by SHE Xiaochen, CHENG Lu

The 2026 Inclusion Conference on the Bund opened in Shanghai on September 10 under the theme "Building the AI Economy Together," bringing together researchers, investors and business leaders to debate how rapid advances in artificial intelligence could translate into broader economic growth.

At the conference's 15,000-square-meter technology exhibition, AI-powered devices were on display alongside robots and robotic dogs from more than 40 embodied-intelligence companies, highlighting how the technology is moving beyond software and into industrial and everyday applications.

But as AI capabilities improve at speed, a harder question is emerging: what does it take for a technological breakthrough to become a lasting source of productivity and economic growth?

At the opening forum, 2025 Nobel economics laureate Philippe Aghion, Princeton University professor Mengdi Wang and Source Code Capital venture partner ZHANG Hongjiang approached the question from different directions — market competition, scientific discovery and corporate organization.

Their arguments converged on a similar point: technological progress alone is not enough.

A breakthrough does not guarantee growth

Aghion began with the potential impact of AI on productivity. Research by him and his colleagues suggests that automation of tasks involved in producing goods and services could add about 0.68 percentage points to annual productivity growth over the next decade. Including AI's potential contribution to generating new ideas could raise the total impact to around 1.08 percentage points.

That would put AI's potential productivity effect on a scale comparable with, or even larger than, the information technology revolution.

But Aghion stressed that this was potential rather than an inevitable outcome.

He pointed to the US IT boom between 1996 and 2005, when rapid technological adoption boosted total factor productivity and helped create a new generation of dominant companies. As industries became more concentrated, however, the rate of new business formation declined and productivity growth subsequently slowed.

The tension goes to the heart of Aghion's work on "creative destruction": successful innovators need sufficient rewards to encourage investment, but dominant incumbents can eventually build barriers that make it harder for the next generation of companies to emerge.

That issue is already relevant to AI. Cloud computing in the US is dominated by a handful of large companies including Amazon, Google and Microsoft, while the market for advanced graphics processors is also highly concentrated.

Whether AI can deliver its full economic potential will therefore depend partly on whether markets remain competitive enough to allow new companies and technologies to emerge, Aghion argued.

The same applies to employment. AI will displace some jobs, while greater productivity and new forms of economic activity could create others. The question is how easily workers can make that transition.

Aghion argued that education should focus less on mastering a fixed body of knowledge and more on teaching people "how to learn." He also suggested preserving parts of the learning process that do not rely on AI, allowing students to continue developing independent reading, calculation and reasoning skills.

If Aghion focused on how AI enters the economy, Princeton's Mengdi Wang turned to an earlier stage of the innovation process: how far AI remains from independently making scientific discoveries.

Large language models have absorbed vast quantities of human knowledge, but Wang said their strength in producing the most statistically likely answer may also limit their ability to find genuinely novel ones, which often sit in the "long tail" of possible outcomes.

In recent research with Princeton sociologist Yu Xie, Wang's team used models including ChatGPT and Claude to simulate individual life trajectories. The models tended to over-predict common outcomes across areas such as marriage age, occupation and gender while underrepresenting less common possibilities.

Wang said this could help explain why large models have advanced particularly quickly in mathematics and coding, where answers are relatively easy to verify, while producing fewer comparable breakthroughs in fundamental sciences such as physics, chemistry and biology.

Scientific research depends not only on information, but also on physical experiments, expert judgment and collaboration across teams. Experiments themselves can also be difficult to reproduce.

As a result, giving AI more information does not necessarily provide it with more useful information. It may instead worsen the signal-to-noise problem.

Wang argued that AI will become more capable of exploring genuinely new scientific possibilities only when physical experiments can be systematically recorded, validated and reproduced, giving models more reliable feedback from the real world.

The challenge changes again when AI enters a company.

Zhang Hongjiang, a venture partner at Source Code Capital and foreign member of the US National Academy of Engineering, argued that AI is evolving from conversational systems into assistants, from reactive tools into proactive ones, and from individual agents into networks of agents.

One possible result is an "agent-to-agent economy", in which autonomous AI agents increasingly interact and transact with one another.

Software could be among the first things to change. Instead of people operating software directly, AI agents may increasingly operate it on their behalf.

But Zhang argued that the deeper transformation will take place inside companies themselves.

Beyond conventional assets, companies will increasingly need to manage proprietary know-how, models, computing resources, high-quality data, systems for coordinating AI agents and workflows that can continuously improve.

"The core competitiveness of companies in the future will depend on whether they can turn their IP, know-how, customer relationships and product capabilities into models, and enable those models to generate emergent capabilities," Zhang said.

Together, the three perspectives point to the same conclusion: faster and more capable AI models do not automatically translate into economic growth. Competition, scientific infrastructure and the ability of companies to reorganize around the technology will all matter.

New questions for the AI economy

A second debate running through the conference was whether massive investment in AI will ultimately deliver productivity gains — and what happens to demand and employment if it does.

Amazon, Microsoft, Alphabet and Meta are together expected to spend more than US$700 billion in capital expenditure in 2026, much of it linked to AI infrastructure. But economists still disagree widely on how much the technology could ultimately contribute to growth.

At a forum on September 9, Shanghai University of Finance and Economics president LIU Yuanchun argued that previous technological revolutions suggest productivity is rarely the first thing to move. Expectations, investment and financial markets often react before measurable gains appear in the wider economy.

Seen that way, the intense discussion, surge in capital spending and speculative activity surrounding AI may themselves be part of an early phase of technological transformation. Liu described bubbles as potentially playing the role of a "midwife" for a new technological paradigm.

That does not mean economists agree on the likely payoff.

Estimates of AI's contribution to annual total factor productivity growth over the next decade range from around 0.07 to 1.3 percentage points, according to Liu, largely because researchers make very different assumptions about how many tasks AI can replace.

Peking University National School of Development dean HUANG Yiping pointed to another uncertainty: timing.

The spread of computers did not immediately show up as faster productivity growth in macroeconomic data, an observation known as the Solow paradox.

"The use of AI does not mean productivity will immediately rise sharply," Huang said. AI could face a similar lag.

Thomas Sargent, the 2011 Nobel economics laureate and a professor at New York University, framed the uncertainty more fundamentally.

He compared the present state of AI with the progression from Johannes Kepler to Isaac Newton: AI has become increasingly good at finding patterns and fitting relationships in enormous datasets, but is still much less capable of understanding why the world works as it does.

That distinction matters when companies and investors are committing vast amounts of capital to models whose ultimate economic output remains difficult to predict.

Large investment does not guarantee equally large returns, Sargent argued, meaning businesses, investors and policymakers need to leave room for uncertainty.

AI is nevertheless already affecting the economy through several channels. Liu pointed to rising chip prices, rapid investment in computing infrastructure and stronger trade in AI-related products.

The longer-term question is what happens once AI meaningfully expands the economy's ability to produce.

If supply rises faster than demand, will consumers have sufficient income to absorb the additional output?

Huang raised the possibility that AI could simultaneously strengthen supply while changing employment and income distribution in ways that weaken demand. If so, what is currently seen in some economies as a cyclical imbalance between strong supply and weak demand could become more structural.

"I don't have an answer," he said.

CICC chief economist MIAO Yanliang raised a similar issue through inflation. In the short term, AI could push up some prices because of shortages in computing power and memory. Over time, however, productivity gains and labor substitution could prove disinflationary.

If a growing share of income flows to capital rather than labor, aggregate demand could also come under pressure.

The labor-market transition may meanwhile be unusually fast.

Morgan Stanley's chief China economist Robin Xing noted that steam power, electrification and the internet unfolded over two or three generations of workers, while the current wave of AI diffusion could take place within roughly a decade.

AI may initially raise the productivity of experienced workers while reducing some entry-level opportunities, creating a period in which new jobs fail to appear as quickly as existing ones are disrupted.

For China, Xing added, there is an additional challenge: it is moving through the AI transition at broadly the same time as the United States, leaving fewer established models to follow.

That brings the debate back to education.

XIONG Hui, associate vice-president at the Hong Kong University of Science and Technology (Guangzhou), argued that machines are best at capabilities that can be standardized, digitized and codified into rules.

Using the analogy of a "knowledge tree," he said education should give greater weight to its trunk and structure rather than spending disproportionate time memorizing individual leaves.

AI may have made knowledge more widely accessible, he argued, but it has not equalized people's ability to use it. When similar models are available to everyone, the differentiating skills become the ability to ask good questions, make judgments, draw on experience and create something from scratch.

Peking University professor SHEN Yan offered a different emphasis, warning against dismissing the value of accumulating broad knowledge. Knowledge and creativity are not mutually exclusive, she argued.

The broader consensus was that the AI era does not make knowledge unnecessary — but knowledge alone will no longer be enough.

Companies face an equivalent adjustment.

JIANG Xiaojuan, a professor at the University of Chinese Academy of Social Sciences and former deputy secretary-general of China's State Council, described AI as an innovation that is "born global."

As more Chinese industries move closer to the technological frontier, companies increasingly have to develop new technologies themselves rather than follow established paths. AI also makes models, data and research tools far easier to reuse, while digital products can reach international users almost as soon as they are launched.

Chinese open-source models have also rapidly expanded their international use, strengthening China's role in the global AI ecosystem.

The pharmaceutical sector illustrates both the opportunity and the bottlenecks. AI can sharply accelerate early-stage processes such as target discovery and molecular design, but clinical trials still operate on much slower timelines.

That mismatch is contributing to changes in the international division of labor in drug development, Jiang argued, as Chinese pharmaceutical companies increasingly license innovative drug assets overseas.

Ultimately, however, an AI economy cannot be sustained by capital spending or corporate cost-cutting alone.

HAN Xinyi, CEO and Executive Director of Ant Group, argued that once AI creates new sources of demand, the next challenge is to turn technological growth into employment and income, establishing a broader cycle between investment, income and consumption.

Huang made the same point from the perspective of distribution: higher productivity does not automatically mean higher wages.

An employment-first approach, he argued, should not mean asking AI to do less, but enabling more people to do more because of AI.

For all the uncertainty surrounding models, productivity and investment returns, that may ultimately be the more important test of the technology's economic impact: whether AI can move beyond technical breakthroughs and generate sustained growth that reaches a broader part of society.