AI Innovation and Inflation

A. Can a Real Technology Still Become a Bubble?

Technology may become cheaper, but does the system required to use that technology necessarily become cheaper as well? And if a technology is socially useful, does that mean every current investment surrounding it is economically justified?

The American debate over artificial intelligence often combines two distinct questions. The first is technological and economic: does AI actually lower production costs and raise labor productivity? The second is financial and investment-related: are the valuations of AI companies and the scale of capital flowing into data centers, semiconductor facilities, and power infrastructure economically reasonable?

These questions are closely related, but they are not the same. The fact that AI is useful does not imply that every AI company will earn high returns. The possibility that AI may increase long-run U.S. productivity does not mean that every data center, power plant, and semiconductor fabrication facility now under development will earn an adequate return on capital.

Recent improvements in AI price-performance have been striking. The direct cost of obtaining a given level of language processing, summarization, code generation, and information analysis has fallen dramatically within only a few years. Productivity gains have also been observed in several real-world tasks.⁶

Viewed in isolation, these findings make AI appear strongly deflationary. If firms can produce the same output with less labor, less time, and lower computing costs, downward pressure on service prices and production costs should follow.

But AI is not merely software. Modern large-scale AI systems are industrial systems that combine data centers, graphics accelerators, high-bandwidth memory, cooling equipment, transformers, generation capacity, and transmission networks. The price of calling a model may fall even as the total cost of supplying the electricity and physical infrastructure needed to operate that model rises.

U.S. electricity demand is now increasing after a long period of relative stagnation. The U.S. Energy Information Administration has identified large computing facilities, including data centers, as a major source of this growth and expects electricity consumption to continue rising through 2026 and 2027. Solar generation is expected to provide the largest share of new generation growth, while natural gas is likely to remain important for reliability and dispatchable supply.⁷

This growth is not geographically uniform. Electricity demand is rising particularly quickly in Northern Virginia and Texas, where large data center clusters are expanding. Commercial electricity sales in Virginia increased by roughly 30 million megawatt-hours between 2019 and 2025, with much of that increase associated with the concentration of data centers. Demand in the ERCOT region of Texas is also rising rapidly as large data centers and other major loads seek interconnection.⁸

Asset prices and investment cycles add another layer of complexity. Companies and investors that expect AI to generate enormous future value are attempting to secure that value before it is fully reflected in current cash flows. They are doing so through investments in data centers, generation capacity, semiconductor factories, and corporate equity. In this process, rational preinvestment can become difficult to distinguish from excessive optimism, and strategic capacity expansion can become difficult to distinguish from duplication and overbuilding.

The Railway Mania and the dot-com bubble help clarify this distinction. Both episodes were built around technologies that genuinely transformed society. Yet the long-term success of the underlying technology did not guarantee the success of the investors who financed it.

The central argument of this article is that the current AI investment boom combines the physical-infrastructure dynamics of the Railway Mania with the valuation and business-model dynamics of the dot-com bubble. AI lowers the unit cost of specific computational and workplace tasks, but deploying it across the United States creates new costs and investment pressures in generation, transmission, distribution, and digital infrastructure.

B. Technology Costs, System Costs, Cost Pass-Through, and Asset Prices

Four analytical layers are necessary to evaluate the economic effects of AI.

The first is the cost of the core technology. This includes model inference prices, semiconductor price-performance, algorithmic efficiency, energy efficiency, and software optimization. If the same level of performance can be achieved with less computing power and at lower cost, the unit cost of AI services should decline.

The second is the cost of complementary infrastructure. Large-scale AI services require data centers, generation capacity, grid infrastructure, cooling equipment, transformers, interconnection, memory, and advanced packaging. If the supply of these inputs is constrained, the total cost of the system may remain elevated even while the core technology becomes cheaper.

The third is cost pass-through. Even when companies reduce costs through AI, the extent to which those savings appear in consumer prices depends on competition, switching costs, market power, and organizational implementation costs. The reverse question is equally important: who pays for the expansion of the grid and generation system required by AI? These costs may be passed through in electricity rates, special tariffs, long-term power contracts, tax incentives, or infrastructure subsidies.

The fourth is asset pricing and investment. Expectations of large future gains can raise stock prices, company valuations, and capital spending before earnings are fully realized. Competition to capture future markets requires some degree of advance investment. But if multiple firms base their plans on the same demand forecast, aggregate investment may produce more data centers and generation capacity than the market ultimately requires.

These four layers can move in opposite directions. AI inference prices can fall while data center construction and electricity procurement costs rise. Workplace productivity can increase while AI company valuations incorporate unrealistically high future profits.

The questions “Is AI productive?” and “Is the current AI investment boom a bubble?” are therefore not mutually exclusive. Productive technologies can become the subject of speculative excess, and speculative excess does not eliminate the long-run value of the underlying technology.

C. Falling AI Inference Costs and Productivity Gains

One of the clearest indicators of declining AI costs is the price of achieving a fixed level of model performance. Stanford HAI’s 2025 AI Index reports that the cost of using a system with performance comparable to GPT-3.5 declined by more than 280-fold between November 2022 and October 2024.⁹

This decline makes it economically feasible for U.S. companies to incorporate AI into routine services. Capabilities that were once restricted to high-value applications can now be used in customer support, search, document drafting, translation, coding, insurance processing, and data analysis.

The same report states that AI hardware price-performance has improved by approximately 30 percent per year and energy efficiency by roughly 40 percent per year.¹⁰ If these improvements compound, the equipment and electricity required for a given amount of computation should become cheaper over time.

The exact rate of decline depends on the benchmark and performance threshold being used. Epoch AI reports that the price of achieving a fixed level of performance with large language models has fallen rapidly, but that estimated annual declines vary substantially across tasks and performance milestones.¹¹

Gundlach and coauthors estimate that the price of achieving a given level of performance on benchmarks involving knowledge, reasoning, mathematics, and software engineering has declined by approximately five- to ten-fold per year. They attribute the decline to a combination of improved hardware, algorithmic efficiency, system optimization, and economic factors.¹²

Lower costs are also beginning to translate into measurable workplace productivity gains. Brynjolfsson, Li, and Raymond examined the introduction of a generative AI assistant in a U.S. customer-support setting. Workers with access to the system resolved 15 percent more issues per hour on average.¹³

Noy and Zhang tested generative AI in professional writing tasks. Participants using AI completed their assignments more quickly and produced higher-quality work on average.¹⁴

Similar results have been reported in software development. Peng and coauthors found that participants using GitHub Copilot completed a JavaScript HTTP server task 55.8 percent faster than the control group.¹⁵

These results demonstrate that AI’s usefulness is not based solely on financial-market expectations. At the same time, technological usefulness does not guarantee financial success for every provider. If model prices continue to decline rapidly, customers may capture much of the benefit while competition compresses the margins of model developers and service providers.

A. The Railway Mania: Physical Infrastructure and Investor Losses

The British Railway Mania of the 1840s is one of the clearest historical examples of a useful technology becoming entangled with financial speculation and large-scale physical investment.

Railways offered faster and higher-capacity transportation than existing land-based alternatives. The commercial success of early rail lines encouraged the belief that railway networks would transform industry, trade, and urban development. These expectations coincided with favorable financing conditions, new investor participation, and a surge in railway company formation.

During the mid-1840s, Parliament received proposals for large numbers of railway projects, and new railway shares were issued to the public. Some shares required investors to pay only part of the face value initially, allowing them to participate in rising prices with an effect similar to leverage. Campbell argues that this financing structure helped amplify the railway investment boom.¹⁶

Railway technology had genuine value, but not every proposed line was economically viable. Competing and duplicative routes were planned, construction costs were underestimated, and expected demand was often overstated. Belief in the future of railways became separated from careful evaluation of the economics of individual lines.

When railway share prices weakened and companies called on investors to provide additional committed capital, the financial burden increased. Projects that lacked viable economics or adequate financing were canceled, delayed, or absorbed by stronger companies. Railway technology succeeded, but many railway companies and investors did not.

Financial losses did not erase all of the physical investment. Some of the capital raised during the boom was converted into tracks, stations, bridges, and other railway infrastructure. Lewis argues that repeated waves of British railway investment left behind substantial infrastructure even as abnormal investor returns disappeared.¹⁷

This distinction shows why social returns and investor returns can diverge. A railway network may reduce transportation time and integrate markets over the long run, while the investors who financed its construction earn disappointing or negative returns.

The comparison with U.S. AI data center investment is direct. AI may prove enormously important over the long term, yet not every data center or power project now under development will necessarily achieve high utilization or acceptable returns. Even if some capacity becomes excessive, power infrastructure, fiber networks, and data center buildings may remain useful for other digital applications. The survival of the infrastructure, however, does not guarantee the success of the original investors.

B. The Dot-Com Bubble: Real Technology and Overvalued Companies

The U.S. dot-com bubble of the late 1990s combined the rise of a genuinely transformative general-purpose technology with excessive valuations of internet companies.

The internet had the potential to reduce the cost of communication, commerce, information distribution, and business operations. As personal computers became more common, telecommunications networks expanded, and web browsers made the internet accessible to a broader public, investors expected established industries to be transformed.

That expectation was not fundamentally wrong. E-commerce, search, digital advertising, cloud computing, and online media later became major parts of the U.S. economy. But the proposition that the internet would transform the economy was not equivalent to the proposition that every publicly traded internet company would generate sustainable profits.

Many dot-com companies received high valuations without having established durable revenue, profits, or customer-retention economics. User growth, website traffic, and the possibility of early market dominance were often treated as substitutes for near-term financial performance. Investors emphasized network effects and potential winner-take-all outcomes while discounting current losses.

Ofek and Richardson argue that when optimistic investors dominate while pessimistic investors face short-selling constraints, internet stock prices can remain detached from underlying cash-flow fundamentals.¹⁸

As concerns about profitability increased and financing conditions tightened around 2000, internet stocks fell sharply. Many companies failed or were acquired, and information-technology investment contracted. The internet itself, however, did not disappear. Information technology became deeply integrated into U.S. business operations and productivity growth.¹⁹

The dot-com bubble did not collapse because the internet was useless. It collapsed in part because the technology’s long-run value had been applied too broadly and too aggressively to the short-run earnings and valuations of individual companies.

The same distinction applies to AI model developers and application companies. Even if AI adoption becomes widespread, declining model prices, competition, open-source alternatives, and customer bargaining power may prevent many individual firms from earning the profits implied by current expectations.

C. Why the AI Investment Boom Resembles Both Bubbles

The current AI investment boom can be understood as a combination of the Railway Mania and the dot-com bubble.

Its first Railway Mania-like feature is the need for large-scale physical infrastructure. AI models may appear to be software products, but they depend on data centers, accelerators, high-bandwidth memory, cooling systems, generation capacity, transmission lines, and transformers.

These assets require long construction lead times and large upfront capital commitments. Companies must invest in anticipation of future demand, which makes it difficult to distinguish underinvestment from overinvestment in real time.

Just as railway companies planned overlapping routes during the Railway Mania, AI companies may build overlapping data centers and model-development capacity. If actual demand does not support every company’s plan, some facilities may experience low utilization and weak returns.

The second dot-com-like feature is the mixture of genuine long-term technological value with uncertain company-level economics. AI may increase productivity and create new markets without allowing every model developer and application provider to earn monopoly profits.

As AI models become more similar in quality and inference prices continue to decline, users may benefit while model providers face margin pressure. The technology’s deflationary effect for customers may translate into intensified competition and weaker profitability for suppliers.

The AI bubble, if one exists, should therefore not be understood as speculation around a worthless technology. It is better understood as the rapid capitalization of a real technology’s future value into current company valuations and physical investment.

A. U.S. Data Center Electricity Demand: The End of Demand Stagnation

For many years, the U.S. power sector operated under the assumption of relatively stagnant electricity demand. Improvements in energy efficiency and changes in the composition of manufacturing kept electricity-sales growth low after the mid-2000s. Data centers, advanced manufacturing, and broader electrification are now challenging that assumption.

Berkeley Lab’s 2024 report estimated that U.S. data centers consumed approximately 176 terawatt-hours of electricity in 2023, equal to about 4.4 percent of total U.S. electricity consumption. The report projected that data center consumption could reach approximately 325 to 580 terawatt-hours by 2028, or 6.7 to 12 percent of U.S. electricity use.²⁰

A 2025 update released in 2026 extended the forecast through 2030. Under the baseline scenario, U.S. data center electricity consumption could reach 649 terawatt-hours, equivalent to approximately 11.8 percent of total U.S. electricity consumption. The broader scenario range is 521 to 843 terawatt-hours, or about 9.5 to 15.3 percent of U.S. electricity use.²¹

The wide range reflects uncertainty about AI chip shipments, server utilization, idle power consumption, replacement cycles, and cooling efficiency. Data center electricity demand is not determined simply by the number of models being developed. It depends on how many chips are installed, how intensively they are used, how much energy they consume when idle, and how quickly they are replaced.

The U.S. Energy Information Administration has identified large computing facilities as a major source of electricity-demand growth. It forecasts that U.S. electricity use will increase by 1 percent in 2026 and 3 percent in 2027, potentially creating the strongest four-year period of demand growth since 2000.²²

The national total is only part of the problem. Data centers are geographically concentrated, and facilities with loads ranging from hundreds of megawatts to more than a gigawatt may request interconnection over relatively short time horizons. Demand growth that appears manageable at the national level can create severe planning and investment pressure for individual utilities and regional transmission organizations.

B. Virginia and Texas: The Cost of Geographic Concentration

Northern Virginia is one of the largest concentrations of data centers in the world. Fiber connectivity, proximity to customers, available land, and established power infrastructure have reinforced the region’s position as a major digital hub.

According to the EIA, commercial electricity sales in Virginia increased by approximately 30 million megawatt-hours between 2019 and 2025. The increase was faster than in any other state except the much larger Texas market, and data center concentration was identified as a major cause. PJM expects the Dominion zone, which includes much of Virginia’s data center region, to experience one of the largest increases in summer peak demand between 2026 and 2030.²³

Texas is experiencing a similar trend. The ERCOT region is attracting data centers, cryptocurrency mining facilities, and advanced manufacturing projects. The EIA has projected that ERCOT electricity demand could increase by 7 percent in 2025 and 14 percent in 2026, with large data centers and other computing facilities among the principal drivers.²⁴

Virginia and Texas illustrate two different models of data center power policy. Virginia operates within PJM, a multistate wholesale power market that coordinates generation capacity, transmission, and reliability across a large region. Texas operates primarily within the separate ERCOT system and must manage rapid load growth largely within its own grid.

The underlying challenge is similar in both regions. Utilities must invest in power plants, transmission lines, substations, and distribution equipment before they know exactly when promised data center loads will begin operating. If projects are delayed or canceled, other customers may be left paying for underused infrastructure. If investment arrives too slowly, data center connections may be delayed and local economic-development opportunities may be lost.

The central policy question is therefore not merely how to produce more electricity. It is how to verify the credibility of large-load requests and allocate the financial risks of grid expansion among data center developers, utilities, and ordinary electricity customers.

C. U.S. Power Generation and the Grid: Who Will Supply AI Electricity?

Rising data center demand is also affecting the U.S. generation mix. Data centers require not only annual energy but also reliable, around-the-clock electricity and sufficient reserve capacity.

The EIA expects solar power to provide the largest share of growth in U.S. electricity generation in 2026 and 2027. Natural gas, however, is expected to remain the dominant generation source, supplying roughly 39 to 40 percent of U.S. electricity. Nuclear power is projected to provide approximately 18 percent, wind 11 to 12 percent, and solar 7 to 9 percent.²⁵

No single generation source is likely to meet AI electricity demand on its own. Solar and wind can be built relatively quickly and can be cost-competitive, but serving continuous data center loads requires storage, transmission, flexible demand, or other dispatchable generation.

Natural gas can respond quickly and benefits from existing infrastructure, making it important for near-term reliability. Expanding gas generation, however, introduces exposure to fuel prices, pipeline constraints, and carbon emissions. Nuclear power offers high capacity factors and low direct carbon emissions, but new large reactors involve long construction periods and high capital costs.

Some technology companies are pursuing nuclear power, advanced reactors, geothermal energy, and long-term power purchase agreements to secure reliable low-carbon electricity. Whether these projects can become operational quickly enough to match data center demand will depend on technology, permitting, regulation, and financing.

The U.S. AI electricity challenge is not only about generation. Even when new power plants and data centers exist, electricity cannot flow without adequate transmission lines, substations, and interconnection capacity.

D. The Global Electricity Perspective

The United States currently accounts for the largest share of global data center electricity consumption, but AI-related electricity demand is a global phenomenon.

The International Energy Agency estimates that data centers consumed approximately 415 terawatt-hours of electricity worldwide in 2024, equal to about 1.5 percent of global electricity consumption. The United States accounted for roughly 45 percent of global data center electricity use, China for approximately 25 percent, and Europe for about 15 percent.²⁶

Under the IEA’s baseline scenario, global data center electricity consumption more than doubles to approximately 945 terawatt-hours by 2030. AI is expected to be the largest driver of this increase, and the United States is projected to account for the largest share of incremental demand. Data centers could represent nearly half of total U.S. electricity-demand growth through 2030.²⁷

Even if data centers account for only about 3 percent of global electricity consumption in 2030, their local effects may be much larger. Data centers tend to cluster in specific metropolitan areas and transmission zones, where they can represent enormous new loads relative to existing grid capacity.

The IEA estimates that new transmission lines in advanced economies typically require four to eight years to develop. It also reports that lead times for critical equipment such as transformers and cables have doubled in recent years. If these constraints are not resolved, approximately 20 percent of the data center capacity planned globally through 2030 could face delays in obtaining grid connections.²⁸

Global AI competition is therefore not merely a competition in model performance. It is also a competition in generation capacity, transmission networks, transformers, cooling water, semiconductor production, and local permitting.

E. Transformers and Grid Interconnection

A data center cannot operate without access to the power grid. Delivering electricity from generators to server halls requires transmission lines, substations, distribution systems, and transformers.

A report from the U.S. National Renewable Energy Laboratory examines the long-term drivers of distribution-transformer demand and the supply-chain challenges affecting the sector.²⁹ It includes industry reports indicating that transformer delivery times for U.S. utilities have extended to as long as two years, that lead times have increased substantially relative to pre-2022 conditions, and that prices for some distribution transformers have risen several-fold in recent years.

Without adequate transformers and interconnection capacity, a completed data center may still be unable to receive power. Longer equipment lead times delay projects, while higher equipment prices and financing costs raise the total capital cost of data center development.

Just as the economics of a railway line depended on land acquisition, bridges, tunnels, and stations, the economics of an AI service depend not only on model prices but also on the availability and cost of the electrical infrastructure required to operate it.

Grid bottlenecks reveal the Railway Mania-like character of the AI investment boom. Infrastructure decisions must be made years before demand is fully known. Underinvestment can cause the loss of economic opportunities, while overinvestment can leave ordinary electricity customers paying for underused assets.

A. High-Bandwidth Memory: Performance Gains and Cost Constraints

Modern AI systems depend not only on computational capacity but also on memory bandwidth. Large models must rapidly read and write enormous numbers of parameters and intermediate calculations. Data movement between processors and memory can therefore limit overall system performance.

High-bandwidth memory stacks multiple memory dies vertically and connects them through a wide interface, allowing data to move at very high speeds. It is important in AI training and inference because it reduces the time processors spend waiting for data.

Samsung Electronics announced in February 2024 that it had developed a 36-gigabyte, 12-layer HBM3E product. The company stated that the device could provide bandwidth of up to 1,280 gigabytes per second and a capacity of 36 gigabytes.³⁰

Micron Technology announced around the same time that it had begun volume production of HBM3E and that its 24-gigabyte, eight-layer product would be used in NVIDIA H200 Tensor Core GPUs.³¹

High-bandwidth memory improves performance but also creates a cost constraint. Xie and coauthors argue that HBM provides high bandwidth and energy efficiency for AI workloads, but that its high cost per bit can limit the scalability of AI infrastructure.³²

HBM requires advanced manufacturing, stacking, and packaging, which makes production capacity difficult to expand quickly. When demand grows faster than supply, price, availability, and allocation can constrain the deployment of AI systems.

In this sense, HBM resembles the rails, locomotives, and construction capacity of the railway era. Critical complementary inputs can determine both the pace and the cost of technological diffusion.

B. The Coexistence of Technology Deflation and Infrastructure Inflation

The evidence suggests that AI may generate technology deflation and infrastructure inflation at the same time.

At the core technology level, costs are clearly falling. The price of achieving a fixed level of model performance has declined rapidly, hardware price-performance has improved, and energy efficiency has increased. Productivity gains have also been observed in customer support, professional writing, and coding tasks.

These improvements can lower the unit cost of work for U.S. companies. In competitive markets, some of the savings may reach consumers through lower prices or higher-quality services.

At the same time, large-scale AI deployment increases demand for data centers, generation capacity, transmission and distribution systems, transformers, high-bandwidth memory, and advanced packaging. Constraints in these sectors can slow deployment and raise capital costs.

Opposing price effects can therefore occur within the same economy. The cost of AI-enabled software and digital services may fall while the cost of electricity infrastructure, land, cooling, natural-gas supply, and grid interconnection rises in regions with concentrated data center development.

The Railway Mania demonstrates how overinvestment in physical infrastructure can leave behind useful social assets while imposing losses on investors. The dot-com bubble demonstrates how long-run technological success can coexist with excessive valuations of individual companies.

AI may combine both paths. Data centers and power infrastructure can become long-lived physical assets resembling railways, while model developers and application companies may be valued according to expectations that resemble those applied to dot-com firms.

A. Implications for U.S. Policy and Investment

U.S. AI policy should not focus only on model performance and semiconductor access. The pace of AI deployment will also depend on generation capacity, transmission lines, transformers, data center siting, cooling, memory, and permitting.

First, large data center interconnection requests should involve stronger financial commitments. If developers submit multiple speculative load requests, utilities may plan infrastructure around demand that never materializes. Deposits, phased capacity allocations, and long-term minimum-demand commitments could help distinguish credible projects from speculative applications.

Second, policymakers must establish clear principles for allocating grid-expansion costs. If the cost of substations and transmission lines built primarily for data centers is shifted to households and small businesses, electricity affordability and fairness become major concerns. At the same time, requiring data centers to pay for all regional transmission upgrades could delay infrastructure that benefits a wider set of users.

Third, policymakers should encourage geographic flexibility. Directing new data centers toward regions with available generation and transmission capacity could reduce congestion in established clusters such as Northern Virginia. Tax incentives and permitting decisions should consider not only job creation but also grid costs, water use, and broader regional economic effects.

Fourth, data center operational flexibility should be integrated into power markets. Some computing tasks can be shifted across time or location, and data centers often possess batteries, backup generation, and redundant capacity. Demand response and flexible operation could reduce grid stress without undermining reliability. The IEA identifies flexible siting and operation as important tools for reducing interconnection delays.³³

Fifth, energy policy should be technologically pragmatic. Renewable energy, storage, natural gas, nuclear power, and geothermal resources have different construction timelines and operational strengths. Meeting the reliability requirements of data centers while managing cost and emissions will require different combinations of resources in different regions.

Investors should distinguish the growth of AI technology from the profitability of individual companies. A rapidly growing market does not guarantee high returns when competition and falling prices compress margins. It is possible for users of a technology to capture most of the economic benefit while suppliers earn modest returns.

Data center investment analysis should therefore include contracted demand, power availability, interconnection status, customer concentration, responsibility for grid costs, equipment obsolescence, and the potential for alternative uses. The persistence of a physical asset does not guarantee that it will generate an attractive return.

B. Conclusion

Artificial intelligence is a genuinely useful technology. The cost of achieving a fixed level of model performance has fallen rapidly, hardware price-performance and energy efficiency have improved, and productivity gains have been observed in U.S. customer support, professional writing, and specific coding tasks.

Yet the validity of AI as a technology does not justify every investment or company valuation associated with it.

The Railway Mania demonstrates that massive advance investment in physical infrastructure can leave behind valuable social assets while producing losses for investors. Railways transformed the economy, but not every railway company or line earned a profit.

The dot-com bubble demonstrates that genuine technological value can coexist with excessive company valuations. The internet transformed the U.S. economy, but many early internet companies failed to develop sustainable business models.

The current AI investment boom combines both features. Like the Railway Mania, it involves large physical investments in data centers, power plants, transmission networks, transformers, cooling systems, and semiconductor production. Like the dot-com bubble, it involves applying expectations about AI’s long-run usefulness to the profitability and valuation of individual companies.

The United States is at the center of this transformation. It currently accounts for the largest share of global data center electricity consumption, and data centers are expected to become one of the main sources of U.S. electricity-demand growth. In Northern Virginia, Texas, and other major clusters, data center demand is already changing generation, transmission, and distribution planning.

Even if some companies and projects fail, data centers, power infrastructure, semiconductor capacity, and AI software ecosystems may remain valuable long-term assets. That possibility does not justify overinvestment, inappropriate cost shifting, or investor losses.

AI will not affect inflation in only one direction. It lowers the unit cost of computation and some forms of work, while increasing demand for electricity, land, equipment, and advanced memory. Technology deflation and infrastructure inflation can occur at the same time.

The central question is therefore not whether AI is real or whether it is a bubble. A technology can be real while the market surrounding it becomes overheated.

How much of AI’s long-term social value will become sustainable corporate cash flow, and how much of today’s investment in data centers, power generation, transmission, and semiconductors will prove to be productive infrastructure rather than excess capacity?

A second question follows for U.S. electricity policy:

How quickly will falling AI unit costs overcome the cost of power grids, transformers, generation capacity, and organizational change, and how will the resulting costs and benefits be distributed among data center developers, utilities, households, businesses, taxpayers, and investors?