. The Trump administration’s consideration of targeted bans on specific Chinese models only sharpens the pitch. If US companies can’t use Kimi or DeepSeek, Nvidia wants the obvious alternative to carry its own brand.

But there’s a real tension here that Nvidia hasn’t fully resolved. Nemotron 4 puts the company in direct competition with its biggest customers — OpenAI, Anthropic, and every other lab spending billions on Nvidia infrastructure. Selling shovels made Nvidia the most valuable company on earth. Deciding to mine for gold too, at a $28 billion scale, changes the relationship. Whether enterprise buyers actually care about the distinction between a 1-trillion-parameter model and a 2.8-trillion-parameter one remains an open question — especially when the scoreboard says the bigger one is already winning.