AI Demands Stretch the Limits of Data Center Retrofits

extract

AI Demands Stretch the Limits of Data Center Retrofits Operators are racing to retrofit aging facilities for AI workloads, but many legacy data centers are running into hard limits around power distribution, cooling, and rack density. The AI infrastructure boom is usually framed as a construction story, focused on the industry,’s need for bigger campuses, larger power commitments, and faster builds. However, operators of existing data centers face a more immediate challenge: determining whether facilities designed for a different era of computing can be adapted to meet the demands of AI. The pressure is building because existing sites already have access to power – an increasingly scarce resource. As utility interconnection timelines stretch and new developments slow in permitting queues, retrofitting older facilities has become one of the fastest ways to bring incremental capacity online. Yet, many of those buildings were designed around rack densities far below what modern AI systems require. In some retrofit scenarios, especially in multi-story buildings, floor-loading limits are emerging as a challenge, with newer high-density AI racks weighing up to 1.5 metric tons even before accounting for additional power and liquid-cooling infrastructure. Traditional enterprise environments typically supported rack densities of 4 to 10 kW, but AI systems now far exceed those levels. This shift is forcing operators to evaluate whether existing electrical systems can be upgraded or if they will become the limiting factor. While new developments offer scalability, they are constrained by permitting, transmission upgrades, and utility interconnection delays. Existing facilities, already energized, hold a key advantage. However, the gap between available utility power and usable rack-level power is widening. Bizo notes that high-density AI deployments are straining power distribution systems originally designed for lower-density environments. According to Uptime Institute analysis, just a few rows of modern AI racks can consume as much power as an entire legacy data hall, and this creates a cascading impact on other critical areas. High-density AI racks can weigh up to 1.5 metric tons, pushing floor-loading limits in retrofitted multi-story data centers. (Image: Alamy) Cooling and Power Distribution Challenges Cooling presents a parallel challenge for data center retrofits. Air-cooled environments, long the baseline for most data centers, reach thermal limits as densities climb. Liquid cooling offers a path forward, but integrating it into existing facilities can require substantial changes to mechanical systems and internal layouts. Uptime Institute survey data shows that while liquid cooling adoption is growing across the industry, most operators continue to rely primarily on traditional air-cooling methods. However, the challenges of retrofitting extend well beyond cooling. Bizo said the first major pressure point is often power distribution infrastructure, particularly as operators push beyond 40 to 50 kW per rack at scale. “At scale, high densities above 40-50 kW per rack average create issues for power distribution much more than cooling,” Bizo told Data Center Knowledge, adding that the availability of large-capacity breakers, busways, and tap-off units has also become a limiting factor as AI infrastructure demand rises. “Getting all that power into relatively few racks concentrated in a small space” becomes increasingly difficult at higher densities, Bizo said. Retrofit, Not Rebuild Rather than full conversions, many operators are pursuing partial data center upgrades – carving out sections of facilities for higher-density deployments or pairing legacy sites with new infrastructure. Alex Cordovil, research director at Dell’Oro Group, said the industry is splitting less along hyperscale vs. enterprise lines and more around workload type. “The more useful dichotomy isn’t ’hyperscaler vs. legacy enterprise,’ but AI factories vs. general-purpose data centers,” Cordovil said. That distinction matters because many traditional workloads – including search, document sharing, and video streaming – continue to run efficiently in lower-density environments. Cordovil said many operators still run facilities in the low kilowatt-per-rack range and will encounter practical ceilings around 20 to 40 kW per rack as they introduce AI infrastructure into existing environments. “The exact limit varies drastically by facility,” he said, adding that self-contained racks and liquid-to-air cooling systems could help reduce the cost and complexity of retrofit projects. New Capacity Dominates Even so, retrofitting alone is unlikely to keep pace with the ongoing surge in data center demand. Primary-market vacancy rates fell as low as 1.4% in late 2025 despite rapid construction growth, according to CBRE, underscoring how quickly AI and hyperscale demand are absorbing new capacity. Synergy Research Group projects hyperscale operators will account for 67% of global data center capacity by 2031, up from 48% at the end of 2025, as AI demand drives another wave of large-scale construction. John Dinsdale, chief analyst at Synergy, said the industry is effectively building past its legacy footprint. Hyperscale capacity in place at the end of 2023 – largely designed for CPU-based workloads – will more than double by the end of 2026, with most of that new capacity built specifically for AI. Another doubling is expected over the following three years, he said. This expansion reframes the role of existing facilities. Rather than being upgraded en masse for AI, legacy data centers are more likely to continue serving traditional cloud workloads. “I don’t see anything as being stranded,” Dinsdale told Data Center Knowledge. The industry’s focus on new construction has obscured the role existing infrastructure will play in the near term. Retrofitting and brownfield redevelopment will never completely negate the need for greenfield development, but they may determine how much capacity can be brought online in the interim. Looking ahead, the focus may shift from the feasibility of retrofitting to how effectively operators can overcome the limitations of legacy facilities to meet the relentless demands of AI.