Here's a number worth sitting with: global data center electricity consumption is on track to hit roughly 565 terawatt-hours in 2026, a jump of more than a quarter from the previous year, and AI-optimized servers are the single biggest reason why (Gartner, via aboutchromebooks.com). That's more electricity than most countries use in a year, flowing into server racks that spend their days predicting the next word in a sentence or the next pixel in an image.

Nobody builds a data center hoping to strain a local power grid or drain a regional aquifer. But that's increasingly the side effect of the AI boom, and it's forcing a question the tech industry would rather answer with a press release than a real trade-off: can the companies building the most powerful AI systems in history also keep their climate promises? The honest answer, based on what's actually showing up in corporate disclosures this year, is "not yet, and not without some hard choices." This piece walks through why AI is so energy-hungry, what's happening to Big Tech's sustainability pledges under that pressure, and where genuine progress is being made.

Every digital convenience has a physical footprint somewhere, and for most of us it's invisible. That footprint starts with the data center itself.

• Energy and cooling. Data centers don't just run servers — they run massive cooling systems to keep those servers from overheating, and cooling can account for a large share of a facility's total power draw. The industry standard metric here is Power Usage Effectiveness (PUE): a facility with a PUE of 1.5 uses 50% more electricity than the servers alone would need, with the rest going to cooling, power conversion, and other overhead. Newer AI-optimized facilities are pushing PUE down through liquid cooling and smarter airflow design, but the overall demand for compute is rising faster than efficiency gains can offset.
• E-waste from hardware turnover. AI's hunger for compute means GPUs and specialized chips get swapped out on aggressive upgrade cycles, often every two to three years, well before the hardware is technically "worn out." That churn generates a growing stream of electronic waste, much of which contains materials that are energy-intensive to mine, refine, and eventually dispose of or recycle. It's a less-discussed cousin of the energy story, but it matters across the full lifecycle of AI infrastructure.
• Water usage — the overlooked one. Cooling towers evaporate enormous volumes of water, and about 80% of the water drawn into a typical evaporative cooling system is lost to evaporation rather than returned to the source (per the NASUCA data center water use brief, citing Bloomberg and Microsoft/Google/Meta sustainability reports; see NASUCA, June 2025). Google's own reporting shows its data center water consumption climbed from about 4.3 billion gallons in 2021 to 6.1 billion gallons in 2024 (MOST Policy Initiative), and the company disclosed a further 34% year-over-year increase in 2025 (GPUSmith). Globally, the IEA estimates data centers consumed around 560 billion liters of water in 2023 when you count both direct cooling water and the water used at power plants to generate the electricity data centers draw (GPUSmith). In water-stressed regions, that's not an abstract statistic — it's a resource being shared with farms and households.

Not all computing is created equal, and generative AI sits near the top of the energy-intensity chart for a few concrete reasons.

• Training is enormous. Training a frontier model isn't a weekend project — it can require running thousands of specialized chips continuously for weeks or months. Estimates put GPT-3's training run at roughly 1,287 megawatt-hours and around 552 tons of CO2 (AIMultiple), while GPT-4-scale training consumed more than 50 gigawatt-hours — enough to power roughly 20,000 U.S. homes for a year (All About AI, 2026). Next-generation frontier models are expected to push past 100 gigawatt-hours per training run before the end of the decade.
• Chip demand is reshaping manufacturing. The race to build ever-larger models has driven explosive demand for GPUs and AI accelerators, and manufacturing that hardware carries its own embedded carbon cost — from semiconductor fabrication (an energy- and chemical-intensive process) to the mining of rare materials. Gartner's 2026 forecast projects AI-optimized servers will draw 175 TWh this year alone and overtake conventional servers as the dominant source of data center power demand by 2027 (aboutchromebooks.com).
• Inference adds up at scale. Training gets the headlines, but inference — actually running the model to answer billions of everyday queries — is a cumulative drain that never stops. Google's own technical disclosures put a median Gemini Apps text prompt at about 0.24 watt-hours of electricity and 0.26 milliliters of water as of mid-2025 (aboutchromebooks.com), which sounds trivial per query — until you multiply it by the billions of prompts sent daily across the industry. One estimate puts daily ChatGPT usage at an annualized energy footprint comparable to Ireland's entire electricity consumption (worldmetrics.org). Small numbers, multiplied by huge scale, stop being small.

Nearly every major tech company has a headline climate commitment: Google pledged to be carbon-neutral, Microsoft to be carbon-negative, Amazon to hit net-zero by 2040, Meta to reach net-zero by 2030. These pledges were made, largely, before the current AI buildout began — and 2026's disclosures show the strain.

Google's total greenhouse gas emissions rose 25% year-over-year, and Amazon's climbed 16%, with most of the growth concentrated in Scope 3 emissions — the supply-chain category covering everything from purchased GPUs to the steel and cement used in new data center construction (Technology.org, July 2026). Amazon's carbon intensity — emissions per dollar of revenue — rose for the first time since the company started tracking it in 2019, meaning emissions are now growing faster than the business itself (Computing.co.uk). Meta reportedly walked away from a pledge to source all its electricity from renewables as it leans on gas-powered generation to keep pace with AI data center demand (Watts Up With That, citing Yahoo Finance).

This is where the greenwashing debate gets sharper. A lot of "carbon-neutral" accounting leans on renewable energy certificates (RECs) — financial instruments that represent renewable generation somewhere on the grid, not necessarily electrons flowing directly into the servers doing the computing (TechRound, Earth Day 2026). Critics argue that without mandatory, standardized disclosure, companies can hit a "net-zero" number on paper while their actual power draw — and the emissions tied to it — keeps climbing. It's also worth noting that greenwashing enforcement has teeth in other corners of finance: Deutsche Bank's asset management arm was fined €25 million in 2025 over misleading environmental claims (Gasilov Group), a reminder that regulators are increasingly willing to test sustainability marketing against the numbers.

That's part of why third-party verification matters. Frameworks like the Science Based Targets initiative (SBTi) give outside reviewers a chance to check a company's stated targets against its actual trajectory, rather than letting companies grade their own homework. Companies that lean on SBTi validation, publish granular Scope 1/2/3 breakdowns, and disclose water and energy data at the facility level are giving outside observers something to actually verify — as opposed to a glossy annual report with a "climate moonshot" framing and no year-over-year comparison.

The good news is that the same companies straining their climate pledges are also the ones funding real efficiency breakthroughs — the question is whether those breakthroughs can outrun demand growth.

• Renewable-powered data centers. Hyperscalers have signed enormous solar and wind power purchase agreements, and Amazon in particular points to projects like its Wind Wall farm in California's Tehachapi Mountains as evidence of direct investment in new renewable capacity, not just credits (GeekWire). The more rigorous version of this idea is "24/7 carbon-free energy" — matching electricity demand with carbon-free supply on the same local grid, hour by hour, rather than netting out annual totals against RECs purchased anywhere in the world (iCert Global).
• More efficient chip architectures. Not every model needs to be trained the same way. DeepSeek-V3 made headlines for reportedly achieving roughly 95% lower training energy use than comparable frontier models while remaining competitive on performance (All About AI) — a sign that architecture choices (like mixture-of-experts designs that activate only part of a model per query) can matter as much as raw compute scale. Custom silicon — Google's TPUs, Amazon's Trainium, Microsoft's Maia chips — is also part of this story, since purpose-built AI accelerators generally deliver more computation per watt than general-purpose GPUs.
• Carbon-aware computing. This is one of the more elegant ideas in the sustainability toolkit: schedule non-urgent compute jobs — batch data processing, model retraining, backups — for times and places where the local grid is running on the most renewable energy. Researchers and cloud providers have been formalizing this into real scheduling systems that shift workloads geographically and temporally to chase clean power (arXiv, Carbon-Aware Computing for Data Centers). It won't help with real-time chatbot queries, which need to run immediately wherever the user is, but for the large share of AI workloads that aren't latency-sensitive, it's a meaningful lever.
• Liquid cooling and other efficiency innovations. Microsoft has reported cutting its average water usage effectiveness to 0.27 liters per kilowatt-hour in 2025, well below the industry average of 0.84 l/kWh, and says its newest AI data centers use no water for cooling during normal operations thanks to closed-loop, chip-level liquid cooling systems that are filled once and then recirculate indefinitely (Trellis; Rinnovabili). Microsoft has also demonstrated microfluidic cooling, etching tiny channels directly into silicon so coolant flows onto the chip itself rather than through a separate heat sink — a much more direct and efficient way to pull heat away from the hottest components.

Voluntary pledges got the industry this far, but 2026 is the year mandatory disclosure starts to bite, especially in Europe.

The EU's revised Energy Efficiency Directive already requires data center operators with 500 kW or more of IT power demand to report energy performance and water footprint data annually to a European database, with the 2025 reporting year due by May 15, 2026 (Gasilov Group). The European Commission is expected to follow with a Data Centre Energy Efficiency Package in 2026, introducing an efficiency rating scheme with the explicit goal of carbon-neutral data centers by 2030 (White & Case).

Germany has gone further on its own, through its national Energy Efficiency Act: new data centers must hit escalating renewable-electricity thresholds (50% now, 100% from January 2027), meet waste-heat reuse targets that climb from 10% to 20% between 2026 and 2028, and comply with hard PUE ceilings for new builds (ModuleEdge).

The regulatory picture in the U.S. is more fragmented. California's SB 253 requires large companies to begin reporting Scope 1 and 2 emissions in 2026, with Scope 3 following later, though it applies only to companies above a $1 billion revenue threshold and doesn't single out AI or computing specifically (ScienceDirect). A more targeted federal bill, the AI Environmental Impacts Act, was introduced in the Senate but has not advanced. That patchwork means a lot of the pressure on U.S. tech giants currently comes from their EU operations, investor scrutiny, and public opinion rather than binding domestic law — though that could change quickly given how much political attention data center construction is now drawing in local communities.

It's worth being fair here: some of the efficiency gains are real, independently verifiable, and worth highlighting without endorsing every claim in a company's annual report.

• Water efficiency engineering. Microsoft's shift toward closed-loop and chip-level liquid cooling is a genuine hardware innovation, not just an accounting trick, and it's being deployed at new facilities in Arizona and Wisconsin (Introl).
• Circular hardware programs. Apple has reported cutting annual cooling water use by roughly 60 million gallons through server upgrades, and is working toward certifying its data centers under the Alliance for Water Stewardship standard (Dgtl Infra).
• Direct renewable investment. Long-term power purchase agreements for new wind and solar capacity — rather than just buying certificates after the fact — add real clean generation to the grid, which is a meaningfully different contribution than offset-only strategies.
• Model efficiency research. Techniques like mixture-of-experts architecture and better inference scheduling (adjusting chip voltage and clock speed based on what phase of a query is running) are already cutting the energy cost of inference by double-digit percentages in research settings (arXiv, Energy-Aware Computing 2026).

The pattern across the data is consistent: efficiency per unit of AI work keeps improving, sometimes dramatically, but total energy and emissions keep rising anyway because demand for AI is growing even faster than efficiency gains can offset. That's not necessarily a story about wasted effort — it's what economists call a rebound effect, and it means efficiency alone won't get the industry to its climate targets. Something has to also constrain absolute growth or fundamentally decarbonize the power supply.

Most of the environmental cost of AI is set upstream, in data center design and corporate energy procurement — but individual choices aren't meaningless.

• Choose efficient cloud providers and regions. Cloud platforms increasingly expose carbon-intensity and water-usage data by region; choosing a region running on a cleaner grid, or a provider with a published low PUE, has a real if modest effect.
• Right-size the model to the task. Using a large, general-purpose model for a narrow task like sentiment classification can consume roughly 30 times more energy than a small model fine-tuned for that specific job (AIMultiple). Developers who default to the biggest available model out of convenience are leaving a lot of energy on the table.
• Optimize code and reduce redundant computation. Caching results, batching requests, and avoiding unnecessary re-runs all reduce the number of inference calls — and therefore the cumulative energy draw — without touching model quality.
• Support transparent, accountable companies. Favor vendors that publish facility-level water and energy data, use third-party verification like SBTi, and disclose Scope 3 emissions rather than companies that lead with vague "climate moonshot" language and no year-over-year comparisons.

None of this substitutes for structural change at the data center and grid level, but it does mean the environmental cost of AI isn't purely something that happens to users — it's also something shaped, in smaller ways, by how AI gets built and used day to day.

Sustainability and AI innovation aren't fundamentally incompatible — but 2026's numbers make clear they aren't automatically compatible either. The same year that brought genuine engineering wins, like Microsoft's zero-water data centers and dramatically more efficient model architectures, also brought some of the largest emissions increases the major cloud providers have ever reported. Voluntary pledges are visibly buckling under the weight of AI-driven demand, and it's mandatory disclosure regimes — starting in the EU and slowly building elsewhere — that are likely to do more to hold companies accountable than another glossy sustainability report.

The trajectory isn't fixed yet. Efficiency gains per query and per training run are real and continuing. Renewable capacity is being built at record pace, even if it isn't keeping up with demand growth. And regulatory pressure is finally starting to require the kind of granular, verifiable disclosure that makes greenwashing harder to get away with. Whether AI and sustainability end up coexisting will depend less on any single breakthrough and more on whether growth in AI capability can be decoupled from growth in AI's environmental footprint — a decoupling that hasn't happened yet, but that the underlying technology, if pointed deliberately in that direction, is capable of achieving.