Why companies are ditching giant tech AI for open models

Hugging Face acts as a massive public digital library where thousands of AI researchers and developers share these open models and vast collections of training data. CEO Clem Delangue reports that roughly half of the Fortune 500 companies now rely on these open versions. The shift is driven by economics: while connecting to a rented model via a connection called an API—short for a digital link that allows one program to use another company's tech—is easy to start, the recurring costs can become massive as a company grows. By instead downloading and customizing their own models, businesses can cut out the middleman and stop paying per-use fees.

Think of a closed model as a professional chef in a private kitchen. If you need a meal, you place an order, pay the fee, and get the result, but you never see the secret recipe and you cannot change the ingredients. An open-source model is like having the recipe, the kitchen, and the staff at your own office. Because you hold the digital blueprint of the model, you can install it on your own servers. This means you do not have to pay a toll every time the model performs a calculation. You also get full control over what the model does, which is vital for businesses that handle sensitive data they cannot risk sending to a third-party company.

The core of the issue is power. When only a handful of enormous companies provide the AI we rely on, they effectively set the rules, the pricing, and the limits of what is possible. Delangue warns that if we move toward a future where a few corporations hold the keys to all intellectual progress, we lose out on the transparency that open standards provide. This becomes especially critical as AI begins to move out of our screens and into our physical lives, such as through home robotics. For these machines, which interact directly with our private homes and family lives, trusting a black-box system controlled by a distant entity is a much higher risk than using transparent, adaptable code we can inspect ourselves.