Что случилось
Google has announced a strategic shift in how it manages the energy consumption of its AI data centers, moving toward a 'grid-interactive' model. According to Tom Garvens, Google's vice president of advanced technology innovation, the company is redesigning its infrastructure from the chip level to the utility interface to mitigate the strain caused by a sevenfold increase in AI demand between May 2025 and May 2026. This strategy involves using battery energy storage systems (BESS) for peak shaving, migrating to 800 VDC power distribution, and implementing software-driven workload shifting to move tasks away from regions experiencing grid stress.
Google reported that it now processes over 3 quadrillion AI tokens per month, a figure that highlights the rapid scaling of its AI infrastructure. To manage this, the company is moving away from centralized Uninterruptible Power Supply (UPS) systems in favor of rack-level battery backup operating at 54 VDC, which provides an approximate 1.5% efficiency gain by reducing power conversion steps.
The company is also implementing 'sidecars'—adjacent enclosures for power distribution—to maximize rack density for high-performance servers. These are intended as a bridge until 800 VDC power distribution becomes broadly available, at which point the space can be repurposed for additional capacity.
Google is collaborating with other major hyperscalers, including Meta, Microsoft, and Nvidia, to define standardized interfaces for communication between data centers and utility providers. This is intended to allow for more predictable load management and better integration with local power grids.
Подробности об источнике: datacenterknowledge.com ↗
Почему это важно
As AI demand reaches gigawatt-scale, traditional power delivery methods are becoming insufficient. By integrating data centers as active grid participants, Google aims to prevent power instability while maintaining the high-density compute required for AI. This shift is critical because it moves data centers from being passive consumers of electricity to active components of utility resilience. The move toward standardized interfaces, developed in collaboration with Meta, Microsoft, and Nvidia, suggests a broader industry trend toward interoperable, grid-friendly infrastructure that could set the standard for future hyperscale AI deployments.
The transition to grid-interactive data centers is a response to the 'speed mismatch' between the rapid deployment of AI hardware and the multiyear timelines required for utility capacity expansion. By using BESS to buffer fluctuating workloads, Google can reduce the reliance on diesel backup generators, which the company notes do not scale effectively to gigawatt-class campuses.
The shift toward closed-loop liquid cooling systems and intelligent software that can shift workloads during heat waves or peak demand periods demonstrates a move toward operational flexibility. This allows Google to maintain service continuity while minimizing the impact on local energy markets, which is increasingly important as data centers become primary employers and power consumers in smaller communities.
Интерактивный механизм: как он на самом деле работает
Изучите технологию, лежащую в основе этой разработки, в интерактивном режиме.
Which component of an AI application is the machine-learning model itself?
Что посмотреть дальше
The industry should monitor the maturation of 800 VDC power distribution and solid-state transformers, which Google identifies as key to future efficiency. Additionally, the effectiveness of the newly defined interface specifications between hyperscalers and utilities remains to be seen. As Google continues to prioritize remote sites and small-town developments to avoid saturated markets, the long-term impact on local utility infrastructure and community relations in these regions will be a significant indicator of the strategy's scalability.
The adoption of the Open Project (OCP) standards for power and facility technology will be a key metric for industry-wide progress. If these standards are widely adopted, they could significantly lower the barrier for efficient, sustainable infrastructure deployment across the sector.
The practical application of workload shifting—moving critical tasks between data centers based on real-time grid conditions—will be tested as AI demand continues to grow. The ability to execute these shifts without impacting service or reliability remains a technical challenge that will define the success of this grid-interactive model.