Is AI Cannibalizing the Global Economy?

All summer long, the Grand Continent will stay in motion. Each day, we will bring you ideas you won’t find anywhere else (but which will be everywhere come autumn), texts that are hard to come by or refreshing with our Sundays. To receive them directly in your inbox and support this momentum, consider subscribing to the magazine

Financial markets have recently faced a few episodes of doubt regarding AI. Some relief arrived after the publication, by the American memory-chip maker Micron, of solid quarterly results: the group had already sold its entire annual production of its most advanced chips, confirming the strength of demand. But, from a macroeconomic standpoint, the question is less about the existence of demand—the AI transformation theme is now ubiquitous across the global economy—than about the ability of supply to respond without triggering excessive price and interest-rate tensions.

More than Micron’s results, perhaps the most significant recent event in the tech sector was Apple’s confirmation that the price of its flagship products will rise due to higher memory costs—analysts estimate a hike of 10 to 20% for upcoming models. This is a concrete illustration of a previously heterodox idea: for now, the AI revolution does not translate into a favorable supply shock capable of lowering the general price level.

The investment effort aimed at building up technological capabilities is already substantial. According to U.S. national accounts, spending on software, IT hardware, and communications equipment has doubled in real terms since 2019. Investments in electrical and telecommunications infrastructure have risen by nearly 30%.

In a broad sense — software, intellectual property, IT and communications equipment, and dedicated facilities (data centers, electrical and telecommunications infrastructure) — the tech-related activities now account for almost two-thirds of private non-residential investment in the United States.

Thus, the capital reallocation extends beyond financial markets, where a growing portion of savings flows toward large tech firms that now constitute an ever-larger share of global market capitalization. It is also physical: the capital stock of the U.S. economy is visibly shifting toward digital technologies.

However, this massive effort is likely only beginning, if one trusts available projections of future computing needs. In a particularly illuminating synthesis, RAND Corporation gathers the main estimates, converging toward a five- to ten-fold increase in computing capacity required in the United States by the end of the decade. It is a genuinely exponential dynamic.

To gauge what continuing the current investment pace would imply, we built two illustrative scenarios.

In both cases, we assume that the growth rates observed in 2024 and 2025 remain unchanged through 2035:

• technological investments continue to grow by 7.5% per year in real terms;
• non-technology investments continue to decline by 2% per year.

In the first scenario, potential GDP continues to grow at its current pace, about 1.75% per year.

In the second scenario, we adopt the hypothesis that AI-related productivity gains lift potential GDP growth by one percentage point, i.e., the midpoint of the 0.8 to 1.3 point range estimated by Philippe Aghion and Simon Bunel in 2024.

One might naturally favor this second scenario, as it seems unlikely that such a substantial investment effort could be sustained without productivity gains. Yet one should not forget that those gains can take a long time to materialize, creating a sizable lag between the investment phase and the expected economic benefits.

Our objective is not to forecast the future with precision but to assess the sustainability of the current trajectory.

In the first scenario, the non-residential private investment ratio would rise from about 15% of GDP today to nearly 21% by 2035. Even in the most favorable scenario, incorporating a sustained acceleration of potential growth, this ratio would still be around 19% of GDP, i.e., about a four-point increase.

In other words, even assuming an almost complete drying up of investment in the traditional economy—down to roughly 3% of GDP in ten years—the United States would have to sustain an investment effort of historic scale.

The starting point is already exceptional: in the first quarter of 2026, the non-residential private investment rate stood at 15.7% of GDP in real terms, versus 11.7% before the 2008 financial crisis. A further four to six point rise would place substantial pressure on the pool of available savings.

A caveat: a large portion of that effort is currently self-financed by the extraordinary cash flows of major technology firms, which cushions near-term pressure on capital markets. Nevertheless, the scale of these programs— the four hyperscalers are expected to spend about $650 billion in 2026, nearly triple their 2024 level—now pushes these groups toward greater reliance on debt via bonds. As financing shifts from equity to debt, the sensitivity of this expansion to interest rate movements grows.

In the most favorable scenario, the rise in funding needs would primarily translate into a widening U.S. current account deficit, without necessarily triggering a sharp rise in rates, provided the rest of the world generates enough saving surpluses to finance these investments.

However, nothing guarantees that such additional saving will appear spontaneously. Europe faces its own investment needs, notably in defense, while Japan confronts similar constraints. As for China, even if it continues to post a savings surplus, it seems unlikely to willingly finance the strengthening of American technological leadership amid escalating geopolitical rivalry.

The Cost of the Transition: When Supply Constraints Reappear

Thus far we have focused on the United States. Yet even if American firms maintain their edge in the race to artificial intelligence, the rest of the world will also be compelled to step up investments. Other major economic powers are already intensifying efforts to build their own capabilities. Consequently, even if they continued funding American investments, a simultaneous global rise in investment would inevitably push up the real equilibrium interest rate.

Faced with an unprecedented investment path, economists typically propose two remedies: a reduction in marginal production costs and technological progress.

Let us start with prices.

For decades, the steady decline in the price of technology equipment facilitated raising its share of total investment. For the same nominal debt or self-financing, firms could acquire ever larger volumes of information technology equipment. This dynamic did not apply to most traditional industrial goods.

That era now appears to be drawing to a close.

The Gulf War recently highlighted the semiconductor industry’s reliance on helium. Yet capacity tensions and price increases existed long before the conflict.

The AI boom is one of the main drivers of this inflation. Data centers now use exceptionally high-performance components that have become scarce, to provide the computing power necessary for training models and for large-scale inference.

The evolution of the cost of computing accelerators is a striking illustration. Each generation of NVIDIA chips — the A100 in 2020, the H100 in 2022, and then the Blackwell generation deployed widely since 2025 — sold for noticeably higher prices than the previous one, with the latest systems trading for tens of thousands of dollars per GPU. More importantly, a growing share of this cost comes from a critical input: high-bandwidth memory (HBM), for which the global production in 2026 was already spoken for at the start of the year.

Why the Cost of Inference Could Continue to Fall

Nevertheless, the hyperscalers’ business model helps absorb part of the rise in the unit cost of data centers.

Efficiency gains remain substantial. Infrastructures are better optimized, thanks in particular to more effective scheduling of inference tasks. Models themselves become more efficient, and economies of scale are fully realized when several thousand GPUs operate within a single cluster.

Thus, the cost per AI query (cost per inference) can keep declining even as the upfront cost of the infrastructure increases. This reduction is enabled by the rapid diffusion of use: firms multiply use cases and gradually grant access to AI tools to a growing number of employees.

For now, technological progress continues to offset rising costs.

Physical Limits Start to Appear

But this logic does have its limits.

The hyperscalers are willing to pay very high prices to obtain the most capable GPUs. When physical supply becomes insufficient, the market cannot simply adjust through price: the quantities available become constrained. In such a context, rationing of computing capacity becomes inevitable. It is the firms building AI applications in the real economy that find their access to computing resources limited.

Constraints are not limited to semiconductors. Operators must also contend with electricity shortages, grid connection issues, and sometimes lengthy regulatory procedures to obtain permission to build new data centers. In addition, local communities increasingly resist the establishment of these facilities.

Spillover Effects on the Rest of the Economy

Even if large technology firms manage to absorb these cost increases, the implications for other sectors could be much harder to manage.

Facing extraordinary demand from AI players, memory manufacturers naturally prioritize their most profitable customers and the most sophisticated components. This strategy reduces the capacity allocated to more standard products used in everyday consumer goods.

The recent Apple announcement is probably the first visible sign of a broader phenomenon.

According to TechInsights estimates, cited by the Wall Street Journal, memory (DRAM and NAND) embedded in the previous iPhone Pro generation accounted for about $50 of cost, out of a total manufacturing cost of roughly $580. For the upcoming generation, that memory alone could approach $200 — about a fourfold increase at constant volume — pushing total manufacturing costs above $700. In other words, the AI transition is not only shifting the distribution of profits between tech firms and traditional firms; it is also affecting consumer-facing tech companies.

Automakers are likewise affected. The amount of semiconductors integrated into vehicles is increasing steadily. While this industry uses somewhat different components than data centers, the reallocation of production capacity toward AI chips exerts pressure on prices and lead times nonetheless.

Today, electronics content represents roughly $1,000 per mid-range internal-combustion vehicle, and more than $2,000 for electric or premium vehicles.

At a moment when electric vehicles still need to close the cost gap with internal combustion engines, higher memory prices clearly add another hurdle.

A Remarkable Industrial Concentration

In the short term, there are few viable substitutes.

Global memory production is extremely concentrated. Three firms—Samsung and SK Hynix of South Korea and Micron of the United States—account for more than 95% of DRAM worldwide. The NAND market is slightly less concentrated, with a handful of players in addition (Japan’s Kioxia, American SanDisk, and China’s YMTC). Barriers to entry are enormous given the colossal investments required.

To this industrial concentration now adds geopolitical fragmentation. Companies cannot always source from the most competitive suppliers. Apple provides an illustration: the group is currently pressuring U.S. authorities to authorize the use of components produced by a Chinese vendor placed under sanctions.

The Risk of a Two-Speed Economy

In the longer run, the economy that is not directly AI-related could simultaneously face two shocks: a sustained rise in prices and a rise in real interest rates.

The combination of squeezed margins and higher capital costs would accelerate the retreat of investment, already visible in traditional sectors of the American economy.

Thus, the risk is not merely that the AI sector grows independently of the rest of the economy, but that this growth ends up hampering the development of other activities.

If traditional firms see their profitability shrink and cut investments, it becomes harder to defend the idea that AI will generate broad diffusion effects across the entire economy—such as creating new jobs to absorb workers displaced by automation.

The bifurcation between a rapidly expanding AI economy and a traditional economy under pressure could thus become more pronounced.

Macro-Economic Implications: A New Challenge for Central Banks

The developments we’ve outlined may seem theoretical and distant. Yet we believe they already exert a tangible influence on the current economic path and deserve the attention of economic policymakers.

A first alert lies in the evolution of U.S. non-technological investment. In the first quarter of 2026, it remained 11% below its 2019 level. In other words, the crowding-out effect we describe could already be at work.

Nonetheless, there is not yet clear evidence that inflationary pressures prompted by the AI transition appear in aggregate statistics.

Despite supply disruptions caused by the Gulf War, whose effects now extend far beyond the energy sector to the entire global manufacturing industry, the prices of core goods (excluding food and energy) in the PCE index have continued to ease in recent months, in contrast to overall inflation, which was boosted by the oil shock associated with the conflict. This decoupling suggests that the tensions we describe are not yet visible in aggregate statistics, but that may not persist indefinitely.

This development also suggests that the 2025 tariff shock—the magnitude and persistence of which were precisely the uncertainty that led Jerome Powell to pause the rate-cutting cycle—is gradually being absorbed by the U.S. economy.

This challenge is particularly unusual given that the Fed’s new chair, Kevin Warsh, publicly argues the opposite view. During his first congressional hearing in July, he argued that the American economy would reap substantial benefits from AI investment, with productivity gains allowing him to lower the policy rate without reigniting inflation. If our analysis is correct, the transition phase could instead generate price pressures in technology goods and push real rates higher. Under such a scenario, monetary policy anchored on the expectation of a favorable supply shock could be misled.

A Transition That Could Be Inflationary

Our analysis may appear pessimistic, but that is not our aim. We do not dispute the profoundly transformative nature of AI or the substantial long-term productivity gains it could deliver.

Our question concerns the transition phase more than the long-run potential.

The economy could indeed converge toward a much more productive equilibrium while enduring a period during which the massive investment requirements exert considerable pressure on capital markets and supply chains.

Unlike prior waves of digitization, the AI revolution demands physically expensive infrastructure: data centers, electrical grids, semiconductor production capacity, cooling equipment, telecommunications networks, and energy resources.

The transition is therefore not limited to software development; it requires a stock of physical capital on an extraordinary scale. In such a setting, the surge in investment demand could outpace the market’s capacity to adjust for a prolonged period.

If that were the case, the macroeconomic consequences would be far from neutral:

• the cost of capital would rise sustainably;
• real interest rates would tend to rise;
• inflationary pressures could reappear in certain industrial segments;
• investments in sectors less directly linked to AI would be gradually crowded out.

In other words, the cost of the transition could be substantial before productivity gains become fully visible.

The Need to Manage Temporary Macroeconomic Imbalances

Taken in isolation, each AI investment project appears perfectly rational. Each firm seeks to improve its competitiveness, cut costs, or protect its market share.

But when all firms pursue this strategy simultaneously, the collective outcome can differ markedly from the sum of individual choices.

The risk is the emergence of a cycle in which the investment boom simultaneously drives up the cost of capital, sustains supply tensions, and fuels inflation in technology goods, at the expense of investment in the rest of the economy.

This scenario does not challenge the potential benefits of AI. It simply underscores that the transition to this new economy could entail higher financing costs, greater price volatility, and different monetary policy challenges than currently anticipated by consensus.

For public authorities, and for central banks in particular, the question is no longer solely how to support a transformative technological revolution, but also how to manage temporary macroeconomic imbalances that it is likely to provoke.