Energy Is Becoming the Real Cost of Compute
The next bottleneck in artificial intelligence may not be chips — but electricity itself
The Illusion of Infinite Compute
As compute scales globally, the limiting factor is quietly shifting from semiconductors toward the physical infrastructure required to power them.
For more than two decades, the digital economy created the illusion that software scales infinitely. A streaming platform could add millions of users with marginal additional cost. A marketplace could expand globally without rebuilding physical infrastructure. Software appeared detached from the constraints that governed industrial economies.
Artificial intelligence is exposing the opposite. Every new model, every AI-generated image, every automated workflow and every large-scale inference request consumes enormous amounts of compute — and therefore enormous amounts of electricity. The more intelligence expands, the more physical infrastructure it requires beneath the surface.
The AI economy is not dematerializing the world. It is re-industrializing it.
Compute Is Becoming an Energy System
The market still talks about AI as if it were primarily a software revolution. Increasingly, it behaves like an energy system. Modern AI workloads operate continuously across hyperscale data centers requiring:
- persistent power availability,
- industrial-scale cooling,
- advanced networking,
- and enormous redundancy capacity.
This changes the nature of compute itself. Data centers are beginning to resemble critical industrial facilities rather than digital office infrastructure. The cloud is no longer simply a virtual abstraction floating above the economy. It is becoming a utility-scale physical system.
Altair Frame: The cloud is increasingly behaving like a power plant with APIs.
This is why the conversation around artificial intelligence is quietly migrating toward:
- grid stability,
- nuclear energy,
- cooling systems,
- transformer shortages,
- and regional electricity capacity.
The next phase of AI will not be constrained solely by chip availability. It will be constrained by the ability to sustain energy-intensive compute at planetary scale.
The Uber Shock: When AI Consumption Becomes Industrial
The most revealing AI stories of 2026 are no longer about model releases. They are about operational costs.
Earlier this year, Uber revealed that it had exhausted its annual AI budget within the first four months of the year after deploying AI coding assistants such as Claude Code and Cursor across its engineering organization.
The adoption was explosive. Roughly 95% of Uber engineers reportedly used AI coding tools monthly, while the majority of committed code was generated or assisted by AI systems. Internal competition even emerged around maximizing AI interactions — a phenomenon now informally referred to inside the industry as “tokenmaxxing”.
The result was immediate. API costs reportedly surged to between hundreds and thousands of dollars per engineer per month. At scale, productivity gains translated directly into compute consumption.
This is the crucial shift.
In the traditional software economy, scaling usage barely changed operational costs. In the AI economy, every additional interaction consumes:
- compute,
- electricity,
- cooling,
- and infrastructure capacity.
AI does not scale like software. It scales like industry.
Booking.com and the Rise of Inference Economics
Booking Holdings illustrates a different side of the same transformation.
The company has reported productivity improvements and customer service efficiencies through generative AI deployment. Yet at the same time, operational costs increasingly reflect rising cloud computing expenses tied to inference workloads.
This is where many organizations miscalculate AI economics. They budget for implementation. They underestimate operation.
Training a model is expensive. But running AI continuously across millions of interactions creates an entirely different cost structure — one dominated by persistent compute demand. Inference is becoming industrial infrastructure. And industrial infrastructure consumes energy continuously.
The Return of Physical Constraints
The early digital economy promised frictionless scale. Artificial intelligence is reintroducing physical limits into the center of economic growth.
Compute requires:
- semiconductors,
- electricity,
- cooling,
- land,
- water,
- transmission capacity,
- and capital-intensive infrastructure.
The implication is profound. The AI economy is not decentralizing the world. It is recentralizing it around regions capable of sustaining massive physical infrastructure loads.
Geography matters again. Energy abundance matters again. Grid resilience matters again. This is why countries are suddenly rediscovering the strategic importance of:
- nuclear energy,
- industrial grids,
- semiconductor sovereignty,
- and domestic infrastructure investment.
The digital economy increasingly depends on systems that look remarkably physical.
The Hidden Winners of the AI Economy
The market still treats utilities as defensive dividend vehicles. That assumption may become outdated.
As AI demand accelerates, value may increasingly accrue not only to chipmakers and cloud providers, but also to the physical systems enabling compute expansion.
This includes:
- electricity producers,
- grid operators,
- cooling infrastructure companies,
- energy transmission networks,
- and industrial utilities capable of supporting hyperscale data centers.
Companies such as NextEra Energy, Constellation Energy and Duke Energy are no longer operating at the edge of the AI economy. They are becoming part of its foundation.
The AI boom may ultimately become one of the largest infrastructure expansion cycles in decades.
The Mispricing of Compute
Markets continue to price AI primarily through:
- software growth,
- model capabilities,
- and productivity narratives.
But the deeper constraint is increasingly physical. The true cost of compute is no longer measured solely in semiconductors. It is measured in megawatts.
This creates a growing disconnect between:
- the perceived scalability of AI,
- and the industrial infrastructure required to sustain it.
The AI economy is beginning to resemble earlier industrial revolutions more than the lightweight internet platforms of the 2010s.
Only this time, the industrial commodity is intelligence itself.
Implication: The Industrialization of Intelligence
Artificial intelligence is often described as the next software wave. That framing is incomplete.
AI is gradually transforming the digital economy into a physical infrastructure system governed by energy availability, compute allocation and industrial capacity.
The future of artificial intelligence may ultimately depend less on algorithms than on who can generate, distribute and sustain power at scale. Because in the emerging economy, intelligence is no longer only computational. It is electrical.
🔻 Series Note
Part of The Substrate Economy, a series examining how compute has become the new constraint — and the new foundation — of economic power.
Caption
What appears lightweight at the surface increasingly depends on massive layers of invisible compute, energy and infrastructure beneath it.
Photo credit:
Photo by Thought Catalog / Unsplash
