Hugging Face commits $10 million in free shared GPUs

Hugging Face, one of the biggest names in machine learning, is committing $10 million in free shared GPUs to help developers create new AI technologies. The goal is to help small developers, academics, and startups counter the centralization of AI advancements.

“We are lucky to be in a position where we can invest in the community,” Hugging Face CEO Clem Delangue told The Verge. Delangue said the investment is possible because Hugging Face is “profitable, or close to profitable” and recently raised $235 million in funding, valuing the company at $4.5 billion.

Delangue is concerned about AI startups’ ability to compete with the tech giants. Most significant advancements in artificial intelligence — like GPT-4, the algorithms behind Google Search, and Tesla’s Full Self-Driving system — remain hidden within the confines of major tech companies. Not only are these corporations financially incentivized to keep their models proprietary, but with billions of dollars at their disposal for computational resources, they can compound those gains and race ahead of competitors, making it impossible for startups to keep up.

Hugging Face aims to make state-of-the-art AI technologies accessible to everyone, not just the tech giants. I spoke with Delangue during Google I/O, the tech giant’s flagship conference, where Google executives unveiled numerous AI features for their proprietary products and even a family of open-source models called Gemma. For Delangue, the proprietary approach is not the future he envisions.

“If you go the open source route, you go towards a world where most companies, most organizations, most nonprofits, policymakers, regulators, can actually do AI too. So, a much more decentralized way without too much concentration of power which, in my opinion, is a better world,” Delangue said.

Access to compute poses a significant challenge in constructing large language models, often favoring companies like OpenAI and Anthropic, which secure deals with cloud providers for substantial computing resources. Hugging Face aims to level the playing field by donating these shared GPUs to the community through a new program called ZeroGPU.

The shared GPUs are accessible to multiple users or applications concurrently, eliminating the need for each user or application to have a dedicated GPU. ZeroGPU will be available via Hugging Face’s Spaces, a hosting platform for publishing apps, which has over 300,000 AI demos created so far on CPU or paid GPU, according to the company.

Access to the shared GPUs is determined by usage, so if a portion of the GPU capacity is not actively utilized, that capacity becomes available for use by someone else. This makes them cost-effective, energy-efficient, and ideal for community-wide utilization. ZeroGPU uses Nvidia A100 GPU devices to power this operation — which offer about half the computation speed of the popular and more expensive H100s.

“It’s very difficult to get enough GPUs from the main cloud providers, and the way to get them—which is creating a high barrier to entry—is to commit on very big numbers for long periods of times,” Delangue said.

Typically, a company would commit to a cloud provider like Amazon Web Services for one or more years to secure GPU resources. This arrangement disadvantages small companies, indie developers, and academics who build on a small scale and can’t predict if their projects will gain traction. Regardless of usage, they still have to pay for the GPUs.

“It’s also a prediction nightmare to know how many GPUs and what kind of budget you need,” Delangue said.

With AI rapidly advancing behind closed doors, the goal of Hugging Face is to allow people to build more AI tech in the open.

“If you end up with a few organizations who are dominating too much, then it’s going to be harder to fight it later on,” Delangue said.

Andrew Reed, a machine learning engineer at Hugging Face, even spun up an app that visualizes the progress of proprietary and open-source LLMs over time as scored by the LMSYS Chatbot Arena, which shows the gap between the two inching closer together.

Over 35,000 variations of Meta’s open-source AI model Llama have been shared on Hugging Face since Meta’s first version a year ago, ranging from “quantized and merged models to specialized models in biology and Mandarin,” according to the company.

“AI should not be held in the hands of the few. With this commitment to open-source developers, we’re excited to see what everyone will cook up next in the spirit of collaboration and transparency,” Delangue said in a press release.