“Every AI company makes a net-zero pledge. Almost none of them have explained how those pledges are compatible with the scale of compute expansion they are simultaneously announcing. Both things cannot be true without a genuinely radical efficiency improvement that has not yet been demonstrated at the pace required. Someone needs to hold these two facts in the same sentence. Nobody in the industry is volunteering to do it.”
Neal Lloyd · Inside The Machine, Day 27Every major AI company has made a public commitment to carbon neutrality or net-zero emissions on a specific timeline. Microsoft committed to being carbon negative by 2030. Google committed to net-zero by 2030. Amazon committed to net-zero by 2040. Meta has committed to net-zero across its value chain by 2030. These are serious, public, board-level commitments made by companies with genuine sustainability programmes and real capital allocated to renewable energy procurement. They are also commitments made by the same companies that are simultaneously announcing the largest data centre construction programmes in history, specifically to serve AI workloads whose total energy demand — as covered in Day 17 — is projected to reach 3-8% of global electricity consumption by 2030. The tension between these two sets of public commitments is not being adequately addressed by any of the companies making them, and it deserves the same analytical rigour that the industry applies to every other quantifiable claim it makes. This is Day 27 of Inside The Machine. Today we hold AI’s environmental commitments to the same evidentiary standard the industry applies to its benchmark scores.
What “Net Zero” Actually Means and Why the Accounting Matters
Corporate carbon accounting operates across three scopes. Scope 1 covers direct emissions from sources a company owns or controls — the diesel generators backing up a data centre, the company vehicle fleet. Scope 2 covers emissions from purchased electricity — the power consumed by data centres, attributed based on the source of that power. Scope 3 covers everything else in the value chain — the emissions embedded in manufacturing the chips, the emissions from the construction of the data centre buildings, the emissions associated with the entire supply chain that produces the hardware AI runs on. The vast majority of corporate net-zero commitments focus most explicitly on Scope 1 and Scope 2, where the company has direct control and where renewable energy procurement can produce dramatic improvements relatively quickly. Scope 3 — which for a chip-intensive industry like AI represents an enormous and rapidly growing share of total emissions — receives comparatively less rigorous public accounting.
The renewable energy matching accounting deserves particular scrutiny. When a company claims its data centre operations are “100% renewable,” this typically means the company has purchased renewable energy certificates (RECs) or signed power purchase agreements (PPAs) equal in volume to its annual electricity consumption — not that the specific electrons powering its servers at any given moment came from renewable sources. A data centre in a region with a fossil-fuel-dominated grid can claim 100% renewable status by purchasing RECs from a solar farm in an entirely different region, even though the actual power flowing into its servers at 2am, when solar generation is zero, comes overwhelmingly from the local grid’s fossil fuel baseload. This accounting framework is not fraudulent — it is the industry-standard methodology, used across most sectors, not just AI — but it produces headline claims that are considerably more reassuring than the underlying physical reality.
The chip manufacturing emissions are the Scope 3 category that receives the least public attention relative to its scale. Producing a single advanced AI chip requires extraordinarily energy- and water-intensive semiconductor fabrication processes. TSMC, which manufactures the chips for Nvidia, AMD, and most advanced AI accelerators, reported that its manufacturing operations consumed significant water and electricity resources, and the carbon intensity of chip fabrication is a meaningful and rapidly growing component of the AI industry’s total footprint that is rarely included in the headline carbon figures that AI companies publicise.
Scope 1 (direct emissions): Generally well-measured and reported. Scope 2 (purchased electricity): Well-measured, but “100% renewable” claims often rely on REC/PPA matching rather than physical electron sourcing. Scope 3 (value chain, including chip manufacturing): Least rigorously measured, largest and fastest-growing category for a chip-intensive industry, least represented in headline sustainability claims. Net effect: a company can accurately claim “net zero in Scope 1 and 2” while its actual total environmental footprint — including the chips, construction materials, and supply chain — is growing substantially. This is not a uniquely dishonest practice; it is the industry-standard accounting framework applied across most sectors. It is also a framework that systematically understates the true environmental cost of rapid AI infrastructure expansion.
The Resource That Gets Less Attention Than Carbon But May Matter More Locally
Data centres consume water for two primary purposes: direct cooling (evaporative cooling systems that use water to dissipate heat) and indirect consumption (the water used to generate the electricity the data centre consumes, particularly for thermoelectric power generation). Microsoft disclosed 1.7 billion gallons of direct water consumption across its global data centres in 2022, with that figure rising substantially as AI workloads expanded through 2023 and 2024. Google’s disclosed water consumption exceeded 5 billion gallons in 2023. These are company-wide figures across all data centre operations, not AI-specific, but the AI-driven expansion of data centre capacity is the primary driver of recent growth in these numbers.
The geographic distribution of this water consumption is the detail that matters most for assessing real-world impact. Data centres are sited based on land availability, power access, and tax incentives — factors that frequently favour regions that are also experiencing water stress. Several major data centre clusters in the western and southwestern United States are located in regions already managing significant water scarcity challenges, where agricultural and municipal water users are competing for an increasingly constrained resource. A data centre’s water consumption, in aggregate national terms, may be a small fraction of total water use. In the specific watershed where it operates, it can represent a meaningful new demand on a resource that local communities and agricultural users were already struggling to manage.
The industry response — closed-loop cooling systems that recycle water rather than consuming it through evaporation, and increasing use of air cooling and liquid cooling technologies that reduce water dependency — is real and improving. Microsoft has committed to water-positive operations by 2030. Google has made similar commitments. The pace of this transition relative to the pace of data centre capacity expansion is the open question: if new capacity is being built faster than the water-efficient retrofit of existing capacity, total water consumption can continue rising even as the per-unit efficiency improves. This is the water-specific version of the Jevons paradox covered in Day 17’s energy discussion, and it applies with similar force.
A data centre’s water consumption may be a rounding error in national water statistics and a significant new burden on the specific watershed where it operates. Aggregate national figures obscure local reality. The communities experiencing water stress because of a nearby data centre are not comforted by a corporate sustainability report showing the company’s global water intensity is improving. They are experiencing a specific, local, physical resource constraint that the global figures do not capture.Neal Lloyd · Inside The Machine, Day 27
The Case That AI Is Net-Positive for Climate, and Where It Holds Up
The case for AI as a net climate positive is not a marketing claim without substance. There is genuine, peer-reviewed evidence that AI is accelerating several climate-relevant technologies in ways that may offset a meaningful share of its own energy footprint. DeepMind’s work on data centre cooling optimisation reduced Google’s own data centre cooling energy use by approximately 40% through AI-driven control systems — a genuine efficiency gain that compounds across every data centre that adopts similar techniques. AI-optimised battery chemistry research is accelerating the development of grid-scale storage technology that is essential for integrating intermittent renewable energy at scale. AI-powered grid management systems are improving the efficiency of electricity transmission and reducing waste in distribution networks. AI weather and climate modelling is improving the precision of renewable energy generation forecasting, reducing the inefficiency that comes from over-provisioning backup capacity for unpredictable wind and solar output.
The materials science applications are potentially the most significant long-term contribution. AI-accelerated discovery of new battery chemistries, more efficient solar cell materials, and carbon capture catalysts represents exactly the kind of scientific acceleration that the AlphaFold precedent (covered in Episode 14’s discussion of John Jumper) suggests AI is genuinely capable of delivering. If AI-accelerated materials science delivers a meaningfully more efficient solar panel, a meaningfully cheaper grid-scale battery, or a viable carbon capture technology years ahead of the timeline that conventional research would have achieved, the climate benefit of that acceleration could exceed AI’s own energy footprint by a significant margin.
The honest caveat: most of these benefits are prospective, not yet realised at scale, and dependent on continued research investment and successful translation from laboratory result to deployed technology — a translation that historically takes years to decades even for promising discoveries. The energy AI consumes today is certain and immediate. The climate benefits AI might deliver are probable but not yet realised, and the timeline for realising them is genuinely uncertain. Comparing a certain, immediate cost against a probable, delayed benefit is not invalid — many worthwhile investments have exactly this profile — but it requires honesty about the asymmetry, which most corporate sustainability communications do not provide.
Certain and immediate: AI energy consumption, today, growing at the pace documented in Day 17. Probable and delayed: AI-accelerated climate technology breakthroughs, on a timeline of years to decades, dependent on successful research translation. Corporate communications routinely present these two facts as though they offset each other in real time. They do not. The energy is being spent now. The climate benefit, if it materialises, arrives later — possibly years later, possibly with a smaller magnitude than current optimistic projections assume. A company can be telling the truth about both facts and still be presenting a deeply misleading overall picture by failing to specify the timing mismatch between them.
The Standards That Would Make AI’s Environmental Claims Actually Verifiable
Time-matched renewable energy claims, not annual matching. The gap between “100% renewable annually” and “100% renewable at the moment of consumption” is the single most consequential accounting distinction in corporate clean energy claims. Google has begun publishing hourly carbon-free energy matching data for some of its data centre regions — a meaningfully more rigorous standard than annual REC matching. This should be the baseline requirement for any company making a public renewable energy claim, not a voluntary enhancement that only the most committed companies adopt.
Full Scope 3 disclosure including chip manufacturing. The carbon and water footprint of chip fabrication is currently the least transparently reported component of AI’s environmental impact, despite being one of the largest and fastest-growing. Requiring AI companies to disclose the embedded emissions of their hardware supply chain — not as an aggregate Scope 3 estimate but with chip-manufacturing-specific detail — would close the largest current gap in environmental accountability.
Local impact disclosure, not just aggregate figures. Water and energy impact assessments at the watershed and local grid level, not just company-wide aggregate figures, would allow the communities actually experiencing the local effects of data centre expansion to assess the real impact on their specific resources, rather than being reassured by global figures that may not reflect their local reality at all.
Honest timing on the climate-benefit offset claim. Companies citing AI’s climate benefits as an offset to its energy consumption should be required to specify the timeline, the probability, and the magnitude of the claimed benefit with the same rigour applied to capability benchmarks. “AI will help solve climate change” is a claim that deserves the same scrutiny as “our model achieves 95% accuracy on this benchmark.” Currently it receives considerably less.
The AI industry applies extraordinary analytical rigour to benchmark scores, capability claims, and competitive positioning. It applies considerably less rigour to its own environmental footprint claims. Both kinds of claims are quantifiable. Both kinds of claims matter enormously to the people relying on them. The asymmetry in scrutiny is not a technical limitation. It is a choice about which claims the industry is comfortable having examined closely.Neal Lloyd · Inside The Machine, Day 27
Inside The Machine, Day 27 · June 30 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs Closed
- Day 15The New Cold War
- Day 16Why AI Lies With Confidence
- Day 17AI Is Eating the Power Grid
- Day 18The Age of AI Agents
- Day 19AI Safety Was Never Just Theory
- Day 20The Surveillance Question
- Day 21AI and the Future of Education
- Day 22AI and Your Health
- Day 23What Is AGI and Are We Close?
- Day 24What Is Work For?
- Day 25AI and Democracy
- Day 26AI and the Future of Money
- Day 27Can the Planet Afford AI?You are here
