What NVIDIA is
NVIDIA is a semiconductor company best known for making GPUs — graphics processing units — that turned out to be exceptionally good at training and running AI. If you've used ChatGPT, Claude, or almost any other major AI product, there's a very good chance NVIDIA hardware was involved somewhere in building or serving it.
But calling NVIDIA a "chip company" undersells what it has become. It now offers a full platform: chips, software to run AI efficiently on those chips (like TensorRT-LLM), ready-to-deploy AI models under its own brand, and enterprise tools for building AI-powered applications. Think of it less like a parts supplier and more like the company that built the roads, the trucks, and some of the cargo.
Why it matters to you
If your organization is evaluating, buying, or building AI tools, NVIDIA's position matters for a simple reason: almost everything runs on its hardware. The major AI labs — OpenAI, Anthropic, Mistral — all train and serve their models on NVIDIA GPUs. When Anthropic signed a deal to access over 220,000 NVIDIA GPUs through SpaceX's Colossus data center, or when Mistral trained its Large 3 model on 3,000 NVIDIA H200 GPUs, those weren't coincidences. They reflect NVIDIA's near-universal presence in serious AI infrastructure.
This means NVIDIA's product decisions — which chips it builds, what software it supports, which partners it favors — ripple through the entire AI industry.
NVIDIA's own AI models
Beyond hardware, NVIDIA has been quietly building a substantial portfolio of its own AI models:
- Nemotron is its family of language and multimodal models. Nemotron 3 Super 120B is a large open-weights model that activates only 12 billion of its 120 billion parameters at a time (a technique called Mixture-of-Experts), making it fast and efficient. Nemotron 3 Nano Omni handles documents, audio, and video. Nemotron 3.5 Content Safety is built specifically for enterprise content moderation.
- Cosmos is NVIDIA's family of models for physical AI — robots and systems that need to understand and act in the real world. Cosmos 3 was released as the first open omni-model targeting physical AI reasoning and action.
- Gated DeltaNet-2 is a research-level architecture that outperforms competing approaches on certain efficiency benchmarks.
- Ising is a family of models for quantum computing calibration, already adopted by institutions like Fermilab and Harvard.
All of these are released as open-weights models — meaning researchers and companies can download and use them freely.
The partnership web
NVIDIA has made itself indispensable by investing in and partnering with the companies that might otherwise be its biggest customers or competitors:
- OpenAI: NVIDIA invested $30 billion in OpenAI's latest funding round and has a separate agreement to deploy 10 gigawatts of AI datacenter capacity together, with the first phase launching in 2026.
- Anthropic: NVIDIA invested up to $10 billion and is co-designing future chip architectures specifically for Anthropic's AI workloads. Claude models run on NVIDIA Grace Blackwell and Vera Rubin systems.
- Mistral AI: NVIDIA is a founding partner of the Nemotron Coalition, a multi-lab initiative to advance open-source AI. Mistral and NVIDIA jointly released Mistral NeMo, and Mistral's models are available as NVIDIA NIM inference microservices — pre-packaged, easy-to-deploy containers.
- Microsoft: Claude models (and by extension NVIDIA compute) are available across Microsoft's Copilot product family and Azure.
Using AI to design its own chips
One of the more striking developments in the events covered here: NVIDIA is using AI to design better NVIDIA chips. At GTC 2025, NVIDIA's chief scientist described several systems in active use:
- NVCell uses reinforcement learning and genetic algorithms to redesign thousands of chip layout cells overnight — work that would otherwise take ten engineer-months.
- PrefixRL designs arithmetic circuits that are 20–30% better than human-designed equivalents.
- ChipNeMo and BugNeMo are AI assistants fine-tuned on internal GPU documentation to help engineers find and fix bugs.
This is a feedback loop: better chips train better AI, which helps design better chips.
The open-weights bet
NVIDIA announced a $26 billion, five-year investment in open-weights AI models. The stated reason is partly strategic: Chinese AI labs have been building capable models on non-NVIDIA hardware, and a thriving open-weights ecosystem that runs best on NVIDIA chips keeps the company central to global AI development — even in markets where its hardware faces export restrictions.
The geopolitical dimension is real. DeepSeek gave Huawei early access to its upcoming V4 model for hardware optimization while blocking NVIDIA and AMD — a signal that China is actively working to reduce dependence on NVIDIA's supply chain.
Where it's heading
The events in this bundle point toward NVIDIA deepening its role on three fronts: as the default infrastructure for frontier AI training and inference, as a model publisher in its own right (especially for physical AI and robotics), and as a strategic investor whose chip roadmap is increasingly co-designed with the labs that use it most. The company is also pushing into enterprise software — NemoClaw, its agentic governance stack, launched with partners including Salesforce, Cisco, and CrowdStrike — suggesting ambitions well beyond selling GPUs.
