Jevons Paradox describes what can happen when technology makes the use of a resource more efficient: rather than reducing total consumption, lower costs stimulate enough additional demand that total consumption actually increases. British economist William Stanley Jevons first described the phenomenon in his 1865 book, The Coal Question.

Is it just my imagination, or are we suddenly seeing frequent references to Jevons Paradox in the popular press and social media?  And yes, I had to look that one up.

The recent surge seems to have started when Microsoft’s Satya Nadella invoked Jevons after DeepSeek challenged prevailing assumptions about how much computing power advanced AI required. More efficient AI doesn’t mean less computing demand, he argued. Cheaper intelligence means we’ll find vastly more things to do with it.

Consider the source.  Microsoft, of course, has a vested interest in the expansion of AI computing even as the price per token falls with newer chips and algorithms. Naturally Nadella forecasts a bright future for demand, more than offsetting the reductions that greater efficiency might otherwise imply.

Put aside the biases of the AI evangelists. We have to acknowledge that Jevons Paradox sure makes enormous sense when it comes to AI.

But wait.

Nadella’s argument assumes cheaper AI and it has certainly become cheaper by many measures. But are today’s rapidly falling AI prices evidence of sustainable unit economics—or are they partly a product of an extraordinary investment race for scale, market share and competitive position? A land grab? We don’t yet know.

Even assuming AI keeps getting cheaper, Jevons speaks to usage, not return.

An employee using AI to deeply interrogate systems requirements and uncover internal control issues that otherwise would have been missed is increased consumption. Someone asking AI to turn three simple broken sentences into six paragraphs of corporate sludge is also increased consumption.

Jevons doesn’t distinguish.

What about the proliferation across enterprises? Cheap experimentation is wonderful. Let people experiment. That’s precisely how valuable new applications are discovered. But eventually experiments begin acquiring dependencies.

Someone else’s process relies on them. They connect to production data. They feed another system. Financial decisions depend upon them.

That’s when experimentation has to encounter the foundation: authoritative data, architecture, process ownership, controls and an economic case.

The secret to governance excellence: Don’t govern curiosity. Govern dependency.

Excel provides an excellent historical analogy. Excel made and makes financial analysis dramatically cheaper and undoubtedly increases productivity. It also produced decades of uncontrolled spreadsheets, shadow systems, duplicate calculations, multiple versions of truth and key-person dependencies. How many analysts have sat red-faced in board meetings, unable to reconcile multiple versions of the numbers—or, worse, discovering that one of them didn’t even foot?

Might AI be Excel in hyper speed?

Further, consider in the case of AI the deeper complexity absent even in the Excel example. The supposedly weightless digital technology turns out to sit on an extraordinarily physical foundation:

• Land
• Concrete
• Chips
• Electricity
• Water
• Capital

Jevons-driven growth puts pressure on those resources. Communities are pushing back. Grid capacity is constrained in many markets. Water has competing uses. Governments respond. Externalities that aren’t reflected in today’s token price begin getting pushed back into the economics.

And those higher costs create incentives for still another generation of innovation—better chips, smaller models, different cooling, new energy sources, different locations.

We don’t stop at the simple conclusion of the Microsoft chairman, “Jevons says consumption rises.” No, the narrative is much richer.

• Efficiency changes demand.
• Demand changes infrastructure.
• Infrastructure creates externalities.
• Externalities change economics.
• Economics drives the next innovation.

The foundation itself evolves.

Where does that leave us in terms of ROI?

Jevons may be right. We may use vastly more AI precisely because it becomes more efficient. But more use isn’t more productivity. More productivity isn’t necessarily higher returns. Thousands of successful experiments don’t necessarily make one successful enterprise. Integration, authoritative data, process ownership and governance still matter.

The faster we learn to build, the more important it becomes to understand what foundation we’re building on.

Jevons may explain why AI use will explode. It tells us almost nothing about whether that explosion creates value—or whether the foundations underneath it can support the growth.