The numbers at a glance: Google’s electricity consumption grew by 37 percent in 2025; since 2019, the increase has totaled more than 250 percent. The data centers alone drew over 42 million megawatt-hours from the grids—a scale that rivals the annual consumption of entire countries such as Denmark or New Zealand. The driver is clearly identified: the expansion of AI infrastructure for training and operating large models. The report acknowledges the core problem with remarkable candor—AI expansion is currently accelerating faster than the power grids are decarbonizing.

Access to Water as a Critical Factor for Operating Large-Scale AI Infrastructure

On paper, Google maintains the facade: for the ninth consecutive year, its total consumption was mathematically covered by renewables, and operational emissions fell by two percent. Behind this lies a procurement program of historic proportions—more than 12 gigawatts of new clean generation contracted for in 2025 alone, including the reactivation of the Duane Arnold nuclear power plant in Iowa, a 3-gigawatt framework agreement for hydropower with Brookfield, and even a power purchase agreement for fusion electricity with Commonwealth Fusion Systems. But the supply chain tells a different story: Indirect emissions rose by 25 percent. The construction of new data centers alone accounted for approximately 2.3 million metric tons of CO2 equivalent, compounded by chip manufacturing at suppliers in Taiwan, Japan, Vietnam, and India, which rely on carbon-intensive power grids. Water consumption also climbed by 34 percent to 10.9 billion gallons—a good 41 million cubic meters—primarily for cooling server fleets. That water is becoming a strategic bottleneck was recently confirmed by a high-profile source: In its IPO prospectus, SpaceX explicitly warned investors that access to water would become a critical factor for operating large-scale AI infrastructure.

The Competition for Grid Connectivity

Why should this matter to a chief manufacturing officer in Stuttgart or Osaka? Because Google is not an isolated case, but rather the spearhead of a trend: Data centers currently account for about 1.5 percent of global electricity consumption, and by 2030, that share is expected to double to about three percent. Alphabet plans to invest up to $185 billion by 2026, with a large portion of that going toward data centers and their power supply. The hyperscalers are thus entering into direct competition with energy-intensive industries—for grid capacity, long-term power supply contracts, and the best locations. Anyone expanding or electrifying a plant in the coming years will be negotiating against opponents with virtually unlimited budgets.

The equipment suppliers’ response: an AI factory built to plan

A reference architecture unveiled in early June—developed by Siemens in collaboration with NVIDIA and storage specialist Fluence, and supplemented by design parameters from cooling provider nVent—shows how the industry is responding to this new load class. The design for NVIDIA’s DSX Vera Rubin NVL72 platform translates the previously rather visionary “AI Factory” concept for the first time into an industrially scalable power, control, and data center architecture: scaled for a total capacity of 136 megawatts, of which 100 megawatts is pure IT load, with a consistent design from the 34.5-kilovolt grid connection through medium-voltage distribution and modular low-voltage power blocks all the way to the rack interface. The hallmarks of classic plant engineering are unmistakable: prefabricated, factory-tested switchgear skids instead of on-site improvisation; maintainability based on Tier III logic—every component can be replaced during operation—and a modular design that allows for expansions from the double-digit to the triple-digit megawatt range without redesign.

Operation Even at Sites with Limited Grid Capacity

Also noteworthy is the role of the integrated Fluence battery storage systems: They smooth out the peaks typical of AI loads and enable operation even at sites with limited grid capacity. For now, this is merely a blueprint, not a completed facility—whether the design delivers on the promises of the spec sheet remains to be seen in the first implemented projects. The message behind it is clear, however: the AI factory is evolving from a one-of-a-kind system into a configurable industrial product—with the same virtues that made mechanical engineering great: standardization, modularity, and repeatability.

What this means for traditional industry

Three consequences stand out. First: Energy strategy is becoming location strategy—the availability of grid connections will play a decisive role in investment decisions going forward, on both sides. Second: Google’s shopping list, ranging from nuclear power to fusion, shows the direction we’re heading in terms of long-term electricity security; power purchase agreements are evolving from a niche instrument to a standard tool, even for small and medium-sized enterprises. Third, this presents a tangible opportunity: efficient cooling technology, grid technology, storage, heat recovery, water treatment—suppliers in these fields are serving the world’s fastest-growing energy consumer, and the Siemens example shows that a market for industrialized turnkey solutions is currently emerging here. AI’s electricity bill has been laid bare. Now it’s being distributed—and those who supply the infrastructure for AI are more likely to be among the senders than the recipients of the bill.