NVIDIA is the GPU and AI infrastructure company whose chips, networking and CUDA software stack sit under most of today’s AI training and inference. For The Signal’s readers it matters in two ways: it sets the hardware ceiling for what you can run locally or rent in the cloud, and it now ships its own open-weight model families, led by Nemotron.
Why NVIDIA matters for open and local AI
Almost every self-hosting decision we cover starts with an NVIDIA question: how much VRAM a card has, whether a model has a CUDA-optimized build, and whether a quantized format fits. Our local AI guide and ComfyUI install guide both assume an NVIDIA GPU as the default path, and the heavier video workflows in our ComfyUI video workflow breakdown are sized against NVIDIA memory tiers.
Hardware platforms
- Blackwell: the current data-center and GeForce RTX 50-series generation, which introduced the 4-bit NVFP4 number format NVIDIA now uses for its own model checkpoints.
- Rubin: the next platform, presented at CES in January 2026 as a six-chip design (Rubin GPU, Vera CPU, NVLink 6, Spectrum-X Ethernet Photonics, ConnectX-9 SuperNIC and BlueField-4 DPU). NVIDIA says Rubin is in full production and claims AI tokens at roughly one-tenth the cost of Blackwell.
- DGX Spark and Jetson: desktop and edge systems that let developers run the same models locally.
Open model families
At CES 2026 NVIDIA grouped its open models into six families: Nemotron (reasoning and agents), Cosmos (robotics and simulation), GR00T (embodied intelligence), Alpamayo (autonomous driving), Clara (healthcare) and Earth-2 (climate).
Nemotron 3, announced in December 2025, uses a hybrid mixture-of-experts design in three sizes: Nano (about 30B total, 3B active), Super (about 100B, 10B active) and Ultra (about 500B, 50B active). NVIDIA also released around three trillion tokens of pre-training, post-training and reinforcement-learning data plus the NeMo Gym, NeMo RL and NeMo Evaluator libraries, which makes Nemotron one of the more open model families by release artifacts, not just weights.
Nemotron 3.5 Lightning followed on August 11, 2026: a 30B-total, 3B-active MoE tuned as a fast execution model for long-running agents (tool calls, validation, sub-agent work). It ships in BF16 and NVFP4 under the OpenMDW-1.1 license on Hugging Face and ModelScope, alongside NeMo Switchyard, a router that sends planning to a frontier model and execution to Lightning.
What to watch
- How quickly Rubin-class capacity reaches cloud GPU rental prices.
- Whether NVFP4 checkpoints become the default format for open models that target a single consumer GPU.
- How Nemotron licenses evolve: OpenMDW-1.1 is permissive, but always check the model card before commercial use.
Facts on this page were last checked on September 23, 2026 against NVIDIA’s newsroom, its CES 2026 keynote recap and the Nemotron 3.5 Lightning technical blog. See our corrections policy if you spot something out of date.