AI carbon transparency gap puts Scope 3 reporting under pressure

Anthropic's missing emissions data forces businesses onto spend-based proxies, exposing a structural gap in corporate AI sustainability reporting.

By Disrupts Editorial Team Wed, 1 July 2026

A brightly lit data center aisle features rows of black server racks with blinking blue indicator lights, connected by overhead cable trays filled with colorful data and power cables.

Anthropic, the San Francisco-based AI safety company behind the Claude family of models, has yet to publish a full corporate sustainability report with audited Scope 1, 2, and 3 emissions data. As businesses embed Claude and competing tools into core workflows, that silence is creating a measurable problem for corporate carbon accountants and sustainability officers trying to meet net-zero commitments.

The issue was surfaced in a research note by UK sustainability consultancy Tunley Environmental, authored by Dr Aaron Yeardley and Ellis Clark. Their analysis does not argue that Claude is inherently more carbon-intensive than Microsoft Copilot or Google Gemini. It argues, more pointedly, that the absence of company-specific data forces organisations into a less precise methodology, and that in a world of tightening Scope 3 disclosure requirements, imprecision carries regulatory and reputational risk.

The numbers behind the transparency deficit

Tunley's worked example is instructive. A business spending £10,000 per year on Anthropic's platform, unable to obtain supplier-specific emissions data, would apply a spend-based proxy of approximately 0.1177 kg CO₂e per pound spent. That produces an annual estimate of roughly 1.18 tonnes CO₂e and a nominal offset cost of around £17.66 at £15 per tonne.

The same spend on Microsoft Copilot, where organisation-level emissions intensity data is publicly available, yields an estimate closer to 0.63 tonnes CO₂e. For Google Gemini, the figure is approximately 0.51 tonnes. The financial difference in offset costs is small: roughly £8 to £10 per year. But the methodological gap matters disproportionately when businesses are required to demonstrate data quality and consistency in sustainability disclosures, not merely report a number.

On a per-query basis, the picture is more nuanced. Tunley cites indicative 2025-2026 estimates of around 0.03 g CO₂e per text query for Google Gemini, 0.13-0.19 g for GPT-4o, and 0.20-0.40 g for Claude standard models. For moderate business use, even tens of thousands of monthly prompts may sum to only a few kilograms of CO₂e annually. At that scale, AI is a marginal emissions line item, well below transport, energy or supply chain activity.

Scale and the second-order problem

The consultancy's more consequential warning concerns trajectory rather than current footprint. AI workloads are not static. Agentic applications, large-scale automation, multi-modal generation and complex reasoning tasks consume substantially more compute than simple text prompts. As enterprises move from experimental chatbot deployments to production-grade AI agents running continuous background tasks, the electricity demand profile changes materially.

This is where the convergence between AI adoption and energy infrastructure becomes strategically significant. Data centre power draw is already a live issue for grid operators across the UK, US and GCC. Hyperscalers and AI labs are competing aggressively for renewable energy power-purchase agreements and nuclear offtake contracts. A company that cannot report its AI supplier's emissions intensity with any granularity has limited visibility into how its Scope 3 footprint will evolve as its own AI consumption scales.

For cross-sector investors and CFOs, the read-across extends to capital allocation. The pressure on AI providers to publish credible sustainability data is unlikely to ease: the EU's Corporate Sustainability Reporting Directive (CSRD) and the UK's growing alignment with ISSB climate standards will progressively tighten what counts as adequate supplier disclosure. AI vendors that cannot satisfy that demand face either reputational friction or exclusion from procurement lists at larger, compliance-heavy enterprise clients.

Anthropic's position as a private company gives it legal latitude that Microsoft and Alphabet do not enjoy, but that latitude is shrinking as AI services become embedded in the supply chains of listed, regulated, and publicly accountable organisations. The transparency gap Tunley identifies today is a compliance gap in formation. Businesses adopting AI at scale would be prudent to build supplier emissions disclosure into their procurement criteria now, before regulators formalise the requirement.