Will single-tenant AI factories become the latest data center trend?

# Colocation data centers are typically designed to accommodate dozens or even hundreds of customers' diverse applications. However, Nvidia offers a unique data center model that is dedicated to running specific applications for a single customer.

The emergence of "artificial intelligence factory"

This new type of data center is different from traditional data centers. It focuses on providing more efficient and flexible infrastructure services . Traditional data centers often host multiple applications and multiple tenants, while new data centers focus more on dynamic allocation and optimization of resources to meet the needs of different applications and tenants. The design of this new data center is more flexible and intelligent, and can adjust resource allocation in real time according to demand, improving overall efficiency and performance. With this innovative design concept, these new data centers are primarily used to host a small number of applications, typically used by a single tenant. They are responsible for processing data, training models, and generating tokens to generate artificial intelligence. We call these new data centers “AI factories.”

Artificial intelligence factories have become a ubiquitous phenomenon. I believe almost every major region will have its own AI cloud, as will every major country. So we're at the beginning of a computing transformation, which is an important inflection point.

At present, this trend is gradually emerging in countries such as India, Sweden, Japan and France. To achieve effective use of artificial intelligence, language and cultural differences between countries must be taken into account. The demand for artificial intelligence also varies from country to country, such as Japan and Sweden. Because of this, AI data centers and single-tenant AI factories tend to be limited to specific countries.

Large cloud service providers such as Amazon and Google and major colocation providers such as Equinix, their data center size is usually comparable Huge, big enough to hold a football field. Given the high power consumption of Nvidia Hopper processors, these AI factories will be comparable in size to McDonald's restaurants.

Data center racks usually budget power consumption between 6kW and 8kW. However, if a server optimized for running LLM is required, the power consumption of a single server is approximately 11kW. This is equivalent to the average power consumption of approximately 14 general-purpose servers.

In this case, only a limited number of GPU servers, such as DGXH100, can be run in a typical data center. If you have a 1MW data center, you can deploy about 50 DGXH100 servers in it. Deploying AI at scale to large numbers of concurrent users will require large clusters of such servers. This means that a typical data center can only serve the needs of a limited number of customers, and most likely only a single customer.

The most cost-effective solution for designing single-purpose GPU environments such as AI factories is to build dedicated data centers with higher density and liquid cooling as the design focus, and positioning it in a location most suitable for artificial intelligence enterprises.

The power consumption of AI clusters will be a limiting factor in data centers with large numbers of servers, and it is likely that some of these data centers will be dedicated to AI. Safety and regulatory frameworks surrounding AI may also drive this trend. The growth of generative and general artificial intelligence raises several security and compliance issues, so enterprises may decide to run such workloads from highly secure, purpose-built facilities.

Since the power density of artificial intelligence is five to ten times that of traditional data centers, the scale of artificial intelligence factories will not reach that of traditional data centers The size of a traditional data center exceeds one million square feet.

Another difference between traditional data centers and AI factories is their location. While giant data centers tend to be built in remote locations next to renewable energy sources, AI factories can be built in city centers or metropolitan areas and in existing facilities with large amounts of available power.

There is a lot of office and retail space that is underutilized at the moment, and what becomes very, very attractive is an abandoned building or underutilized urban space, or part of an old warehouse in the middle of nowhere, They already have power, you can put some AI equipment in there, some liquid cooling and plug it in.

While it’s impossible to predict the future of the data center industry, the rapid growth of artificial intelligence suggests that AI factories may soon become a necessity as digital infrastructure operators scramble to meet growing demand.