By: Uma Vatsa and Shailesh Mishra
The global artificial intelligence (AI) competition is increasingly being treated as a race to build more: more chips, more data centers, more power plants, more transmission lines, and more capital-intensive infrastructure. AI spending is moving at an extraordinary pace, with one recent estimate putting 2026 hyperscaler capital expenditure at around $725 billion. Alphabet has committed $180-$190 billion in 2026 capital expenditure, largely to meet unprecedented AI compute demand, while Meta has raised its 2026 capital expenditure forecast to $125-$145 billion as it expands AI infrastructure. This raises a basic but important question: is the AI economy being built through intelligence, or through brute force?
There is no doubt that AI will be transformative. The question is whether the current path is the most efficient way to get there. The data center buildout is moving so quickly that electricity, not just chips or talent, is becoming one of the binding constraints on AI’s growth. The International Energy Agency estimates that global data center electricity consumption was around 415 TWh in 2024 and could more than double to around 945 TWh by 2030. In the United States, Berkeley Lab estimates that data centers consumed about 176 TWh in 2023, or 4.4% of total U.S. electricity, and could rise to 325-580 TWh by 2028.
Meeting this demand will require major investment, new generation capacity, transmission expansion, and time. But it also raises a harder question: will long-term AI demand be strong enough to justify the scale of infrastructure now being built, or could some of this investment become underutilized, as happened with fiber-optic networks after the dot-com boom? Investment in AI infrastructure is not inherently wasteful. Some of it will be necessary, and some may support genuinely new capabilities. But the scale and speed of spending invite a more disciplined question: could similar progress be achieved through a less capital-intensive path?
The comparison with the fiber-optic boom of the 1990s is useful. Telecom companies spent heavily on fiber networks on the assumption that internet demand would rise endlessly and immediately. The technology eventually became indispensable, but demand did not mature as quickly as investors expected, and much of that capacity remained unused for years. Some estimates suggest that 85-95% of fiber laid in the 1990s remained dark after the bubble burst. Several companies collapsed under the weight of debt taken on to build networks ahead of actual demand. Fiber-optic infrastructure ultimately proved immensely valuable, but the timing and scale of capital expenditure proved fatal for many companies.
Today’s AI data center boom carries a similar risk. The industry is making large, long-lived bets on future demand, even as the technology itself is becoming more efficient. If AI systems become cheaper to run, more efficient to train, or less dependent on massively centralized compute, some of today’s infrastructure may prove oversized, poorly located, or economically fragile.
Algorithmic progress in the large language model (LLM) world has already delivered major efficiency gains. Mixture-of-experts models activate only part of a model for a given token: Google’s GLaM reportedly used one-third of the energy used to train GPT-3 while requiring half the inference computation. QLoRA shows how 4-bit quantization can enable fine-tuning of a 65-billion-parameter model on a single 48GB GPU while preserving performance. These are not marginal technical details. They directly affect electricity demand. Stanford’s AI Index found that the cost of querying a model performing at GPT-3.5 level fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold reduction in approximately 18 months. This means the future energy footprint of AI will not be determined only by how many data centers are built; it will also depend on how quickly models become more efficient.
That should make policymakers, utilities, and investors cautious. Building energy infrastructure for AI based only on today’s compute intensity may lead to overbuilding tomorrow. In the first half of 2025, utilities requested a record $29 billion in rate increases, more than double the level in the first half of 2024, affecting around 40 million customers. Data center demand is not the only driver, but it is an important contributor.
There is also a natural resource competition dynamic. Large data centers require substantial electricity, land, cooling systems, and water. Enterprise data centers can consume 300,000-500,000 gallons of water per day, while large hyperscale centers can consume one to five million gallons per day, roughly equivalent to the water use of a town of 10,000-50,000 residents. In water-stressed regions, this can create local tensions. If new AI demand is met through fossil fuel generation or gas turbines, it can also increase local air pollution.
The investment risk is not limited to Big Tech. A large ecosystem has grown around data center expansion: builders and operators such as Equinix, Digital Realty, and Iron Mountain; compute-first companies such as CoreWeave and Crusoe; and utilities, contractors, equipment suppliers, financiers, asset managers, and lenders. If AI demand slows or efficiency gains reduce the need for new capacity, the shock will not stay contained within Google, Meta, Microsoft, Amazon, Oracle, or OpenAI. It could ripple through data center companies, construction firms, financiers, local energy systems, and communities.
This does not mean governments should block data center growth. It means they should treat AI infrastructure as energy infrastructure in addition to digital infrastructure. Data centers should be encouraged where the grid can support them, where they use clean energy, recover waste heat, deploy advanced cooling, and participate in demand response. Regulators should require greater transparency on projected power demand, water use, backup generation, and emissions impact.
The fiber-optic boom was not wrong about the future of the Internet. It was wrong about timing, scale, and capital discipline. AI may follow a similar path. The technology may become indispensable, while some of today’s infrastructure bets still prove excessive. The challenge is not to slow AI innovation. It is to make the AI buildout smarter. A smarter AI economy will not be built just by throwing unlimited power, capital, and concrete at the problem. It will be built by advancing algorithms, improving efficiency, and aligning AI ambition with energy reality.
Uma Vatsa is the Head of Artificial Intelligence and Data Science at Nelumbium Capital and Shailesh Mishra is a Climate Justice Fellow at New York State Energy Research and Development Authority.