AI · · Updated: July 15, 2026 · ⏱ 9 min read

The hidden cost of reasoning: thinking has a bill

Reasoning models consume 6 to 13 times more energy than standard queries. What data engineering is still not measuring.

The industry has decided the future of AI is reasoning. Models that think before answering, that explore hypotheses, that deliberate. What almost never appears in that narrative is a figure: each of those thoughts has a measurable energy cost. And that cost is not a detail. It is a design variable.

A lit bulb whose cable trails off frame
The bulb turns on. The cable that feeds it trails off frame.

The starting point: a metric pipelines still do not measure

Whoever designs data infrastructure has been watching certain numbers for decades. Latency. Throughput. Error rate. Cost per query. They are the vital signs of any production system, and there are mature tools to observe them.

There is a new vital sign that almost no pipeline is measuring yet: energy consumed per unit of reasoning.

For years it was not needed. A model received a prompt, generated a response and the energy cost, though real, was stable and predictable. It could be ignored without consequences. That era ended when reasoning models stopped being the exception and started being the default path.

And here appears the problem this article’s title names: reasoning has a hidden cost. Hidden not because it is secret, but because most of the systems consuming it are not observing it.

How much thinking costs: the numbers on the table

It is worth starting by dismantling two opposite exaggerations, because both circulate and both get in the way.

The first exaggeration says that every AI query is an environmental disaster. It is not. A recent study published in ScienceDirect estimates that, under realistic conditions of large-scale deployment, a standard query to a frontier model consumes a median of 0.31 Wh —the equivalent of a few seconds of a laptop running. Google declared similar figures for Gemini: around 0.24 Wh per text prompt. A normal query is not the problem.

The second exaggeration says the opposite: that efficiency improves so much the problem solves itself. Also untrue. And this is where the nuance matters.

Because the standard query is not the mode the industry is pushing. The mode becoming standard is reasoning. And the numbers change by orders of magnitude:

  • The same ScienceDirect study calculates that a reasoning query with sequences 15 times longer than usual raises the median 13 times, up to 3.91 Wh.
  • A benchmarking of 30 models cited by Earth911 found that models like o3 and DeepSeek-R1 consumed over 33 Wh in a single long prompt — more than 70 times what a lightweight model spends on the same task.
  • While standard models generate on average 37.7 additional tokens per question, reasoning models generate 543.5 additional tokens — even for simple multiple-choice questions.

The datum that best summarises the situation comes from that same study: if reasoning queries came to represent only 10% of daily traffic, that 10% would be enough to more than double total system energy consumption.

Ten per cent of the volume. One hundred per cent of additional consumption. That is the asymmetry a pipeline needs to understand before enabling reasoning by default.

The paradox of the two curves

There is an apparent contradiction in the data worth resolving, because it generates confusing headlines in both directions.

On one hand, energy per query is going down. Mixture-of-Experts architectures, which activate only a fraction of the model per query, and cache optimisations have reduced consumption per query by around 15% over two years, even though model capacity nearly doubled in the same period.

On the other hand, absolute consumption is rising forcefully — around 38% year over year.

Both things are true at once. And they do not contradict: each query is more efficient, but many more queries are made, and a growing portion of them are expensive reasoning queries. Efficiency per unit improves while total volume shoots up. Volume wins.

For the engineer, the lesson is uncomfortable but clear: provider efficiency does not exempt the system designer from responsibility. That Google made Gemini more efficient does not mean a pipeline that invokes it ten thousand times a day in deep mode is efficient. Model optimisation and usage optimisation are two different problems. The second is ours.

What this changes in data engineering

If the cost of reasoning is real and measurable, it stops being a corporate sustainability matter and becomes a pipeline design matter. Four concrete decisions appear.

1. Reasoning is a resource, not a default setting

The temptation is to enable reasoning mode across the pipeline “just in case.” More reasoning, better answers, right? The numbers say that decision, multiplied by volume, is ruinous.

Reasoning must be treated like any expensive resource: it is allocated to whoever needs it. A format validation does not need to deliberate. A routine classification does not either. Deep reasoning is reserved for what truly requires it: complex planning, ambiguity resolution, multi-step problems.

Thinking is expensive. And expensive things are managed.

2. The new metric: energy and cost per useful decision

A modern pipeline needs to instrument a metric that until now did not exist in its dashboards: how much energy —and how much money— each useful decision the system delivers costs.

Not raw tokens consumed. Not latency. Useful decision: each conclusion that reaches the business and gets used. That is the unit. And measuring it requires crossing three sources that usually live separately: token count, inference mode used, and actual business outcome.

Without that metric, optimising reasoning cost is impossible, because you do not even know where you are spending.

3. Complexity-based routing stops being optional

If a reasoning query costs between 6 and 13 times more than a standard one, routing well is not a marginal improvement: it is the difference between a viable pipeline and one bleeding money.

The reasonable pattern is tiered. A first light layer resolves the trivial and classifies the rest. Only what that layer flags as genuinely complex ascends to reasoning mode. It is the old principle of tiered systems —do not send to the expensive specialist what the cheap generalist can solve— applied to inference.

The pipeline’s intelligence is no longer only in the model. It is in knowing when not to call it in its most expensive mode.

4. Where and when it runs also counts

There is a lever data engineering knows well and that applies directly here: context-aware scheduling. The same compute load does not have the same cost —or the same footprint— depending on which region and which moment it runs in, because the grid’s energy mix varies. Studies estimate that scheduling loads according to renewable energy availability can reduce associated consumption by 10 to 20%.

For a pipeline with non-urgent reasoning tasks —batch processing, overnight analytics, retraining— this is money and carbon on the table. What we call scheduling in databases becomes here an environmental decision too.

What the numbers do not say

An honesty note before continuing. Almost all figures in this article are estimates, not official measurements.

The reason is simple: providers do not publish the data. As Earth911 gathers, of the major players in the sector only OpenAI and Google have offered any per-query consumption figure, often without detailed methodology. The rest operate with closed models. Researchers at Lawrence Berkeley National Laboratory have been blunt about it: the available information is not enough to make reasonable projections about upcoming energy demand.

That leaves whoever designs infrastructure in an uncomfortable position: making consumption decisions about a resource whose real cost the provider does not disclose. And it is worth naming, because it is part of the problem. You cannot optimise well what you cannot measure well.

The other side: the bill that does not appear on the bill

So far, engineering. But the cost of reasoning has a dimension no dashboard captures, and it is worth looking at head-on.

When a model “thinks,” it generates hundreds or thousands of internal tokens the user never sees. That hidden deliberation has real physical consumption: electricity, and with it cooling water and emissions. The user receives an elegant, silent response. They do not see the process. And above all, they do not see its cost.

That is the true meaning of the title. The cost is hidden on two planes. It is hidden in the pipeline, because few measure it. And it is hidden in the experience, because the interface is designed to make thinking look free.

On apparent gratuity

An AI that deliberates for several seconds and delivers a sophisticated conclusion transmits a specific sensation: that quality reasoning costs nothing. Just ask for it.

But it does cost. It costs energy, it costs water, it costs emissions. What happens is that this cost has been displaced out of sight of whoever causes it: to a distant data centre, to a bill someone else pays, to a diffuse environmental externality that appears nowhere in the workflow.

And when a cost becomes invisible, it is consumed without criterion. Not from bad faith, but because nothing in the experience invites moderation. The interface does not say “this query has been equivalent to charging two phones.” It just responds.

On the sector’s direction

What makes this moment particularly relevant is the direction of the journey. The industry is not moving towards more restrained reasoning use. It is moving in the opposite direction: towards models that reason by default, towards agents that deliberate among themselves, towards systems that —in the words of some sector leaders— aspire to “think for hours, days, even weeks.”

Each of those steps multiplies the hidden cost. And it multiplies it silently, because the improvement of efficiency per query serves as the perfect alibi: it lets the story be told that everything is fine while absolute consumption shoots up.

On the responsibility of who builds

Here appears the discomfort that closes the loop. Whoever designs a pipeline takes, in every architecture decision, an energy decision that affects third parties not in the room. Default inference mode, routing policy, execution region: all of it has a footprint, and that footprint is paid by a shared power grid, watershed and atmosphere.

It is not about renouncing reasoning. It is a genuinely valuable capacity and, well used, transforms what a system can do. It is about using it with the awareness that it has a bill — even if that bill arrives at a different mailbox.

Open questions

  • If the cost of reasoning is invisible to whoever consumes it, what incentive is left to moderate it?
  • Should an AI interface show the energy cost of each query, like a car shows instant fuel consumption?
  • Does it still make sense to talk about “efficiency” per query while absolute consumption grows unchecked?
  • When providers do not publish the data, can we talk about a truly informed engineering decision?
  • Are we designing systems that think better, or just systems that think more?

None has a closed answer. But there is a certainty worth taking away: in a system where thinking is expensive and that cost is hidden, sobriety stops being an environmental virtue and becomes an engineering competence.

References