TL;DR

  • Structured Workload and Power Planning: Sustainable growth requires a disciplined seven-step sequence that aligns specific computing needs directly with grid operators and utility interconnection plans.
  • Maximizing Facilities and Hardware: Operators can optimize existing capacity by repurposing suitable industrial buildings, adopting direct-to-chip liquid cooling for dense racks, and extending server lifespans to curb electronic waste.
  • Transparent Community Impact: Long-term viability relies on publishing measurable environmental data regarding power, water, and noise usage, alongside formal community benefit agreements

# # #

By Jeffrey Gregor, General Manager, OVHcloud US

AI is increasing the need for computing capacity faster than many power systems, supply chains and local permitting processes can accommodate. The 2024 United States Data Center Energy Usage Report, prepared by Lawrence Berkeley National Laboratory for the Department of Energy, estimated that data centers consumed 176 terawatt-hours of electricity in 2023, or 4.4% of all U.S. electricity. Its scenarios put the 2028 total between 325 and 580 terawatt-hours, equal to 6.7% to 12% of U.S. consumption. A June 2026 update from Berkeley Lab extends the outlook to 2030, with a central estimate of 649 terawatt-hours, or 11.8% of U.S. electricity use, and a scenario range of 521 to 843 terawatt-hours (9.5% to 15.3%).

Meeting that demand calls for a disciplined seven-step sequence: define the workload, build a credible power plan, recover capacity in existing facilities, consider adaptable buildings, address cooling and water early, extend hardware life, and make local impacts measurable and public.

The first step is to define the demand precisely. Training a frontier model, running inference and supporting enterprise applications create different requirements for power density, latency, resiliency and location. The 2024 Berkeley Lab report estimated that servers allocated to inference accounted for nearly 60% of AI server energy use in 2023, while training server energy is projected to surpass inference by 2028 as additional higher-power GPUs are allocated to training. Because inference accounts for a large share of AI server energy use and many inference workloads are latency-sensitive, much of that work can run closer to users rather than only in giant centralized campuses. A distributed footprint can also spread demand across markets.

The second step is to make power a siting and operating plan. The Federal Energy Regulatory Commission staff’s 2025 State of the Markets report identifies size, speed of interconnection and load flexibility as defining features of data center demand. Before committing to a site, developers should work with utilities and grid operators to test interconnection queue timing, transmission and substation upgrades, rate design and cost allocation. Where on-site generation or storage is proposed, it should be evaluated with the utility for reliability, emissions, fuel supply, permitting and consumer-cost impacts. Flexible workloads can support phased energization or demand-response agreements, but service commitments and operating controls must be defined up front.

The third step is to inventory what is already available. The Uptime Institute’s 2026 Global Data Center Survey says concerns about capacity forecasting and power availability are growing, while legacy infrastructure and cooling constraints continue to slow efficiency gains. Operators should measure actual IT load against UPS and cooling capacity, identify stranded electrical and thermal headroom, and validate redundancy and network limits before assuming that available floor space can support denser systems.

The fourth step is to consider whether an existing building can do the job. A building with suitable power access, fiber, structural strength and industrial zoning may offer a sound alternative to ground-up construction. Former factories, warehouses and government facilities can sometimes be adapted for data center use, reducing demand for land and building materials. At OVHcloud, our Beauharnois campus in Quebec occupies a converted former aluminum smelter and houses one of our server-production plants. Each candidate still requires a rigorous assessment of floor loads, cooling, backup generation, security and expansion options. Reuse should be treated as a site-selection strategy, with economics and engineering determining where it is feasible.

The fifth step is to design cooling and water use with the site. Uptime Institute’s 2026 survey notes average rack densities continue to rise, and more operators now report peak rack densities of 30 kW or above. Direct-to-chip water cooling carries heat away close to high-density CPUs and GPUs and can support denser racks with less cooling energy. Water efficiency needs a specific design target: a 2026 paper published by the National Academy of Engineering, citing Berkeley Lab’s 2024 report, puts U.S. data center water consumption at 140 to 280 billion liters in 2028 (about 37 to 74 billion gallons). Closed-loop systems can recirculate water and use dry coolers for much of their heat rejection, limiting ongoing water consumption. Cooling architecture also affects energy use, noise, building design and the computing capacity that can fit within an existing footprint, so these choices belong in the first site assessment.

The sixth step is to extend hardware life. The Global E-waste Monitor 2024 found that the world generated 62 million metric tons of electronic waste in 2022; only 22.3% was documented as formally collected and recycled. Extending server life, reconfiguring machines for lighter workloads and harvesting usable components reduce pressure on supply chains and waste systems. Servers designed for disassembly, repair and redeployment can serve several roles over their useful lives. Procurement teams should therefore evaluate serviceability and reuse alongside performance, price and energy efficiency.

The seventh step is to make local impacts measurable and public. Residents and local officials reasonably want to know how a facility will affect electricity, water, traffic, noise, views and public finances. Before approval, developers should publish projected electricity and water use, noise, backup-generator testing and traffic impacts, then report actual site performance after opening. Community benefit agreements can fund local priorities such as workforce training or infrastructure, while heat-reuse projects can serve nearby buildings or industry where a stable off-taker exists. Site selection should account for nearby homes, zoning and visual and acoustic mitigation. Engagement should continue after construction.

Meeting AI demand will take more than new data centers. We need to get more from the power, buildings and equipment already in place, make thoughtful choices about cooling and water, and be candid with communities about each project’s impact. If we do that well, we can add capacity without pushing the costs onto the grid, the environment or the people who live nearby.

# # #

About the Author

Jeffrey Gregor is General Manager of OVHcloud US and President of the Board of Managers of OVH US LLC. He joined the company in 2017 after serving for 20 years as a commissioned officer in the U.S. Navy. He holds a master’s degree in naval architecture and marine engineering from MIT and an MBA from the UCLA Anderson School of Management.