The Complex Costs of AI No One Is Budgeting For

The cost of AI is not just money

While Artificial Intelligence has been a field of study and in use for decades, it burst into public consciousness on November 30, 2022 with the release of ChatGPT 3.5. Since then, AI has remade industries and upended businesses and lives around the world. Artificial Intelligence has demonstrated its ability to accelerate scientific discovery and unlock trillions in productivity gains. But beneath the optimism lies a set of costs that are staggering in scale, growing at speed, and poorly understood. The true price of AI is not measured only in dollars. It is measured in terawatt-hours, billions of gallons of water, and the livelihoods of individuals whose jobs have changed or disappeared.

‍The financial cost: A trillion-dollar model development & infrastructure race

The capital being deployed to develop and implement AI models and build AI infrastructure is unprecedented. According to Gartner, worldwide AI spending totaled $1.76 trillion in 2025. By 2027, that figure is expected to double to $3.49 trillion. Global data center CapEx spending reached approximately $430 billion in 2024 and is projected to more than double to $1.1 trillion by 2029, according to the Dell'Oro Group. A single gigawatt-scale AI data center now requires roughly $38 billion in upfront capital expenditure and $0.9 billion in annual operating expenses, according to Epoch AI. Hardware costs for leading AI data centers increased 1.9x per year between 2019 and 2025, while power requirements climbed 2x annually over the same period. Georgetown, Epoch AI, and RAND researchers project that within six years, building a single leading-edge AI data center could cost as much as $200 billion.  

These investments have impacts throughout society. Businesses, governments, utilities and the public are all absorbing escalating AI-related expenditures.  

The resource cost: water, land, and a straining grid

AI's physical resource footprint is straining natural resources around the world. Global AI data centers consumed approximately 415 TWh of electricity in 2024, and consumption could reach 945 TWh annually by 2030, according to a June 2026 UN University (UNU) report. For every kilowatt-hour of electricity consumed by a data center, an average of 2 liters of water is needed for cooling. According to a Cal State Sacramento study, Google's data centers consumed roughly 5.6 billion gallons of water in 2022; a 20% increase largely associated with AI workloads. Microsoft consumed 1.7 billion gallons, a 34% increase driven primarily by the expansion of AI infrastructure.  

In the United States and much of the world, the increased energy consumption is straining both power generation and distribution. Data centers are expected to drive half of all U.S. electricity demand growth through the end of the decade. Connecting new generation capacity and new data centers to an already overtaxed power grid will be a challenge. Current wait time on the interconnection queue for a project to get on the grid varies from as low as 3.8 years in the Northeast to an estimated nine years in California, and the cost of an interconnection has risen 88% over the last decade.  

There are multiple obstacles to bringing new power onto the grid as well. The lead time for large power transformers is now up to five years, with switchgear and high voltage cable lead times increasing as well. The lead time for gas turbines has increased to between five and seven years. Next-generation nuclear power at scale is years away, and the economics of new coal-fired power plants don’t pencil out. Grid-scale renewables and storage currently face significant political headwinds in the United States.

Data center siting is becoming increasingly problematic, as data center developers seeking locations with abundant power and water increasingly find themselves in conflict with local residents upset about rising power prices and impacts on water supplies.

AI’s high energy and water consumption, as well as its dependence on the grid, are solvable problems. Data centers can be built with behind-the-meter renewable power. A range of options exists depending on location, including wind, solar, geothermal and clean ammonia and hydrogen. These can be augmented with energy storage to ensure a steady supply of dispatchable power without taxing the local grid or passing costs on to ratepayers. Behind-the-grid power also increases site flexibility by eliminating the need to find a site with an existing grid connection and abundant local power. When data centers with behind-the-meter power have a grid connection, they can be good neighbors by serving as grid stabilizers during periods of particularly high or low demand.  

Water consumption can be reduced by using novel cooling technologies. Direct-to-chip cooling and immersion cooling are the most widely adopted, reducing water consumption by up to 90% and energy consumption by up to 50%.

The human cost: value creation and the human cost of AI

Hyperscalers expanding infrastructure and AI users betting that investments in AI tools will reduce costs and improve productivity are laying off workers to fund their AI investments. Bloomberg reports that in the first five months of 2026, the tech sector cut 123,653 jobs, with over two-thirds of the cuts attributed to AI-related investments.  

Development teams are spending massive amounts of money on AI inference. Uber spent its entire 2026 AI coding budget in four months, and a single Anthropic employee spent $150,000 on Claude Code use in a single month. While industry is making massive investments in AI, business value hasn’t materialized for most. According to an MIT study, 95% of generative AI projects fail to show measurable financial returns within six months, and 75% of companies report low to zero gains from the use of AI so far. AI has been so expensive for corporations that two-thirds of companies that conducted AI-related layoffs are rehiring some or all the roles laid off. Companies learned that poorly planned AI implementation can’t replace institutional knowledge or provide customer satisfaction.

Most companies will agree with generative AI’s current assignment to Gartner’s trough of disillusionment. However, a well-designed AI strategy, integrated into the company’s strategy and operations, will eventually result in most companies reaching the plateau of productivity and delivering measurable value. For those who don’t find value in AI, the result will likely be bankruptcy.

The cost of inference will decline over time as data centers become more resource-efficient and models become more powerful and efficient. Solutions that intelligently route each prompt to the best model will become a part of the fabric of AI inference as well. However, that will not drive an overall reduction in AI costs. Jevons paradox states that when the cost of something falls people will use more of it. So, as the cost of inference declines, companies will use more inference to add more value at a cost that will no longer require them to meter their use of AI. This effect has already been seen in radiology. Instead of AI reducing radiology jobs, the demand for radiologists has increased because cheaper image interpretation has unlocked demand for more scans.

While AI value creation has been slow to materialize, AI angst arrived like a freight train. A Goldman Sachs Research study estimates that AI could displace 6–7% of the U.S. workforce if widely adopted. Large groups of employees fear that their jobs will disappear and be replaced by AI agents. Most likely, a small percentage of those jobs will be eliminated. Still, most will be redesigned, creating the potential for an increase in labor productivity far exceeding that of the internet boom of the early 2000s.  

These productivity gains will drive increased profitability and employment as organizations grow, and new companies create additional jobs. Anthropic’s engineers are already reporting productivity gains in excess of 100%, and roughly 27% of Claude-assisted tasks are work that wouldn’t have been done otherwise because it was too labor-intensive to make sense to perform.  

AI adoption will be disruptive to the labor market. In most roles, employees won’t be replaced by AI. They will be replaced by employees who use AI better than they do, making workforce upskilling an imperative for every company and individual as employees learn to leverage AI to increase their effectiveness. However, just as in every major technological change since the Industrial Revolution, it is a given that some roles will be eliminated as those functions are completely replaced by AI. New jobs that haven’t yet been imagined will be created.  

Each of those jobs eliminated in the short term represents a person whose career has been interrupted, with real consequences for that individual and those who rely on them. Both corporations and governments need to develop programs to support displaced workers in training for new opportunities. AI has the potential to create the greatest productivity boom the world has ever seen. Everyone should have the opportunity to share in the rewards.

The future is complex

AI is changing our world in dramatic ways. It is genuinely powerful, and its benefits are real. But not all costs are well understood, and costs that are hidden are costs that are not managed. The financial outlays, the resource demands, and the human toll are accumulating simultaneously. While this article covered key financial, environmental and social costs, it hasn’t addressed some of the more complicated impacts of AI: market distortions and crowding out of other investments by AI, model bias, deep fakes and intellectual property rights, to name a few, all of which come with their own additional costs. Companies, individual AI users, and governments must deal with these costs sooner rather than later. Addressing the cost of AI is not anti-innovation. It is the condition under which sustainable innovation can occur.

‍

Ann Dunkin is a Distinguished Professor of the Practice at the Georgia Institute of Technology and CEO of Dunkin Global Advisors, where she provides strategic advice to organizations navigating complex technology decisions and evolving technology environments. She served as CIO of the U.S. Department of Energy under the Biden-Harris administration, managing a $5 billion IT portfolio, and as CIO of the U.S. Environmental Protection Agency under the Obama administration. Earlier in her career, she held leadership roles at Dell Technologies, the County of Santa Clara, and Hewlett-Packard. She is the author of Industrial Digital Transformation and a licensed professional engineer.