Open-source large language models (LLMs) such as ChatGPT are revolutionising the way we live our lives and do business, and there is no shortage of brilliant minds working to hone and adapt AI for specific uses from creating smart cities to lower carbon energy production.
But the computing power needed to convert a great idea into a functional, scalable application does not come cheap; the cost of computing power required to train the models can run into tens of millions of dollars, even in the relatively early stages of development.
That price tag risks restricting future AI breakthroughs to the handful of deep-pocketed big-tech players, like Meta, Google and Microsoft, who can stump up the development costs – and afford to fail. Meanwhile, the smaller innovative players in the open-source community, which have been the lifeblood of the sector, struggle to raise adequate finance, pay their bills and contend with the associated risks.
Stability AI, a UK-based start-up, recently entered discussions with investors to raise more money after it lost more than $30 million in the first quarter, while owing almost $100m in outstanding bills to cloud computing providers and others, according to media reports. This is not an isolated case.
ZeroGPU launch
In an effort to ensure that creativity and independent thinking is not lost to the industry, Hugging Face, which provides computing space and tools for open-source developers, has said it will provide some $10m-worth of free GPU – graphics processing unit – computing power to developers.
The company’s ZeroGPU initiative, to which AI builders gained early access in May 2024, is a shared infrastructure for independent and academic AI builders to run AI demos on Hugging Face’s machine learning platform via its Spaces hosting application.
The shared GPUs will be made available to multiple applications at the same time, so, rather than needing a dedicated GPU, a user can simply use the precise amount of GPU power they need, with the remaining capacity being available for other users. The company says this approach will give users the freedom to pursue their work without concern for computing costs.
Hugging Face has infrastructure that now plays host to more than 4m AI builders. In addition to an open-source AI community that includes academic labs, startups, and independent hobbyists, those who have released open models and datasets on the Hugging Face platform include heavy hitters such as Meta, Apple, NVIDIA, Bytedance, Snowflake, Databricks, Microsoft and Google.
ZeroGPU leverages Hugging Face's experience in hosting and serving more than 100 petabytes monthly from its Hub, which hosts over 1 million models that have been downloaded over a billion times.
Many of these are better than proprietary application programming interfaces (APIs), according to Hugging Face CEO Clément Delangue .
“For example, more than 35,000 variation models of Llama have been shared on Hugging Face since Meta’s first version a year ago—including more than 7,000 based on Llama-3—ranging from quantized and merged models to specialized models in biology and Mandarin, to name a few,” he said on announcing the ZeroGPU initiative.
The Hugging Face platform is regarded as a valuable tool across the industry. In mid-2023, the company raised $235 million in a funding round that included investment from Amazon, AMD, Google, IBM, Intel, Nvidia and Qualcomm and placed an estimated valuation on the company of $4.5bn.
The company was founded in 2016 by a group of French entrepreneurs working in New York, where it is based. It is named after an emoji.
(Image: Shutterstock)
