Is the future of AI green? What can innovation diffusion models say about generative AI's environmental impact?
This paper applies the classic Abernathy-Utterback (A-U) innovation diffusion model to generative AI's environmental impact. The authors argue that alarmist predictions about GAI's carbon footprint often ignore how innovation diffusion drives process optimization and efficiency gains. They forecast that the GAI industry is transitioning from the 'fluid' A-U:1 phase to the 'transitional' A-U:2 phase, where dominant designs will emerge. The paper predicts two main business models: large generalist platforms serving mass audiences, and smaller specialized models targeting specific use cases. Their core argument is that GAI 'will never be green, but its impact may not be as problematic as is sometimes claimed' depending on which business model dominates.
The paper offers a valuable reframing of the GAI environmental debate through an established industrial economics lens, but suffers from speculative leaps in its financial modeling and an overly optimistic view of open-source solutions. The A-U model application is legitimate and the efficiency trends cited are empirically grounded. However, the claim that GAI will reach A-U:3 'specific phase' commoditization seems premature given that even A-U:2 dominant designs haven't stabilized. The simulation of OpenAI's break-even at 'over 900 million WAU' relies on unverified assumptions about conversion rates and costs that the authors themselves describe as 'simplified'. The paper acknowledges this limitation but still uses it to draw conclusions about platform viability.
The paper's strongest contribution is connecting established innovation economics to GAI environmental analysis—a genuinely novel framing. The efficiency gains cited are empirically supported: the Ho et al. finding that compute requirements halve every 8 months is accurately cited, and the MoE efficiency claims about DeepSeek are consistent with contemporaneous reports. The distinction between training (20-40% of energy) and inference (60-70%) is crucial and often overlooked. The classification of two business models—mass platforms versus targeted SLMs—is a useful analytical framework grounded in software economics literature. The paper correctly identifies the Jevons paradox risk: 'if each model requires less training, more models will be needed' could offset efficiency gains.
Several foundational premises deserve scrutiny. The claim that '400 LLMs' were evaluated in Ho et al. is incorrect—the paper actually studied 'over 200' models. More seriously, the authors' economic model assumes inference costs decrease with user volume, yet Table 2 explicitly notes 'The inference cost is probably decreasing slightly with the number of users' without empirical basis. The 5% conversion rate assumption for OpenAI is speculative; the source is cited as blog posts rather than audited financial statements. The authors advocate for FOSS models like LUCIE-7B as environmental solutions but provide no quantitative comparison of training energy versus proprietary alternatives, nor do they address whether shared infrastructure amortization truly reduces carbon intensity per unit of capability. The DeepSeek efficiency gains are promising but geopolitically contingent; treating them as inevitable market trends ignores that such efficiencies may be strategically suppressed by dominant players seeking competitive moats. The optimistic forecast of 'a limited number of FOSS foundation models' appears to underweight the fragmentation incentives in open ecosystems.
The evidence for rapid efficiency improvements is well-founded—the 8-month compute halving claim comes from a rigorous analysis on established benchmarks (Wikitext, Penn Treebank) with proper confidence intervals. The Lu et al. survey of 70+ SLMs supports claims about their task-specific competitiveness. However, comparisons to mature technologies like Google Search are rhetorically effective but analytically weak; the paper correctly questions such comparisons but then proceeds to make similar apples-to-oranges assumptions about future GAI efficiency. The comparison to other industries undergoing A-U transitions could be stronger—there's no discussion of whether software actually follows the same commoditization patterns as physical goods, where manufacturing standardization is qualitatively different from model inference. The paper acknowledges but downplays the 'rebound effect' literature, which is robust in environmental economics and directly challenges the optimistic trajectory assumed.
Reproducing this work is problematic. While the A-U framework itself is standard, the OpenAI financial simulation relies on multiple blog sources (Nickie Louise, George Hammond/Cristina Criddle) that lack methodological transparency and may not be contemporaneous. The exact derivation of the '$32.5 + $10' cost figures in Table 2 is not shown—readers cannot verify how fixed costs were allocated. Dates of data retrieval ('retrieved in October 2025') postdate the paper's March 2026 submission, suggesting inconsistent documentation. The Google '44x reduction' claim cites an IEA report but isn't directly verifiable from the provided citation. Crucially, there is no code, dataset, or supplementary materials provided that would allow independent reproduction of the break-even analysis. The IEA energy breakdown percentages (10% development, 20-40% training, 60-70% inference) are sourced to agency estimates but the underlying measurement methodology isn't critiqued or made verifiable.
The rise of generative artificial intelligence (GAI) has led to alarming predictions about its environmental impact. However, these predictions often overlook the fact that the diffusion of innovation is accompanied by the evolution of products and the optimization of their performance, primarily for economic reasons. This can also reduce their environmental impact. By analyzing the GAI ecosystem using the classic A-U innovation diffusion model, we can forecast this industry's structure and how its environmental impact will evolve. While GAI will never be green, its impact may not be as problematic as is sometimes claimed. However, this depends on which business model becomes dominant.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.