Executive Roundtable: Evolution of Data Center SLAs in the Age of AI

Danielle Rossi, Trane:  With the increase of AI, we are seeing drastic rises in density. 

Supporting high densities during outages and maintenance requires extra redundancy and planning. 

Going forward, we will see more stringent plans for servicing, including connected and virtual services, as well as long-term vendor-provided scheduled maintenance. 

Customers have already begun specifying longer-range service plans and more detailed spare parts kits. 

Along with these requirements, requests for vendor service capabilities are increasing. 

Global technician coverage and area-specific availability have always been a consideration but are increasing in priority. 

Sean Farney, JLL:  I've managed to SLAs across all levels of infrastructure for products ranging from low-latency trading to natural gas to Happy Meal toys (seriously!), and it's quite operationally complex.  

On top of the complexity, appetite for technical downtime is plunging as connectedness becomes more pervasive and we trust digital infrastructure to deliver more mission-critical services related to health, safety, security, finance, autonomous vehicles and more.  

I think service level expectations and requirements will continue to creep upward. 

To satisfy this need and reduce the brand risk and costs of violation, I see a surge in reliability engineering. 

This means taking a systemic approach to designing, building and operating technical infrastructure to higher uptime levels instead of cobbling together services for individual mechanical, electrical and plumbing (MEP) and information technology (IT) components. 

JLL's Reliability Engineering practice is inundated with requests to solve these big, impactful problems.  

The long-term, holistic approach leads to better performance, lower cost and smoother operations.  

Harry Handlin, ABB:  Innovation will be the key driver in the evolution of service level agreements for infrastructure equipment.  

AI presents many challenges for service organizations.  

First, the rapid growth of the AI market coupled with the increased scale of data centers has created a shortage of qualified service engineers and technicians.  

In addition, AI data center locations are not constrained by latency requirements. This has results in many data centers being in built in areas that are unlikely to be supported by a four-hour response time.  

For some remote sites, the location is more than four hours of travel from the closet field service location.

Josh Claman, Accelsius:  With AI and large learning models, we frequently hear that it doesn’t really matter if there’s downtime. 

No one talks about five-nines or other traditional IT metrics of availability when it comes to LLMs. 

The problem, however, is that the assets you’ve deployed can’t sit idle–because, once idle, your ROI for those assets is now muted.

Sure, for inference engines, high-frequency trading and HPCs, we should aim for that 99.999% availability. 

But for large learning models, the math needs to change. 

You need to ensure those assets are fully utilized, which is why a lot of our clients move to liquid cooling. 

They want to avoid the throttling events that occur with traditional air cooling. 

Futuristic AI data center rendering. Does this image envision some future prospect where AI travels holographically for SLA via the visible light spectrum? Probably not. And yet, gazing into the GPUs' illustrated self-conception, such questions arise.