The Spiralling Costs of AI: Balancing Innovation, Economics, and Sustainability

Artificial Intelligence (AI) today is no longer just a technology; it is an entire industry woven into the fabric of possibly every country, company, and society. Once celebrated purely as a symbol of innovation, AI now represents growth, power, and unprecedented revenue. But beneath the promise of limitless potential lies an uncomfortable truth: the cost (monetary, environmental, and social) of building and running AI systems is spiralling faster than most imagined.

When Jensen Huang unveiled NVLink in May this year, a breakthrough that enables massive internet-scale data transfer in machines weighing less than 60 pounds, it demonstrated how AI hardware can dramatically boost reasoning and performance, bringing us closer to the future. But these leaps come at a price: energy-hungry chips, costly data centres, and a race for computing power that very few economies can afford to keep up with.

A 2024 United States Data Centre Energy Usage Report estimates that data centres account for 4.4% of the nation’s total electricity consumption. Projections indicate that by 2028, half of this demand will be driven by AI-related workloads, a level of usage equivalent to the annual electricity consumption of approximately 22% of U.S. households. It is implied that other global economies may experience similar conditions.

ChatGPT also recently become the fifth most-visited website in the world. OpenAI said that it receives over one billion messages every day, which translates to 109 gigawatt-hours of electricity for an entire year. The number is big enough to power 10,500 homes. And that is just one website.

Vishal Karungulam, Clinical Assistant Professor (Teaching), Information Systems, ISB says, “These instances prove that AI is gradually increasing its footprint, and understanding its economics is quickly becoming as important as the technology itself. Understanding why costs are surging and how to mitigate them is critical for countries, organisations, and leaders that are seeking sustainable adoption. Equally important is recognising the broader environmental and social consequences of AI’s rapid development.”
 

Cost 1 – Investment in AI Development vs. Realised Value from an Organisation’s Perspective

According to a recent Gartner study of 451 senior technology leaders, 57% of Chief Information Officers (CIOs) globally reported having been tasked with developing an AI strategy. Yet, many face significant hurdles that limit their ability to innovate quickly and deliver tangible value. Given the pace at which the field is evolving, leaders often describe feeling as though they are “constantly chasing the hype”. What may be a relevant element of an organisation’s AI strategy today can quickly become obsolete, leaving those responsible for change struggling to justify investments already made.

There are multiple layers to this value challenge:

Layer 1: Benefits that do not always match the investment

For AI to deliver meaningful value, it must be embedded deeply and used consistently across the organisation. A global survey of 5,000 digital workers found that AI tools helped them save an average of 3.6 hours per week. While this appears positive, in the context of a 48-hour workweek, the gain represents only 7.5% efficiency — a modest improvement given the excessive cost of AI development and deployment.

To be fair, productivity gains can compound over time. Organisations that introduce AI in a measured, phased manner — while continuously mapping its usage against measurable value — are more likely to achieve substantial returns eventually.
 

Layer 2: Strategic Volatility and Cost Uncertainty

The AI space is highly volatile, with thousands of new developments emerging each year. This rapid change leaves AI leaders uncertain about which strategy will endure, and which technologies will become obsolete. 

Extending to cost forecasting as well, organisations can face budget miscalculations of 500% to 1,000%, potentially derailing entire projects and disrupting core operations.
 

Layer 3: Data Proliferation and Security Risks

As AI adoption spreads, so does the creation, processing, and storage of vast amounts of organisational data, often distributed across internal and external environments. This proliferation introduces heightened cybersecurity risks, with potential for severe monetary impact in the event of a breach. Data centre security, compliance with evolving regulations, and robust governance frameworks demand substantial and ongoing investment.
 

Cost 2 – Environmental Costs

Even when AI projects succeed in delivering measurable business value, the very infrastructure that underpins AI’s capabilities — including vast data centres, high-performance computing clusters, and always-on inference engines — consumes enormous quantities of electricity, water, and rare earth minerals.

According to Climatiq, emissions from digital technologies and data centres now exceed those from commercial aviation. Digital technology is responsible for an estimated 3.7% of global CO₂ emissions, compared with around 2.5% each for aviation and shipping. Ironically, the much-used term “cloud” belies its environmental reality: far from being weightless and immaterial, the cloud is rooted in an energy-hungry, carbon-intensive infrastructure. 

Large Language Models (LLMs), which sit at the heart of modern AI projects, demand immense computational capacity to store, process, and iterate over vast volumes of data. To prevent overheating, the data centres that power these models require vast amounts of water for evaporative cooling. A study led estimated that training GPT-3 in Microsoft’s state-of-the-art U.S. data centres could have consumed up to 7,00,000 litres (approximately 185,000 gallons) of water. Open AI’s Sam Altman recently revealed that each AI query on ChatGPT consumes about 0.34 watt-hours and around 0.000085 gallons of water, equal to one fifteenth of a teaspoon, highlighting the hidden but tangible energy and water costs of everyday AI use.

Yet, there are signs of progress. Google’s recent research on its Gemini applications shows how optimisation and clean energy procurement can dramatically reduce per-prompt impact. Between May 2024 and May 2025, Google cut the energy required to serve a single Gemini text prompt by 33x, bringing it down to about 0.24 watt-hours, roughly the energy of keeping a low-power LED bulb on for a few seconds. Carbon emissions per prompt decreased by 44 times, to 0.03 grams of CO₂ equivalent, while water consumption was reduced to approximately 0.26 ml, equivalent to about five drops. These improvements stemmed from model efficiency gains, improved hardware utilisation, and cleaner energy sources, demonstrating that while the footprint of AI is undeniably large, it can be managed and mitigated through deliberate innovation.
 

Cost 3 – Social and People Costs

For employees, AI often represents both opportunity and anxiety. On one hand, it promises efficiency, skill augmentation, and relief from repetitive tasks. On the other hand, it raises fears of redundancy, job displacement, and the erosion of roles that once defined careers. This duality means millions of workers face an uncertain transition, where reskilling and adaptability become survival skills rather than optional upskilling.

A 2024 PwC survey revealed that 44% of employees worry AI will change their jobs beyond recognition within five years, fuelling rising workplace stress and mental health concerns. The pressure to adapt to constant technological change is also reflected at the leadership level. According to ServiceNow’s Enterprise AI Maturity Index 2025, 82% of respondents expect to increase AI investment, but only 28% of executives say they are “very familiar” with agentic AI, often citing a lack of clarity on organisational readiness, ethical guardrails, and long-term strategy.

Unequal access to AI tools also risks deepening the digital divide, particularly between advanced economies and developing ones. UNESCO estimates that half the world’s population still lacks meaningful internet access. This gap could worsen as AI systems, tools, and benefits cluster around well-connected regions, industries, and socioeconomic groups. Closer home, the NASSCOM AI Adoption Index (2024) for India found that while large enterprises are accelerating AI deployment, small and mid-sized businesses remain largely excluded, widening competitiveness gaps and potentially limiting inclusive growth.
 

Conclusion

The AI footprint on our planet and our lives today is the smallest it will ever be. Ultimately, the spiralling cost of AI lies not in the technology itself, but in how societies manage its adoption. If handled responsibly, AI can empower workers, democratise opportunities, and drive human progress. If not, it risks exacerbating inequality in the society and impact global sustainability.
 

Image Credit: robert_hrovat from Pixabay