AI + climate tech: promise, payback, and the procurement gap

• Stern-led, peer-reviewed study estimates AI could abate 3.2–5.4 Gt CO₂e annually by 2035 across power, transport and food — potentially net-positive even after data-centre growth.

• There are real-world wins (e.g., DeepMind’s wind forecasting lifting value c. 20%), but deployment depends on buyers like grid operators — not just Big Tech.

• My bias: the opportunity is investable where there’s measurable system savings and procurement clarity (ESO, TSOs/DSOs, retailers). Risk sits in reproducibility, incentives, and access to operational data.

Context & data

The FT spotlighted research led by Lord Nicholas Stern arguing that targeted AI can cut emissions meaningfully in power, transport (light road vehicles) and food — while warning that lab results often stumble in the field and commercial incentives skew AI toward fossil-fuel optimisation.

• Abatement potential: The Stern team (npj Climate Action) estimates 3.2–5.4 Gt CO₂e/yr abatement by 2035 from AI applications in power, mobility and food. They argue these savings exceed AI/data-centre emissions growth if we scale the right use cases with policy support. (23 June 2025).

• Power-sector proof point: DeepMind reports its ML forecasting has boosted the value of wind energy by ~20% vs. no time-based commitments, by improving day-ahead predictability. (26 Feb 2019).

• UK deployment signal: Britain’s system operator (now NESO) ran a Solar NowCasting project with Open Climate Fix (OCF) to increase the amount of solar the grid can handle — i.e., lower balancing costs and emissions. (19 Jul 2024).

My take (PE lens)

I’m cautiously optimistic — with caveats on who pays and unit economics. Where a TSO/DSO or retailer bears balancing costs, AI forecasts and optimisation can have crisp ROIs: fewer reserve procurements, tighter intraday positions, better curtailment management. That supports ARR-style contracts priced off measured system savings (e.g., % of avoided balancing costs) rather than vague “AI licences”. Vendors with live integrations, high GM (software-heavy) and short payback can scale.

Moat is less algorithms, more data access + deployment muscle: signed data-sharing with grid operators, robust MLOps, and credible SLA/SOC2. I’d underwrite businesses showing (i) verifiable uplift vs. incumbent forecasts, (ii) recurring revenues with grid/retailer customers, and (iii) pathways into adjacent optimisation (storage dispatch, EV flexibility). The Stern paper’s “active state” point matters: without procurement reforms and outcome-based tenders, the best lab models won’t cross the chasm.

Risks & watch-list

• Reproducibility & scientific rigour: recent criticism of AI-discovered carbon-removal materials shows how lab claims can unravel on replication — watch for independent validation before underwriting climate impact. (July 2025).

• Incentive mismatch: AI also improves fossil-fuel operations; capital and talent may chase higher-margin O&G workflows over public-good grid use cases. (Theme highlighted by the FT column).

• Data and IP access: grid/commercial data sharing, model IP ownership and vendor lock-in can stall deployments — structure contracts carefully (data rights, auditability).

• Operationalisation risk: moving from pilots to 24/7 control-room use requires reliability, security, and governance; buyers need clear KPIs (forecast error, avoided curtailment, reserve costs).