Why Nvidia and Stripe Are Spending Billions on 'Free' AI

Key points

• Nvidia is reportedly in talks to acquire Hugging Face, a platform for sharing freely downloadable AI models, for around $13 billion.
• Two weeks earlier, payments company Stripe bought OpenRouter, a service that routes businesses to open AI models, for more than $7 billion.
• Nvidia previously struck a $6 billion deal for Poolside, a company that builds open-weight AI models.
• Just 6% of companies currently use open-weight models, according to spending data from Ramp, a corporate finance platform.
• The push reflects fear among hardware and software giants that the biggest AI labs, OpenAI and Google, are building their own chips and cutting out middlemen.

Three deals. More than $26 billion. All for software that is, in a very real sense, free.

Nvidia, the company behind the specialised chips that power most of the world's AI, is reportedly close to paying around $13 billion for Hugging Face, first reported by TechCrunch AI. Hugging Face is best described as a library where researchers and companies can upload and download AI models, the trained programs that do things like answer questions or write code, at no charge. Think of it as GitHub, the free code-sharing site used by millions of developers, but specifically for AI.

Why would anyone pay $13 billion for something free?

Because the people using the free thing are enormously valuable customers. Nvidia sells the chips those developers run models on. Owning the platform is a way to steer millions of AI builders straight toward Nvidia hardware.

There is a second, more defensive reason. OpenAI and Google are now designing their own AI chips, threatening Nvidia's core business. Nvidia already makes its own family of open models, called Nemotron, but uptake has been modest. Buying Hugging Face would hand it a ready-made community of developers it could push toward its own products and technical standards.

What are open-weight models, and why do businesses care?

An open-weight model is an AI program whose internal settings, called weights, are published openly so anyone can download, modify and run it on their own computers. That is the opposite of models from OpenAI or Anthropic, which you can only access by paying for each use through their online services.

Running a model yourself is cheaper at scale. Nik Albarran, AI product lead at developer-tools company Jellyfish, told TechCrunch AI that businesses doing millions of repetitive tasks, such as answering customer service questions, can fine-tune an open model to handle those specific questions at a fraction of the cost of a premium service.

Still, adoption is modest. Only 6% of companies use open-weight models today, per Ramp spending data. Just 2% of software engineers surveyed by Jellyfish use them regularly.

For complex tasks like coding or multi-step reasoning, frontier models from the big labs still tend to win. But Albarran notes that as companies mature their AI workflows, switching to a cheaper open model becomes more practical. "If the prices continue to go up from the frontier labs, more and more companies will be forced to at least consider it," he said.

What does this mean for ordinary people?

For most people, nothing changes immediately. But if businesses find cheaper ways to run AI, the cost savings could translate into lower prices for AI-powered products, or simply allow smaller companies to compete with tech giants that currently rely on expensive cloud AI services.

Lin Qiao, CEO of Fireworks, an open-model hosting service for businesses, says her platform processes 40 trillion tokens per day, more than either Google's Gemini or OpenAI's API. A token is roughly three-quarters of a word, the basic unit AI services charge for. Her prediction: every company will eventually train its own specialised AI model for its own needs, the way companies today maintain their own websites.

The dominance of OpenAI and Anthropic is real. It is not guaranteed to last.