AI's Hidden Emissions: The Sustainability Puzzle

AI's growing presence in our daily lives comes with a less talked about, yet significant, side effect: emissions. Here's the thing, while companies are keen to showcase their AI-powered tools, there's a blind spot in how these services are accounted for in corporate greenhouse gas (GHG) inventories.

The Methodology Gap

Think of it this way: AI inference services, like API subscriptions and enterprise chat tools, are now officially under scrutiny thanks to the Corporate Sustainability Reporting Directive (CSRD), effective from 2024. However, the problem isn't that companies are ignoring these emissions. It's that there's no standardized way to measure them.

Most current practices either skip this category or use a broad economic input-output (EEIO) factor, which lumps AI with the entire ICT sector. This approach can overstate AI emissions by a staggering 10 to 40 times compared to more precise, physically-derived estimates. If you've ever trained a model, you know the computational demands aren't trivial, but they're not quite the carbon monsters these estimates suggest either.

A Better Blueprint

Enter a new four-tier framework. This approach offers precision matched to the availability of data. It ranges from direct, token-based physical estimation, think GPU energy benchmarks and regional grid carbon intensities, to a fallback EEIO method when usage data is nonexistent. This isn't just a theoretical exercise. Applied to a 200-person firm in Europe, the framework reveals emissions under 1 tCO2e. In essence, it's not the size of the emissions that's the challenge, but the way we measure them.

Now, what about the water-carbon trade-off? This is where it gets intriguing. Sweden, with its hydro-centric grid, ranks low in carbon intensity but high in water footprint. So, companies might need to rethink data center locations. Are the trade-offs worth it? That's a question worth pondering.

Why This Matters

Here's why this matters for everyone, not just researchers. If AI services are to be integrated responsibly into corporate operations, companies need to know what they're dealing with. The analogy I keep coming back to is driving a car without knowing its fuel efficiency. To be truly sustainable, businesses must understand the real impact of their AI tools, both carbon and water.

So, as firms gear up to comply with the CSRD, they face a choice: continue with broad, inaccurate estimates or push for precise, data-backed methods. Betting on the latter couldn't only transform corporate sustainability efforts but also redefine how we perceive AI's environmental footprint.