Most people recognize that AI can’t happen without data centers. But not everyone knows that what we call “AI” isn’t just one process happening in one place. It’s a series of interrelated workloads distributed across different types of data centers in different locations, and there’s often nuance required when it comes to determining which workloads should go where.
These AI workloads can be divided into three main categories:
- AI model training: How organizations develop their models. This involves processing massive volumes of data in order to establish pattern recognition.
- AI inference: How organizations apply their models. This involves making predictions based on the pattern recognition established during the training phase.
- AI data in motion: How organizations move data to support both training and inference. Training a model is an iterative process, meaning that datasets will need to cycle between locations consistently. Organizations need excellent connectivity between those locations to get AI data where it needs to go without delay.
These workloads have different requirements, and different kinds of data centers can help meet them. Instead of trying to adapt their legacy on-premises data centers for the AI era, businesses often turn to colocation data center providers to help meet the diverse needs of their AI workloads.
Colocation providers help customers unlock AI-ready infrastructure
Leading colocation providers have been consistently investing in AI-ready infrastructure for years. This includes scalable, powerful hardware to help customers keep up with the rapid growth of AI datasets. It also includes the advanced cooling capabilities and reliable power supply needed to keep that hardware running consistently. When customers deploy in a colocation data center, they instantly benefit from those investments.
There are different varieties of colocation data centers that appeal to different customers for different purposes:
- Traditional colocation data centers are used to support workloads with midsized capacity requirements. They’re typically distributed across many different locations to help customers deploy in proximity to their applications and end users, thereby keeping latency low.
- Hyperscale data centers offer a solution for customers that need to support very large workloads. Cost-efficient space and power are top concerns for these customers. That’s why hyperscale data centers are often found in remote areas where energy and real estate prices tend to be lower and vast land parcels are more readily available to support multiple-megawatt, campus-style developments.
Let’s take a closer look at how both hyperscale and colocation data centers are enabling the future of AI.
Different varieties of colocation for different AI workloads
AI model training requires a lot of compute capacity, which in turn needs a lot of power. Hyperscale data centers are better suited to provide that capacity and power. Also, model training isn’t particularly sensitive to network latency, so customers have the flexibility to take advantage of hyperscale data centers in many different locations—even those far away from their data sources.
In contrast, inference workloads are very sensitive to latency, whether it’s users looking for answers from a chatbot or radiologists using AI-driven image classification to streamline X-ray diagnosis, as in the case of Equinix customer harrison.ai. When a model makes predictions based on inference data, the recency of that data directly impacts the accuracy of the predictions. This means it’s essential for organizations to avoid delays in their inference workloads. They can achieve this with highly distributed colocation data centers that help them keep latency low. Inference datasets are also significantly smaller than training datasets, so they wouldn’t need the high capacity that hyperscale data centers provide.
While it’s essentially true that training is a good fit for hyperscale data centers and inference is a good fit for colocation data centers, there may be further nuance required. For instance, who’s deploying the AI workloads also matters. Service providers have different AI requirements than enterprises do, so it’s no surprise that these two groups use data centers differently. Training workloads may end up in hyperscale or colocation data centers, depending on who’s deploying them.
Hyperscale data centers are built with large service providers in mind
When cloud and SaaS providers are developing large language models (LLMs) for their customers’ use, they tend to work with especially large datasets pulled from many different data sources. Therefore, hyperscale data centers are a particularly good fit for their model training needs.
Also, service providers tend to be very technologically savvy, so they’re well positioned to make the most of what hyperscale data centers offer. Compared to colocation, hyperscale data centers allow customers to take more of a hands-on role in defining and operating their deployments.
The ideal hyperscale customer is an organization that’s capable of running their own data centers. At the same time, they may have situations where they’d prefer to let a partner handle the heavy lifting of building the facility, installing advanced cooling capabilities and sourcing the necessary energy. Large cloud and SaaS providers fall into this category.
Enterprises need flexible AI infrastructure
While service providers often have architectures that span multiple types of data centers, enterprises typically want out of the data center business altogether. Some enterprises use public cloud services to help them start on their AI journeys quickly. This means they could be indirectly using hyperscale data centers for their model training without even knowing it.
Enterprises that are further along in their digital maturity may choose a hybrid approach to AI: They may use the public cloud for certain AI use cases, while building their own private training environments for others. They may want to do this in order to protect their sensitive proprietary datasets using a private AI model.
A colocation data center is an ideal place for enterprises to build their private training environments. They won’t need to run their own data center environment like a service provider would inside a hyperscale data center. And since the model is for their private use only, they likely won’t need anywhere near as much capacity as the typical hyperscale customer would.
With colocation, enterprises can work with the colocation provider to build a data center environment that meets their needs. After that, the hard work of running that environment would fall entirely on the colocation provider. This frees enterprises to focus on more important things—like putting their AI-driven insights into action.
Not all hyperscale providers are equal
As AI becomes increasingly important for service providers and enterprises, Equinix is responding to keep up with the growing demand for AI-ready data centers. For instance, AI is one key driver behind the continued expansion of our Equinix xScale® portfolio of hyperscale data centers. Earlier this month, we announced a joint venture that intends to raise more than $15 billion in capital to fund new AI-focused xScale data centers in the U.S.
Providers that specialize in hyperscale data centers don’t always offer colocation data centers, or may offer them only in limited markets. In contrast, Equinix offers a comprehensive platform of both hyperscale and colocation data centers to support the entire AI lifecycle. In addition to our existing xScale data centers, we have Equinix IBX® colocation data centers in 70+ global metros. We also support AI data in motion: Equinix Fabric®, our virtual networking solution, makes it quick and easy for customers to connect their AI workloads in different locations and connect to partners and service providers in their AI ecosystem.
In short, whatever combination of data centers and network infrastructure you need to support your diverse AI workloads, you can find it on Platform Equinix®. To learn more, read the IDC Vendor Profile: Equinix Experiences Strong Growth Driven by AI, Hyperscale, and Digital Infrastructure.[1]
[1] Courtney Munroe and Avinash Naga, Equinix Experiences Strong Growth Driven by AI, Hyperscale, and Digital Infrastructure, IDC Vendor Profile, May 2024, IDC #US50186623.
Subscribe to the Equinix Blog