Nvidia Ships Nemotron 3 Ultra -- a 550B Open Model to Seed Its Own Ecosystem

Nvidia released Nemotron 3 Ultra, a roughly 550-billion-parameter open model, extending its strategy of giving away strong models to drive demand for the hardware that runs them. As the open-weights frontier heats up, Nvidia is using openness as a moat around its silicon rather than its software.

By the Numbers

~550B
Parameters
Open weights
License
Nemotron 3
Family
Trace Cohen
By Trace Cohen · Value Add Pulse AI Desk
Founder & Editor-in-chief, Value Add VC
1 min read
Nvidia Ships Nemotron 3 Ultra -- a 550B Open Model to Seed Its Own Ecosystem: key figures
ShareXLinkedInEmail

THE RUNDOWN

1

Nvidia gives models away to sell GPUs -- open weights are a demand-generation tool, not a product line

2

A 550B open model raises the bar for the open ecosystem just as enterprises seek self-hostable options

3

Nvidia partnering with Mistral and shipping its own open models hedges against any single lab's dominance

The VC Read

Value Add VC analysis

Nvidia giving away a 550B model looks generous until you remember the business model: every open weight it ships sells more GPUs. Openness is a complement to hardware, not a competing product -- and that's a structurally unbeatable position. The smart read for founders is that the open-weights frontier now has the world's most powerful company actively feeding it, which makes self-hosting a more credible strategy than ever. Watch whether Nemotron and the Mistral partnership pull developers onto the Nvidia stack by default.

Analysis

Nvidia released Nemotron 3 Ultra, a roughly 550-billion-parameter open model, the latest in its Nemotron family of openly available models. The release continues a deliberate Nvidia strategy: seed the ecosystem with strong, freely available models so that more developers and enterprises build on -- and ultimately buy -- the GPUs required to train and serve them.

The move lands as open weights become one of the most important themes in enterprise AI. Buyers increasingly want models they can self-host to hedge against vendor and regulatory risk, and Nvidia is happy to supply that demand because every self-hosted deployment is more compute sold. Openness, for Nvidia, is a complement to hardware rather than a competing product.

“The move lands as open weights become one of the most important themes in enterprise AI.”

Nvidia has also partnered with Mistral to accelerate a new family of open models, reinforcing a portfolio approach: rather than betting on any single lab, Nvidia spreads its influence across the open ecosystem to ensure the next wave of AI -- wherever it's built -- runs on its chips.

ShareXLinkedInEmail

More on

Nvidia →

More from the Pulse archive: Artificial Intelligence coverage → · All topics

Key Sources

1 source

THE WIRE in your inbox— Tech, startup & VC news with The VC Read, a few times a week. Free to subscribe, no spam.

No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗

Read Next

AI
Anthropic's Revenue Run Rate Tops $30B as It Locks Up 3.5GW of Google TPUs
Anthropic's Revenue Run Rate Tops $30B as It Locks Up 3.5GW of Google TPUs logo
AI$30B run rate

Anthropic's Revenue Run Rate Tops $30B as It Locks Up 3.5GW of Google TPUs

Anthropic said its annualized revenue run rate has passed $30B -- up from roughly $9B at the end of 2025 -- as it expands its Google/Broadcom partnership to secure 3.5 gigawatts of next-generation TPU capacity from 2027, on top of 1GW already coming online in 2026. It's one of the steepest revenue ramps and largest compute commitments in AI history.

AI
Anthropic Locks 3.5-Gigawatt Google TPU Deal as It Passes OpenAI on Revenue
Anthropic Locks 3.5-Gigawatt Google TPU Deal as It Passes OpenAI on Revenue logo
AI

Anthropic Locks 3.5-Gigawatt Google TPU Deal as It Passes OpenAI on Revenue

Anthropic committed to roughly 3.5 gigawatts of Google TPU capacity through 2031 via Broadcom -- one of the largest disclosed AI compute contracts ever -- as the company's revenue run rate hit a reported $30 billion annualized, up from $14 billion in February, overtaking OpenAI. The pairing of an enormous non-Nvidia compute deal with surging revenue marks a power shift in the model race.

AI
Huawei-Led Team Post-Trains DeepSeek's 1.6-Trillion-Parameter Model on Ascend Chips, Not Nvidia
Illustration for: Huawei-Led Team Post-Trains DeepSeek's 1.6-Trillion-Parameter Model on Ascend Chips, Not Nvidia
AI

Huawei-Led Team Post-Trains DeepSeek's 1.6-Trillion-Parameter Model on Ascend Chips, Not Nvidia

A research group including Huawei completed full-parameter post-training of DeepSeek's 1.6-trillion-parameter V4-Pro model using a cluster of at least 1,000 Huawei Ascend 910C chips -- demonstrating that a frontier-scale Chinese model can be trained end-to-end on domestic silicon. It's a milestone in China's drive to break free of US chips and the CUDA ecosystem.