(David Paul Morris/Bloomberg via Getty Images)
Mark Zuckerberg plans on acquiring 350,000 Nvidia H100 GPUs to help Meta build a next-generation AI that possesses human-like intelligence.
Zuckerberg mentioned the figure today as he announced his company’s long-term effort to develop an artificial general intelligence (AGI), or an AI that can learn and be used to perform a variety of tasks.
Meta's CEO envisions the AGI powering a wave of cutting-edge services and devices, such as more powerful digital assistants and augmented reality glasses. “Building the best AI assistants, AIs for creators, AIs for businesses and more, that needs advances in every area of AI,” he said in a video on Instagram.
But to get there, Meta is going to need Nvidia’s H100, an enterprise GPU that’s adept at training large language models. “We’re building an absolutely massive amount of infrastructure to support this,” Zuckerberg said. “By the end of this year, we’re going to have around 350,000 Nvidia H100s. Or around 600,000 H100 equivalents of compute if you include other GPUs.”
The 350,000 number is staggering, and it’ll also cost Meta a small fortune to acquire. Each H100 can cost around $30,000, meaning Zuckerberg’s company needs to pay an estimated $10.5 billion just to buy the computing power, not to mention paying all the electricity costs.
The statement also provides a glimpse into how far the leading tech companies will go to develop new AI models. Other tech giants, including Microsoft, Google, and Amazon, also likely bought between 50,000 to 150,000 Nvidia H100s last year, according to research firm Omdia.
In his Instagram post, Zuckerberg indicated Meta is already using the GPUs to train Llama 3, the company’s answer to OpenAI’s ChatGPT tech. But Meta’s own AI models won’t remain siloed away in the company’s servers. Instead, the company plans on open-sourcing them, like it did with Llama 2.
“This technology is so important and the opportunities are so great that we should open source and make it as widely available as we responsibly can,” he added.