AI’s productivity promise and the US-China race

How will Artificial Intelligence shape economic productivity, growth and employment in the decade ahead? We expect it to deliver a boost to the US economy that will more than offset the negative impact of changing demographics and policy headwinds. We also explore the implications of differing US-China strategies, and the path towards artificial general intelligence.

Global spending on AI is rising rapidly, driven by heavy investment in data centres, chips and software. As the main focus of AI investment, led by a handful of its largest tech firms, we expect the US economy to be reshaped by AI in the decade ahead, as it gets a productivity and potential growth boost.

‘Potential’ growth describes how much an economy can grow in the absence of shocks; it can be split into growth in the labour force and productivity growth. AI should increase the latter considerably. On average, we expect AI to raise US productivity growth from 1.60% to 1.85% annually over the next ten years. The benefits should be felt mainly later in the decade as AI adoption rises throughout the service sector.

AI should revive US productivity

This may sound like a small increase, but it would bring productivity back in line with, and even slightly above, levels experienced through the rise of the internet and personal computers in the early 2000s. It should also build on the recent increase in US productivity seen post pandemic. AI will have both direct and indirect effects on productivity. Direct impacts include taking on tasks currently done by humans and doing them in a more cost-effective way, raising ‘total factor productivity’ (TFP).

We expect these impacts to average 0.15 percentage points per year over the coming decade

We expect these impacts to average 0.15 percentage points per year over the coming decade. Indirect impacts include increasing capital per worker – see the ‘capital deepening’ illustrated in chart 1 – and helping generate new ideas and hasten technological progress, including assisting scientists, engineers or medical researchers. We estimate these indirect effects to be around 0.1 percentage points per year. Here, however, we acknowledge there is significant uncertainty, and this impact could be higher if some form of artificial general intelligence (AGI) – in line with that of humans across a wide range of tasks – were achieved soon.

AI should not drive broad unemployment

Despite the potential for AI to assume human tasks, we expect its impact on unemployment to be contained and gradual. AI has already been responsible for some layoffs in the services industry, notably at tech firms, which are now doing more with fewer people, as well as a tougher job market for recent graduates.

However, unlike in a recession where the lowest income workers tend to suffer, AI has greater implications for educated, middle and upper-middle income workers. Such individuals often have skills that can either allow them to focus on different tasks as some are automated by AI, or ultimately help them find alternative occupations as the economy adjusts. We therefore only expect a modest rise in ‘frictional’ unemployment that takes place gradually over the coming decade as AI adoption rises and the economy adjusts. This impact should also be softened by other labour force trends, which we explore below.

related content.

investment insightsUHNWI

investment insightsUHNWI

Insurers and private assets’ marriage of convenience

December 23, 2025

investment solutions.

learn more

AI should more than offset lacklustre demographics

Of course, the AI productivity boost is not happening in isolation. Broader US economic trends, including changing policies and demographics, are less favourable. We expect the former – including efficiency losses from tariff-related supply chain disruption, productivity losses from lower-skilled immigration, and threats to the credibility of US institutions – to shave -0.15% off potential growth in the coming decade. However, most of these effects will be felt in the short term and may even be reversed under later US administrations.

Demographics are also going to be much less favourable. Since 2019, high immigration helped the US labour force expand, contributing almost a full percentage point to US growth each year. However, an ageing population and sharply falling immigration should slow this labour force growth considerably.

…we expect potential growth in the US to average 2.1%

The combined impact of an AI boost, negative policy trends and less favourable demographics means we expect potential growth in the US to average 2.1% over the next decade, with the benefits from AI raising US potential growth as high as 2.4% by 2035.

Could AI deliver a massive boost to growth?

We see three reasons why AI will not have a massive macroeconomic impact and will fail to deliver the double-digit growth that is sometimes forecast by evangelists for the technology. Firstly, even with exponential progress, as computing power becomes cheaper, AI services could become a progressively smaller share of GDP. Secondly, physical constraints such as energy supply, or supply chain bottlenecks, may limit AI’s potential in the physical world, even if the process of scientific discovery is made significantly easier. Finally, our current AI architecture may also have limits. LLMs such as ChatGPT, Gemini and DeepSeek learn by predicting language patterns. That makes them excellent at drafting, summarising and coding, but does not necessarily give them a stable grasp of the real world - objects, cause and effect, and the accumulation of actions over time.

Some argue that the missing element… is an ‘agent’ with an explicitly-learned world model

Other AI systems, often called diffusion models, can generate strikingly realistic images and video, but realism is not the same as understanding. Some researchers argue that the missing element for truly general intelligence is an ‘agent’ with an explicitly-learned world model, trained on data including video, audio and other signals. The goal is AI that can anticipate outcomes and plan before acting, with language as one interface among others – rather than it forming the entire engine.

Progress with LLMs and AI more broadly will also depend on when, and whether, we can achieve AGI, or even artificial superintelligence (ASI) - the capacity to exceed human capabilities across several domains.

If LLMs can achieve artificial general intelligence soon, risks to our estimates are to the upside. On the other hand, if they prove to be the wrong path and instead a proper agent with an explicit learned ‘world model’ is necessary, then the most substantial benefits from AI may be further down the road. That said, even if progress at the frontier of LLMs slows, productivity enhancements in line with those we estimated above in the next few years should still materialise as AI adoption spreads and companies learn how to leverage it.

… the predominant approach in the US is mostly ‘closed’ models… China is focusing mostly on optimising and commoditising LLM models

China and the US are running different races

There would also be geopolitical implications if LLMs prove to be the ‘wrong’ architecture to make major advances towards AGI/ASI, and LLMs’ performance progress eventually slows. US tech firms are investing huge sums to stay at the cutting edge of LLM development, with a few US firms beginning to explore world models in parallel. With some exceptions, the predominant approach in the US is mostly ‘closed’ models where companies can build and maintain their advantages. Some US firms may hope to build entire proprietary AI ecosystems around these cutting-edge models – analogous to Apple’s IoS or Microsoft Windows in prior tech eras – that end up dominating the global market. Conversely, China is focusing mostly on optimising and commoditising LLM models, employing ‘open weight’ strategies, or models that can be locally downloaded, to improve accessibility.

Chinese authorities are also focused on embedding AI in companies and industrial processes. For now, we think there is space for both approaches. Each of these paths plays to respective advantages – frontier innovation in the US versus scaling and wider adoption in China. Enduring geopolitical tensions suggest both of these AI ecosystems will continue to develop in parallel.