In August 2025, Google published the first detailed measurement from a major tech company of how much energy a single query costs: a median text prompt to Gemini uses 0.24 watt-hours of energy and 0.26 millilitres of water. For comparison, that is less than nine seconds of watching television.
The number looks small the first time you read it. But for whoever owns the sustainability (ESG) report, the equation changes once that question is multiplied by tens of thousands of users, every day, across every model and every department. Months earlier, in April 2025, the International Energy Agency (IEA) had already published its own projection for data centre electricity use: it is set to more than double between 2024 and 2030, reaching 945 terawatt-hours a year — equivalent to Japan's current electricity demand.
The turning point is not the absolute quantity. It is that, for the first time, AI consumption is visible and measurable. And once something is measurable, it becomes obvious: the team that owns ESG reporting can decide question by question, model by model, reuse by reuse.
The invisible cost of a question no one reused
Not every company uses Gemini. But every company running any AI model at scale — whether built into an internal system, bought as SaaS, or running inside corporate assistants — faces the same dilemma. A question asked today consumes energy. The same question asked tomorrow consumes it again, unless there is a mechanism to reuse the answer that was already computed.
Electricity consumption from data centres is set to double by 2030, and power use from those focused on AI is poised to triple.
A corporate sustainability report has to answer questions IT rarely asks: which department is using the heaviest models most? How many times was the same question asked across different weeks? Which routine task could run on a smaller model without losing quality? Why was the same question sent to the AI three times in one month?
Those answers do not come from a personal ChatGPT or Gemini account. Nor from an agent running in parallel with no record of it. They come from a place where consumption is logged and attributed by person, group and purpose — which is exactly why what stays invisible is always what grows the most.
The latest numbers show the trend has not slowed down. According to a Gartner forecast released in June 2026, global data centre electricity consumption is set to grow 26.4% this year, reaching 565 terawatt-hours, up from 447 terawatt-hours in 2025. AI-optimized servers account for the largest share of that jump: they are expected to use 175 terawatt-hours in 2026, an 84.2% increase over the prior year, and already represent 31% of all data centre energy consumption worldwide.
For Brazilian and Latin American companies reporting on ESG targets, that growth lands at an awkward moment: energy used by AI gets added to the rest of the operation, and the electricity bill does not distinguish between a necessary query and one repeated for lack of a reuse mechanism. When the sustainability team cannot tell the two apart, the emissions report grows for a reason nobody actually decided on.
Why usage caps do not cut consumption
The obvious response looks like restriction: cap how many questions each person can ask, or ban expensive models for small tasks. That has been the strategy at many companies over the past two years. And it has not worked.
When there is no official place where the task is easier to do, people find another place — even outside the controlled perimeter. Evidence from other areas of corporate technology shows that a ban with no alternative cuts visibility, not consumption. The same person who cannot use the licensed AI will use a personal account instead, and this time the company keeps no trail of what happened — no data, no outcome, no lesson learned.
Picture a routine common to almost any office: the procurement team asks an AI model, every month, to summarize similar supplier-contract clauses for different vendors. Without a reuse mechanism, each summary gets recomputed from scratch, even when the contract changed very little from one vendor to the next. Multiplied across dozens of departments asking the same kind of repeated question, the energy waste stops being a technical detail and becomes a line item the sustainability team cannot explain.
The fix is not a harder limit. It is an environment where doing it the right way is easier than the alternative — and, by design, better controlled.
What has to be in place
Corporate identity on every question. Who asked it? Which department? Which session? The answer comes from the company directory: every query is attributed, traceable, and disappears once the person leaves.
The right-sized AI model for each task. A routine task does not need the same processing power as a complex diagnosis. If the support team can resolve 85% of questions with a lighter model, that is where those questions should go. If R&D needs deep reasoning for 10% of its cases, that model handles that 10%.
Reuse with permissions. If the same question has been asked five times in the last three months, the sixth time should not go back to computation — it should return the answer that already exists, as long as the person and their department have permission to access it. The first question carries a cost; the following ones carry only the cost of access.
Consumption visible by department. Not "AI cost USD 100,000 this month," but "AI in operations cost USD 24,000, AI in analytics cost USD 36,000, AI in R&D cost USD 40,000." Once cost is visible, the conversation changes — and the incentive to pick the right model returns to whoever is using it.
An audit trail for every sensitive action. Approve a document? Logged. Change a permission? Logged. Reuse an answer? Logged. In an audit or a compliance investigation, that is the difference between reconstructing what happened in minutes and not being able to reconstruct it at all.
That is how Skyller was designed: corporate identity on every query, the model selected to match the task, and consumption visible to whoever pays for it. That combination answers the question the ESG team actually needs answered — not just how much energy is being used, but where, why, and whether there is room to optimize.
The governance and rediscovery payoff
The most obvious benefit of visibility is control. Once consumption is measured by department, it becomes clear where the efficiency opportunities are — and where usage is growing with no justification behind it.
But there is a second gain, less obvious but just as important: whatever works well stops being personal. When an analyst finds a script that cuts 40% off the time spent on a reporting task, that script can be made available to other analysts — as long as the permissions and controls are in place. The company does not start from zero every time the task comes up again. The second person builds on what the first one learned.
The same logic applies to the upside of that gain. In a support center running different shifts, a response script validated in the morning does not need to be rediscovered by the afternoon team — as long as it is available with the right permission. That saves the person's time and the system's energy at once, because the same question, computed once, stops being recalculated on every shift.
Three questions to bring to the CFO and IT leadership
Before setting hard limits, it is worth answering these three questions with numbers:
-
What is our energy consumption by department, today? If the answer is "we don't know," the first step is not to limit usage — it is to measure it. Without visibility, there is no optimization, only guesswork.
-
How many times was the same question asked in the last 30 days? If the number is high, there is an opportunity for reuse. Every repeated question is energy that went unsaved for lack of a mechanism.
-
What share of questions left the controlled perimeter? If the answer is "we don't know," real usage is higher than what gets measured — and more expensive than it looks.
