Google이 저압 DC(LVDC) 전력 시스템, 배터리 에너지 저장 시스템(BESS), 액체 냉각을 AI 데이터센터 포트폴리오에 통합하여 백업 전력에서 능동적 그리드 파트너십으로 전환하고 있습니다.

AI infrastructure is expanding at an unprecedented pace. Google now processes more than 3 quadrillion AI tokens per month, representing a sevenfold increase in compute demand from May 2025 to May 2026. Meeting this velocity requires innovation in energy efficiency and power delivery.

Tom Garvens, vice president of advanced technology innovation at Google, outlined this challenge during a keynote at Data Center World Power 2026 in Houston, held September 21-23. He described Google's full-stack approach to scaling AI, positioning data centers as active partners in stabilizing local grids.

To bridge the speed mismatch between rapid AI rollout and multiyear utility capacity expansion timelines, Google has developed energy strategies that transform backup assets into grid-supportive resources, streamline operations during peak demand, enable more intelligent load shaping, and prevent disruptive power swings while supporting utility resilience.

Google is reviewing and redesigning the entire AI data center ecosystem—from chip to utility—optimizing for liquid cooling, low-voltage direct current (LVDC) architectures, intelligent algorithms, and facility design. A core pillar of this work is publishing standardized power and facility technology specifications through communities such as the Open Compute Project, enabling the industry to build interoperable, efficient, and sustainable infrastructure.

"Google has tightly coupled vertical integration in AI along with close partnerships with energy providers," Garvens said. "We develop and design our own data centers, chips, servers, and rack designs. By having control of our entire hardware and software stack, we can optimize these things together for sustainability and energy efficiency."

Google is investing alongside utilities to add grid capacity—vital as a gigawatt-class data center campus can comprise hundreds of thousands of interconnected chips running a single workload. The company's approach is twofold: work with utility partners to streamline power delivery and adjust its own hardware and software to be more grid-friendly.

"Utilities are setting different requirements that we must demonstrate that we can meet," Garvels noted. "Google, Meta, Microsoft, and Nvidia have come together to define an interface spec on how utilities and hyperscalers can interface with each other."

Battery energy storage systems have become critical in Google's campus designs for onsite energy storage and buffering rapidly fluctuating workloads. Buffering smooths power and voltage transitions. Backup generators remain in use, but diesel does not scale well to campuses at 1 GW and beyond. BESS are increasingly used to provide standby power, supported by demand-response agreements with utilities for peak shaving.

"When peak needs are beyond what we've got power for, we can get that from local storage," Garvens explained.

Rather than competing for constrained capacity in saturated markets, Google has identified remote sites with ample land and available power. The company is also considering campus developments in small towns where the data center would be a primary employer.

Google has moved away from centralized Uninterruptible Power Supply systems, favoring rack-level battery backup operating around 54 VDC. Eliminating an extra AC-DC-AC conversion step yields approximately a 1.5% efficiency gain.

"We do all of our power conversion at the top of our racks, where we take our AC down to 54 volts, and then we run DC through our racks; our battery backup also runs at that voltage," Garvens said.

Google has introduced changes at the chip, rack, and facility levels, supported by software that intermixes workloads to smooth overall load and make it more predictable. Rack density remains a major challenge, with roadmaps anticipating racks approaching 1 MW each over the next year or two—a requirement that necessitates liquid cooling for at least the hotter elements.

Delivering the required current at lower voltages would demand bus bars and cabling with significantly more copper, calling for higher distribution voltage. Google is pursuing a gradual migration toward 800 VDC for power distribution, expected to take a couple of years to mature.

In the interim, Google has implemented "sidecars" alongside its racks—adjacent power-distribution enclosures housing cabling and power gear—reserving the racks for high-density servers. The tradeoff is floor space. When 800 VDC becomes broadly available, these sidecars can be converted to additional AI server racks.

"We're working with the industry on bringing solid-state transformers to maturity," Garvens said. "We are also continuing to improve our power usage effectiveness. We can't ever get to 1.0, but we continue to challenge ourselves to get as close to it as possible."

With dozens of data centers in operation and many more in development, Google can be fluid about where it runs workloads. "If there's a heat wave in Ohio and we were running a big workload, we've got agreements in place whereby with the right amount of foresight we can shut certain workloads down or move them elsewhere if they're critical," Garvens said. "We have a portfolio where we can shift things around."

Water strategy is another focus area. Google is investing in closed-loop cooling systems—sealed liquid circuits where the working fluid remains within the system and makeup water is minimized—marking a shift from designs that consumed large quantities of water. There is no single approach; depending on altitude, humidity, water table, drought risk, and community needs, Google works with partners to determine the appropriate power capacity and optimal cooling design.

"We work with our partners in terms of what's available and figure out what is the right amount of power we can get and what water usage or closed-loop design is most appropriate," Garvens noted.

Each new generation of silicon tends to draw more power, and multiple chips arrive every year. "The most precious thing we have is a watt: you don't want to ever strand a watt," Garvens said. "So, making our data center and our designs as crisp as we can for as far out as we can see in these roadmaps, it's why our partnerships with our silicon and GPU design teams are so critical."