Chapter 5 · Part 3
Layer 4: Models
The fourth layer is what most people think AI is: the model. A model is the trained system that takes an input (a question, an image, a protein sequence, a camera feed) and produces an output. The three layers below exist to train models and run them.
Not just chatbots
When people say "AI model" today, they usually mean a large language model like the ones behind ChatGPT, Claude or Gemini. Huang is careful to say those are one category among many. In his essay, the models layer covers systems that understand language, but also biology, chemistry, physics, medicine and robotics.
Language models are the kind everyone knows: they write, code and answer questions.
Models for science and the physical world are where Huang sees some of the biggest payoffs: drug discovery, new materials, and what NVIDIA calls physical AI, meaning robots and self-driving cars that have to understand and act in the real world.
The three breakthroughs
At Davos, Huang named three changes that, in his view, moved models from impressive demos to useful tools:
- Grounding. Models can now research, check sources and reason through problems, so they make things up far less often than early chatbots did.
- Open reasoning models. He singled out DeepSeek-R1, released in early 2025, which he described as the first open reasoning model. Suddenly a model that could "think" step by step was available for anyone to download.
- Physical intelligence. Models that understand proteins, chemistry and physics, not just words.
Why open models matter to the whole cake
Some models are closed: you can only use them through the company's own app or API. Others are open: the trained weights are published, so any company, university or government can download the model, run it on its own hardware and adapt it.
Huang argues that open models are good for every layer of the cake. When DeepSeek-R1 came out, he says, it "activated demand across the entire stack." Thousands of organizations that would never have built a reasoning model themselves could now run one. Each of them needed infrastructure, which needed chips, which needed power.
It's worth noticing who is making this argument. NVIDIA sells into every organization that runs a model, open or closed, so more models running anywhere is good for NVIDIA. That doesn't make the point wrong, since open models really have spread AI much more widely. But it's a useful habit to ask which layer a speaker makes money from.
A model, however clever, doesn't earn anything until somebody uses it for something. That's the top layer.