A Credible Sustainability Ledger for AI and Data Centers

A data center can improve its efficiency ratio while total electricity use rises. It can sign a clean energy contract while consuming power during hours when the local grid is fossil intensive. It can reduce onsite water use while shifting water demand to electricity generation or equipment manufacturing. Each statement may be accurate within its boundary, yet the combined phrase "sustainable AI" can still mislead. A credible assessment begins with a ledger, not a badge. The ledger keeps operational electricity, greenhouse gas emissions, water, and embodied impacts in separate columns; defines the boundary and functional unit for each; reports both absolute totals and useful intensity measures; and records uncertainty. Only then can efficiency, workload scheduling, infrastructure design, procurement, and hardware life be compared without turning unlike impacts into one score. Ledger one: report useful work and total electricity together Measure facility electricity and, where possible, the portion delivered to computing equipment. Power usage effectiveness can reveal overhead from cooling and power conditioning, but it does not measure whether the computing is useful, whether servers are well utilized, or whether total demand is falling. Report the metric with its averaging period, facility boundary, climate conditions, and treatment of onsite generation. Pair facility metrics with workload measures that match the service: completed training run, thousand inferences at a defined quality and latency, processed tokens, virtual core hour, or another reproducible unit. Preserve model version, hardware, precision, batching, utilization, and software configuration. A per query number without those conditions is not portable across systems. Then publish absolute electricity consumption. Efficiency per task can improve while traffic, model size, generated media, or agentic workloads grow faster. The IEA's energy analysis clearly treats efficiency, adoption, and changing use cases as separate drivers. A credible program therefore sets an absolute budget or forecast alongside intensity targets and explains material changes in workload mix. Operational improvements include selecting the smallest system that meets the quality requirement, raising accelerator utilization, batching compatible requests, caching reusable outputs, routing simple tasks away from expensive models, and retiring idle capacity. Measure the whole service path so savings in one component are not merely displaced to storage, networking, or repeated failed calls. Ledger two: distinguish physical electricity from emissions accounting Kilowatt hours and carbon dioxide equivalent answer different questions. Electricity use is measured at the facility; operational emissions depend on generation, location, and time. Report Scope 1 emissions from owned or controlled combustion and refrigerant losses separately from Scope 2 emissions associated with purchased electricity. The GHG Protocol Scope 2 Guidance requires location based reporting and, where applicable, market based reporting. The location based result reflects average emissions factors for the grids where consumption occurs. The market based result reflects qualifying contractual instruments and supplier information. Showing both prevents a procurement claim from erasing the physical context of where and when electricity was consumed. For decision making, add time resolved data when credible marginal or average grid factors are available. Flexible training, batch analytics, and other delay tolerant work may be shifted to periods or regions with lower emissions, subject to latency, reliability, data governance, and network constraints. Do not claim a reduction from scheduling unless the baseline, signal, displaced generation assumption, and rebound effects are documented. Moving work can change emissions; it does not make the electricity disappear. Keep offsets and avoided emissions claims outside the gross inventory totals. Report purchased credits, removals, or claimed benefits in separate lines with project, vintage, methodology, retirement, and double counting controls. Products that may help others avoid emissions do not cancel emissions from operating the data center. Ledger three: show water by source, place, and pathway Water accounting needs more than one global total. Report withdrawal, discharge, and consumption separately, identifying municipal, surface, groundwater, reclaimed, and other sources. Show basin or facility location, season, cooling method, and performance during water stress. A liter consumed in a water abundant basin is not operationally equivalent to a liter consumed during local scarcity, although both still belong in the inventory. Separate direct onsite water from indirect water associated with electricity generation. Peer reviewed research in Nature Sustainability models both pathways and shows why location and grid composition can materially affect results. The figures are scenario estimates for U.S. AI servers, not universal factors for every query or facility. Operators should replace generic factors with facility and supplier data where available and disclose gaps. Cooling changes can create trade offs. A design that lowers cooling electricity may use more water, while a water saving approach may increase electricity or embodied material. Report water usage effectiveness only with its boundary and denominator, and pair it with absolute consumption and local stress. Water replenishment projects should be reported separately from consumption, with the basin, timing, volume method, durability, and verification stated; replenishment is not the same physical event as avoiding a withdrawal. Ledger four: include construction, chips, servers, and end of life Operational reporting misses emissions and resource use from buildings, concrete, steel, semiconductors, servers, networking equipment, batteries, cooling systems, transport, and disposal. The GHG Protocol Scope 3 Standard places purchased goods and capital goods within the value chain inventory. ISO 14040 provides the framework for life cycle assessment: define goal and scope, compile the inventory, assess impacts, interpret results, disclose limitations, and apply review appropriate to the claim. Choose a functional unit before comparing designs. A server, rack, megawatt of installed capacity, completed workload, or year of service can produce different rankings. State the study period, geography, allocation method, expected equipment life, replacement assumptions, recycling credits, and excluded components. Do not compare one vendor's cradle to gate product footprint with another operator's full data center inventory as if the boundaries match. A 2025 peer reviewed Nature study of data center cooling evaluated raw material acquisition, production, transport, operation, and end of life. It found that as operational electricity is decarbonized, embodied impacts can become relatively more important. That does not mean hardware always dominates. It means procurement, utilization, repair, refurbishment, and service life should remain visible instead of disappearing behind a low operational emissions claim. Procurement claims need an evidence chain A power purchase agreement, renewable energy certificate, onsite generator, and round the clock clean energy match are different instruments. For every claim, identify the facility or load covered, technology, grid region, contract start, operating date, volume, matching interval, certificate ownership, and treatment of curtailed output. Distinguish signed capacity from projects that are operational and delivering attributes. Additionality is not a universal yes or no label. Explain the mechanism by which procurement is expected to support new supply, such as long term revenue certainty, financing, or direct investment, and disclose counterfactual uncertainty. Also show residual electricity and unmatched hours. Google's 2025 Environmental Report is useful as an operator disclosure because it separately presents data center energy emissions, contracted clean energy capacity, water replenishment, and hardware efficiency; its company figures should be read with its methodology and footnotes, not generalized to the sector. Supplier and operator disclosures should be auditable to contracts, meter data, certificate registries, emissions factors, water bills, bills of materials, and life cycle models. Where commercial confidentiality prevents publication, an independent assurance statement can describe the tested boundary and controls without exposing security sensitive details. Publish a ledger that can survive scrutiny For each reporting year, release four tables: electricity and useful work totals; Scope 1 and dual method Scope 2 emissions; direct and indirect water by location; and material Scope 3 or life cycle impacts. Add a fifth table for procurement instruments and claims. Show prior year values, restatements, estimation methods, uncertainty ranges, exclusions, and responsible reviewers. Do not aggregate the tables into a single sustainability score unless the weighting is public and the underlying results remain visible. A project can reduce one impact while increasing another. Decision makers need the trade off, not a green label that conceals it. The strongest sustainability claim is specific and bounded: what changed, compared with what baseline, during which period, at which facilities, using which method, and with what remaining impact. That discipline does not guarantee that AI growth is sustainable. It makes progress and failure measurable, procurement claims testable, and resource choices available for informed review. Sources Energy supply for AI https://www.iea.org/reports/energy and ai/energy supply for ai GHG Protocol Scope 2 Guidance https://ghgprotocol.org/sites/default/files/standards/Scope%202%20Guidance Final Sept26.pdf Corporate Value Chain Scope 3 Standard…