Key points.
- In the absence of exogenous shocks, AI-related productivity gains are expected to help boost potential US growth to an average of 2.1% over the next decade, reaching 2.4% by the mid-2030s.
- AI is expected to have only a limited and gradual impact on unemployment, as it will primarily affect high-income earners with a greater capacity for retraining.
- Within the American economy, AI is expected to offset the effects of political obstacles and less favorable demographic trends over the next ten years.
- The ability of “large language models (LLMs)” to achieve artificial general intelligence will be a determining factor in the success of the AI race. The United States and China are pursuing different paths: while American companies are primarily developing closed, proprietary models, China favors a more open and standardized strategy.
How will artificial intelligence shape economic productivity, growth, and employment over the next decade? We believe it will boost the US economy, largely offsetting the negative effects of demographic shifts and political obstacles. We also explore the impact of the divergence between US and Chinese strategies, as well as the path toward artificial general intelligence.
Global spending on artificial intelligence (AI) is growing rapidly, driven by massive investments in data centers, semiconductors, and software. As the primary AI investment hub, dominated by a handful of very large technology companies, the United States is expected to see its economy reshaped by AI over the next decade, thanks to productivity gains and a boost in potential growth.
Potential growth refers to an economy’s capacity to expand in the absence of shocks. It is broken down into labor force growth and productivity growth. AI is expected to significantly boost the latter. On average, we anticipate that AI will drive US productivity growth from 1.60% to 1.85% per year over the next ten years . The beneficial effects of this acceleration should be felt primarily toward the end of the decade, as the rate of AI adoption across the services sector increases.
AI is expected to boost productivity in the United States
This increase may seem modest, but it could bring productivity back to levels comparable to—or even slightly higher than—those seen during the rise of the internet and personal computers in the early 2000s. It would also extend the recent acceleration in US productivity observed after the pandemic. AI will have both direct and indirect effects on productivity. Direct effects include taking over tasks currently performed by humans more cost-efficiently, thereby increasing total factor productivity.
We estimate that these effects will translate into an average of 0.15 percentage points per year over the next decade.
We estimate that these effects will translate into an average of 0.15 percentage points per year over the next decade. Indirect effects include an increase in capital per worker—see the capital intensity shown in Figure 1—as well as support for generating new ideas and accelerating technological progress, particularly through assistance to scientists, engineers, and medical researchers. We estimate these indirect effects at around 0.1 percentage points per year. However, there is considerable uncertainty surrounding this, and this impact could be greater if a form of artificial general intelligence (AGI)—comparable to human intelligence across a wide range of tasks—were achieved in the near future.
AI is not expected to lead to a widespread increase in unemployment.
Despite AI’s ability to perform tasks currently reserved for humans, we believe its impact on unemployment will be limited and gradual. Admittedly, AI has already led to layoffs in the service sector, particularly in technology companies, which now have to accomplish more with fewer staff, and has made the job market more challenging for recent graduates.
However, unlike a recession, where the lowest-income workers are typically hardest hit, AI has a greater impact on educated, middle- and upper-class workers. These individuals often possess the skills to focus on tasks different from those automated by AI, or to find alternative employment as the economy evolves. For this reason, we anticipate a moderate rise in frictional unemployment that will occur gradually over the next decade as AI adoption increases and the economy adjusts. This effect should also be mitigated by other trends affecting the workforce, which we examine below.
AI should more than compensate for a gloomy demographic situation.
The productivity gains associated with AI are certainly not an isolated phenomenon. Key trends in the US economy, including evolving policies and demographics, are less favorable. We believe the first category—which includes efficiency losses from supply chain disruptions caused by tariffs , productivity losses from lower-skilled immigration, and threats to the credibility of US institutions—will reduce potential growth by 0.15% over the next decade. However, most of these effects are expected to be felt in the short term and could even reverse under future US administrations.
The demographic outlook is also much less favorable. Since 2019, high immigration has contributed to the expansion of the US labor force, adding nearly one percentage point per year to the country’s growth. However, population aging and a sharp decline in immigration are expected to significantly slow labor force growth.
We project potential growth for the United States of 2.1% on average over the next decade.
Given the combined effects of the AI drive, negative political trends and a less favorable demographic situation, we project potential US growth of 2.1% on average over the next decade, with AI-related gains potentially pushing that growth up to 2.4% by 2035.
Can AI massively revive growth?
We have identified three reasons why AI will not have a massive macroeconomic impact and will fail to deliver the double-digit growth touted by fervent tech evangelists. First, even with exponential technological progress, AI services may represent a decreasing share of GDP as computing costs fall. Second, physical constraints such as energy supply or supply chain bottlenecks could limit AI’s potential in the physical world, even if the scientific research process is significantly facilitated. Finally, the current architecture of AI may also have its limitations. Large Language Models (LLMs) such as ChatGPT, Gemini, or DeepSeek learn by predicting linguistic patterns. This makes them particularly effective for writing, summarizing, or programming, but does not necessarily give them a stable understanding of the real world – of objects, cause-and-effect relationships, and the accumulation of actions over time.
Some researchers believe that achieving truly general intelligence requires an “agent” with an explicitly acquired model of the world.
Other AI systems, often called broadcast models, can generate strikingly realistic images and videos. However, realism does not equate to understanding. Some researchers believe that achieving true general intelligence requires an “agent” with an explicitly acquired model of the world, using data that includes video, audio, and other signals. The goal is to create an AI capable of anticipating outcomes and planning its actions, with language serving as just one interface among many, rather than constituting the entire reasoning engine.
More generally, the progress made with LLM and AI will also depend on the timing and possibility of achieving AGI, or even artificial superintelligence (ASI), capable of surpassing human performance in several areas.
If LLM were to soon achieve artificial general intelligence, the risks to our forecasts would be tilted upwards. Conversely, if this path proved unsuccessful and a true agent with an explicitly learned “world model” were required, then the most substantial benefits of AI might materialize later. That said, even if progress at the cutting edge of LLM slows, the productivity gains consistent with our estimates above should still materialize in the coming years as AI adoption becomes more widespread and businesses learn how to leverage it.
The dominant approach in the United States relies on “closed” models… China, on the other hand, focuses on optimizing and standardizing LLM models
China and the United States are not playing on the same field
Geopolitical repercussions could also be possible if LLMs prove to be the “wrong” architecture for making major advances toward AGI and ASI, and if their performance gains were to slow. US tech companies are investing heavily to stay at the forefront of LLM development, with some beginning to explore models from around the world in parallel. With a few exceptions, the dominant approach in the US relies on “closed” models, allowing companies to retain and expand their competitive advantages. Some hope to build complete proprietary AI ecosystems around these leading models—much like Apple did with iOS or Microsoft with Windows in previous tech eras—that will eventually dominate the global market. China, on the other hand, is focusing on optimizing and standardizing LLM models through “open-weight” strategies, meaning models that can be downloaded locally to improve their accessibility.
Chinese authorities are also focused on integrating AI into businesses and industrial processes. For now, we believe both approaches have their place. Each trajectory leverages its respective advantages: cutting-edge innovation in the United States versus scaling up and broader adoption in China. Ongoing geopolitical tensions suggest that these two AI ecosystems will continue to develop in parallel.
