1.8 Gt power-sector AI case; one cited program in the hundreds of kilotonnes

On 9 November 2026, MIT Technology Review reported an AI-assisted review of public records from large actors in power, meat and dairy, and vehicles. Few were pursuing the emissions routes in a peer-reviewed Grantham Research Institute and Systemiq study. That study put AI-enabled cuts at 3.2 to 5.4 billion tonnes of CO2-equivalent a year by 2035, after subtracting AI’s own energy use. The power slice alone was up to 1.8 billion tonnes, from lifting renewable delivery from already-built wind and solar by as much as 20 percent.

Reviewers found one working program among operators checked in China, the United States, and Europe: National Grid ESO with Open Climate Fix. The review credits it with halving solar-forecast error and avoiding hundreds of thousands of tonnes of CO2 a year, three orders of magnitude under the 1.8 billion-tonne power case. NotCo and Kraft Heinz’s Giuseppe venture formulates plant proteins; it has not published herd or land-use cuts. IEA’s Energy and AI report put data-centre electricity near 415 TWh in 2024 and near 945 TWh by 2030 in the base case.

The missed variable is the inventory, not the model. Sam Altman’s 23 September 2024 essay, The Intelligence Age, named fixing the climate as an expected outcome. A 2023 op-ed by Google’s head of sustainability put a related claim at a 5 to 10 percent global cut by 2030. Those are scenarios. Google’s 2024 Environmental Report logged 14.3 million tCO2e in 2023, up 48 percent from 2019. Microsoft reported emissions about 29 percent above its 2020 baseline as data-centre construction expanded. Abatement stayed prospective while measured corporate totals rose.

Antalya’s host call, that AI be designed, powered, and deployed for climate action, does not create a test. A test is published forecast-error series, metered curtailment avoided, and third-party tonnes, by operator, against a pre-AI baseline. Until those series are filed, the 3.2 to 5.4 billion-tonne range is an upper bound on technical potential, not an expected value. Vehicle and protein pathways in the same frame have thinner public result sets. Deployment is the measurement.