The climate case for AI has not been made

A review of public information found few major power, meat and dairy, or transportation companies are using AI in ways that demonstrate the emissions cuts projected by a study: 3.2–5.4 billion tons of CO₂ annually by 2035. Reported examples include UK grid forecasting that halved solar-forecast error, CATL cutting battery-material screening time by two orders of magnitude, and Didi reducing per-trip emissions by 16%; cultivated-meat venture funding fell from $989 million in 2021 to about $83 million last year. The article also warns that AI-driven data-center emissions could exceed 300 million tons annually by 2035 under the IEA base forecast.

Analysis

The investable signal is not a near-term climate-tech revenue pool; it is a widening gap between AI’s power demand and its still-unproven ability to reduce emissions at scale. For Alphabet (GOOG) and Microsoft (MSFT), that raises a second-order risk: data-center load can increase procurement costs and scrutiny of climate commitments before AI efficiency gains are measurable. The equity impact is more likely to arrive through power availability, cost, and permitting constraints than through direct climate penalties.

Grid forecasting has a credible mechanism, but deployment and grid rules—not model capability—are the bottlenecks. The UK example supports a case for operational tools, not a material earnings conclusion for National Grid (NGG); verify whether savings or avoided curtailment translate into regulated returns. For Kraft Heinz (KHC), the NotCo partnership is product-development optionality, not evidence of scaled demand or a near-term earnings driver. Consumer adoption, unit economics, and repeat purchases remain the key tests.

Over 1–3 months, climate-conference rhetoric is unlikely to change fundamentals absent binding policy or company guidance. Over 6–18 months, watch data-center power contracts, utility capex recovery, and whether AI-driven load accelerates gas generation or delays renewable integration. The contrarian risk to the negative framing is that improved forecasting and dispatch unlock existing renewable capacity faster than new generation can be built; the counter-risk is rebound—efficiency lowers costs and induces more compute, offsetting emissions savings. The thesis weakens if tech firms show falling emissions per unit of compute alongside credible absolute-emissions progress, or if grid operators document scaled, independently verified avoided emissions.

Key Decisions for Investors

• Do not trade the conference narrative alone. Treat GOOG and MSFT climate claims as a monitoring risk, and check quarterly disclosures for absolute emissions, emissions per unit of compute, data-center power sourcing, and contracted electricity costs.
• Use a relative-value watch rather than an outright NGG long: consider exposure only if filings show AI-enabled grid efficiencies are recognized in regulated investment or returns. The UK operating example alone does not establish parent-level earnings materiality.
• Keep KHC’s NotCo collaboration in the optionality bucket; require evidence of scaled launches, repeat purchase, and competitive cost/taste before underwriting earnings or multiple expansion.
• Track utility interconnection queues, renewable curtailment, and new gas build decisions as the practical test of whether AI-related load is enabling or crowding out decarbonization. A rise in load without faster renewable integration would falsify the bullish grid-efficiency pathway.