The AI Diffusion Challenge

The development of increasingly powerful models is central to the unfolding AI revolution. But this revolution has a second, equally important component: the adaptation and adoption of AI models across the economy, both to lower the cost of existing products and services, and to create new or improved products and services capable of advancing economic and social development.

Overall, AI will follow a J-curve pattern. At first, there is a huge amount of investment – in areas like physical infrastructure, software, business-model adaptation, data consolidation, and human-capital development – which does not yield immediate benefits. During this period, there is downward pressure on productivity, broadly defined to include benefits not measured by conventional national income accounts.

Then the technology’s value-creation potential kicks in, and the curve slopes upward. Since we haven’t yet reached this point, it is impossible to say exactly what this upswing will look like – the J-curve’s height and slope. By and large, investors seem to be betting on a massive payoff, but a distinct sense of uncertainty still permeates discussions about AI, and some predict that the technology will fall short of expectations, leading to a bust. Who turns out to be right will depend far more on diffusion than development.

So far, AI diffusion has been uneven, with some sectors (especially technology, finance, and professional services) embracing the technology, and others (including large-employment sectors like health care and construction) lagging behind. While such disparities are not surprising at this point, their persistence would lead to a flatter J-curve, representing muted returns on today’s investments and delays in growth and productivity gains. Put differently, whether or not we currently have an AI investment bubble will be largely determined by the pattern and speed of diffusion in the next few years.

Diffusion happens through multiple channels, the fastest of which is arguably software-as-a-service (SaaS) providers. Providers such as Google Search, Microsoft Office, Copilot by Notion, Salesforce, and Adobe are already embedding AI into their offerings. AI can also be incorporated into scientific processes relatively quickly. And with the major developers of large language and multimodal models providing application programming interfaces (APIs) that allow for the quick creation of tailored AI models, progress may pick up in other areas.

Open-source models – so far seen more in China than in the US – create even more opportunity, because they enable increased specialization and competition, including from smaller firms and countries that lack the massive computing infrastructure needed for the largest models. But there are still barriers to entry: a reliable electricity supply, robust computing capacity, and accessible mobile-internet connectivity are prerequisites to broad adoption.

Trade – especially of inputs like advanced semiconductors – also makes a difference. So does human capital: from advanced AI engineering and high-level strategic management to user-related skills, an economy needs to ensure access to an array of capabilities through education, reskilling, and labor mobility. The final piece of the puzzle is data. Where data systems are fragmented, incomplete, inaccurate, or inaccessible, training effective models will be slow, at best.

While AI diffusion depends significantly on private-sector initiatives, policy frameworks and regulatory structures also matter. China’s leaders understand this. As Huawei founder Ren Zhengfei recently observed, China has adopted a practical approach aimed at using AI to address real-world development and economic challenges. So, while developing increasingly capable large models is a high priority, so is deploying AI broadly, in order to secure the rapid gains in service quality, efficiency, and productivity that will be needed to offset the effects of rapid population aging.

China’s government is actively directing innovators toward these outcomes. Beyond encouraging the large tech platforms to build open-source models, China’s government has tasked them with developing or enabling applications in specific sectors, such as autonomous driving, health care, robotics (in manufacturing and logistics), supply-chain management, and green technologies. China’s government also regularly sponsors developer conferences and competitions.

Such efforts have paid off. For example, China accounts for over 30% of total global manufacturing output. In 2024, China accounted for 54% of all robot installations globally. The country now boasts almost half the world’s installed robots – at just over two million.

Relative to the US, China’s policy framework is much more engaged and geared toward providing direction with respect to applications and adoption across sectors in the economy. By contrast, US tech giants and well-funded AI startups are pushing the boundaries of large models, often in pursuit of artificial general intelligence and artificial superintelligence. While diffusion channels are open, their use is being left largely up to the private sector.

That may work in a few sectors, like tech, finance, and professional services, with the resources and know-how to experiment and then adopt. But private actors alone are unlikely to address the factors inhibiting AI adoption in specific sectors, such as data fragmentation, capacity deficiencies, regulatory hurdles, and scale problems. The likely – and unnecessary – result is a two-speed pattern of diffusion, leading to subpar economic growth, negative distributional outcomes, and the erosion of the economic underpinnings of national security.

When it comes to defense, the US government has long recognized that some state guidance is appropriate to ensure that private-sector innovation advances public goals. AI diffusion demands a similar approach. Something like this hybrid, active, pragmatic, and sector-specific approach is needed across a wide swath of the economy. Failure to do so will result in subpar economic growth, problematic distributional outcomes, and a weakening of the economic underpinnings of national security.

When it comes to diffusion, watching, waiting, and hoping is not a strategy.