Hyperscalers such as Google, Meta, Microsoft, and OpenAI dominate headlines about new data center construction, building or contracting for campuses up to 1 GW to run AI workloads.
But the rest of the business world—the likes of airlines, banks, hospitals, or carmakers—still depends on data centers to run day-to-day operations. These enterprises also need new capacity to train and run new AI capabilities. And the scramble for data center space has made finding the necessary capacity and equipment more difficult for them.
"For every one hyperscale user out there, there are probably 100 large enterprise users," said Kirk Killian, president at Partners National Mission Critical Facilities, during his conference presentation at Data Center World 2026. "They may not be adding 100 MW at a time—it might be a quarter meg, 1 meg, or 5 megs—but it's still critically important."
Killian, a data center planning consultant who has completed over 400 projects since 1999, outlined the critical differences in what enterprises look for in a data center compared with hyperscalers. As more enterprises run their own AI applications and agents, finding data center capacity becomes critical to companies that drive the everyday economy.
1. Enterprises Deploy a Wider Mix of Infrastructure Than Hyperscalers
Perhaps the starkest difference lies in how enterprises and hyperscalers deploy their infrastructure. A hyperscaler may take 40 MW of capacity in a new AI factory and deploy identical hardware across almost every cabinet in the room. "Whereas the enterprise, in 50 different cabinets, may have 20 different load profiles based on the type of compute they're doing," Killian said.
This diversity creates unique planning challenges. Enterprises must accommodate traditional computing for operations, finance, and HR alongside AI training and inference workloads—all within the same facility. So, while many enterprises are procuring new data centers to deploy mostly 20 to 35 kW per rack today, Killian said, "they want to pick facilities where they can clearly grow that over time to accommodate the 50s, the 80s, maybe some cabinets at 100+ kilowatts."
The challenge, he noted, is "the scalability without stranding a bunch of capacity."
2. AI Inference Raises Enterprise Concerns About Data Security and Control
When enterprises train an AI model, they're often comfortable outsourcing the data center computing to an AI factory that specializes in training—if it doesn't involve live company data. Running AI inference is different.
"Once enterprises are ready to deploy AI inference at scale, they're going to be loading their crown jewels corporate data into those models," Killian said. "Those models then are going to be customer-facing, revenue-producing, or expense-controlling, and those are going to be of vital importance to enterprise users."
Many enterprises will want to run those AI inference workloads in an on-premises data center, colocation facility, or a trusted public cloud where they're highly confident in the security, internal controls, and performance. Speed might be the priority in AI training, but with AI inference, "the number one concern for many large enterprises is data security," he said.
3. Enterprises Have Tighter Location and Connectivity Requirements
While hyperscalers increasingly build massive AI factories far from major cities, enterprises have typically relied on suburban data centers 10 to 25 miles from city centers. In terms of connectivity, rural AI factory locations may have two to four lit redundant fiber providers, whereas even smaller facilities in suburban locations often have eight to 10 different fiber providers.
For AI inference applications requiring low latency, connectivity and proximity become critical. "There's considerable push now to consider edge deployment placement in order to get the very low latency that some of these AI inference applications require," Killian said.
4. Enterprises Need Smaller Capacity That's Getting Squeezed Out
Most of today's enterprise-class colocation data center capacity was built between 2015 and 2022, when "a 20 MW data center used to be a pretty big project," Killian said. A typical enterprise might lease a 2 to 4 MW data hall where they could have full control inside a colocation provider.
Today, 90% of new capacity being built is in the 200 MW to 1 GW campus range, focused on hyperscaler needs. Meanwhile, data center vacancy rates have hit an all-time low, running about 2% across the U.S. "Some markets like Northern Virginia, Phoenix, and Portland are under 1%. 85% of data centers under construction now are pre-leased," Killian said.
That makes it hard for enterprises to find the space they need. Killian used to advise enterprises to start searching 12 to 18 months before they would need data center capacity. Now he recommends starting 24 to 36 months out.
5. Enterprises and Hyperscalers Share One Challenge: Predicting AI Demand
Enterprises are fired up about AI's potential. 74% of respondents to the AFCOM State of the Data Center report say they plan to deploy AI-capable solutions in their data centers. 72% expect AI workloads to increase capacity requirements, with 48% anticipating the impact will be significant. Yet only 34% of companies believe their IT infrastructure is fully adaptable and scalable for AI projects, Cisco's 2025 AI Readiness Index finds.
Enterprises and hyperscalers face a common problem: the AI future is unpredictable. Will a new AI capability emerge that becomes a must-have? Will a new swath of employees within a company need access to AI features? Every question about AI use rolls back to data center capacity, and no one has a clear picture of AI demand more than a few months out. Yet enterprises and hyperscalers alike need to plan data center capacity more than a year out, given how long it takes to secure and operationalize capacity.
"We recommend that enterprises come up with a low growth profile, a most likely growth profile, and then a high growth profile," Killian said.