NVIDIA's Nemotron 3 Ultra tops benchmarks as Blackwell lead times stay tight

Nemotron 3 Ultra is NVIDIA's pitch that its open model, orchestrated with LangChain's Deep Agents, can match or beat top closed models on price-performance — a play to keep developers building agentic systems inside the NVIDIA ecosystem rather than defaulting to OpenAI or Anthropic APIs. It reinforces NVIDIA's strategy of selling not just chips but the full model-and-software stack.

The supply signal matters as much as the model: H200 and Blackwell lead times remain 12–20 weeks for large orders, evidence that GPU demand still vastly outstrips supply. Grok 4.5 trained on tens of thousands of GB300s and Samsung's memory-driven profit surge both underline how tight the hardware market is. NVIDIA is also expanding revenue-sharing financing for emerging GPU cloud providers, effectively bankrolling its own demand ecosystem.

The threats are real, though. This week ZML (endorsed by Yann LeCun) released LLMD, a free inference server letting open-source LLMs run efficiently across NVIDIA, AMD, Google TPU, Apple Metal, and Intel Arc — an explicit attack on vendor lock-in — and DeepSeek and Meta are pursuing custom silicon. NVIDIA also partnered with Hugging Face to bring Cosmos 3 and new frameworks to LeRobot, extending into robotics. The strategic question is whether software abstraction layers like ZML erode NVIDIA's CUDA moat faster than its stack strategy can deepen it.