As artificial intelligence becomes a core part of business operations, its environmental footprint is attracting growing attention. With AI adoption accelerating across industries, businesses face a new responsibility: reducing the environmental harm of the models they use. Practices such as selecting more energy-efficient architectures, optimising deployment, and demanding better supplier transparency can lower impact.
Not all providers are equally transparent about their impact, which makes it difficult for organisations to make well-informed choices. In some cases, that’s less a matter of AI companies hiding wrongdoing than of the current rules simply not asking them to disclose it.
The Transparency Gap
Published evidence supports caution around per-prompt figures. Google reports that a median Gemini Apps text prompt uses 0.24 Wh of electricity, emits 0.03 gCO₂e, and consumes 0.26 mL of water. However, the methodology excludes some lifecycle categories such as model training, user devices, external network energy, and data storage. These exclusions underline a wider problem: the sector cannot credibly claim environmental sustainability without consistent full-lifecycle measurement.
The gap becomes clearer once you look at what AI companies are required to disclose. Anthropic, for example, publishes a genuinely detailed Transparency Hub, including dedicated reports under the EU’s Digital Services Act (DSA) covering content moderation on Claude.ai granular enough to break down dozens of user reports by category. But the DSA governs illegal content and platform accountability, not carbon or water. A company can be fully compliant with it while disclosing almost nothing verifiable about its environmental footprint, because no current law requires that of it. Sustainability reporting and platform-safety reporting are simply different regimes, and right now the former is much less mature than the latter.
EU AI Act
Regulatory expectations are moving, but unevenly. The EU AI Act requires providers of general-purpose AI (GPAI) models to maintain technical documentation. The European Commission consultation notes that known or estimated energy consumption is part of those transparency obligations. However, the obligation is narrower than it sounds, and the Act itself draws the line explicitly.
Under Article 53, GPAI providers owe two different documentation packages to two different audiences: Annex XI, the technical documentation reserved for the EU’s AI Office and national authorities, does include “known or estimated energy consumption of the model” as a required field. Annex XII, the separate documentation providers must give to downstream developers integrating the model into their own products, has no energy or environmental field at all, it covers architecture, parameter count, input/output formats, and training-data provenance, but stops there.
In practice, this means a developer building an application on top of a GPAI model has no legal entitlement to the energy figure the provider is required to hold on file for regulators. Despite the EU AI Act’s GPAI transparency framework, downstream AI providers still lack the detailed environmental data needed to calculate the carbon footprint of the products they build.
CSRD
Meanwhile the Corporate Sustainability Reporting Directive (CSRD), the law that was meant to force large companies to report environmental impact has been significantly narrowed. The waves of companies that would have captured most large tech firms have had their start dates pushed from 2026 to 2028, with thresholds relaxed; only the largest non-EU parents generating over €450 million in EU turnover remain clearly in scope in the near term.
In practice, this means audited, comparable environmental disclosure for AI companies is still several years away for most of the industry, and what fills the gap in the meantime is a patchwork of voluntary blog posts, academic lifecycle studies, and estimates that all use different assumptions.
Jevons Paradox
All of this raises an uncomfortable question: if per-query efficiency keeps improving this fast, why isn’t AI’s total environmental footprint falling?
In the 1860s, the economist William Stanley Jevons noted that making steam engines more coal-efficient didn’t reduce Britain’s coal consumption. It increased it, because efficiency made coal power cheaper and more attractive to use, and total demand grew faster than the savings per unit. The same applies to AI. Even if each query becomes 20× more efficient, if AI usage grows 100×, not an unreasonable prediction given current adoption curves for chatbots, coding assistants, and agentic tools, total electricity demand still rises.
For businesses, the practical implication is that “our AI usage got more efficient” and “our AI usage’s total footprint fell” are two different claims, and only the second one should be treated as evidence of genuine sustainability progress.
Models Reducing Impact, Not Solving Sustainability
Some challengers are publishing more credible environmental data or adopting lower-compute model designs. This is progress, but it should be framed as impact reduction rather than proof that AI models are environmentally sustainable.
- Mistral AI has gone further than most by releasing a lifecycle assessment of Mistral Large 2, reporting 20.4 ktCO₂e, 281,000 m³ of water consumed, and 660 kg Sb eq after 18 months of use, plus marginal inference impacts of 1.14 gCO₂e and 45 mL of water for a 400-token Le Chat response. Notably, hardware manufacturing and end-of-life accounted for 61% of the model’s total materials footprint despite being a small share of its emissions, a reminder that lifecycle assessments catch embedded hardware impact that simple energy-per-query figures miss entirely.
- Microsoft’s BitNet research points to a promising route for lower-power AI through 1.58-bit model weights and CPU-friendly inference. The research literature describes substantial efficiency gains, while wider enterprise impact will depend on workload fit, deployment maturity, and hardware support.
- Hugging Face is democratising AI with open-source models and tools, while also providing transparency and encouraging efficiency-first designs across its community-driven platform. Open-weight models also carry a subtler sustainability benefit: when a developer fine-tunes or distils an existing open model rather than training a new one from scratch, the embodied energy cost of that original training run gets repaid across far more downstream use.
- A 2025 academic study, Small is Sufficient, put a number on what right-sizing could achieve across the industry rather than one model at a time. Modelling real-world AI task distributions, the researchers estimated that simply routing each task to the smallest adequate model, rather than defaulting to the largest, could cut global AI energy consumption by 27.8%, saving roughly 31.9 TWh in 2025 alone, equivalent to the annual output of five nuclear reactors. It’s a reminder that model selection, not just model design, is an underused lever.
Comparing Chatbot Options: LLaMA vs BitNet vs MoE
For companies evaluating chatbots, the sustainability trade-offs become clearer when comparing these different approaches:
LLaMA. Open-weight LLaMA models can be attractive where organisations need control, customisation, or on-premises deployment, but per-query energy depends heavily on model size, hardware, quantisation, batching, and context length. Rather than relying on a single universal watt-hour figure, companies should benchmark representative prompts on their intended infrastructure. Worth flagging: Llama 4 (Scout and Maverick, released April 2025) is, as of mid-2026, still the newest open-weight Llama, Meta’s newest frontier model, Muse Spark, launched closed-weight in April 2026 under its new Meta Superintelligence Labs division, reversing four generations of Meta handing over the weights.
Organisations betting on “the next open Llama” for sustainability or cost reasons should treat that as an assumption to revisit, not a given. It’s also worth noting that Llama 4 itself runs on a Mixture-of-Experts architecture, so it isn’t strictly a separate category from the MoE approach below, it’s an example of it.
BitNet. Microsoft’s BitNet introduces a more energy-efficient architecture by employing 1.58-bit weights in its computations. Microsoft’s research reports improved latency, memory, throughput, and energy characteristics compared with full-precision transformers of similar scale, and the official BitNet framework now reports energy reductions of up to 82.2% on x86 CPUs and 70.0% on ARM. However, results depend on model size, implementation, and hardware support. Trade-offs include:
Accuracy may be slightly lower on tasks requiring fine-grained language understanding or numerical reasoning.
Its ecosystem is smaller, with fewer pre-trained models and community resources.
Hardware and framework compatibility can affect energy savings.
BitNet is more mature than it was: Microsoft has shipped an official 2B-parameter model on Hugging Face, released GPU inference kernels, and continued regular framework updates into 2026. It’s still a narrower ecosystem than mainstream open-weight models, and best suited to CPU-constrained or edge deployments rather than a drop-in replacement for very large or highly specialised models.
Mixture of Experts (MoE). MoE models activate only a subset of parameters during inference, improving compute efficiency without activating the full model for every token. The approach was proven early by Google’s GLaM paper (2021), which found its largest MoE model used about one-third of GPT-3’s training energy and about half the inference FLOPs while achieving strong benchmark results. It’s since gone from research curiosity to industry default: Llama 4, DeepSeek V3, Mixtral, and Gemini all use MoE architectures. So it’s now the load-bearing design choice behind most frontier-scale models shipping in 2026.
For sustainability alone, BitNet is highly attractive, and MoE-based models also offer significant efficiency gains over dense architectures. However, practical considerations such as performance, ecosystem maturity, licensing (open vs. closed-weight), and hardware compatibility will often lead organisations toward a mainstream MoE model like Llama 4 or a competitor over a narrower framework like BitNet.
Why It Matters for Business
AI adoption is no longer only about what a model can do, it is also about the environmental cost of doing it. For businesses committed to ESG goals, model choice directly affects their carbon footprint. Smaller models, quantised architectures, transparent vendors, and efficient infrastructure can reduce environmental impact, but they do not remove it, and given Jevons Paradox, rising usage can outpace those gains at the organisational level too. Businesses should therefore treat AI sustainability as a risk-management and reduction challenge, not as a solved problem. Practical considerations include:
- Right-sizing models: Choosing a smaller or task-specific model can reduce emissions by up to 90% while still meeting business needs.
- Vendor selection: Working with providers who disclose lifecycle data, not just per-query efficiency claims, offers more confidence and accountability, precisely because per-query figures alone can mask rising total footprint.
- Operational efficiency: Practices like batching queries, caching frequent responses, or scheduling compute during low-carbon grid times can further cut emissions.
- Hardware and infrastructure choices: Cloud infrastructure choices matter. Providers’ renewable-energy procurement, data-centre efficiency, and regional electricity mix can all affect reported emissions, so sustainability claims should be checked against the provider’s methodology and the workload’s actual deployment region.
- Open-source innovation: Platforms like Hugging Face encourage efficient, reusable models that reduce redundant compute.
Looking Ahead at Regulation
Governments are beginning to recognise the hidden costs of AI, but the regulatory picture is still mixed, and it moved again in July 2026 in ways that mostly confirm a “narrowing, not disappearing” trend.
The EU AI Act requires transparency from providers of general-purpose AI models, and the CSRD was meant to complement this with full corporate sustainability reporting. But in July, the Commission further narrowed the requirements, adopting revised ESRS standards that cut mandatory data points by over 60% and the standard for non-EU parent companies above €450 million in EU turnover. And existing regimes AI companies already comply with, like the DSA, simply weren’t built to capture environmental impact at all.
At the same time, consumer expectations are shifting. As climate awareness grows, end-users increasingly want to know whether the digital services they use align with sustainable practices.
Key Takeaways
- Big tech isn’t sharing enough: The largest LLM providers don’t publish full lifecycle emissions data, and current digital-regulation don’t require them to.
- Efficiency gains are real but don’t equal sustainability: Studies now show per-query energy use has fallen 4–20× versus earlier estimates, with further gains expected. But a 20× efficiency gain paired with 100× usage growth still means total demand rises, so efficiency reporting alone cannot answer whether AI’s overall footprint is shrinking.
- Transparency is becoming a competitive advantage: Pioneers like Mistral AI and open-source communities are setting a higher bar for accountability, but standardised comparisons across vendors won’t emerge.
- Efficiency is achievable at the design level too: approaches such as quantisation, smaller task-specific models, and MoE architectures show that lower-compute AI can still provide strong performance when matched to the right use case, cutting energy use by up to 90% in some cases.
- Sustainability claims need caution: current AI models still consume enormous amounts of electricity, water, and hardware resources, so businesses should focus on reducing impact rather than claiming AI is environmentally sustainable.
- Adaptation will be rewarded: early movers that prioritise transparency, right-sized models, and lower-carbon infrastructure are better positioned to manage regulatory, reputational, and environmental risks.
