From Retro Selfies to Rising Kilowatts: The Sustainability Cost CIOs Can No Longer Ignore

Every couple of months, the internet grows obsessed with a new trend created by the combined effort of people using Gen AI and Instagram, sharing their creations without a second thought. The current AI trend of choice is the 1980s-style portrait.

It’s delightful, eye-catching, and entirely shareable – much like the Ghibli image-pocalypse of early 2025. But lurking behind the nostalgia is a less charming warning for the technologists. The trend is a consumer-scaled analogue of Shadow AI: frictionless, high-engagement, and discreetly expensive when it comes to data risk, energy, and water.

For CIOs, it’s both proof that generative AI is part of the new ordinary behavior and a warning of what’s to come. An unrestrained, ungoverned use of AI is now a tangible sustainability and cost issue.

The Hidden Cost of a Viral Prompt

The International Energy Agency estimates that the use of electricity by data-centers could grow from 485 TWh in 2025 to 950 TWh in 2030, constituting nearly 3% of global electricity consumption. The growth in data-centers demand will come from AI-optimized data centers, with the electricity demand predicted to more than quadruple by 2030.

Water’s story is comparable. In a 2023 research publish by researchers at the University of California, Riverside and the University of Texas at Arlington estimated that it might take roughly 700,000 litres of freshwater to prepare the GPT-3 mannequin. The researchers additional modeled the water consumed as the mannequin was operating.

However, there’s an important caveat. The viral claim that ‘one ChatGPT prompt uses 500 ml of water’ is not what the study actually found. Its estimated water use was about 500 ml for every 10-50 medium-length responses, varying by location and how it was operated. That was a modelled estimate for GPT-3, not a generic unit for all AI models today.

However, the magnitude of image generation is now no longer marginal. The UNU approximates that within the two years, greater than 15 billion photos have been generated utilizing text-to-image programs – practically 34 million AI photos are generated every single day of 2023.

When the Selfie Becomes an Enterprise Risk

Hundreds of millions of users – and even staff – are sending selfies to third-party AI companies for retro portraits, apparently not concerned that the data makes its way to somewhere you don’t have visibility into, and that you have no idea how much one image costs in compute cycles and data burned. That simple one click is already happening at your organization: Staff inserting consumer-facing chatbots and image generators into slides, social media, and internal chats.

This is not a matter of cultural vanity for CIOs. Instead, this is the confluence of three enterprise risks: forbidden AI usage, biometric and privacy compromise, and inference-driven energy costs rising along with every instantaneously viral prompt.

But the parallel is instructive. A consumer trend might last two weeks, but the behavior behind it- employees turning to convenient AI solutions without going through the right governance channels, won’t. Without safe, governed options, Shadow AI will take over.

From Tokenmaxxing to Responsible AI

 Early on in the era of enterprise AI, most companies rewarded volume. They promoted their staff to trial and often praised their top end users. “It’s the tokenmaxxing era,” says Ganesh Narasimhadevara, Director, Solutions Consulting at New Relic.

But that is changing. As AI application costs start to rise and pressure on the environment increases, organizations are in the process of moving beyond the volume of usage, and straight to value, and then to the environment.

The problem is not AI adoption. It is unmanaged adoption: redundant prompts, duplicated agent workflows, duplicated calls to models and compute work that don’t merit the energy they consume. Every cycle of wasted compute is wasted energy.

Observability is a tool for sustainability. Prompts, tool calls, model usage, and agent interaction data all in real time – so CIOs can see which AI projects bring value and which just drive up costs. Clearly visible data becomes the leader in waste reduction.

The Efficiency Gains are Slow, but Demand Grows Faster

The IEA explains that the electricity used for each AI task is rapidly declining due to improved efficiency of hardware and software, but demand is still increasing due to an accelerating adoption of AI by people and businesses and due to the higher energy intensity of novel use cases such as reasoning models, video generation and autonomous agents.

“Existing data centers serving AI run on the amount of electricity used by 100,000 homes, and a handful of the largest proposed data centers would use 20 times that.”

It’s not just efficiency. Instead, CIOs need their very own Green Ops playbook: measure their AI, turn off those big-but-slow workloads, rightsize the infrastructure, and incorporate sustainability into their daily operational management of AI.

New Relic officially adopted its climate strategy in FY23 and set a goal of net zero emissions by 2030. Having moved large parts of workloads to greener servers, it delivered $18m in FY26 cost savings by its Green Ops program – ahead of its target of $15m, while decreasing cloud carbon emissions by 10-12% per unit of work.

The lesson: energy efficiency and high performance can go hand in hand.

What CIOs Should do?

CIOs would be wise to track AI’s energy impact as a leading indicator rather than as something to be offset after the fact. In beginning to monitor AI usage as it happens – prompts, tokens, model calls, agent activations, cloud spend – is the first step toward a future where CIOs can make AI sustainable. These measurements should be associated to value-based KPIs, which measure value/impact, cost and energy as each business value unit is being produced.

CIOs need to define and eliminate waste by eliminating the underperforming workloads, redundant agents and obsolete trials. They should optimize workloads by shifting workloads to more efficient deployment methodologies, infrastructure, and processor types where possible. To factor in the energy consumed, cost and carbon emissions from AI applications and equipment at all levels of the lifecycle – from AI inception and procurement to deployment and maintenance – Green Ops should be integrated into AI governance. 

Narasimhadevara says it’s all simple, but it does need self-discipline: Companies have to get the fact that they know where the compute is going, and their wasting it and it can be done a lot cheaper.

Manufacturing offers a practical blueprint

This is most evident in the manufacturing sector, where energy, waste, inventory and logistics have a direct impact on both profit and carbon emissions.

ESG is a cause of concern for every business and we consider AI to be a very pragmatic lever for ESG improvement in the following four areas.

AI can also help improve the sustainability of supply chains and production by enhancing capabilities. For instance, energy monitoring in manufacturing could (i) optimize the use of assets by reducing peak loads, (ii) combine demand forecasts with correlating shop-floor systems that can identify underperforming machines.

Likewise, finding quality issues early in the manufacturing process avoids making the necessary repairs to the defective part or creating unnecessary scrap, which could later require additional parts to be purchased or longer cycle times.

Demand forecasting can help optimize stocks and prevent overshipping throughout the supply chain, while route optimization can save fuel consumption and autonomous vehicles can improve the consistency of rail or shipping routes. These are a few instances of how AI may help with sustainability by using fewer resources and energy.

Blue Star has successfully implemented its energy-management strategies, tested computer vision for quality inspection, and is reviewing route-planning software. Its aspirational goal is within reach: a minimum 5% saving in energy consumption each year over the next three years.

For Pahwa, however, the starting point is not the technology. It is the question of the business: Why are you trying to make use of this AI use case? What problem are you trying to solve, and what difference will it make?