The environmental cost of datacentres is rising. Is it time to quit AI?

Datacentre power demand is growing ~4x faster than other sectors globally and is on track to exceed Japan's electricity use by 2030; in Australia datacentre energy demand is expected to triple within five years and surpass electricity used by the EV fleet by 2030. One study estimates AI's 2025 footprint at 32.6–79.7m tonnes CO2 and 312.5–764.6bn litres of water, prompting calls from coalitions (Clean Energy Council, ACF, unions) for binding public-interest principles requiring on-site renewables and water recycling. The article highlights rising local community pushback and an opt-out/ boycott movement (QuitGPT), indicating growing reputational and regulatory risk for big tech and operators of large datacentres.

Analysis

Hyperscalers will face a two-front margin squeeze: rising embedded datacentre costs (capex for renewables/water recycling and higher PPA prices) and reputational/regulatory liabilities that can slow deployments. Expect permitting and community pushback to add 6–24 months to new campus builds in water- or grid-constrained regions, increasing module-level build costs by an estimated 10–20% versus current internal models. That timing friction creates an asymmetric window where demand for third-party edge/offline models and on-device inference could meaningfully accelerate, benefiting firms that can productize lower-power LLMs.

Second-order supply-chain effects matter: pressure to decarbonize will shift OEM demand from general-purpose GPUs to more power-efficient accelerators and liquid-cooled racks, concentrating supplier power in a handful of chip and cooling vendors and pushing lead-times out by 3–9 months. That concentration raises tactical risks for companies that haven’t pre-paid or secured multi-year supply — expect episodic service slowdowns and higher spot pricing for cloud instances during peak new-model rollouts. Meanwhile, utilities in tight markets gain bargaining power to extract capacity payments or renegotiate industrial tariffs, creating an earnings tailwind for regulated utilities and a cost headwind for cloud-native services.

Regulatory catalysts are the near-term trigger: public-interest datacentre rules (renewables + water recycling mandates) or mandatory AI energy disclosures in 12–24 months would force accelerated capital deployment or slower feature rollouts. The consensus underprices the flexibility of hyperscalers to squeeze model efficiency (smaller models, batching, sparsity) — that provides an upside hedge, but not enough to offset a structural uplift in infrastructure opex if local rules mandate embedded renewables at build time.

Key Decisions for Investors

• Pair trade (6–12 months): Short META vs Long MSFT, equal dollar exposure. Rationale: META is most exposed to consumer generative workloads and local permitting backlash; MSFT's enterprise contracts, Azure sustainability offerings and ability to monetize on-prem/edge AI provide better pass-through. Target 20–30% relative outperformance; stop-loss if pair underperforms by 12%.
• Long GOOG (12–24 months) via LEAPS call spread (buy Jan-2027 calls, sell further OTM Jan-2027 calls) to cap premium. Rationale: Google’s TPU + model-efficiency roadmap and Search monetization create convex upside if efficiency wins out; limited cost if regulatory/energy headwinds persist. Risk: premium loss; reward: 2–3x on strong AI monetization scenarios.
• Tactical hedge (3–9 months): Buy out-of-the-money puts on META expiring 6–9 months if local datacentre moratoria or Australia/EU regulatory announcements occur. Rationale: quickest path to re-rating is a headline/regulatory shock that directly impacts build timelines and capex; put premium is insurance against a ~25% downside shock.
• Event-driven idea: Monitor Australian and EU rule announcements (12–24 month window). If mandates require embedded renewables at build, consider short exposure to hyperscalers' capital-intensive projects and rotate into utilities/renewable contractors; take profits or reverse within 6 months of rule publication.