Backcasting 2035 AI-Induced Emission Reduction Projections (with Claude Code)

by Jeremy Tamanini, Founder, Dual Citizen LLC

In June 2025, a team of researchers from the London School of Economics (Grantham Research Institute) and Systemiq published a novel, peer-reviewed study in Nature. The article “Green and Intelligent: the role of AI in the climate transition” estimates the potential for GHG emission reductions in the power, meat & dairy, and light vehicle sectors (collectively contributing around 50% of global emissions). Their finding is that through efficiency gains, adoption of greener product alternatives, or a combination of both, AI-induced emission reductions across these three sectors could reach between 3.2 – 5.2 GtCO2e in 2035, against the business-as-usual scenario.

Inspired by learning from a Kith Climate workshop this summer, I partnered with Claude Code to perform a backcasting analysis, using the Nature study as an anchor. For each sector, I asked Claude to identify the most significant market actors (in China, the United States and the European Union) required to integrate AI in order to realize the study’s 2035 projections.

For example, in the power sector Claude identified 27 grid operators whose AI integration would be critical to reaching the efficiency gains projected in the study. Similarly, in the Meat & Dairy sector, Claude found the key 14 alternative-protein market actors whose AI-integration could compel consumers to adopt different consumption patterns, thereby reducing sector emissions. Lastly, 8 light vehicle firms were assessed on their current state of AI integration. This analysis, and the study, does not account for any spillover effects from these efficiency and adoption gains (more on that in the Methodology).

This backcasting analysis revealed a troubling trend. Despite the ubiquity of AI promotion on earnings calls and corporate marketing, few market actors analyzed had integrated AI to their business models, based upon publicly available information. Each market actor was ranked by 1 (AI deployed & quantified); 2 (real but partial or adjacent); or 3 (no evidence found, or deprioritized). Across the three sectors, only 6 of the 49 total market actors ranked 1, meaning AI deployed & quantified. Jump to the full sector analysis by clicking below:

Jump to Power – Jump to Alternative Proteins – Jump to Light Vehicles

In September 2026, concrete case studies demonstrating how AI of any kind can reduce emissions at scale remain scarce. The Nature study anchoring this backcasting analysis is one of only a few efforts to model what future AI-induced emission reductions may look like. The value of this backcasting analysis is to demonstrate the current state of AI integration among market actors critical to realizing these 2035 emission reductions from the Nature study.

This illuminates opportunities for stakeholder pressure to accelerate progress: Could shareholder resolutions pressure these companies to hasten AI adoption for efficiency, much like they have around climate change? Could activists engage market actors (as they have so successfully around data center development) to accelerate consumer adoption of alternative proteins? Could the market actors who are integrating AI more quickly today serve as early adopter case studies that diffuse within their sectors? Please use this analysis to stimulate that conversation, and contact me here with your feedback.

Power

Of the 27 largest grid operators assessed for AI integration status (China: 2 of 2 grid companies; USA: 12 of 12 in-scope operators – 7 RTOs/ISOs + 5 non-RTO utilities/federal authorities; EU: 13 of ~26 originally-listed TSOs), only one demonstrated clear evidence of AI integration for grid forecasting/DER management.

Alternative Proteins

Unlike Power, the Meat & Dairy sector’s 2035 emission reductions are driven by adoption whereby AI improves alternative-protein taste/texture and lowers cost, which is modeled to shift consumer demand away from meat/dairy. Also unlike Power, the sector exhibits several strong, commercially deployed AI integrations across the 14 market actors assessed. However, two of the sector’s most prominent, best-funded cultivated-meat companies (Believer Meats, Meatable) ceased operations entirely in December 2025 and several established fermentation leaders (Perfect Day, The Every Company) show no confirmed AI-specific claim despite being well-funded.

Light Vehicles

The Light Road Vehicles sector looks at 4 battery makers and 4 mobility platforms, reflecting the study’s modeling of emission reduction scenarios based on both efficiency from consumers moving to shared mobility platforms and adoption from cheaper batteries with greater charging infrastructure.

Contact Jeremy Tamanini to continue the conversation on this topic, as well as reading these related recent insights from the practice:

Applied AI for Sustainability: Opportunities for Integration (link here)

AI Everywhere: Tangible Applications for Sustainability Teams (link here)

The AI Elephant in the Room (link here)

AI x Sustainability in Trump 2.0 (link here)

Remarks to the National Sustainability Society (link here)

AI in Building & Construction: Tangible Applications for Sustainability Teams (link here)

How to Work with Satellite-Based Sustainability Data (link here)

Methodology

This analysis is a backcasting exercise: instead of cataloguing AI-for-sustainability case studies as they are reported, we started from a credible, peer-reviewed forward-looking emissions-reduction target for 2035, and worked backward to identify which real-world market actors would need to act – and how far they currently are from doing so.

A cataloguing approach (collecting AI case studies as they appear in press releases, industry reports, and platforms like the ITU AI Playbook or IEA Observatory) answers the question ‘what has been publicized?’ It does not answer ‘is this enough, and who specifically needs to move?’ Backcasting from a quantified target reframes the question around accountability and gap identification: for each sector, we identify (a) the specific mechanism a credible model says AI must deliver, (b) the real companies/operators who control that mechanism, and (c) how much public evidence exists that they are actually doing it. This produces a genuinely different kind of output than a case-study list: a gap analysis, sector by sector and actor by actor, that shows where the anchor scenario’s assumptions are – and are not – being borne out today.

Stern et al. (LSE Grantham Research Institute + Systemiq), ‘Green and intelligent: the role of AI in the climate transition,’ npj Climate Action (2025) – chosen because it is peer-reviewed, discloses a full technical annex and methodology, and explicitly nets out AI’s own emissions footprint (estimated 0.4-1.6 GtCO2e) against the reductions it credits AI with enabling. The paper is deliberately narrow: it models only 3 sectors – Power, Meat & Dairy, and Light Road Vehicles – together representing an estimated 3.2-5.4 GtCO2e/year in AI-driven emissions reduction potential by 2035, equivalent to roughly 36% of the gap between a business-as-usual and an ambitious global emissions trajectory.

For each of the paper’s 3 sectors, we: (1) read the paper’s own stated mechanism for how AI achieves that sector’s reduction – not just the headline number; (2) identified the real-world category of actor that mechanism implies (which is not always obvious – see sector notes below); (3) built a roster of those actors, prioritized where the full set was too large to check exhaustively; (4) searched for public evidence of each actor’s AI adoption specifically on that mechanism (general AI use, e.g. chatbots or unrelated automation, was explicitly excluded); and (5) recorded findings with source attribution and an honest confidence/verification tier, so a self-reported vendor claim is never presented with the same weight as a primary operator disclosure or peer-reviewed result.

In every sector, at least one major, credible, well-resourced actor was found to be explicitly moving AWAY from the anchor scenario’s assumptions, not merely slow to adopt: Southern Company (Power) deprioritized AI-for-grid-efficiency for physical buildout; the cultivated-meat industry (Meat & Dairy) saw two prominent companies cease operations entirely; and evidence for the vehicles’ efficiency mechanism had to be sourced from outside academia rather than from the mobility platforms’ own strategic communications. This recurring pattern – genuine headwinds and misalignments, not just an evidence gap – is arguably the most important cross-sector finding of the whole exercise, and a more defensible, differentiated basis for a platform than presenting only encouraging examples.

Roster coverage is not exhaustive in any sector (notably: ~13 lower-priority EU TSOs, Panasonic and Lyft were only lightly checked, and no company beyond the priority lists was assessed). Nearly all ‘deployed’ findings rest on self-reported claims by the organizations involved, not independent audits – the exceptions being the npj paper itself and a small number of peer-reviewed academic studies (e.g., the Didi ridesplitting research). Absence of public evidence for an actor should be read as ‘not found,’ not as ‘confirmed absent’ – this is a materially weaker claim for entities with limited English-language press coverage (e.g., several EU TSOs, US federal power authorities) than for entities that explicitly discuss the topic in their own materials (e.g., CAISO’s own issue papers).

Each of the 3 roster tabs was subsequently extended with two further columns: an AI Integration Rank (1 = most advanced, 3 = least, based on the overall strength of AI-integration evidence found for that actor – a broader test than the strict on-mechanism bar used for the adoption-status column itself) and a Market Volume figure appropriate to the sector (grid operators: annual electricity demand/throughput or peak load; alt-protein companies: annual revenue or nearest available proxy; vehicle-sector actors: annual battery production or gross bookings/ride volume). A subset of these figures were confirmed via primary or named-research-firm sources this session and marked VERIFIED (e.g. CATL/BYD/LG/Panasonic’s 2025 GWh installations, Terna’s 311.3 TWh annual demand, NotCo’s and Perfect Day’s revenue, Uber’s gross bookings, Waymo’s ride volume); the remainder are order-of-magnitude estimates from general industry knowledge, explicitly marked as approximate and not independently verified this session. This distinction should be preserved in any downstream use.

The following is explicitly speculative, not evidence-based like the rest of this analysis – it was not derived from search or from npj’s own modeling (which does not address cross-sector spillovers), but is offered as a reasoned extension of the mechanisms already identified, to help think through what else might follow if the 3 anchor scenarios materialize as modeled by 2035. It should be clearly distinguished from the sourced findings elsewhere in this workbook in any external write-up.

If grid operators achieve the load-factor/forecasting gains npj models, more predictable, higher-utilization renewable supply would likely lower and stabilize wholesale electricity prices – a direct input cost for adjacent sectors. Cheaper, more reliable clean power could accelerate industrial electrification and improve the economics of green hydrogen production (which depends heavily on low-cost, low-variability renewable electricity). It could also ease the EV-charging-infrastructure siting problem for Light Road Vehicles’ adoption sub-mechanism, since a more AI-optimized grid better accommodates distributed EV charging load without new physical buildout.

If AI-driven alternative proteins close the taste/cost gap and reach npj’s adoption range, the most direct spillover would be reduced demand for livestock farming, freeing agricultural land and reducing land-use-change emissions beyond what npj’s own accounting captures. It could also pull investment and AI adoption into adjacent food-tech categories not covered here – e.g. AI-driven crop optimization for the feedstocks (soy, peas, mycelium substrates) alternative-protein production depends on. Conversely, the sector’s documented 2025 investment collapse could have a chilling spillover effect on adjacent climate-tech venture funding categories competing for the same investor base.

AI-driven battery-materials breakthroughs are directly transferable to grid-scale battery storage – itself a component of Power’s own load-factor/forecasting mechanism, since better, cheaper storage makes renewable output easier to dispatch predictably. This creates a plausible reinforcing loop between the Vehicles and Power mechanisms that npj’s sector-siloed modeling does not itself capture. Similarly, if AI-driven shared-mobility efficiency gains scale up, reduced personal-vehicle-ownership patterns could have knock-on effects for automotive manufacturing emissions – a sector entirely outside npj’s 3-sector scope.

In every sector, the strongest-evidenced actor (National Grid ESO/NESO, NotCo, CATL) has a core AI capability that is at least partially transferable to a different sector’s mechanism – suggesting the realistic 2035 outcome may look less like 3 independently-progressing sectors and more like a smaller number of general-purpose AI/forecasting/materials-discovery capabilities diffusing across sector boundaries faster than sector-specific investment trends alone would predict. This is a genuinely speculative extrapolation, offered as a hypothesis worth testing further, not a verified finding.