Hugging Face, a small start-up with an emoji as its logo, has become one of the most important players in artificial intelligence. And not just because Nvidia is buying it for $12.9 billion.

The developer platform is a hub for open-weight AI models, which are a cheaper, more customizable alternative to leading-edge models from OpenAI and Anthropic. It is also an example of what AI may ultimately look like.

One vision of AI: Anthropic and OpenAI have surged to lead of the AI race with proprietary models that users aren't able to fundamentally alter. Businesses access these models through OpenAI and Anthropic's own interfaces or public clouds. These businesses pay through subscriptions or by the token, which is the standard measure of AI inputs and outputs consumed.

Both companies are able to charge a premium for their tokens, in large part because of the raw power of their models-and the trust they have built among customers. It appears to be working. Anthropic's estimated annualized revenue run rate sits at a staggering $65 billion ahead of its initial public offering, according to CNBC.

Hugging Face presents a more decentralized vision for AI built on a different type of model.

Open-weight developers release parts of their algorithms-called the weights-to the public, allowing users to customize models for their specifications. Customized models can respond in different conversation styles, learn from their users in different ways, and more. Companies can host these open-weight models on their own servers or private cloud, and developers don't directly receive royalties.

Per-token costs for open-weight models are often much cheaper. Claude's most expensive model, Fable, costs $50 per output token, according to Artificial Analysis. Open models can range from $15 down to 50 cents.

More than 3 million models are available on Hugging Face. Users are able to browse by model size, potential data-center and software providers, and ideal tasks, ranging from text generation to spreadsheets. Hugging Face sells both individual and business subscriptions that give users storage capacity and compute credits, and lists hourly prices for Nvidia processors.

Hugging Face also allows individual users to post their own models-or just post blogs The platform is, at its most basic level, like Reddit for computer nerds.

Many of the most powerful open-weight models come from Chinese labs. They tend to attract entrepreneurs and developers who are constricted more by cost than by regulatory or national-security concerns, explains Jeff McMillan, an AI consultant and former AI lead at Morgan Stanley.

"If I'm a 23-year-old researcher at Stanford, of course, I'm going to be using these open weight models," McMillan says. "I can't afford anything else."

But the gap between OpenAI and Anthropic and their open-weight competitors may be closing. Kimi-K3, an open model released by Chinese company Moonshot AI in July, outperformed its most advanced peers in certain evaluations. Microsoft introduced Kimi-K3 as an option for its AI infrastructure users.

In the U.S., meanwhile, Nvidia is investing in a family of open-weight models it calls Nemotron. Nvidia is the single largest contributor of models and data on Hugging Face, according to Nvidia CEO Jensen Huang.

The potential concern for Anthropic and OpenAI investors is that open-weight models that are "good enough," or even rival proprietary models, could cause token prices to fall across the industry-and put pressure on profit margins.

Meanwhile, Nvidia is preparing for either possibility: An AI industry centered around Anthropic and OpenAI. Or one where developers customize and ship models back and forth through platforms like Hugging Face.