Environmental Impact of AI Products: What PMs Should Consider

AI features have a real, measurable environmental cost — electricity for training and every inference call, water for data-center cooling, and emissions baked into the hardware itself. PMs can't eliminate that footprint, but they can measure it, right-size model choices to the job, and disclose trade-offs honestly, the same way they already weigh cost, latency, and privacy.

Quick Answer: AI's environmental footprint comes from three places — training compute, per-query inference at scale, and the water and hardware behind data centers. PMs influence it most through model-size choices, caching, and honest disclosure, not by waiting for a perfect carbon dashboard that doesn't exist yet.

Why This Belongs on a Product Roadmap, Not Just in Ops

Environmental impact is a product decision because the choices that drive AI's footprint — which model powers a feature, how often it's called, whether a smaller model would do the job — sit squarely inside a PM's scope. Engineering can optimize infrastructure efficiency, but only product decides what gets built, how often it runs, and for whom.

Consider a common pattern: a team defaults to the largest, most capable frontier model for every AI feature "to be safe," then never revisits that choice as the feature scales to millions of calls. That's not a technical oversight — it's an unreviewed product decision with a compounding environmental and cost impact.

This is also part of a broader shift already underway in how PMs are expected to think about AI. Environmental footprint sits alongside fairness, transparency, and manipulation risk as one more dimension of a responsible AI practice, not a separate sustainability checkbox bolted on afterward. Our complete guide to responsible AI and product ethics covers how these dimensions fit together across the product lifecycle.

Treating environmental cost as an ops-only problem also invites a subtler risk: greenwashing. A vague "we care about sustainability" line in a product FAQ, unsupported by any actual model-choice discipline, is its own kind of manipulation — the same family of issue explored in our piece on dark patterns and ethical product design.

Where the Footprint Actually Comes From: Training, Inference, and the Data Center Itself

AI's environmental cost breaks into three distinct phases with very different economics: a one-time (but large) training cost per model version, a per-query inference cost that repeats and scales with every user request, and the underlying data-center infrastructure — cooling, backup power, and hardware manufacturing — that supports both. Most PM attention goes to training, but inference is usually the bigger long-run number.

The practical takeaway: training is a headline number, inference is the recurring bill. A model trained once but queried a billion times a month has an inference footprint that dwarfs its training footprint within weeks — which is exactly why per-query design choices matter more than most roadmaps currently reflect.

Water is the footprint dimension PMs think about least, and it deserves more attention. Researchers led by Shaolei Ren at UC Riverside have estimated that training a large language model can consume hundreds of thousands of liters of fresh water for evaporative cooling, and that aggregate everyday usage adds up meaningfully at scale — even though no single chat message feels like it should.

What Regulators and Big Tech Are Already Disclosing

Environmental disclosure for AI is no longer hypothetical: the EU AI Act now requires energy-consumption reporting for general-purpose models, California has passed a corporate emissions-disclosure law that reaches major cloud and AI providers, and the largest AI companies are already reporting emissions increases tied directly to AI compute in their own sustainability filings.

The EU AI Act (Regulation (EU) 2024/1689) requires providers of general-purpose AI models to document known or estimated energy consumption as part of their mandatory technical documentation, particularly for models deemed to carry systemic risk. The Act separately encourages voluntary codes of conduct that explicitly list environmental sustainability — including resource-efficient training and model reuse — among the commitments providers can make.

In the United States, California's SB 253 (the Climate Corporate Data Accountability Act, signed into law in October 2023) requires companies with over $1 billion in revenue doing business in California — a bar most major cloud and AI vendors clear — to publicly disclose Scope 1 and Scope 2 emissions starting in 2026, with Scope 3 supply-chain emissions (which would capture compute infrastructure) following in 2027.

Big tech's own numbers back up why this scrutiny is increasing:

• Google's 2024 Environmental Report disclosed that the company's total greenhouse-gas emissions had risen roughly 48% relative to five years earlier, attributing a significant share of that increase to the energy demands of AI computing and data-center expansion.
• Microsoft's 2024 sustainability report showed overall emissions running nearly 30% above its 2020 baseline, driven largely by the construction and operation of AI data centers — working directly against its public pledge to be carbon-negative by 2030.
• Associated Press reporting has documented spikes in water use tied to Microsoft's West Des Moines, Iowa data-center campus, which supports OpenAI's model training, during periods of local drought stress — a concrete example of infrastructure decisions creating community-level externalities, not just a line item in a sustainability report.

None of this means every AI feature is an environmental crisis. It means the disclosure expectations PMs will face are tightening faster than most product roadmaps currently account for.

A Framework: Five Questions to Ask Before You Ship an AI Feature

You don't need a carbon-accounting team to make better decisions — you need to ask the same five questions consistently, before an AI feature ships, not after it's already at scale. Each question maps to a decision you already own.

1. What job is this actually solving, and does it need a frontier model to do it? Grounding the feature in the underlying customer job — the way a Jobs to Be Done analysis forces you to — often reveals that a smaller, cheaper, lower-footprint model handles the job just as well as the flagship one.
2. How often will this be called, and is each additional call cheap or expensive? A feature triggered on every keystroke has a wildly different footprint profile than one triggered on explicit user request.
3. Can caching, batching, or a distilled model cover the majority of cases? Reserve the largest model for the tail of genuinely hard queries rather than routing everything through it by default.
4. Are we disclosing model choice and its trade-offs the way we'd disclose data handling? Users increasingly expect the same transparency about how an AI feature works that they expect about what happens to their data — a theme covered in depth in our guide to AI transparency and explainability in product UX.
5. Where in the customer's path does this AI feature actually sit, and is it justified at every touchpoint? Mapping AI touchpoints against a customer journey often surfaces redundant AI calls stacked at the same moment — three separate "smart" suggestions firing on one screen, each with its own inference cost.

Environmental cost is also worth pairing with the other responsible-AI review a feature should get before launch — the same review cadence that should already be checking for bias in AI-driven product decisions. Treating footprint and fairness as one combined pre-launch pass, rather than two separate afterthoughts, tends to catch more real issues than either check run alone.

Design Levers That Actually Move the Needle

Most of the environmental cost in an AI feature comes down to five design decisions PMs can influence directly, each with a real trade-off attached — there's no lever here that's free.

Read across the table, the pattern is consistent: the environmentally lighter option is usually also the leaner product decision — narrower scope, more deliberate defaults, less "always-on" behavior. That's a useful heuristic even for teams that never build a formal carbon model: if a lever cuts energy use, it's very often also cutting unnecessary complexity or unreviewed defaults.

Building the Question Into Your Workflow, Not Just Your Conscience

The hardest part of any responsible-AI dimension — environmental impact included — isn't agreeing it matters. It's remembering to ask the question at the moment a spec is being written, not three sprints after a feature has already shipped at scale.

Key Takeaways

• Environmental footprint is a product decision, not an infrastructure afterthought — model choice, call frequency, and defaults are all things a PM directly controls.
• Inference, not training, is usually the bigger long-run cost once a feature reaches meaningful scale, because it repeats with every single call.
• Regulation is already catching up: the EU AI Act requires energy-consumption documentation for general-purpose models, and California's SB 253 will force large AI and cloud vendors to disclose emissions starting in 2026.
• Big tech's own disclosures show the trend: both Google and Microsoft reported meaningfully higher emissions in 2024, attributed in significant part to AI-driven data-center growth.
• Water use is underweighted in most environmental conversations but is well documented by researchers and by real reporting on data-center operations during drought conditions.
• The highest-leverage move is right-sizing the model to the actual job, using a Jobs to Be Done lens rather than defaulting to the largest available model.
• Disclosure and transparency compound trust — treating AI's footprint the way you'd treat a data-handling disclosure avoids both regulatory risk and greenwashing accusations.