Can AI be carbon neutral?

AI can be described as carbon neutral within a clearly defined accounting boundary, usually after reducing emissions, procuring lower-carbon electricity and balancing residual emissions with credible carbon removals or credits. It cannot be literally impact-free. Chips were manufactured, data centres were built and electricity was consumed.

The phrase is only useful when it answers five questions: which product or organisation, which emissions scopes, which time period, which electricity-accounting method and which offsets or removals? “Carbon-neutral AI” without those details can mean anything from annually matched operational electricity to a full value-chain claim. Those are not equivalent. A strong climate strategy prioritises avoiding and reducing emissions, then deals transparently with the remainder rather than using a neutral label as the starting point.

Neutral is an accounting result, not no emissions

Carbon neutrality generally means counted greenhouse-gas emissions are balanced by counted reductions or removals over a stated period. The servers still draw power while the claim is in force. A credit does not travel backwards through the cable and erase generation.

That does not make carbon accounting pointless. A defined inventory helps organisations find emissions, set targets and fund climate action. The problem is using the net result to hide the gross footprint or omit inconvenient sources.

For AI, the boundary might cover one inference service, a data centre, all company operations or the full supply chain. A reader cannot compare neutral claims until that scope is visible.

Clean electricity comes in different forms

A company can buy enough renewable electricity or certificates over a year to match annual consumption. Yet its data centres still use the local grid at night or during periods when the contracted generation is elsewhere. Hourly, location-matched carbon-free energy sets a more demanding standard.

The International Energy Agency’s data-centre energy-supply analysis uses the physical generation mix rather than operators’ contractual procurement. It expects renewables to meet a large share of demand growth, while natural gas and coal also contribute materially in the near term.

This is one reason a company can report renewable matching while a system-level study still attributes fossil generation to data-centre demand. The methods answer different questions.

Scope 3 is difficult to neutralise

Operational electricity sits mainly in direct and purchased-energy inventories. Manufacturing semiconductors, producing steel and concrete, constructing facilities and buying equipment commonly appear in Scope 3 value-chain emissions. These can dominate even when market-based operational electricity looks low.

Microsoft’s 2025 Environmental Sustainability Report said its total Scope 1, 2 and 3 emissions were 23.4% above its 2020 baseline in the reported year, with AI and cloud growth among the contributing factors. It also reported progress reducing Scope 1 and 2 emissions and contracting carbon-free electricity and removals.

The mixed result is more informative than a single label: operational improvements can coexist with a growing value-chain footprint.

Offsets and removals are not interchangeable

Avoided-emission credits claim that emissions elsewhere were prevented relative to a baseline. Carbon removals take CO2 from the atmosphere and store it in biological or geological systems. Their durability, additionality, leakage risk, monitoring and timing vary.

A tonne of fossil CO2 has a long atmospheric effect. A tree can store carbon while it grows, but fire, disease or land-use change can reverse that storage. Geological removal aims for greater durability but currently differs in availability and cost. Treating every tonne-labelled credit as identical hides these differences.

Credible climate action can fund nature and removals without claiming that either makes inefficient computation harmless. Gross energy and emissions should remain visible beside any balancing action.

Meta’s 2025 Sustainability Report says it continues to match 100% of its electricity use with clean and renewable energy and aims for net zero across its value chain in 2030. The future value-chain target is separate from its current electricity claim.

Google’s 2025 Environmental Report describes clean-energy procurement and progress reducing data-centre energy emissions while pursuing carbon-free energy around the clock. Microsoft describes a 2030 carbon-negative goal alongside current gross-emissions challenges.

These reports show serious work and remaining gaps. They are company-level disclosures, not proof that any particular AI response has zero emissions. Boundaries, baselines and market instruments must travel with the headline.

A better order of operations

First, avoid unnecessary computation: do not use a generative model where ordinary software works. Second, reduce energy per useful task through smaller capable models, efficient hardware, sensible output limits and high utilisation. Third, supply the remaining electricity with additional low-carbon generation that is well matched in place and time.

Then reduce embodied emissions through longer hardware life, lower-carbon materials, supplier requirements and reuse. Finally, address hard-to-eliminate residual emissions with high-quality, transparently reported removals. This order protects the neutral claim from becoming permission to expand wastefully.

AI training versus inference emissions helps locate reductions in the model life cycle, while AI carbon footprint explained sets out the operational and embodied boundaries.

How to test a carbon-neutral AI claim

Look for a named product and reporting year, gross emissions before offsets, Scopes 1–3, location- and market-based electricity results, the share reduced directly, and a registry or project record for credits. Check whether removals are durable and independently verified. Ask whether model training and hardware manufacture are included.

If the evidence is only a tree counter or annual renewable certificate total, the claim may cover a worthwhile contribution but not the full AI life cycle. If estimates are used, their uncertainty should be stated.

Treechat does not claim that its messages are carbon neutral. It estimates operational energy and carbon using disclosed assumptions, and reports tree funding as a separate action. It does not subtract trees from the per-message receipt or present planting as erasure. The current approach and limits are on our methodology page.

So yes, AI can meet a defined carbon-neutral accounting claim. The more important question is whether that definition drives real gross reductions, cleaner infrastructure and credible removals—or merely produces a cleaner-sounding label.