Preparing data centers for AI means rethinking capacity, efficiency, and cost in a rapidly growing market
By 2034, the global AI data center market will reach approximately $157.3 billion. That’s a significant increase from the $14.3 billion in 2024. While hyper-increased demand for computing power is needed for large language models, AI integration into data centers is still in its early and bumpy stages.
Last year, North America led the AI Data Center Market with more than 40% market share and $5.7 billion in revenues. This was driven, for the most part, by advanced infrastructure and major technology investments, including in the Graphics Processing Units (GPUs) — effective in parallel processing for AI models — with more than 55% share.
Key considerations for data center readiness for AI integration
Data center operators must assess a few critical points before jumping on the AI bandwagon.
Top of this critical points list is power availability, cooling design, capacity, and network services — both internal and external.
First, consider your goals before integrating AI into your data center. Determine how this will align with your other business services already in operation. Most importantly, determine if hosting these systems internally can be profitable, or if using external applications would be a better alternative.
When it comes to the ROI of integrating an AI solution, you must determine whether upgrading in-house systems may outweigh your assumed ROI. If you decide to upgrade, work with a qualified team to help you review and determine whether your existing infrastructure capacity can support an AI workload. This requires:
- A review of utility, back-up generation, UPS Systems, battery backup, and rack-level distribution
- On the cooling side, determine if your current system or plant has the cooling capacity for an AI workload
- Most importantly, does your facility have the capacity to compensate for rapid increases in demand from AI applications
AI integration poses hyper-increased power usage. A simple AI search uses about 10 times more energy than a standard internet search and can further strain energy grids and environmental impact. While traditional data centers manage around 12 kW per rack, integrating AI applications would dramatically increase energy usage as the ultra-high-density racks consume 85 kW per cabinet. Some AI hardware deployments in tightly packed server racks can exceed 120kW per rack. Future projections suggest an increase between 200 to 250 kW per rack as demand for AI workloads increases.
At the network level, your current internal and external connections must have sufficient bandwidth and speed to handle the throughput required for the AI system’s deployment. Finally, you must assess if the work can be done without disrupting your current data center operations–or more importantly, risk costly downtime.
Legacy data centers’ key upgrade considerations
Upgrading legacy data center design, with hundreds of watts per square foot, may not be possible, or may not be able to support the thousands of watts needed for today’s AI systems. Legacy power systems were originally designed for smaller power loads and lower power density. At the rack level, delivering power for AI applications requires larger circuits and distribution. This may require removing and rewiring all the data center circuits feeding the racks.
Another major concern in the legacy data centers is the cooling systems. These may have to be redesigned to ensure proper airflow and containment. With AI workloads generating extreme heat, reliable cooling is paramount to respond to a rapid increase in load or power outages. With most legacy systems lacking UPS-backed pumps or fans, there can be an increased potential for thermal events during a power outage. Higher heat loads may cause cooling systems to restart too slowly, preventing proper cooling of equipment.
Financial considerations, substantial investment costs
Potential costs of retrofitting an AI-ready data center vary widely, and can range between $4 to $8 million per megawatt, excluding hardware costs. Cost variances are based on the extent of the upgrade needed, the specific AI workloads (training or inference), and whether the retrofit is for a hybrid or colocation setup. The need for higher power density, advanced cooling, and reinforced infrastructure are just a few things to consider as you grasp the depth of the upgrade.
Here are some of the top financial costs to consider:
- Equipment costs, which can fluctuate daily based on market conditions
- Cost of hiring top, skilled experts
- Total shutdown of a data center during an upgrade can be more cost-effective since you don’t have to cover overtime expense, coordination with stakeholders for downtime of other applications, and potential equipment damage, among others.
An AI-workload-ready data center upgrade could take up to 18 months. As a data center operator, you must determine if you can afford the wait. Finding another, more suitable data center space in a shorter time may be another, affordable option.
As with any project, your first goal is to define your goals. Then benchmark what success would mean for your specific data center business. Assess the current state and capacity of your data center. If you do not have the required capacity today, lay out a high-level upgrade plan, including a timeline of your goals, to help you determine the feasibility of an upgrade. Working closely with all your business stakeholders and trusted partners is essential in helping you reach a final decision.
Considering all the critical capabilities required for AI integration, you may ultimately decide that a new site, specifically purpose-built to meet the demands of AI applications’ higher power and cooling requirements, may be the most cost-effective. As demand for AI capabilities rises, preparing your data center for the future is crucial for staying competitive in today’s evolving landscape.
Table of Contents
