AI for Investors Has to Be Sustainable

Should the AI that investors use be sustainable? Yes. For investors this is a question of portfolio risk, responsible investment and governance, not just values.

Investors meet AI twice. As owners, they hold the companies building it and carry the risks of that build-out. As users, they are putting AI inside research, screening and reporting. On both sides, investors need evidence: what environmental impacts AI creates, how reliably it performs, and how its risks are governed.

As owners: AI is now a portfolio risk

The AI build-out is one of the largest capital programmes in the market. The IEA puts capital spending by the five largest technology firms above US$400 billion in 2025, and expects it to rise by about 75% in 2026.

That spending is arriving in emissions. Google's reported emissions remain substantially above its 2019 baseline, while Microsoft's total reported Scope 1, 2 and 3 emissions rose about 25% year on year, from 16.2 to 20.29 million tonnes CO₂e in its latest report. Both companies have 2030 climate commitments, raising questions about how their AI infrastructure expansion will affect progress toward those targets.

Shareholders are already asking about it. In 2026, As You Sow and other investors filed proposals at Amazon, Meta and Alphabet on how AI's electricity demand fits their climate commitments. Per Glass Lewis, support reached 18.4% at Amazon, 7.4% at Alphabet and 6.9% at Meta.

Support varied across the three companies, but the proposals show that some investors are explicitly questioning how AI infrastructure expansion affects climate commitments. An investor who asks Microsoft how its AI squares with its climate plan will be asked the same about the AI in their own process.

As users: the AI in your process is part of your process

An investment decision is only as defensible as the evidence under it. When AI reads the filings, scores the company or drafts the assessment, its errors become part of the decision record.

The errors are measurable. One study reports that in a traceability audit of ten extracted claims, eight verified against their source documents, with the two failures attributed to leakage from examples used in the system's prompting. Ten claims is too small a sample to read as an 80% accuracy rate. It shows why extraction accuracy and source verification have to be evaluated rather than assumed, and why outputs still need validation.

There is a second failure that accuracy scores miss. Research at ACL 2025 found that language models tend to repeat exaggerated sustainability claims as if they were performance. A tool can extract a disclosure perfectly and still pass greenwashing straight into a portfolio.

Traceability is one safeguard among several. Tying every answer to its source passage lets a reviewer check it, but does not make the answer correct, so accuracy testing, expert review and ongoing monitoring are needed alongside it. Regulators are moving toward the same layered approach: OSFI's Guideline E-23 takes effect on 1 May 2027 and brings AI models, including vendor models, under model risk management.

The AI you buy sits in your own footprint

AI services can contribute to an organization's Scope 3 greenhouse gas emissions, depending on how the services are purchased and accounted for. For investors and other organizations reporting under applicable sustainability disclosure requirements, the challenge is to understand whether these emissions are material and how they can be measured consistently.

A separate footprint for each AI assessment is not universally required. However, reliable information about the environmental impact of AI services helps organizations improve their emissions estimates, assess vendors and understand the consequences of increasing AI use.

That number is harder to get than it should be. Google reports company-measured figures for a median Gemini Apps text prompt: 0.24 Wh of energy, 0.03 g CO₂e and 0.26 mL of water, under its published methodology. Mistral's lifecycle assessment is a modelled estimate: 1.14 g CO₂e and 45 mL of water for a 400-token response. The two use different boundaries, models and workloads, so the figures cannot be compared directly.

Accounting methods also matter. A September 2024 Guardian investigation estimated that the location-based emissions of in-house data centres run by four large technology companies were about 7.62 times their officially reported figures for 2020–2022. Location-based accounting counts the electricity actually drawn from the grid, while market-based reporting lets renewable energy certificates offset it.

The footprint is also a design choice. Research on AI model energy use found substantial differences in energy use across models and tasks, pointing to efficiency gains from smaller models where they suit the application. The practical lesson is to choose a model that meets the task's requirements without using more compute than necessary. An investor can ask a vendor which models it uses, and why.

Eight questions to ask any AI vendor

Investors already know how to do this. It is the due diligence they run on portfolio companies, pointed at their own tools. Sustainable AI is AI you can trust and afford to run, so the questions cover reliability, footprint and accountability, and data protection.

A vendor that cannot answer these is asking the investor to carry the risk instead.

The same standard, turned inward

Investors ask companies to measure their impact, disclose it and show the evidence. Sustainable AI asks the same of the tools investors use to make those judgments.

We think this is where AI for investors is heading. Better analysis will matter, and so will being able to show where each conclusion came from and what it cost to produce. See how Lucid Axon works for investors.

If you want to know more about governance in sustainable finance, read our companion article, Who Is Accountable for the AI?