The 662% Data-Center Gap Was Historical, Not an AI Emissions Audit

The 662% figure was not a new measurement of AI’s environmental footprint. The Guardian’s September 2024 analysis estimated that location-based emissions from company-owned data centers operated by Google, Microsoft, Meta and Apple during 2020–2022 were 662% higher than their aggregated market-based total; Amazon was excluded because comparable data-center figures could not be isolated.

What has changed is the scale and governance of the problem. Electricity demand from data centers is projected to grow sharply, while the framework used for corporate electricity emissions is being reconsidered. A July 29, 2026 GHG Protocol update said multiple reporting approaches remained under consideration and placed an integrated public consultation on a consolidated corporate standard in the second quarter of 2027.

What the 662% comparison measured

The comparison concerned Scope 2 emissions: greenhouse gases associated with electricity purchased by a company. It contrasted location-based accounting, which applies emissions factors from the grids supplying facilities, with market-based accounting, which reflects qualifying electricity contracts and energy-attribute instruments claimed by the buyer.

Those methods can produce very different totals without describing different quantities of electricity. A data center may consume power from a regional grid that includes fossil-fuel generation while its operator separately acquires renewable energy certificates or contracts for low-carbon power. The location-based result reflects the emissions profile around physical consumption; the market-based result reflects the contractual attributes assigned to that consumption.

The historical percentage therefore identified an accounting gap, not a direct measurement error at every facility. It also did not prove that the companies fabricated their inventories: market-based accounting is an established reporting method. The central transparency issue is whether a low contractual total is presented alongside the location-based figure and with enough explanation for readers to understand what each number represents.

The comparison had narrower boundaries than the phrase “AI emissions” suggests. It covered selected company-owned data centers, not every colocation facility or third-party cloud resource used by the companies. It also did not encompass the full supply chain, including server manufacturing, construction and other impacts normally treated as Scope 3 emissions.

The accounting rules are still unsettled

Proposed reforms have focused on making contractual claims correspond more closely to operating conditions on the grid. Hourly matching would connect clean-energy attributes to the periods when electricity is consumed, while deliverability requirements would limit claims to generation capable of serving the relevant electricity market.

This addresses a weakness in broad annual matching. Renewable generation available during one part of the day can otherwise be used contractually against consumption occurring at another time, including hours when the supplying grid relies more heavily on fossil fuels. Greater temporal and geographic precision would narrow that mismatch, although it would not eliminate market-based reporting.

The latest official status is less settled than the rejected idea of a finished rule change. Consultation feedback produced divergent views, and further proposals still require technical development and independent review. Hourly matching and deliverability should therefore be described as reform options under consideration, not requirements already imposed on technology companies.

The planned consolidation of corporate carbon-accounting standards also broadens the process beyond one amendment to Scope 2 guidance. That may improve consistency across reporting systems, but it means current corporate disclosures must still be read under the rules and methodological choices applicable to their reporting periods.

Electricity demand makes the distinction more important

The strongest newer evidence is an energy forecast rather than a replacement for the historical percentage. The International Energy Agency’s 2025 base case projects electricity generation for data centers to rise from 460 terawatt-hours in 2024 to more than 1,000 TWh in 2030, with renewables meeting nearly half of the increase but natural gas and coal together supplying more than 40% of the additional demand.

That model considers the fuel mix of electricity physically consumed at facilities rather than operators’ contractual portfolios. It therefore addresses the side of the accounting divide that the location-based method is intended to illuminate: what generation is serving data-center load where the facilities operate.

The forecast does not mean data centers will become the dominant source of global emissions. It does show why an expanding computing load can increase physical electricity-sector emissions even when companies acquire more renewable-energy attributes. Procurement can support cleaner generation, but a contractual claim alone does not establish what supplied a facility at a particular place and time.

Efficiency also needs to be separated from absolute impact. A model or data center can use less energy for each unit of computation while total consumption rises because far more computation is performed. Conversely, higher location-based emissions do not establish that every AI task has become less efficient.

Why the figure cannot be assigned to AI alone

Company data centers run cloud services, search, advertising systems, storage, communications and many other workloads alongside artificial intelligence. Unless an operator publishes a verified allocation by workload, its facility-level electricity emissions cannot be attributed entirely to model training or inference.

The same limitation applies to estimates for an individual prompt, generated image or video render. A credible allocation would need information about the model, hardware, utilization, cooling, regional grid conditions and shared infrastructure. Dividing a corporate emissions total by an estimated volume of outputs creates apparent precision without resolving those variables.

Environmental claims from AI providers are more informative when they disclose several dimensions together: absolute electricity consumption, both versions of Scope 2 emissions, the share attributable to data centers, geographic and temporal matching of clean-energy purchases, and relevant Scope 3 impacts. No single efficiency metric or renewable-power claim substitutes for that inventory.

What remains true

The defensible conclusion is narrower than the original premise but still significant. A historical analysis found a large difference between location-based and market-based electricity emissions for a defined group of company-owned data centers. It did not measure the complete environmental cost of artificial intelligence, establish a current emissions gap or cover the entire data-center industry.

Newer evidence strengthens the reason to examine that distinction rather than extending the old percentage into the present. Physical electricity demand is expected to rise substantially, fossil fuels are projected to supply part of the increase, and the accounting framework remains under revision. The 662% figure should stay attached to its original companies, period and methodology; the transparency problem it exposed is still unresolved.

Also read:

• AI Adoption Leads to Job Growth, Not Mass Layoffs: New Study Finds Companies Hiring More After Implementing AI
• Texas pauses data-center grid approvals as requests hit 474 GW
• Why Was Data Mesh Architecture Introduced In The First Place?