Amazon’s 2025 sustainability report highlights data center water usage as the company strives to become “water-positive.” Its cloud division, AWS, has met 75% of its target for this segment and plans to reach 100% by 2030, a goal requiring the company to return more water to local ecosystems than its infrastructure consumes.
However, the report also notes that Amazon’s server farms consumed 9.5 million cubic meters of fresh water last year. The glossy document does not explain exactly how the company intends to return more water to nature than it extracts.
Data Center Water: Beavers and Chemical Wastewater
Amazon has no plans to bottle wastewater for drought-stricken farmers. Nor does Jeff Bezos plan to hire a crew of beavers to dig new lakes in the Swiss Alps. Water entering data center cooling systems either evaporates as steam in giant cooling towers or becomes industrial runoff laden with salts and pipe-cleaning chemicals. Returning this wastewater to municipal water supplies is both technically impossible and legally prohibited.
The Secret of “Water Positivity” and Virtual Accounting
To claim this eco-friendly status, Amazon uses a system of virtual offsets, similar to carbon credits. The scheme is simple: an AWS data center in arid Virginia pumps millions of liters of drinking water from local reserves to cool processors. Instead of returning it to the source, Amazon sponsors a third-party environmental project elsewhere. For instance, it might fund leaky pipe repairs in a neighboring state or pay to cut down trees that consume too much moisture. On paper, environmentalists calculate the saved liters, the corporation subtracts one figure from another, and announces: “Congratulations, we have returned more water to nature than we consumed!”
Geographical Accounting Tricks
The core manipulation lies in geographical mismatch. Water is a local resource; it cannot be stored in a bank or sent over the internet. If an Amazon data center depletes local wells in Oregon, funding wetland restoration in rainy Ireland does nothing to help Oregon. Residents of the arid region lose their groundwater immediately, while the PR department in Seattle celebrates its global achievements. Amid the artificial intelligence boom, these reports look increasingly cynical. AI search queries and training new models require four to five times more computing power, which demands vastly more cooling.
The Secret Geography of Amazon’s Servers
Amazon keeps the exact number of its data centers strictly confidential for commercial and national security reasons. Officially, the company acknowledges only 39 geographic regions divided into 123 availability zones. However, leaked documents analyzed by Bloomberg reveal that the company’s total number of facilities worldwide has surpassed 900. This figure includes massive proprietary mega-campuses, colocation spaces leased in third-party data centers, and smaller network edge nodes.
Where Our Data Physically Resides
| Macro-region | Major Hubs | Key Features |
| North America | Northern Virginia, Oregon, Ohio, California | Virginia is the world capital of data centers, routing up to 70% of global internet traffic. |
| Europe | Ireland (Dublin), Germany (Frankfurt), United Kingdom, France | Ireland is favored for tax incentives, while Frankfurt serves as the main EU financial hub. |
| Asia-Pacific | Japan (Tokyo), Singapore, Australia, South Korea, India | Tens of billions of dollars are currently pouring into this region. |
| Middle East & Africa | UAE, Bahrain, South Africa | Local hubs built to comply with local data sovereignty laws. |
What About Google?
Google uses the same public relations tactics. In its latest report, the company pledges to replenish 120% of the water consumed by its offices and data centers by 2030. It claims to have already virtually “replenished” millions of cubic meters of water through charitable projects, such as installing smart irrigation sensors for Indigenous farmlands in Arizona and restoring peatlands in Ireland.
Google Consumes More Water Than Amazon
When PR claims are compared with engineering metrics, Google appears even more resource-intensive than its competitor. Analysts estimate that Google’s data centers consume an average of 1.15 liters of water per kilowatt-hour of energy, whereas AWS uses less water by relying on direct evaporative or air cooling. Driven by the AI race, Google’s physical water consumption has nearly doubled in recent years, reaching unprecedented levels.
The True Audience: City Halls and Wall Street
If the reality is widely understood, who is the intended audience for these reports? There are two key targets.
Every 100 words generated by a large language model physically consume about 500 ml of water due to intense processor heating. Residents in arid states have protested upon learning that neighboring server farms will draw from their local water supply. To secure permits for new AI clusters, tech giants must placate local municipalities by promising to fund infrastructure improvements, such as repairing school pipelines.
The more critical audience consists of major investment funds like BlackRock, Vanguard, and Fidelity. Today, billions of dollars are allocated based on ESG indexes (environmental, social, and governance standards). If a company like Amazon loses its sustainable business certification, these funds are legally required to divest. A single-digit percentage drop in market capitalization can cost tech executives tens of billions of dollars. Publishing a report about salmon conservation is simply a minor cost of keeping institutional investors from dumping tech stocks.
The Engineering Reality Check in Silicon Valley
The market has realized that building endless server warehouses and cooling them with scarce water is no longer sustainable as physical limits and local resource constraints push back. To avoid soaring utility bills, tech giants are adopting four actual technical solutions rather than relying purely on PR:
- Direct-to-Chip Liquid Cooling: Next-generation chips like the Nvidia Blackwell generate too much heat for traditional air cooling. Instead, a plate with liquid coolant is attached directly to the silicon. This closed-loop system operates like a car radiator, circulating water for years and reducing ongoing consumption to nearly zero.
- Custom Silicon (ASICs): Purchasing general-purpose GPUs from Nvidia has become prohibitively expensive. Google (TPU v6) and Amazon (Trainium) are shifting to specialized processors. Designed specifically for AI matrix multiplication, these chips require 30% to 40% less power and generate significantly less heat.
- Nuclear Energy Adoption: Recognizing that municipal power grids are reaching capacity, Microsoft contracted to restart a reactor at the shuttered Three Mile Island nuclear plant, while Amazon purchased a data center directly adjacent to a nuclear reactor in Pennsylvania. Future plans include deploying small modular reactors (SMRs) directly on-site.
- Algorithmic Efficiency: Engineers are optimizing the software itself. The Mixture of Experts (MoE) architecture activates only the necessary parts of a model during a query, leaving the remaining 90% of the chip idle. Additionally, quantization compresses data, significantly reducing processor workloads.
The Jevons Paradox: Why Optimization Will Not Save the Planet
These engineering solutions all aim to make computing cheaper and more energy-efficient. However, they are subject to a classic economic law known as the Jevons Paradox, which states that as a resource becomes more efficient and affordable to use, overall demand for it increases rather than decreases.
If training a model required one million liters of water in 2025, and new technologies cut that requirement to 500,000 liters in 2026, Amazon and Google will not pocket the savings. Instead, they will train ten times as many models. Total water and energy consumption will continue to grow exponentially, offsetting efficiency gains and increasing pressure on ecosystems.
Global Data Center Water Resilience
Ultimately, tech giants are not pursuing sustainability out of environmental altruism. Energy and water have become the primary constraints on their profit margins. If a chip melts or a company loses its water quota, it falls behind in the competitive landscape. This is a matter of market survival: optimize resources or fail.
NEWSROOM IN utilized artificial intelligence in preparing this article, calculating the actual water footprint of its generation. Based on the formula of 100 words per 500 ml of water, generating this text consumed approximately 6.15 liters of fresh water—equivalent to a standard six-pack of one-liter water bottles.