Artificial intelligence represents a fundamental shift in how human knowledge is produced. Historically, cognitive accumulation relied on individuals absorbing education and experience through years of learning and interaction. This traditional approach was highly personalized, inefficient, and subject to diseconomies of scale. Large language models have now industrialized the production of cognition, enabling scalable intelligence generation. This transformation gives rise to two pivotal economic questions: how does this new form of production organize its inputs and outputs, and how is the value it creates distributed, particularly in distinguishing value generation from value capture? These questions stem from the fact that cognition is a primary input into all economic activity, and its mass production is reshaping the division of labor and the fundamental structure of economic organization. The focus is on establishing a framework for analyzing these dynamics.

China and the United States currently lead the globe in the development of large language models. Social cognitive capacity can now expand by increasing the scale of computational power, data, and digital infrastructure. When asked about Europe's lack of AI progress compared to these leaders, one prominent US official attributed it to strict regulatory environments. However, this explanation falls short of clarifying why AI development in other non-EU nations also lags behind. The massive capital expenditure surge in AI has also triggered debates about potential stock market bubbles and whether future returns justify present optimism. Examining the economic principles governing these forces is essential to answering these broader questions about value distribution.

Where to begin

The technological principle of scaling laws dictates that consistent performance improvements demand continuous injections of computational power, data, electrical energy, and increasing parameter counts. Diminishing marginal returns set in once a model's size crosses a specific threshold, meaning each extra resource unit yields a progressively smaller performance gain. Economically speaking, this aligns with the concept of diminishing returns to scale, which carries crucial implications for investment behavior.

This technological framework transforms AI into a "heavy-asset" industry, representing a new phase of digitally industrialized production. Developing and training frontier models requires an enormous upfront expenditure in GPU clusters, specialized computing infrastructure, electricity, and top-tier scientific talent. The rapid technological iteration cycle in AI hardware means much of this fixed investment becomes a sunk cost. Furthermore, a significant share of the cost profile is variable. Every single inference, every model query, and every user interaction consumes real-time computing power and energy. Continuous 24/7 operation accelerates physical chip degradation. Though these marginal and variable costs are substantially lower than those in traditional manufacturing or services, they are significantly higher compared with the near-zero marginal costs of the "Internet 1.0" platform companies.

The heavy-asset nature of AI sets an exceptionally high barrier to entry. Only large corporations and vast economies can marshal sufficient resources to develop frontier models and pursue technological leadership, let alone aim for Artificial General Intelligence. The investment landscape underlines a stark differential between the US and China. While US AI capital expenditure accounted for over 40% of the country's GDP growth and more than 60% of total demand growth through the first quarter, its contribution to China's GDP growth during the first half of this year was a modest 0.1 to 0.2 percentage points. This stark contrast in investment impact can be traced to a few fundamental factors.

Investment demand is fundamentally driven by expectations of returns and the cost of capital. The extraordinary pace of US investment persists even in a higher interest rate environment, reflecting extremely optimistic corporate and market expectations. These expectations are mirrored in US stock valuations where the margin of safety over long-term government bonds has compressed to near zero. In contrast, investors in China's stock market still demand a premium of around 3 percentage points over the risk-free rate. Beyond financial considerations, a physical constraint plays a major role. Due to US restrictions on GPU exports, even when Chinese entities have the willingness to invest heavily in compute-intensive projects like data centers, they face hard procurement challenges that directly cap capital expenditure on AI.

From a macroeconomic perspective, total demand encompasses domestic output (GDP) plus imports. Therefore, AI capital spending in the US not only pulls in domestic consumption and investment, but also boosts global growth through increased imports. In return, other nations provide real resources that help suppress inflationary pressures within the US and help contain the upward trajectory of interest rates, which in turn sustains the AI investment cycle. This dynamic manifests itself as a trade deficit for the US and a corresponding rise in foreign net claims against the US economy. This pattern is reflected in financial markets, where a substantial portion of US equities is now held by overseas investors, often concentrated in AI-related stocks. On the surface, this creates a virtuous loop where other nations supply physical and capital resources to the US, thereby participating in the future income streams of the AI industry. Yet, this loop brings to the fore two central issues: how the current massive capital expenditure will translate into economic value and be allocated, and whether the resource endowment of large economies outside the US and China can realistically be mobilized to compete in foundational AI research.

A separate but critical concept, economic scale effects, helps answer how current investment boosts future supply. As a general-purpose technology, AI's dynamic evolution is inherently tied to economies of scale: scale promotes technological progress, which in turn enables higher-level economies of scale. Large economies benefit from deeper and wider trial-and-error processes, generating more reliable algorithmic innovation. This advantage is not solely internal to firms but also emanates from external economies of scale—the synergies from supply chain coordination, shared infrastructure, and pooled talent pools. Whereas the US holds a distinct lead in access to raw compute power, China has a substantial edge in AI talent. As of 2025, China accounts for 51% of top AI talent measured by undergraduate institution origin and 37% when categorized by primary employer, compared to 32% in the US. These comparative advantages mean China is positioned to progress via algorithmic efficiency, while the US can evolve with its hardware and compute dominance.

When concerning application, AI's economic penetration occurs on three major fronts. Firstly, it mass-produces certain human cognitive functions, thereby supplying a new type of intelligent labor that reframes factor matching. Secondly, it reshapes how economic organizations and markets operate, altering micro-level corporate behavior and necessitating novel public governance mechanisms. Thirdly, it acts as a powerful instrument for scientific and technological innovation. In all these dimensions, the scale benefits of a large economy are vastly amplified. Consider AI agents; the more enterprises and users that adopt them, the more societal value they generate. This involves the internal scale economies of a growing firm and the external network feedbacks stemming from interconnected data, application scenarios, and technology spillovers.

Combining the barriers to entry from scaling laws with the advantages of scale economies provides an explanation for why large economies like China and the US are at the AI frontier. Large economies can afford the heavy upfront costs, and because they are vast, they can maximize the value created from application. This offers a clear motivation for investment. Yet, this rationale still does not fully explain the lag in other substantial economies like France, Germany, Japan, or even India. A likely explanation is the "path dependency" of the digital era. Foundational model development depends heavily on the ecosystem built by platform behemoths from the "1.0" era of the internet. These giants possess the massive data, cloud infrastructure, and stable cash flows needed to absorb the high sunk costs of AI training, providing an unparalleled foundation that newer entrants in other economies struggle to replicate.

Why just 10 ASX 200 shares?

The debate over "winner-take-all" dynamics in the tech sector is central when moving from the Internet's 1.0 era to AI's 2.0. The market power of digital platform giants has long been scrutinized, and whether AI leaders will follow the same monopolistic path depends on their distinct value creation and capture mechanisms. From a supply-side perspective, platforms are a "light-asset" structure, whereas AI models are a "heavy-asset" model. This makes it nearly impossible for AI developers to mimic the strategy of offering free products at-scale to build market share. A more fundamental difference is in user stickiness and network effects. Platforms possess powerful network effects, where increasing user bases boost the whole ecosystem's value, creating huge switching costs and high "lock-in." In contrast, a large language model is a technical utility with incredibly low switching costs; users can easily migrate to another model if a competitor offers faster speed, better capability, or lower prices, leading to sudden and dramatic shifts in market share.

This structural difference implies that internet platforms with their strong network effects can solidify monopoly positions, internalize value, and sustain profitability. In contrast, the AI market is shaping up to be a dynamic oligopoly, where upstream firms suffer under the burden of "heavy assets" and downstream competitors are locked in brutal price wars. Because of the weak network effects associated with large language models, the value they generate is more accessible, spread across the entire economy and externalized to society—this is the nature of their external scale economies. A pivotal question concerns the link between AI models and platform giants. AI is a technology tool, while platforms are commercial entities. But they are interdependent. Artificial intelligence provides platforms with enhanced tools and infrastructure to expand their business, while simultaneously lowering barriers to entry in other segments, thereby increasing contestability in the market. Whether AI ultimately strengthens or weakens the market power of existing platforms remains to be seen.

The operational models chosen by AI developers also reveal key insights. In the US, companies like OpenAI and Google treat their top models as core strategic assets, often keeping them closed-source or restricting access to paid application programming interfaces to protect their competitive edge and commercial returns. In contrast, many Chinese developers, notably DeepSeek, have adopted an open-source approach, allowing a global community to fine-tune, iterate, and develop tools. Closed-source models benefit from internal economies of scope, reusing proprietary knowledge assets like code libraries and accumulated R&D across different projects to lower costs. Open-source models, conversely, rely on external scale economies, dramatically reducing the societal marginal cost of deploying and customizing AI. Thus, the closed model invigorates frontier innovation, while the open model undermines monopoly rents, creating what appears to be a healthy hybrid ecosystem that balances cutting-edge advances with broad-based development for the wider economy.

This analysis reinforces the fundamental question of whether current investor expectations for AI capital expenditures are overly optimistic. The economic characteristics of innovation suggest a powerful divergence. The benefits of inventions are broadly shared across society, but the risks and costs are shouldered by individuals and firms. This mismatch of "externalized benefits with internalized costs" is a major market failure that deters risk capital from flowing into pioneering, uncertain ventures. Historically, equity market bubbles have inadvertently played a corrective role by assigning extraordinary valuations to venturesome companies. This high valuation environment crowds in massive pools of social capital, overcoming the chronic undersupply of high-risk innovation funding. Bubbles inevitably deflate, and a portion of investors absorb those losses, thereby subsidizing the vast technological leap that benefits society as a whole. Meanwhile, the primary risk in the Chinese model is underfunding, which could lead to a technological gap widening at the frontier. In China, where equity markets are less fervent in funding heavy AI R&D, state-led industrial policy is arguably indispensable to priming the pump and directing resources to core strategic technologies.

Given these underlying constraints—the hefty financial requirements and the necessity of scale—it is unsurprising that China and the US dominate the frontier of foundational AI. Both nations can sustain the heavy capital load while enjoying the distinct benefits that massive scale provides for both R&D and mass market deployment. This position allows them to earn "Schumpeterian profits"—innovation rents earned from a temporary monopoly. But this advantage is not permanent. Human history demonstrates the strongly diffusive nature of knowledge and the difficulty in maintaining long-term excludability. In contrast to the monopolistic walling-off characteristic of internet platforms, AI's tech infrastructure relies more on mutual learning and ecosystem collaboration, making it inherently harder to restrict. The enormous public and private investments that the US and China are making will benefit their own people first, but they are equally, if not ultimately, providing a new fundamental basis for global economic growth and technological advancement for all of humanity.