AI Data Centers Competing for Power: Why U.S. Manufacturing Energy Costs Keep Soaring

The Hidden Cost of the AI Boom: Electricity Prices

When we discuss the explosion of generative AI, the conversation tends to focus on compute power, GPU supply, and the iterative advancement of large models. Yet a more fundamental—and more easily overlooked—resource is becoming the critical bottleneck for AI industry expansion: electricity.

According to reports cited on Hacker News, U.S. manufacturers are facing a predicament of sharply rising energy costs, and a major driver behind this is the enormous electricity demand from AI data centers. This phenomenon reveals a little-known chain of transmission behind the AI boom: the compute arms race is directly driving up the operating costs of the real economy.

Data Centers: The Invisible Power Black Hole

Staggering Energy Consumption Scale

The energy consumption of modern AI data centers is on an entirely different order of magnitude than that of traditional data centers. This high energy consumption has deep technical roots: generative AI models widely adopt the Transformer architecture—first introduced in the landmark 2017 paper Attention Is All You Need by Google's research team, which completely replaced the previously dominant serial RNN/LSTM structures with a revolutionary self-attention mechanism. Through the "Self-Attention" mechanism, the model must perform computations between each token and all other tokens in the sequence, causing computational complexity to grow quadratically with sequence length.

From a mathematical standpoint, the computational complexity of self-attention is O(n²·d), where n is the sequence length and d is the model dimension. This characteristic means that if the processed context length is extended from 1,000 tokens to 4,000 tokens, the amount of computation surges by 16 times, not 4 times. The context window of ultra-large models like GPT-4 has expanded to 128K tokens, with a theoretical computational load approximately 16,384 times greater than that of a 1K context. Even when engineering teams introduce optimization techniques such as Flash Attention (an efficient attention algorithm that reduces memory access through block-wise computation) and sliding window attention, the magnitude of computational demand still far exceeds that of traditional NLP models. When large models with hundreds of billions of parameters perform inference, each word generated requires hundreds of billions of floating-point operations (FLOP), and this process must be executed continuously at extremely high frequency on GPU clusters. The training phase is even more intensive—backpropagation requires storing intermediate activations and traversing the entire dataset multiple times, consuming roughly hundreds of times more energy than the inference phase. Training and running large language models requires thousands of high-performance GPUs running continuously, and the electricity consumption of a single large AI data center can rival that of a small-to-medium-sized city. As major tech companies race to expand their compute infrastructure, the pressure on regional power grids has grown exponentially.

This is by no means an abstract metaphor. According to a report released by the International Energy Agency (IEA) in 2024, global data centers consumed approximately 240–340 terawatt-hours (TWh) of electricity in 2022, and with the explosion of generative AI, projections indicate this figure could exceed 1,000 TWh by 2026. A single training run of a GPT-4-class large model is estimated to consume over 50 gigawatt-hours (GWh) of electricity, equivalent to the annual electricity use of about 5,000 American households. A single NVIDIA H100 GPU consumes up to 700 watts of power, and a supercomputing cluster equipped with 50,000 H100s can have a total power draw of 35 megawatts (MW)—already equivalent to the electricity demand of a small city.

It is worth noting that electricity consumption does not end with computation itself. Another hidden dimension facing AI data centers is the energy pressure from cooling systems. Traditional data centers use air-cooling solutions, where cooling energy consumption typically accounts for 30–40% of total electricity use. Measured by the PUE (Power Usage Effectiveness) metric, traditional data centers have a PUE of about 1.5–2.0, meaning that for every kilowatt-hour of IT equipment power consumed, an additional 0.5–1 kilowatt-hour is needed for cooling. However, the thermal density of high-power GPUs like the NVIDIA H100 is approaching the physical limits of air-cooling systems—the heat generated by a single GPU is equivalent to that of a high-power electric heater running continuously. This is driving the industry to accelerate its migration toward liquid-cooling technologies, including direct liquid cooling (DLC, where coolant piping makes direct contact with the chip's backplate) and immersion liquid cooling (where the entire server is submerged in an insulating coolant). The latter can reduce PUE to an extremely low level of 1.03–1.1, but the initial infrastructure retrofit and specialized coolant costs are extremely high. The energy consumption of cooling systems is likewise counted in the data center's overall demand on the grid, pushing its actual electricity usage further beyond estimates based on compute alone.

More critically, the pace of this demand growth far exceeds the ability of power supply to expand. Building power plants, laying transmission lines, and upgrading grids are all measured in years, whereas the deployment cadence of AI data centers is measured in months or even weeks. The direct result of this supply-demand imbalance is an overall rise in electricity prices.

Manufacturing Bears the Brunt

Electricity is a regional, shared resource. When data centers absorb large amounts of power supply in a given region, other users on the same grid—especially manufacturing enterprises that are themselves major electricity consumers—will inevitably bear higher electricity bills.

This phenomenon has a precise mechanism in economics. In regional electricity markets, prices are typically determined by the Marginal Cost Pricing mechanism. Understanding this mechanism is crucial: grid operators follow the principle of "economic dispatch," calling upon generating units starting from those with the lowest generation costs—first nuclear and renewables with near-zero fuel costs, then combined-cycle gas units, and finally the least efficient gas-turbine peaking units. The market clearing price is determined by the bid of the last unit called upon (the "marginal unit"), and all generators are settled at this price. When load surges, the marginal cost of the last peaking unit deployed can be 10–20 times that of baseload nuclear power, and this high price becomes the settlement benchmark for the entire market, producing a significant "pricing spillover" effect.

It is worth noting that the U.S. electricity market is not monolithic, but is operated in fragments by multiple regional Independent System Operators (ISO/RTO). Among these, PJM covers 13 states in the eastern U.S. plus the District of Columbia and is the world's largest wholesale electricity market; ERCOT covers nearly all of Texas and has a unique market structure due to its high degree of isolation from the federal grid. In these markets, real-time electricity prices can surge within just a few minutes from $20 per megawatt-hour to hundreds or even thousands of dollars—this extreme volatility is precisely the typical manifestation of the marginal pricing mechanism under extreme supply-demand imbalance. When the large-scale entry of data centers shifts the total demand curve to the right, more expensive peaking units (such as gas turbines) must be called upon, and the marginal electricity price rises accordingly, affecting all electricity consumers. When AI data centers with high willingness to pay flood in, they are willing to pay a premium to lock in power supply, which drives up the marginal electricity price across the entire region. Manufacturing users, even if their electricity usage behavior has not changed, must pay at the higher market price. This phenomenon has already appeared in U.S. electricity markets such as PJM and ERCOT—in data-center-dense regions like Northern Virginia and Texas, industrial electricity price increases have been markedly higher than the national average.

A deeper impact lies in the large-scale use of long-term Power Purchase Agreements (PPAs). A PPA is a long-term electricity procurement contract signed between a buyer (typically a large enterprise) and a power generator, usually with a term of 10–20 years, locking in a specified volume of electricity at a fixed price. For tech companies, PPAs serve both as a financial tool to hedge against electricity price volatility and as the primary means of declaring "100% renewable energy" in their ESG reports. However, this large-scale capacity pre-locking behavior creates a significant "queuing effect" in the electricity market: when giants like Microsoft, Google, and Meta sign gigawatt-scale PPAs, the sellable capacity of independent power producers shrinks dramatically, and small-to-medium manufacturing enterprises can only face scarcer and more expensive residual supply on the spot market, further compressing the electricity quota available to manufacturing.

For energy-intensive industries such as steel, chemicals, and aluminum smelting, electricity costs account for a considerable proportion of total costs. Rising electricity prices directly erode profit margins and weaken product price competitiveness. This effectively creates an energy "crowding-out effect": the high-profit, high-growth AI industry is squeezing out the energy resources on which traditional manufacturing depends for survival.

A Game of Priorities

Who Should Have Priority Access to Power?

Behind this question lies a profound policy and economic proposition: given a limited electricity supply, how should society allocate this scarce resource?

AI is seen as a core element of national competitiveness, and tech giants have deep financial resources, able to withstand—or even actively drive up—electricity prices to lock in supply. Meanwhile, manufacturing is the cornerstone of employment, supply chain security, and the real economy, and its competitiveness bears on the livelihoods of countless ordinary workers.

When the market mechanism allocates electricity purely on a "highest bidder wins" basis, AI enterprises with greater capital naturally gain the advantage, which may lead to structural economic imbalance. This is precisely why this topic is gradually entering the sphere of public policy discussion.

The Lag of Grid Infrastructure

One root of the problem lies in the aging and long-term underinvestment in U.S. grid infrastructure, which has deep historical roots. The core backbone of the existing U.S. transmission and distribution network was largely built in the 1950s–1970s, with a typical design lifespan of 40–50 years, meaning that a great deal of infrastructure has already exceeded its design life. In its 2021 Infrastructure Report Card, the American Society of Civil Engineers (ASCE) gave U.S. energy infrastructure a grade of C- (near failing). The grid collapse caused by the 2021 Texas cold snap was a concentrated exposure of this systemic fragility. Even more serious, the interconnection queue time for new generation projects has extended from less than 2 years in the 2000s to an average of over 5 years today, with a large number of renewable energy projects stalled due to interconnection approval backlogs. Although the Biden administration allocated $65 billion for grid modernization through the Infrastructure Investment and Jobs Act, this remains a drop in the bucket compared to the estimated trillions of dollars in total investment required.

In the short term, expanding the grid faces multiple obstacles including permitting, siting, environmental assessment, and enormous capital requirements. This means the situation of tight electricity supply is unlikely to fundamentally ease in the foreseeable future, and the cost pressure borne by manufacturing may persist over the long term.

Reflections on the Sustainability of the AI Industry

Energy Efficiency Will Become a Core Competitive Advantage

The above phenomenon is also forcing the AI industry to re-examine its own energy efficiency. Over the past few years, the growth in model scale has been a near cost-blind "arms race," but when electricity becomes a hard constraint, how to achieve better results with less compute will rise from a technical optimization problem to a strategic proposition.

Knowledge Distillation is one of the key techniques for addressing this challenge. This method was first systematically proposed by Geoffrey Hinton and others in their 2015 paper Distilling the Knowledge in a Neural Network. Its core innovation lies in introducing a "Temperature" parameter to soften the probability distribution output by the teacher model—the higher the temperature, the more the probabilities of various classes tend toward equality, carrying richer information about the relative similarity between classes. The student model learns not only the true hard labels but, more importantly, absorbs from these soft labels the teacher model's implicit understanding of data structure. In the field of large language models, distillation techniques have further evolved into various forms such as sequence-level distillation (having the student model directly learn the teacher's complete generated output), intermediate-layer feature distillation (aligning internal representation vectors), and implicit distillation (incorporating teacher model preferences within the RLHF framework). Efficient models such as DeepSeek^Ad-R1, Meta's LLaMA series, and Google's Gemma have made extensive use of this approach, significantly reducing the parameter count and computational requirements while retaining powerful capabilities.

Model Quantization compresses model parameters from 32-bit floating-point numbers down to 8-bit or even 4-bit integer representations. The latest post-training quantization (PTQ) algorithms such as GPTQ and AWQ can even compress models to 4-bit (INT4) with almost no loss of precision, enabling a 70B-parameter model that originally required 80GB of VRAM to run on consumer-grade graphics cards, reducing overall computation and VRAM usage by more than 75%.

In addition, the Mixture of Experts (MoE) architecture is also an important direction for reducing inference energy consumption, and its engineering logic is worth understanding in depth. MoE replaces the feed-forward network (FFN) layers in a traditional Transformer with multiple parallel "expert" sub-networks and introduces a lightweight "Router" network that dynamically selects 2–4 of these experts to activate for each input token (sparse activation). This means the model's total parameter count (which determines knowledge capacity) far exceeds the number of parameters actually involved in computation during each inference (which determines computational cost). Take Mixtral 8x7B released by Mistral AI as an example: its total parameters are about 46.7B, but each forward pass only activates about 12.9B parameters, reducing inference computation by about 60% compared to a dense model of equivalent performance. Models like GPT-4 and Mixtral have already validated their efficiency potential in production environments: each inference only activates a small fraction of the total parameters, thereby substantially lowering the energy consumption of a single computation.

On the other hand, tech companies' investment in self-built power sources such as nuclear and renewable energy is also accelerating, with landmark progress already emerging. Microsoft signed an agreement with Constellation Energy to restart the Three Mile Island nuclear power plant, the first commercial nuclear restart case in the U.S. in decades; Google signed a contract with Kairos Power to procure power from Small Modular Reactors (SMRs)—SMRs represent the fourth-generation evolutionary direction of nuclear technology, with representative designs including NuScale's VOYGR, TerraPower's Natrium, and Kairos Power's fluoride salt-cooled high-temperature reactor. SMRs have a single-unit capacity between 50 and 300 megawatts, and the entire reactor system can be modularly prefabricated in a factory before being transported to the selected site for assembly, compressing the construction cycle from the 10–15 years of traditional nuclear power to 3–5 years. Some SMR designs also use molten salt or high-temperature gas coolants, operating at higher temperatures, which both improves thermal efficiency and enables deep integration with hydrogen production and industrial heat supply. Even more critically, SMRs, through "passive safety" designs—which can achieve core cooling relying on physical mechanisms such as gravity and natural convection even in the extreme accident scenario of a total loss of external power—fundamentally avoid Fukushima-style meltdown risks, making them an ideal solution for data centers to obtain near-zero-carbon stable baseload power. Amazon AWS, meanwhile, is directly building data center campuses near nuclear power plants. These initiatives show that leading tech companies have come to regard energy security as a strategic asset just as important as compute supply, and self-built reliable power sources are becoming the new norm in the industry.

Policy Coordination: The Key to Resolving the Game

Relying on spontaneous market adjustment alone is difficult to resolve the deep-seated conflict between AI and manufacturing over electricity. A reasonable policy framework—including accelerating grid modernization investment, encouraging clean energy development, and formulating differentiated electricity allocation mechanisms for different industries—will be the core lever for balancing the interests of all parties.

Conclusion

The AI boom is reshaping the real economy in ways we have yet to fully comprehend. The rise in U.S. manufacturing energy costs is merely one microcosm of this profound transformation. It reminds us that no technological revolution occurs out of thin air—behind it lies the real consumption of energy, capital, and infrastructure.

While we marvel at AI's capabilities, we must equally confront the resource allocation challenges it brings. How to promote technological progress while safeguarding the healthy operation of the real economy will be a challenge that policymakers and industry must face together in the years to come.

Key Takeaways