Artificial intelligence has long been more than just a technological issue. It is transforming investment flows, productivity, labor markets, global trade, and financial markets—and thus precisely those factors that are of particular importance to central banks. In his speech at the Global Fintech Fest in Mumbai, Pablo Hernández de Cos, General Manager of the Bank for International Settlements (BIS) and former governor of the Bank of Spain, therefore describes this development as a macroeconomic challenge. Hernández de Cos has headed the BIS since July 2025 and previously served, among other roles, as a member of the ECB Governing Council and Chair of the Basel Committee on Banking Supervision. Three questions take center stage: How is AI changing the global economy? Why do its effects vary depending on the country and economic structure? And what does this mean for monetary policy and financial stability?
An investment boom on a macroeconomic scale
The most visible impact of AI is currently taking place on the demand side. Data centers, specialized semiconductors, cloud infrastructure, and high-performance hardware require enormous investments. According to BIS data cited by Hernández de Cos, the five largest Big Tech companies alone plan to collectively invest more than one trillion U.S. dollars in AI-related projects in 2025 and 2026. Industry representatives expect global AI investment to rise from the current level of around 500 billion U.S. dollars to three to four trillion U.S. dollars by 2030.
As a result, AI is reaching a scale that can no longer be ignored from an economic perspective. The valuations, capital expenditures, and revenues of large AI companies now influence entire economies. At the same time, the financing of this boom is changing. Investments are increasingly being financed not only from current profits but also through debt—including private credit. It is precisely here that the BIS later identifies a potential risk to financial stability.
The momentum extends beyond the financial markets. Global trade is also changing. Prices for AI-related goods have risen significantly, while export prices for many other goods have stagnated or fallen. Whether an economy benefits from this therefore increasingly depends on where it is positioned in the global AI value chain. Countries such as Taiwan, Korea, Malaysia, and Singapore benefit as upstream exporters. But even there, the profits are partly concentrated among just a few companies. In Korea, according to the figures cited in the presentation, about 43 percent of export revenues in the first quarter of 2026 came from just five companies—up from 27 percent two years earlier.
Productivity: Impressive on a Small Scale, Uncertain on a Large Scale
At the corporate level and for individual tasks, the results are sometimes spectacular. Hernández de Cos cites empirical studies showing that generative AI can enable productivity gains of between 10 and 65 percent for certain tasks. Measured time savings range from approximately 20 to 50 percent, depending on the study. So far, such effects have been particularly evident in programming, consulting tasks, and professional writing.
Another finding is interesting: For narrowly defined tasks, less experienced employees often benefit more than their more experienced colleagues. AI could thus offset part of the experience and knowledge gap—at least in areas where tasks can be effectively supported by AI.
However, a 40 percent time savings in a single task does not automatically translate into a 40 percent jump in productivity for the entire economy. This is precisely where one of the key uncertainties lies. The decisive factors are how quickly companies adapt their processes, how capital and labor are reallocated, and what organizational barriers slow down the adoption of AI. The estimates discussed in the literature vary widely; the median cited by Hernández de Cos corresponds roughly to an additional 0.5 percentage points of total factor productivity growth per year. At the same time, initial cross-country results suggest that countries that were better prepared for AI as early as 2023 have since recorded higher labor productivity growth on average.
The key message is therefore this: the microeconomic evidence is substantial—but how much of it actually translates into macroeconomic growth is much harder to predict.
AI Complements Work—and Replaces It
For the labor market as well, the presentation does not paint a simple success story or a straightforward doomsday scenario. Two forces are at work simultaneously.
On the one hand, AI can complement human labor. It increases productivity, especially when judgment, experience, and human decision-making are still required. On the other hand, it can replace workers in routine cognitive tasks. Which effect predominates depends on the occupation, industry, and national economy.
So far, the actual displacement of jobs has been limited, according to Hernández de Cos. Many companies are experimenting with AI and are still taking a wait-and-see approach in light of regulatory uncertainty, costs, and questions regarding reliability. However, initial displacement effects are already evident in areas such as customer service, programming, and administrative tasks. It is also worth taking a look at corporate conferences: According to the earnings calls analyzed by the BIS, nearly 80 percent of companies are discussing plans to automate production processes and to increase the substitution of human labor.
The BIS chief's conclusion is explicitly not to halt the technology. Rather, continuing education and retraining are crucial. Whether AI complements or displaces humans will also depend on how well workers are prepared for technological change.
Not every country starts from the same point
The question of the international distribution of AI gains is particularly relevant. According to Hernández de Cos, two factors in particular determine how much an economy can benefit: the ability to use AI productively and its position in AI production.
When it comes to deploying the technology, economic structure and "AI preparedness" play a key role. Generative AI has so far achieved particularly strong productivity gains in financial services, professional services, and information-based occupations—that is, in sectors where many cognitive and information-processing tasks are performed. Such sectors often account for a larger share of the economy in developed countries than in emerging and developing economies, where economic output is more heavily influenced by agriculture or traditional manufacturing.
Other factors include digital infrastructure, human capital, innovation capacity, and an appropriate regulatory framework. A BIS study cited by Hernández de Cos, covering 56 economies and 16 sectors, shows: Sectors particularly exposed to AI grow more rapidly in countries that are better prepared for the use of artificial intelligence. This raises the risk that AI may initially widen rather than narrow existing productivity gaps between countries. At the same time, the presentation warns against overly broad categorizations: there are also enormous differences among emerging economies.
Five Levels of a New Value Chain
On the production side, the BIS describes AI not as a single product, but as a complex global value chain. It ranges from semiconductors to cloud infrastructure, training data, and foundation models, all the way to end-user applications.
Many of these levels are characterized by high fixed costs as well as significant economies of scale and scope. This favors large providers. The global AI giants have therefore so far concentrated on a few countries and regions, particularly the United States, China, Taiwan, Korea, and the Netherlands. Some of the largest companies even cover several levels simultaneously—from chip design to cloud platforms to models and applications.
For smaller economies, however, this does not mean they must build a complete domestic AI industry. Hernández de Cos instead identifies a classic economic trade-off: On the one hand, building their own capabilities can increase resilience. On the other hand, it would be inefficient to duplicate the same enormous fixed costs everywhere. Opportunities may therefore also lie in specialized applications, locally adapted models, languages, or domain-specific data.
Three Possible Futures for AI
It will be particularly challenging for central banks because no one knows for certain how significantly AI will actually alter long-term growth. The BIS Annual Economic Report 2026 therefore distinguishes between three fundamentally different scenarios, which Hernández de Cos addresses in his speech.
In the first scenario, AI delivers a limited but sustained boost to productivity. The growth rate shifts upward without altering the underlying dynamics of the economy. The second scenario is far more radical: with transformative AI, the technology increasingly improves itself, which could give rise to a self-reinforcing growth path. The BIS refers to the third scenario as a "demand bottleneck". Automation increasingly shifts income from labor toward capital and further AI investments. If many workers are replaced, consumers—and thus demand—are lost at the same time. Productivity gains could then, paradoxically, eventually come to naught: the limiting factor would not be technical capacity, but rather insufficient demand.
These scenarios would have entirely different effects on the natural interest rate and inflation. This is precisely why the task facing central banks is not getting any easier.
The mandate remains—but the world is becoming harder to read
AI does not change monetary policy objectives. Central banks are still expected to ensure price and financial stability. However, the conditions under which they pursue these objectives could become significantly more complicated.
AI simultaneously affects demand, supply, investment, productivity, and financial markets. This increases uncertainty regarding key variables that are not directly observable anyway—such as production potential or the natural interest rate. A structural technological leap could lead traditional models to misjudge actual production potential, thereby distorting the measured output gap.
The challenge, therefore, is not so much that central banks need a completely new mandate. Rather, they must navigate an economy whose structure is changing more rapidly and whose key parameters are consequently more difficult to estimate.
When the AI Boom Becomes a Financial Risk
The presentation's focus on risk becomes particularly clear when discussing financial stability. A primary concern is cybersecurity. More powerful AI models benefit both defenders and attackers. Hernández de Cos, however, points out an asymmetry: in a worst-case scenario, the attacker need only find a single vulnerability, while the defender must protect the entire system. AI could shift this structural imbalance in favor of the attackers.
A second set of risks arises directly from the investment boom. Stock market valuations are heavily concentrated in a few AI companies and are based on high expectations for future profits. At the same time, the investments of some large companies are increasingly exceeding their cash flows. Debt financing and private credit are gaining in importance.
Added to this are complex interdependencies that are difficult to fathom. In so-called "circular Financing", for example, chip manufacturers or hyperscalers invest in AI companies, while those companies simultaneously commit to purchasing chips or computing power. This can interlink financing, demand, and company valuations, making the actual risk exposure harder to identify.
History has seen similar bouts of euphoria
The presentation deliberately draws parallels to earlier waves of technological investment: the canal mania of the 1830s, the British railroad boom of the 1840s, the electrification of the 1920s, and the dot-com boom of the late 1990s. In all cases, the euphoria was driven by a real, economically significant innovation. Nevertheless, at times more capital flowed into the new technology than the subsequent returns could justify.
This is precisely where an important distinction lies: a technology can be revolutionary and yet still be overvalued. If the economic returns from AI fall short of expectations, today's investment boom could turn into a significant correction. According to the BIS's assessment, the consequences could even be more severe than in previous episodes. Households today hold a larger portion of their wealth in stocks, so that price declines could have a greater impact on consumption. At the same time, U.S. stocks carry enormous weight in global capital markets, which means a correction could spread internationally.
Hernández de Cos expressly does not predict that such a development must occur. However, the speed and scale of investment, as well as high expectations for future returns, warrant increased attention.
Technology alone is not the deciding factor
Ultimately, the message of the presentation is remarkably sober. The economic potential of artificial intelligence is real. However, whether this leads to broadly shared growth or new economic imbalances is not determined solely by the performance of the models.
What matters are investments in infrastructure and skills, competition, data governance, and effective institutions. The crucial question, therefore, is not only how powerful AI will become, but also who will benefit from that power.
For central banks, this means that while their mandate remains the same, uncertainty is increasing. According to the BIS, measurement, economic judgment, policy adaptability, and international cooperation will therefore become more important. Or to put it another way: AI could forecast many things more accurately—but at the same time, it may initially make the economy that central banks must forecast more difficult to understand.
Bibliography:
- Pablo Hernández de Cos (2026): Artificial intelligence, growth, and financial stability: challenges for central banks. Speech at the Global Fintech Fest 2026, Mumbai, September 10, 2026, Bank for International Settlements (BIS). Internet
