Single-Tenant 'AI Factories' Break Colocation Data Center Mold

Before the emergence of generative AI, a primary business model for the data center industry involved colocation data centers that support hundreds of customers running different applications at the same time.

The explosive growth of Gen AI, as well as the voracious appetite of AI data-processing servers for electricity, is ushering in the advent of small “AI factories” that are roughly the size of a McDonald’s fast-food outlet.

Jensen Huang, CEO of AI chipmaker Nvidia, said during a recent earnings call that this new class of data centers is aiming to serve a single customer running a handful, or even just one, AI application, Data Center Knowledge reported.

“These new data centers [host] very few applications, if not one application, used by basically one tenant. It processes data, it trains models and then generates tokens and generates AI. And we call these new data centers ‘AI factories,’” Huang said.

According to Huang, AI factories are proliferating in India, Sweden, Japan and France. The enormous power needs of AI could soon bring AI factories to the U.S. A server utilizing Nvidia’s AI chips consumes 11 kilowatts, comparable to the power needs of up to 14 general-process servers, could soon bring AI factories to the U.S.

Single-purpose AI factories will have specialized designs featuring higher power density and liquid cooling. They will be located in proximity to AI companies, including in urban centers. The size of a single-purpose AI factory could be as small as 10K SF, the report said.

Hyperscale data center facilities of cloud service providers including Amazon and Google as well as major colocation providers have been massive campuses as large as 1M SF.

The rapid expansion of U.S. data center capacity to support the demand for artificial intelligence will double the size of the industry by 2030, measured in gigawatts. The U.S. data center footprint will absorb 35 gigawatts by 2030, more than twice the 17GW total in 2022, according to a new U.S. Data Center Market report from Newmark.

While data center footprints, power and cooling requirements will be adjusted in similar ways across global markets, the next phase of the development of generative AI bot training technology is likely to be much more provincial.

Because the training of AI bots is a reflection of the culture and language in which it takes place, the next generation of AI products are being designed to permit significant customization “that’s going to make a lot of people uncomfortable,” Open AI CEO Sam Altman told Axios in an interview at Davos this week.

AI bots trained for end-users in different countries with different languages and cultures may give different answers to the same questions, he said, based on “values preference.”

Altman also offered what may be a preview of OpenAI’s defense to a copyright infringement lawsuit filed by the New York Times, which says its content was used without permission to train ChatGPT.

Altman told Axios that OpenAI doesn’t need NY Times content to build successful AI models and would “respect an opt-out” by the publisher to the use of its content, adding “but NYT content has been copied and not attributed all over the web” and OpenAI can’t avoid training on that.