In-House or Colocation for AI?

Park Place Technologies

Paul Mercina September 12, 2019

One-size-fits-all rarely works in technology.

We’ve been talking about artificial intelligence lately, but then, who hasn’t? It’s the biggest thing on the IT block. And it’s a topic on which we’ve built up some knowledge, based on our forays into machine learning, AI’s precursor.

In the past, we’ve also offered a consolidated comparison of in-house vs. colocation options for data centers, and that post remains one of the most frequently consulted on our website. So here’s a question, has the emergence of AI changed that equation?

Facilities Issues Are Challenging

One key factor for AI, it’s compute-intensive. That means it’s power hungry unto itself, and it generates a lot of heat in need of dissipation. For example, when testing Tensor Flow 3.0, Google found its existing cooling systems were no longer up to the task. And while few IT organizations are this deep in AI yet, the same problems are on the horizon for much of the industry.

With densities increasing, enterprises are looking at a transition from 7 kW per-rack maximums to 30 kW and above. In most cases, liquid cooling—which can require expensive data center retrofitting—is the best way to chill the hardware. For organizations that lack the capital, expertise, or interest to build out an AI-ready, high-density data center, a colocation can provide a great alternative.

Colocations can also deliver the value-add of leveraging the best emerging technologies in AI-focused facilities management. Enterprises may find it difficult and risky to research, test, and deploy newer approaches, but colocation providers stand to save mightily on any such innovation, if successful, and are often more game to serve as early adopters.

When added to the other colocation advantages—including the potential for volume discounts on electricity and the enhanced redundancy and disaster recovery capabilities—colocation can turn out to be the right investment.

Latency Remains a Barrier

On the other hand, AI is facing some of the same “edge arguments” as IoT. By this, we mean that many applications run best nearest to where the data is created. Consider a facial recognition system in an airport. One probably wouldn’t want to send all the data it generates to a faraway colocation facility, as latency would become an issue in delivering timely security alerts.

Although this is a public sector example, many use cases in which AI must “examine” real-time events and provide immediate “answers,” informed by reams of historical data, are going to be latency-intolerant. Just as IoT is likely to drive developments at the edge, some experts predict AI will compel installation of small, high-density data centers more widely distributed across the landscape. This may be how colocation providers and other third parties get in on the AI action, but for now, low-latency applications will generally require on-premises infrastructure.

What’s the Answer?

One-size-fits-all rarely works in technology. Just as there are use cases for doing AI in the cloud, there are times when a colocation provider’s ability to translate a capital expense into an operating one, apply their facilities expertise to this very new field, and offer energy savings will make it the best-fit option for AI. For enterprises that expect to take a deep dive into AI as a central, differentiating technology, however, it would be difficult to avoid considering on-premises infrastructure for lower latency and the better long-term ROI associated with asset ownership.

About the Author

Paul Mercina, Director, Product Management Marketing
Paul Mercina brings over 20 years of experience in IT center project management to Park Place Technologies, where he oversees the product roadmap, growing the services portfolio, end-to-end development and release of new services to the market. His work is informed by 10+ years at Diebold Nixdorf, where he worked closely with software development teams to introduce new service design, supporting implementation of direct operations in a number of countries across the Americas, Asia and Europe.