“The cost of using AI models has declined at about 94 percent per year, reflecting falling prices for given AI capability levels.” – “Where is AI in GDP statistics?” – May 2026 – Anton Korinek (PIIE) and Patrick McKelvey (Bank of Canada)
Economic transformations rarely begin with booming revenues; they begin when the cost of a critical input collapses. In digital industries, that input is often computation, and over the past few years the price of delivering a fixed level of frontier AI capability has fallen at a pace that is extreme even by the standards of information technology.1 The resulting disconnect between the scale of technological change and the apparent modesty of measured output growth exposes structural blind spots in national accounts and complicates how policymakers, firms, and workers interpret what is happening in the economy.
From compute to capability: what is actually getting cheaper?
The headline figure that the cost of using AI models has declined at roughly 94 percent per year refers not to a vague notion of “AI” becoming cheaper, but to an explicitly defined unit of capability.1,4 Korinek and McKelvey treat AI as a production sector whose output is task performance, such as benchmark scores or specific application-level competencies, and then ask: how much must a user spend in a given year to buy a fixed quantum of that performance, delivered via current models and infrastructure?1,5
This cost falls because three forces compound on one another.4,5
- First, data centre capacity devoted to AI is expanding rapidly, so fixed costs are spread across more inference and training tasks.4,5
- Second, hardware improves: newer accelerators deliver more raw compute per dollar, continuing a variant of the long run efficiency gains associated with semiconductors.4,5
- Third, and most powerfully, algorithms become more efficient. For a given performance level, the amount of compute required falls sharply, so the same hardware can deliver much more capability.2,5
When combined, these factors imply that the effective price of a given performance level falls far faster than nominal AI spending grows. Korinek and McKelvey estimate that nominal AI compute spending in the United States grew at more than 140 percent per year in 2024 and 2025, while raw compute capacity used for AI grew at over 200 percent.5,6 Yet after adjusting for algorithmic progress and falling API prices at fixed performance, the quality-adjusted output of the AI sector rises at more than 2 000 percent per year.4,5,6 A 94 percent annual price decline is simply the mirror image of that explosive quantity growth: if the quality-adjusted quantity of AI services rises by a factor of roughly 20, but total spending rises far more modestly, the implied price per unit must fall towards zero.
Invisible boom: why GDP barely moves
Such dynamics create a statistical paradox. On one hand, AI production, understood as capability delivered, is exploding. On the other hand, conventional GDP statistics register only a relatively small sector with nominal revenues around USD 250 billion in 2025 in the United States, a figure that, while large, remains modest relative to the overall economy.4 Measured in this way, the AI sector looks like a fast growing but not dominant niche.
The paradox arises because GDP records market transactions, not physical or functional quantities. A national accountant records how much was spent on cloud services, AI chips, and model access, weighted by prevailing prices. When those prices fall by nearly 94 percent per year for a given capability, revenue growth is capped even if the real quantity of AI services explodes.2,4 The situation echoes the history of semiconductors, where each generation of chips became dramatically cheaper per unit of performance, so the sector’s GDP share remained modest while the underlying computational capacity grew exponentially.2
This has two implications. First, the AI sector’s footprint in GDP will systematically understate its contribution to productive capacity when prices fall faster than quantities rise in monetary terms. Second, any attempt to infer AI’s real economic weight from observed revenues will be misleading if quality-adjusted output is not explicitly constructed, as Korinek and McKelvey attempt to do through an “AI GDP” satellite account.1,4,5
Hedonics, quality adjustment, and their limits
Economic statisticians have tools to deal with products whose quality improves rapidly while prices decline, most notably hedonic price indices. In such approaches, statisticians regress prices on measurable characteristics of goods, effectively asking what portion of a price change is due to quality improvements rather than pure inflation or deflation. For AI, the relevant “characteristic” is model capability, such as benchmark scores or error rates on defined tasks. By comparing the cost of delivering a fixed performance bundle over time, one can infer the true price index for AI services.2,4
Korinek and McKelvey’s work can be read as a sectoral application of this logic: they construct quality-adjusted AI output by combining data on spending, compute, and algorithmic progress, effectively backing out an AI-specific price index that falls at the extraordinary rate implied by their 94 percent figure.1,5 However, as commentators have noted, such hedonic adjustments, while powerful for describing production-side changes, do not automatically capture the value users derive from those improvements.2
GDP is built around market transactions. Even if the price of a unit of AI capability falls close to zero, a firm might use it to create new products, automate internal processes, or enable services that were previously impossible. The welfare gain from such applications can be vast, but unless it translates into higher measured revenues or explicit quality adjustments in downstream sectors, the national accounts will register only incremental changes. Hedonic indices for AI chip performance or API capability therefore bridge only part of the gap between technical progress and measured welfare.2,4
Inside the 94 percent: compounding forces and simple formalisation
To understand the backstory more formally, consider a simplified relationship among spending, capability, and price. Let Q_t denote the quality-adjusted quantity of AI services delivered in period t, such as a standardised unit of benchmark-equivalent inference. Let P_t be the price per unit of this quality-adjusted output, and E_t the total expenditure on AI services in that period. In a standard national accounts identity, one can write E_t = P_t \times Q_t.
Suppose quality-adjusted output grows at a gross factor g_Q each year, while expenditure grows at g_E. Then the implied gross factor for prices g_P satisfies g_E = g_P \times g_Q. If Q_t rises by more than 2 000 percent per year, which corresponds to a factor of around 21, and E_t grows by roughly 2,4 (consistent with 140 percent spending growth), then g_P\approx 2,4/21, implying a price decline of about 89 percent in this stylised example. Korinek and McKelvey’s more granular estimates, incorporating detailed compute and algorithmic data, yield an even steeper effective decline, around 94 percent per year.4,5
Crucially, this decline is not driven by a single mechanism. Hardware efficiency enters mainly through more operations per joule and per dollar, while algorithmic progress reduces the operations needed for a given performance. If H_t denotes hardware efficiency and A_t algorithmic efficiency, the total effective cost per unit capability can be conceptualised as proportional to 1/(H_t \times A_t). Rapid growth in both H_t and A_t yields multiplicative declines in unit cost, helping explain how price reductions of this magnitude are possible within a few years.5,6
The AI sector as a “producer of producers”
In their policy brief and working paper, Korinek and McKelvey argue that AI should be treated as a distinct production sector whose main output is an intermediate input into other activities, somewhat analogous to electricity or cloud infrastructure.1,4,5 The falling price of AI capability therefore works its way through the economy not primarily via consumer purchases, but via firms embedding AI into their internal processes and products.
This framing underscores the strategic significance of the 94 percent figure. For a firm, AI capability is a cost line item: the expense of integrating a model into a workflow, running inference at scale, or fine-tuning for specific tasks. When that cost collapses year on year, projects that previously failed a cost-benefit test suddenly become feasible. A workflow that was uneconomic to automate at last year’s prices becomes attractive today; a product feature that once required a human-intensive process becomes a near-zero marginal cost AI service.
The competitive dynamics that follow resemble those of earlier general-purpose technologies. Early adopters may pay high prices for frontier models and custom integration, effectively subsidising the development and scaling that will later be available much more cheaply to others. Over time, as per-unit costs fall steeply, the barrier to adoption shifts from capital expense towards organisational capability: data readiness, process redesign, and capability to oversee AI systems become the binding constraints, not access to the technology itself.
Why national accounts struggle with nearly-free capability
The more rapidly AI capability becomes cheap, the more it undermines certain implicit assumptions in national accounts. Much of the national accounts architecture was built for an economy where major productivity-enhancing inputs were physical capital or labour time, both of which have relatively stable per-unit costs and clear ownership. When a firm invests in a new factory, that spending appears directly as gross fixed capital formation. When it hires workers, the resulting wage bill enters compensation of employees.
By contrast, a large share of AI’s impact may come from non-market uses of near-zero-price tools within organisations. An employee who previously spent hours on routine drafting can now use an internal AI system to complete the same work in minutes. The value of the time freed up may be high, but if no explicit price is charged internally for those AI interactions, the national accounts see only unchanged wages and modest cloud spending. The dramatic effective price decline for AI capability therefore amplifies an existing challenge in measuring intangible capital and internal process improvements, pushing more of the productivity action into statistical shadows.
Korinek and McKelvey’s proposed AI satellite accounts reflect a recognition that, in such a setting, standard aggregates like GDP need complementary measures that explicitly track the production and diffusion of AI capability.1,4 By treating AI as a distinct sector and constructing quality-adjusted output measures, their work provides an alternative lens through which the 94 percent price decline is not an oddity, but a central organising fact about the emerging AI economy.
Debates and objections: is the price decline overstated?
Despite the methodological care in the underlying work, several lines of debate arise around such extreme figures. One concern is that quality measurement for AI is inherently difficult. Benchmark performance may not map neatly onto real-world usefulness; models can be highly capable on standard tests while performing unevenly in complex, context-rich tasks. If the chosen capability metric overstates the functional improvement users actually experience, then the implied price decline for “true” capability may be overstated.
Another debate concerns the representativeness of the data. API prices for leading providers and benchmarked model performance may describe the frontier of commercially accessible AI, but not the entire distribution of tools used by firms and individuals. Smaller models, on-premises deployments, and open-source systems may follow different cost and capability trajectories. If most economic activity uses non-frontier systems, then the average price decline experienced in practice might be less dramatic than the headline 94 percent.
There is also a conceptual objection grounded in welfare economics. Even if the cost of achieving a given level of benchmarked performance falls dramatically, the welfare value of moving from one performance level to another may be highly non-linear. Many tasks exhibit threshold behaviours: a model that fails 20 percent of the time may be unusable, while reducing the failure rate to 3 percent suddenly unlocks a wide range of applications. In such cases, the relationship between price per benchmark unit and welfare per benchmark unit is complex. Critics may argue that focusing on cost per benchmark ignores the relationship between capability and the set of economically meaningful tasks that the model can undertake.
Korinek and McKelvey’s framework does not claim to solve these normative issues; rather, it aims to construct a consistent, production-side measure of AI output and prices. For that purpose, using observable prices for defined bundles of API access at fixed performance, together with estimates of algorithmic efficiency, is a defensible approach, even if it cannot fully capture user surplus or task threshold effects.1,4,5
The strategic tension: deflationary technology, inflationary narratives
Beyond statistical debates, the notion of a 94 percent annual decline in AI costs exposes a strategic tension. Public and policy discourse often frames AI as a booming sector commanding enormous valuations, attracting massive investment, and concentrating economic power. On this view, AI appears as a source of “inflation” in technological rents: a handful of firms supplying indispensable capabilities to the rest of the economy at high mark-ups.
Yet the production-side arithmetic tells a different story. If prices for a fixed capability bundle fall almost completely each year, then any static snapshot of mark-ups risks being misleading. The surplus is shared not only via producer profits but also via rapidly improving terms of trade for AI users. A firm that signs a long-term contract for model access today may find, within a year or two, that the same or better capability is available at a fraction of the cost, forcing providers into continuous price cutting or capability upgrades merely to sustain revenue.2,4
This dynamic complicates policy responses. On one hand, concerns about market power, data advantages, and lock-in remain salient, especially where switching costs and proprietary models create barriers to entry. On the other hand, the technology’s internal deflationary pressure is intense, pushing prices downward independent of regulatory intervention. Policymakers must therefore distinguish between static concentration in market structure and dynamic competition driven by rapid cost declines and algorithmic innovation.
Why it matters for productivity, labour, and policy
For productivity analysis, the core implication is that GDP-based measures will lag and understate AI’s true impact if they fail to incorporate quality-adjusted AI output or downstream quality improvements in AI-intensive sectors. If a firm maintains constant prices but uses cheaper AI capability to improve product quality or reduce internal costs, standard productivity metrics may capture only a fraction of the gains. This echoes earlier episodes with digital technologies, but the speed and magnitude of AI price changes make the divergence more acute.
For labour markets, the falling cost of capability increases the economic incentive to substitute or augment labour with AI across more tasks. As unit costs fall, it becomes profitable to apply AI to lower-value tasks that previously did not justify automation. This may accelerate the reallocation of tasks within occupations: routine cognitive work is increasingly shifted onto models, while human workers focus on supervision, exception handling, and tasks that remain resistant to automation. The distributional consequences of such a shift depend heavily on institutions, bargaining power, and the pace at which new task categories and roles emerge.
For macroeconomic policy, the presence of a rapidly deflating technology sector complicates inflation measurement and interpretation. If AI services become dramatically cheaper while their use spreads through the economy, properly constructed price indices would show strong deflationary pressure in AI-intensive categories. Without explicit quality adjustments, much of this effect may be misclassified or missed, leading policymakers to misread the balance between demand-driven inflation and supply-driven cost reductions.
Finally, for industrial strategy, acknowledging the 94 percent annual price decline is crucial. It suggests that national comparative advantage may hinge less on temporary access to a scarce, expensive technology and more on the ability to build complementary assets: high-quality data, organisational know-how, regulatory frameworks that permit experimentation, and human capital capable of reconfiguring processes around abundant AI capability. Governments that focus narrowly on subsidising AI infrastructure risk chasing a moving target, while those that invest in absorptive capacity and measurement infrastructure may be better positioned to harness AI’s true economic potential.
In this light, the dramatic fall in the cost of AI capability is not just a technical statistic. It is a sign that a general-purpose technology is approaching a regime of near-ubiquity, where the marginal cost of deploying intelligence-like services in software approaches that of bandwidth or storage. How societies measure, govern, and adapt to that shift will shape the extent to which AI’s productive possibilities become visible not only in GDP but in lived economic outcomes.
References
1. Where is AI in GDP statistics? | PIIE – 2026-05-18 – https://www.piie.com/publications/policy-briefs/2026/where-ai-gdp-statistics
2. Measuring the AI Economy Before GDP Can See It – Zenodo – 2026-06-01 – https://zenodo.org/records/20500501
3. Is GDP Failing to Capture AI? – Conversable Economist – 2026-06-04 – https://conversableeconomist.com/2026/06/04/is-gdp-failing-to-capture-ai/
4. Where is AI in GDP statistics? – IDEAS/RePEc – 2026-02-02 – https://ideas.repec.org/p/iie/pbrief/pb26-7.html
5. Measuring the AI economy | PIIE – 2026-05-18 – https://www.piie.com/publications/working-papers/2026/measuring-ai-economy
6. Measuring the AI economy – IDEAS/RePEc – 2026-02-02 – https://ideas.repec.org/p/iie/wpaper/wp26-9.html
7. New Research Finds AI Sector Growing Over 2000% Annually Yet … – 2026-05-25 – https://digg.com/ai/x2jl06ob
8. Measuring the AI Economy by Anton Korinek, Patrick McKelvey – 2026-05-26 – https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6813982
