The Price of Intelligence Is Collapsing: What Is Driving AI’s Extraordinary Economic Curve?

Artificial intelligence may be experiencing one of the fastest declines in performance-adjusted cost ever measured for a transformative technology. That finding has consequences extending far beyond the economics of large language models. If sustained even at substantially slower rates, declining AI costs could change who can afford advanced computational intelligence, accelerate its diffusion throughout the economy, increase both worker productivity and organizational performance, and transform AI from an expensive frontier technology into increasingly ordinary economic infrastructure.

Two recent analyses provide an unusually useful perspective on what is happening.

The first is The Plunging Price of Thought, a September 22, 2026 report by Luke Emberson and David Roodman of Epoch AI. Contrary to an understandable assumption, this is not a peer-reviewed scientific-journal article. Epoch classifies it as a research report published online. Its methodology, underlying data, code, interactive plots, and tables are publicly available, however, making the analysis unusually transparent and reproducible. (Epoch AI)

Epoch AI — The Plunging Price of Thought

The second analysis was conducted independently by economist Karthik Tadepalli. Rather than asking how rapidly AI prices are declining with time, Tadepalli asks how rapidly they are declining relative to the enormous quantity of research and development resources being devoted to AI. He published the analysis, methodology, data sources, and code in a public GitHub repository titled Price Declines per Unit of R&D. (t.co)

Karthik Tadepalli — Price Declines per Unit of R&D

Together, these analyses suggest a fascinating possibility: AI’s extraordinary progress may result not from a single technological breakthrough or an unprecedented underlying rate of return on research, but from enormous investment interacting with multiple simultaneous technological improvement curves.

Measuring the Declining Price of Artificial Intelligence

Measuring AI cost is considerably more difficult than measuring the price of electricity, steel, batteries, or computer memory.

A model costing $1 per million tokens is not necessarily less expensive in any economically meaningful sense than another model costing $5 per million tokens. If the second system requires fewer tokens, solves much harder problems, makes fewer errors, or produces dramatically better results, price per token tells us little about the actual cost of obtaining useful work.

Epoch therefore attempts to measure something more economically meaningful: the cost of obtaining a fixed level of AI performance.

The researchers assembled model performance and pricing information across five principal benchmarks covering mathematics, hard sciences, and games of skill. Instead of asking how much an AI token costs, they effectively ask:

What is the least expensive available AI system capable of achieving a specified level of performance?

Using that framework, Emberson and Roodman estimate that since 2023 the cost of obtaining a given level of AI performance has fallen approximately 47% every quarter.

Compounded annually, that corresponds to approximately a 13-fold reduction in cost each year. (Epoch AI)

The decline differs by domain. Epoch estimates quarterly cost reductions of roughly 39–43% for game-oriented puzzles and approximately 50–52% for mathematical problems. (Epoch AI)

The comparison with previous transformative technologies is even more remarkable.

Measured in logarithmic rates of price decline, Epoch estimates that AI performance-adjusted prices have been falling approximately:

4 times faster than DNA sequencing,
6 times faster than general-purpose computing,
18 times faster than lithium-ion batteries, and
54 times faster than U.S. electricity during the historical period studied. (Epoch AI)

Epoch reports historical annual price-improvement factors of approximately 1.05× for electricity, 1.16× for lithium-ion batteries, 1.51× for computing, and 1.84× for DNA sequencing, compared with approximately 13× for contemporary LLM inference. (Epoch AI)

This does not mean AI is literally comparable to a kilowatt-hour. Epoch explicitly recognizes the conceptual limitations of describing benchmark performance as the “price of thought.” Nevertheless, performance-adjusted cost provides a far better economic measure than token price alone.

Yesterday’s Frontier Becomes Tomorrow’s Commodity

One of Epoch’s most important findings may not be the average 47% quarterly decline.

It is what happens immediately after a new performance level first becomes possible.

Across Epoch’s five principal benchmarks, the cost of performance that has just reached the state of the art declines on average approximately 66% per quarter—equivalent to roughly a 75-fold annual rate of decline. Two years later, the rate slows to approximately 32% per quarter, still equivalent to roughly a 4.7-fold annual decline. (Epoch AI)

That suggests an important economic mechanism.

A newly achieved capability begins as an expensive frontier capability. Engineering improvements, algorithmic advances, smaller models, better hardware utilization, competition and new architectures then progressively reduce the resources required to reproduce it.

The frontier moves forward while the old frontier becomes inexpensive.

That distinction is essential because the enormous expense of developing frontier AI and the collapsing price of consuming previously frontier capabilities are not contradictory.

The frontier can become increasingly expensive to create while previously achieved intelligence becomes increasingly inexpensive to use.

This is potentially one of the defining economic characteristics of artificial intelligence.

From Scarce Intelligence to Widely Distributed Capability

The implications for AI diffusion are substantial.

A large enterprise can justify multimillion-dollar AI infrastructure investments that a 20-person company cannot. A medium-sized manufacturer may justify a sophisticated enterprise subscription that an independent contractor cannot. A family cannot purchase a data center.

But they may ultimately consume capabilities originally developed in those data centers for dollars—or cents.

This is how declining performance-adjusted cost can progressively move AI through the economic structure:

Frontier laboratories → hyperscale enterprises → conventional corporations → midsized businesses → small businesses → individual professionals → consumers and families.

Evidence from the U.S. Census Bureau shows that this diffusion is already underway but remains incomplete. During November 2025 through January 2026, approximately 18% of U.S. firms reported using AI in at least one business function, representing 32% when weighted by employment. Among very large firms in information, professional services and finance, adoption reached roughly 50–60%. Yet 57% of adopting businesses still used AI in three or fewer functions. (National Bureau of Economic Research)

Cost therefore remains only one constraint. Organizational readiness, applicable use cases, expertise, data infrastructure, security, governance and integration all influence diffusion. Research on U.S. manufacturing similarly identifies cost and lack of an applicable use case among the most commonly reported barriers to adoption. (TopCat)

Continued reductions in AI cost remove one of those barriers.

That matters particularly for smaller organizations, where a fixed technology expenditure represents a much greater proportion of revenue.

Productivity Is Only Half of the Economic Story

AI’s economic impact should not be evaluated solely through productivity.

Productivity asks, approximately:

How much output can be produced from a given quantity of inputs?

Performance asks a different question:

How well can the task be performed?

AI can affect both.

Suppose an employee previously required four hours to produce an acceptable analysis and can now produce it in two. That is a productivity improvement.

But suppose the employee still spends four hours and, with AI assistance, produces a substantially more comprehensive analysis, considers more alternatives, identifies previously overlooked risks and produces a better decision.

That is primarily a performance improvement.

The distinction matters enormously.

A technology that merely substitutes for labor creates one economic trajectory. A technology that simultaneously increases productivity, improves quality, expands capabilities and enables previously uneconomic activities can create a very different one.

There is already empirical evidence for both productivity and performance effects. A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, subsequently published in the peer-reviewed Quarterly Journal of Economics, examined more than 5,000 customer-support agents. AI assistance increased productivity approximately 15% on average, with particularly substantial improvements among less-experienced workers. The research also found effects involving quality, worker learning and customer interactions. (National Bureau of Economic Research)

This illustrates a mechanism with potentially enormous economic importance: AI can help distribute expertise.

Knowledge that previously resided primarily in highly experienced workers can increasingly become available to less-experienced workers through AI-mediated assistance.

Productivity, Performance and Employment

This also makes the employment question considerably more complicated than the frequently presented proposition that “AI eliminates jobs.”

Productivity improvements can reduce the amount of human labor required to produce a fixed quantity of output.

But economies do not necessarily maintain fixed quantities of output.

Lower costs can increase demand. Improved products can create new markets. New capabilities can create entirely new products and services. New businesses can form around activities that were previously economically impossible. Existing employees can move toward higher-value tasks.

The available evidence does not yet establish that AI-created employment is offsetting AI-displaced employment across the entire economy. That question remains unresolved and deserves substantial longitudinal research.

The evidence does, however, show why simple displacement calculations are inadequate.

The 2026 Census-based analysis of U.S. businesses found that 66% of AI-adopting firms were using AI solely for task augmentation, while only about 2% reported AI-related employment reductions. At the same time, deeper functional AI integration and operational investment were associated with employment decreases in some firms. (National Bureau of Economic Research)

Meanwhile, Epoch AI and Ipsos found that one in five employed U.S. adults reported using AI for at least one task that had previously been delegated to another person. Epoch appropriately cautions that task substitution does not necessarily imply worker displacement. (Epoch AI)

We therefore appear to be observing several processes simultaneously:

AI substitutes for some tasks.

AI augments other tasks.

AI increases the productivity of some workers.

AI increases the quality and performance of other workers.

AI creates new capabilities, products, services and infrastructure requirements.

And AI changes the skills and occupations demanded by the economy.

Determining the net employment consequence requires measuring all of these effects rather than counting only the jobs or tasks that AI can automate.

Then Comes the Second Question: Why Is This Happening So Fast?

Epoch establishes the remarkable rate of price decline.

Karthik Tadepalli asks a different question:

Why is the rate so extraordinary?

His insight is to change the horizontal axis.

Epoch compares technological price improvement primarily against time. Tadepalli replots the historical price series against cumulative R&D investment.

This changes the interpretation dramatically.

Tadepalli estimates AI inference prices declining approximately 12.7-fold annually while the associated R&D stock grows approximately 2.5-fold annually. He calculates a central price-to-R&D elasticity of approximately 2.7, with sensitivity estimates ranging from 2.5 to 5.4. (t.co)

For comparison, his central elasticity estimates are approximately 5.9 for DNA sequencing, 4.1 for computing, 1.2 for lithium-ion batteries and 0.5 for electricity. (t.co)

Suddenly AI no longer looks like the extreme technological outlier.

It falls somewhere in the middle.

Tadepalli therefore proposes a provocative interpretation: perhaps AI’s unprecedented improvement per year largely reflects an unprecedented quantity of research resources being compressed into each year.

In other words, the technological machine may not be extraordinarily efficient per research dollar.

We may simply be running the machine extraordinarily fast.

Capital Is an Accelerator, Not an Explanation by Itself

That interpretation is important, but it should not be taken too far.

Tadepalli explicitly characterizes his elasticity estimates as descriptive rather than causal. His calculation cannot cleanly separate formal R&D expenditure from economies of scale, learning-by-doing, spillovers and other mechanisms contributing to technological improvement. (t.co)

Money alone does not make AI improve.

Capital purchases the resources that allow improvement mechanisms to operate.

Those mechanisms include increasingly powerful accelerators; improvements in semiconductor manufacturing; larger and faster memory systems; higher-bandwidth interconnects; improved data-center architectures; better model architectures; mixture-of-experts techniques; quantization; distillation; caching; speculative decoding; improved inference engines; algorithmic efficiency; better training methods; synthetic data; increasingly automated experimentation; competition among model developers; open-weight model ecosystems; and the accumulation and diffusion of engineering knowledge.

These mechanisms interact.

That produces something closer to a technological feedback system:

Investment → compute and human talent → experimentation → better algorithms and architectures → better hardware utilization → higher capability → lower performance-adjusted cost → greater adoption → larger markets → additional revenue and investment → further R&D.

The extraordinary phenomenon may therefore be convergence.

Capital is accelerating semiconductor improvement.

Semiconductor improvement enables larger experiments.

Larger experiments reveal better architectures.

Better architectures improve compute efficiency.

Better inference engineering lowers deployment cost.

Lower deployment cost increases adoption.

Greater adoption produces larger markets.

Larger markets justify additional infrastructure and R&D.

And increasingly capable AI itself can assist researchers and engineers conducting the next generation of research.

These improvement loops can compound.

The Scale of the Investment Is Historically Important

The financial resources entering AI are unquestionably extraordinary, although they should not be confused with Tadepalli’s R&D-stock measure.

Stanford’s 2026 AI Index reports that global corporate AI investment reached approximately $581.7 billion in 2025, up about 130% from the previous year. Global private AI investment reached approximately $344.7 billion. U.S. private AI investment alone reached approximately $285.9 billion, compared with $12.4 billion in China, although Stanford cautions that private-investment comparisons understate Chinese government-directed investment. (Stanford HAI)

These numbers do not prove that investment causes AI’s cost curve.

They demonstrate the extraordinary scale of resources being mobilized around the technology.

That distinction is critical.

Why Falling AI Costs Could Matter to Civilization

The greatest consequence of declining AI cost may ultimately be diffusion rather than the achievement of any particular benchmark.

Computers became civilization-changing technologies not merely because the fastest computer became faster. Computing became transformative because computation became sufficiently inexpensive to place inside factories, automobiles, offices, homes, telephones, medical equipment, appliances and eventually billions of devices.

Electricity followed a similar pattern. Its importance came not merely from generating electricity but from distributing affordable electrical power throughout an economy and allowing thousands of complementary innovations to develop around it.

AI may follow an analogous path.

If increasingly sophisticated reasoning, analysis, translation, programming, tutoring, design, scientific assistance and decision support continue becoming cheaper, AI can move from scarce capital equipment toward ubiquitous intellectual infrastructure.

A large corporation can use it to optimize global logistics.

A midsized manufacturer can use it for engineering and quality control.

A small business can obtain analytical capabilities previously requiring specialized staff.

An entrepreneur can perform market research, financial modeling and product development that previously required a team.

A physician or scientist can use specialized systems to examine larger bodies of evidence.

A student can receive individualized tutoring.

A family can obtain assistance navigating information, education, finances and everyday decisions.

The economic importance comes from multiplying human capability throughout society.

America’s Economic Interest in AI Leadership

The United States currently possesses a substantial advantage in private AI investment. Stanford estimates U.S. private AI investment during 2025 at approximately $285.9 billion—more than 23 times the corresponding measured private investment in China—while also noting important limitations in comparing private investment across different economic systems. (Stanford HAI)

The long-term economic significance of maintaining a strong American AI ecosystem is broader than competition with any individual country.

The important question is whether the United States remains one of the principal places where AI research, semiconductor technology, data-center infrastructure, software platforms, startups, scientific applications, skilled workers and new AI-enabled industries develop and scale.

Technological leadership can affect productivity, capital formation, entrepreneurship, scientific capability, wages, exports and the creation of entirely new industries.

But invention alone is insufficient.

Diffusion matters.

The economic return from AI will depend on whether its capabilities move beyond frontier laboratories and hyperscale technology companies into manufacturers, hospitals, universities, professional firms, small businesses, entrepreneurs and households.

That makes the Epoch cost curve particularly significant.

A rapidly declining performance-adjusted cost lowers the economic threshold for adoption throughout that hierarchy.

The Investment Question

Tadepalli’s analysis adds an important strategic dimension.

If AI’s extraordinary calendar-time progress partly reflects extraordinary research intensity, then the rate of future progress may depend significantly on whether economically sustainable investment in AI research, infrastructure and commercialization continues.

That does not imply that every AI investment is economically justified, nor that more investment automatically produces proportionally more progress. Capital can be misallocated. Infrastructure can be overbuilt. Technical bottlenecks can emerge in energy, semiconductor manufacturing, data, networking or algorithmic progress. Returns can diminish. Financial leverage introduces additional risk.

The economically important objective is therefore not simply maximum expenditure.

It is maintaining a productive innovation system capable of converting capital into scientific knowledge, computational capability, useful products, lower costs and ultimately measurable improvements in human and economic performance.

If the Epoch and Tadepalli analyses continue to hold as additional evidence accumulates, they suggest something remarkable.

We may be witnessing a technological system in which enormous capital investment is being multiplied by simultaneous advances in hardware, algorithms, architectures, infrastructure, competition and accumulated knowledge.

That combination produces greater capability.

Greater capability produces greater performance.

Engineering improvement lowers the cost of reproducing that capability.

Lower cost accelerates diffusion.

Diffusion increases economic utilization.

Utilization creates additional incentives for investment and innovation.

The resulting cycle can simultaneously increase productivity—doing more with available resources—and performance—doing things better than they could previously be done.

That distinction may ultimately prove more important than the headline 47% quarterly price decline.

If artificial intelligence merely allowed civilization to perform today’s activities with fewer labor hours, its economic consequences would already be substantial.

If it simultaneously allows people and organizations to perform existing activities better, solve previously uneconomic problems, distribute expertise more widely, create new products and services, accelerate scientific discovery and make increasingly sophisticated intellectual capabilities available at progressively lower cost, the effect is much larger.

The most consequential curve in artificial intelligence may therefore not be the growth of model parameters or even benchmark intelligence.

It may be the rapidly declining cost of useful intelligence—and the expanding portion of civilization that can consequently afford to use it.

References

1. Emberson, Luke, and David Roodman. “The Plunging Price of Thought.” Epoch AI, September 22, 2026. Research report examining the decline in the cost required to achieve fixed levels of AI performance across five principal benchmarks. The report estimates an approximately 47% quarterly decline since 2023—roughly a 13-fold annual reduction—and compares the rate with historical declines in the costs of computing, DNA sequencing, lithium-ion batteries, and electricity. The report includes methodology, interactive figures, underlying data, and code.
    https://epoch.ai/publications/the-plunging-price-of-thought
2. Tadepalli, Karthik. “Price Declines per Unit of R&D.” 2026. Independent economic analysis extending the Epoch AI comparison by examining technological price declines relative to cumulative R&D investment rather than simply elapsed calendar time. Tadepalli calculates price-to-R&D elasticities for AI inference, DNA sequencing, computing, lithium-ion batteries, and electricity and provides supporting methodology and computational materials.
    https://github.com/KTadepalli/rd-price-declines
3. Epoch AI. “How Much Does It Cost to Train Frontier AI Models?” Research examining the increasing development costs of frontier AI systems, including expenditures associated with computing hardware, energy, technical personnel, and other development resources. This work provides important context for the simultaneous increase in frontier-development costs and decrease in the cost of reproducing previously achieved AI capabilities.
    https://epoch.ai/publications/how-much-does-it-cost-to-train-frontier-ai-models
4. Stanford Institute for Human-Centered Artificial Intelligence. AI Index Report 2026. Stanford University, 2026. Comprehensive annual assessment of artificial intelligence research, investment, industry development, adoption, technical performance, economic effects, and public policy. The 2026 report provides the investment figures discussed in the article, including global corporate AI investment, global private AI investment, and U.S. private AI investment.
    https://hai.stanford.edu/ai-index/2026-ai-index-report
5. Stanford Institute for Human-Centered Artificial Intelligence. “Economy.” AI Index Report 2026. Stanford University, 2026. Economic chapter of the AI Index examining private investment, corporate investment, business adoption, labor-market developments, and other measures of AI’s expanding role in the economy.
    https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
6. Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. “Generative AI at Work.” The Quarterly Journal of Economics, Vol. 140, No. 2, 2025, pp. 889–942. Large-scale empirical study of generative-AI assistance among customer-support workers. The research reports substantial productivity improvements, particularly among less-experienced workers, and examines effects on worker learning, quality, and customer interactions. An earlier version was published as NBER Working Paper No. 31161.
    https://www.nber.org/papers/w31161
7. McElheran, Kristina, Erik Brynjolfsson, Adam Looney, and related researchers. “The Rise of Artificial Intelligence: Evidence from the U.S. Census Bureau.” National Bureau of Economic Research, 2026. Census-based research examining AI adoption across U.S. businesses, including differences by firm size, industry, business function, task augmentation, operational investment, and reported employment effects. The evidence is important for distinguishing AI adoption and task augmentation from economy-wide employment displacement.
    https://www.nber.org/papers/w35141
8. Epoch AI and Ipsos. “One in Five Workers Delegate Work to AI.” Epoch AI, 2026. Survey research examining the extent to which employed U.S. adults report using artificial intelligence for activities that previously would have been delegated to another person. Epoch cautions that substitution for individual tasks should not automatically be interpreted as displacement of workers or entire occupations.
    https://epoch.ai/publications/one-in-five-workers-delegate-work-to-ai
9. American Economic Association. Research on AI Adoption in U.S. Manufacturing. AEA Papers and Proceedings, 2026. Research examining artificial-intelligence adoption and barriers to adoption among U.S. manufacturers, including cost, availability of applicable use cases, organizational readiness, and related constraints affecting diffusion of AI technologies through the industrial economy.
    https://doi.org/10.1257/pandp.20261033
10. Epoch AI. Research and Data on Trends in Artificial Intelligence. Epoch AI maintains datasets and research concerning AI model development, computational requirements, training costs, hardware, algorithmic progress, inference economics, and related technological trends. These datasets provide broader technical context for the cost and investment trends discussed in the article.
    https://epoch.ai/