Forget the standoff over green data centres. The path to sustainable AI lies in corporate discipline and recognising that the cheapest choice and the greenest choice are exactly the same, says Paulo Ferreira, Chief Technology Officer at KYND
The debate about AI’s environmental impact has settled into a familiar standoff. On one side, alarming numbers about the electricity and water that data centres consume. On the other, reassurance that innovation will eventually make all of it cleaner. Both are partly right, and neither helps a business decide what to do on Monday morning.
The more useful question is quieter, and it applies to the whole length of the chain. Never mind the net-zero pledge or which provider has the greenest credentials. Are we using the technology efficiently at all, both where the demand is created and where it is met?
Efficiency is the rare spot where the commercial and the environmental cases start agreeing. Get the same result from less compute, and you spend less while removing demand which someone, somewhere, has to build a data centre to serve. The green choice and the cheap choice turn out, conveniently, to be the same choice.
Cost is easy to miss, because AI models hide it so well
Most companies will never train a frontier model. Their footprint comes from how they consume AI: bloated prompts, agents left running unattended, and the steady habit of confusing how much AI you use with how much value it creates. The cost is easy to miss, because models hide it so well, until something that served a handful of users at almost no cost lands in a real workflow and the monthly bill stops resembling the demo.
And it is rarely only the bill that suffers. Feed a model more context than the task needs and the output often gets worse, not better, as the signal drowns in the noise. The waste shows up twice, in cost and in quality.
The ‘tokenmaxxing’ myth
There is a fashionable piece of advice in AI circles: embrace the exponentials. The trouble is that the world the technology runs in is stubbornly not exponential. Electricity, water and capital are finite, and so is the patience of the people meant to absorb all this cleverness. A system that burns ten times the compute should be worth ten times more, not merely look better in a demo. Otherwise “embracing the exponentials” is just a polished way of describing waste at scale.
We have, if anything, got rather good at dressing that waste up as a virtue. The habit even has its own slightly ridiculous name now: tokenmaxxing, judging an engineer’s productivity by the size of their model bill. Nvidia CEO Jensen Huang has said he would be alarmed if an engineer on half a million dollars a year wasn’t getting through at least a quarter of a million dollars’ worth of tokens. It is the old lines-of-code metric in fancy dress. The question was never how much AI you used, but what it produced.
All that unnecessary demand lands on infrastructure already under strain
The better systems are usually the restrained ones. They reach for a model where genuine language, ambiguity or judgement is in play, and fall back on ordinary software when a rule or a workflow would do the job more cheaply. The opposite is already everywhere: a model asked to make a clean, deterministic decision a few lines of code would handle better, or three agents made to “collaborate” on what one well-scoped request would have settled.
All that unnecessary demand has to land somewhere physical, and it lands on infrastructure already under strain. Data centres drink enormous quantities of water and lean hard on local grids; by Bloomberg’s estimate, two-thirds of new US data centres built or in development since 2022 are in areas already short of water, and in some markets the cost is already turning up on other people’s energy bills. Providers carry real responsibility for how their sites are cooled and powered. But they will not move on virtue alone. They will be pushed, by cost, regulation and a public mood that is turning faster than the industry expects, towards less thirsty cooling and towards making their own power, which is why the hyperscalers are suddenly so taken with nuclear.
Learning to be disciplined
None of this requires turning every CTO into a carbon accountant. It asks only that AI gets the same discipline we already apply to cloud spend and reliability: be clear about the problem before reaching for a model, choose the least complex thing that solves it, and measure the result. For agentic systems, boundaries matter as much as capability. The idea that autonomy is inherently sophisticated is a trap; the smarter design is the one that knows when not to make the call at all.
For agentic systems, boundaries matter as much as capability
This is the uncomfortable truth beneath ESG. Businesses rarely act on sustainability out of pure virtue; they act when the incentives line up, and with AI they are starting to. Reputation, regulation and cost are beginning to point the same way, at both ends of the chain. So sustainable AI will not arrive because the technology quietly greens itself. It will come from businesses being more disciplined about when and how they use it, and from providers being forced to build the infrastructure behind it more responsibly. AI may well be exponential. The world it runs in is not. The organisations that come out of this looking clever will be those that used AI where it earned its place, gave it only what it needed, and never mistook burning more of it for doing something new.
Forget the standoff over green data centres. The path to sustainable AI lies in corporate discipline and recognising that the cheapest choice and the greenest choice are exactly the same, says Paulo Ferreira, Chief Technology Officer at KYND
The debate about AI’s environmental impact has settled into a familiar standoff. On one side, alarming numbers about the electricity and water that data centres consume. On the other, reassurance that innovation will eventually make all of it cleaner. Both are partly right, and neither helps a business decide what to do on Monday morning.
The more useful question is quieter, and it applies to the whole length of the chain. Never mind the net-zero pledge or which provider has the greenest credentials. Are we using the technology efficiently at all, both where the demand is created and where it is met?
