Green data center for AI: why water and power matter more than the algorithm

When a Mexican company decides to invest in generative AI for its operations, the conversation usually revolves around the algorithm: which model, how many parameters, which provider. But that conversation omits two physical constraints that matter more: the water consumed by the data center that trains or serves that model, and the electricity required to keep GPUs running 24/7. A data center designed for AI training consumes between 10 and 50 times more energy per rack than a traditional data center, and requires liquid cooling systems that evaporate water in volumes already stressing entire watersheds in arid regions of Mexico.

The question is not whether your company should invest in AI, but where that compute is physically operating. If the AI you use runs in a Mexican data center powered by CFE and cooled by water from the local aquifer, the hidden environmental and operating costs are yours even if the algorithm belongs to Microsoft or Google. If it runs in a hyperscaler campus in Querétaro with a renewables contract and is water-positive, they are yours as well. This article explains why water and electricity matter more than the algorithm when you evaluate an AI project, and which concrete metrics to review before signing.

The figure that opens the conversation: according to EESI and TechRepublic’s coverage, AI data centers are consuming water in regions where water scarcity is already structural. Lei et al. published in June 2025 a review on the determinants of water consumption in data center workloads that confirms liquid cooling systems can consume between 1 and 5 million liters of water per MW per year, depending on climate and architecture. For a Mexican company that wants to run AI in a national data center, this means the environmental and regulatory cost of compute is quickly becoming a business variable, not just a compliance one.

The water that evaporates when you train a model

Liquid cooling for AI servers works like this: cold water circulates through cold plates attached to the chip, absorbs heat, and is carried to a cooling tower where part of it evaporates. The evaporated portion is lost. In arid or semi-arid regions—which include much of northern and central Mexico—that evaporated water leaves the local hydrological cycle and must be replaced with aquifer extraction or desalination, both of which are costly and regulatorily sensitive. Three figures to size the problem:

• A 100 MW hyperscale campus with evaporative cooling in a warm climate consumes between 1.2 and 1.8 million liters of water per day, according to benchmarks published by operators.
• Training a large AI model can emit between 200 and 500 tons of CO2 equivalent, according to data cited in peer-reviewed literature; the water consumed during that training is of the same order of magnitude as that of a Mexican family for 50 years.
• Cooling efficiency (measured as WUE, Water Usage Effectiveness) ranges from 0 to 2.0 liters per kWh of IT load: a data center with WUE 0.2 consumes ten times less water than one with WUE 2.0, but both can operate in Mexico today without different regulatory consequences.

For your company, this means that two identical AI providers can have radically different water footprints depending on where they operate. Before signing a contract, ask the AI provider what the WUE is of the data center serving your workload, in which watershed it operates, and whether it has a verifiable water replenishment contract. If it cannot answer with numbers, assume the worst.

The electricity spent keeping GPUs on

A traditional server consumes between 300 and 800 watts. A server with a GPU for AI training consumes between 2 and 10 kilowatts. That is between 4 and 30 times more per unit. In a 42U rack, density moves from a typical 8 to 15 kW to 40 to 80 kW with GPUs. This changes three things your CFO should understand before approving a project:

• Electricity cost per inference: each call to a large model consumes between 0.001 and 0.01 kWh depending on model size. An SMB making 100,000 calls per month to a medium-sized model pays between 200 and 1,500 MXN of additional electricity at the hosting data center, which is passed through in the colocation invoice.
• Contracted electrical capacity: if your company is going to train models or run heavy inference, the kW contracted with the data center increases 3 to 5 times. This directly impacts the capacity contract and the fixed monthly cost, not only the variable consumption.
• Grid constraint: in Mexican cities with a saturated electrical grid (CDMX, Monterrey, Guadalajara), increasing rack density can trigger CFE interconnection requirements that take between 6 and 24 months to approve, according to reports from local operators.

Electricity is the most expensive variable in an AI decision. An inference that looks cheap on paper can be expensive on the invoice if the data center does not have renewables contracted at a fixed price. Always ask: what percentage of your data center’s electricity mix is renewable, what is the actual PUE (Power Usage Effectiveness, which measures the electrical overhead of cooling), and whether it has a peak-coverage contract with the local grid.

How to choose a green data center for AI with verifiable criteria

Before committing to an AI provider or a data center for training models, demand five verifiable data points:

1. WUE published in the last 12 months: the operator must provide an audited WUE, not an estimate. If it does not have one public, assume it is high.
2. Annual average PUE broken down by season: PUE in Mexican warm climates can vary between 1.3 in winter and 1.6 in summer. The difference is thousands of pesos per month in electricity.
3. Verifiable electricity mix with I-REC certification or equivalent: the operator must hold purchased and retired Renewable Energy Certificates, not just a signed PPA.
4. Watershed of the cooling water and replenishment plan: if the data center operates in an overexploited watershed per CONAGUA, future regulatory risk is high. Confirm whether there is a zero-discharge or replenishment contract in place.
5. Rack density capacity up to 80 kW: if your AI project requires dense GPUs, the data center must be able to deliver 40–80 kW per rack without re-engineering. If it caps at 15 kW, it does not fit AI.

Mexican SMBs that sign AI contracts without verifying these numbers usually discover the hidden costs 12–18 months later, when the electricity bill arrives with an overage of between 30% and 50% over what was projected and the environmental authority starts asking about water consumption.

Verdict for Mexico

The algorithm matters less than your AI provider wants you to believe. The real cost is in the water and electricity that compute consumes, and those costs are determined by engineering and location decisions the data center operator already made. Your job, before signing, is to audit those numbers with public data and demand third-party certifications. If the provider does not have them, operate with the footprint it has: probably higher than the brochure says.

Sources

1. EESI — Environmental and Energy Study Institute: Data Centers and Water Consumption. Quantitative analysis of water consumption in AI data centers. — https://www.eesi.org/articles/data-centers-and-water-consumption
2. Uptime Institute — Tier Topology standard and Tier I to Tier IV certifications to evaluate redundancy and efficiency in data centers. — https://uptimeinstitute.com/
3. ASHRAE — Thermal Guidelines for Data Processing Environments. Reference standard for operating ranges and thermal efficiency. — https://www.ashrae.org/
4. TIA Online — TIA-942 standard for telecommunications infrastructure in data centers, the basis for Rated-1 to Rated-4. — https://tiaonline.org/