Peng Wensheng: Industrial capital dominates AI innovation financing, US private credit plays an important role

Main text

By Peng Wensheng

Over the past few years, we have been discussing technology finance and how finance can support innovation. But let us consider: as a technological advance of epoch-making revolutionary significance, what kind of interactive relationship does the AI large model have with finance? What role has finance played in the development of AI large models, and conversely, in what ways might AI change the paradigm of technology finance? This helps us, from a broader level, explore the experience and insights of finance empowering innovation.

At present, China and the United States lead the AI large model track. One fact worth noting is that US AI capital expenditure is far ahead of China's. This has given rise to two concerns in the market: first, will high US AI capital expenditure breed an asset bubble? US stocks have recently seen a marked pullback. Second, will insufficient AI capital investment in China cause large model R&D to lag and widen the gap between China and the United States? This is precisely where finance needs to enter the discussion. AI capital expenditure is essentially investment in future supply. How should we understand the economic mechanism behind it, and what role does finance play?

When it comes to AI, the most intuitive feeling is that you can ask AI questions, which is not as troublesome as using a search engine to find information in the past. The disruptive innovation of AI large models lies in the scaled production of cognition. In the past, accumulating knowledge and building cognition depended on study, work, and interpersonal interaction, often requiring decades of effort; even so, each person's cognition remained limited, mostly confined to one's own professional and work domain, and the cost of acquiring cognition across domains was very high. AI large models have to some extent overcome this difficulty. In an economic sense, we can call this the scaled production of cognition. This brings up two questions: first, what is the input-output efficiency of this scaled production of cognition itself; second, how much value can the cognition produced at scale create at the economic and social level. Two concepts help us understand this issue.

The first is the scaling law. With AI large models, to improve model performance, one must continuously invest in computing power, data, and electricity. But as investment continues to increase, the pace of model performance improvement slows, which is diminishing returns to scale in economics. What does diminishing returns to scale mean? If improving model performance is the primary goal, large and sustained investment is needed. Unlike past soft-tech innovation, AI large models belong to a heavy-asset industry. The characteristics of heavy assets require capital expenditure to be front-loaded, and the US case is very representative. Can such investment achieve social returns? This leads to the second concept, economies of scale, namely how much value the produced cognitive capabilities can create when put into economic activity, which is related to scale. The larger the economic scale, the larger the population, the broader the industrial coverage, and the richer the application scenarios, the greater the economic value that the cognition produced by large models can create.

Putting these two concepts together, it is no surprise that China and the United States are in the first tier in the global AI field. The scaling law tells us that as the world's two largest economies, China and the United States are best able to mobilize resources, cross the investment threshold, and continuously invest in AI large model R&D; and economies of scale mean that China and the United States have the greatest potential to create value through AI large model applications, and therefore also the strongest incentive to invest in related innovation. Sufficient innovation resources and strong innovation incentives, combined, create the economic foundation for China and the United States to lead.

There is another important question here: even with resources and incentives, how should innovation activities be organized, or is the innovation mechanism effective? First, who organizes innovation? The integration and implementation of innovation resources, including financial support, cannot be separated from organizers. To borrow Schumpeter's words, it cannot be separated from entrepreneurship. In AI large model R&D, who organizes and mobilizes various resources? Is it independent venture capital, industrial capital, or some form of government force? Second, innovation has positive externalities: risks are borne by the innovating entities, but the benefits are shared by society as a whole. Relying solely on market mechanisms tends to lead to insufficient innovation investment. How can this problem be overcome? Generally speaking, internalizing externalities often involves a role for public policy, so as to better leverage the role of market mechanisms in allocating resources. Then how is this reflected in the development of AI large models? Has the government played a role? In what form has it played a role?

Industrial capital (CVC) is the dominant player in AI innovation financing

In the internet platforms of the digital economy 1.0 era, their startup and rise stages were mainly driven by financial capital (independent venture capital, IVC); whereas in today's AI foundation large model R&D, the main investors have shifted to industrial capital (CVC). Early internet platforms were an asset-light model, with controllable capital expenditure per project. During the startup phase of Sohu, Sina, Tencent, and Alibaba, the core investors were independent venture capital firms such as IDG, Sequoia, SoftBank, and Goldman Sachs. The investment goal was simply to bet on the company growing large, exiting via IPO or M&A, with financial return as the top priority. The large-scale rise of industrial capital CVC happened precisely after the first batch of internet companies successfully went public and held huge cash flows in hand.

The main vehicle for AI large-model R&D is large technology companies, not independent venture capital. The shift in the capital paradigm is essentially a shift from financial capital supporting business-model innovation to industrial capital supporting underlying heavy-asset technological innovation. Large technology companies benefit from the path dependence of digital economy development. In the digital economy 1.0 stage, leading internet platforms had already accumulated computing power and data resources, while mature business models brought stable cash flow, so they naturally had a first-mover advantage in large-model R&D. This also partly explains why other relatively large economies such as Europe, Japan, and India lag behind China and the United States in AI: these economies did not build up large domestic platform companies in the digital economy 1.0 era.

Looking at the data, although China's overall AI capital expenditure is lower than that of the United States, the two countries show similar structural characteristics: large technology companies, represented by hyperscale cloud service providers, account for the vast majority of AI capital expenditure in their respective markets, and capital investment such as computing power is mainly borne by such entities. Of course, this is not only companies' own funds; it also includes CVC (corporate venture capital), through which they invest in startup teams. The reason such industrial capital has an investment advantage is that it is rooted in the digital industry itself and possesses industry experience, technological accumulation, and supporting data and computing power resources.

Comparing China and the United States, we can observe a notable change: in earlier years, the scale of venture capital investment by large domestic internet technology companies once exceeded that of the United States; but in recent years, that is, after the rise of generative large models, this trend has reversed, and we have fallen behind. One possible reason is US export restrictions on advanced semiconductors to China. Another possible factor is that after the platform antitrust policy was introduced in 2021, CVC (corporate venture capital) activity by leading domestic technology companies contracted markedly, and the scale of their funding for startup projects fell significantly. Antitrust regulation is intended to address the excessively concentrated market structure of internet platforms and is a matter of fairness. The United States also has many controversies over the market power of platform companies, and how to balance innovation and fairness is an important issue in digital economy governance.

This raises a new question: do we need to worry that the current dominance of large-model innovation by large technology companies with first-mover advantages will create a new monopoly? The key difference here is that platform companies in the digital economy 1.0 era had an asset-light model and also possessed extremely strong network effects. Take social and short-video platforms as examples: once users form usage habits, it is very difficult for new similar products to replace them, and the advantages of the original platforms are hard to break. But large models are different: they are a heavy-asset model, network effects are relatively weak, and user switching costs are low. Today one may use Doubao, tomorrow Kimi, and the day after switch to Qwen or an overseas model; user loyalty is not high, and their network effects are far weaker than those of platform companies in the digital economy 1.0 stage.

Based on these business-model characteristics, our concern about monopoly risk in the large-model field should be lower than in the digital economy 1.0 era. But this also raises an important proposition: since massive resource investment is needed for large-model R&D, while it is at the same time difficult to form solid commercial barriers, how can commercial entities achieve profitability? This involves how to view the high valuations currently given by capital markets to AI-related targets.

That is the first analytical dimension I want to share, and it differs somewhat from conventional understanding: in the past, when we thought about science and technology innovation investment and "investing early and small," the main players were often independent venture capital, that is, financial capital; in this round of the AI wave, the leaders are industrial capital and large technology companies.

Lessons from US private credit

The second dimension may likewise differ from everyone's ingrained intuition. In the past, when we talked about science and technology innovation, investing early and small, and laying out hard technology, the first thing we thought of was often equity investment. But in this round of the AI wave, credit, especially US private credit, has played a very important role. The logic by which credit can play a role is not complicated: AI involves heavy-asset investment, essentially fixed-asset investment such as computing power infrastructure. Heavy assets have a certain collateral attribute; although their collateral effectiveness is less than that of land, they are markedly different from purely software intangible assets, and therefore naturally have a basis for obtaining credit support. Private credit has provided important support for the construction of data centers by large US technology companies.

At the same time, private credit serves small and medium-sized growth-stage science and technology innovation companies in the United States. For such companies, the prominent advantage of debt financing is that it does not dilute the founders' equity. But innovation activities naturally have high-risk, high-return characteristics, which are incompatible with the underlying logic of traditional bank credit. US private credit funds come from institutional capital, have higher risk tolerance and longer investment cycles, and feature highly customized contractual terms; they conduct valuation and pricing based on computing power assets, long-term orders, and expected future cash flows, showing high-risk, high-return characteristics. By contrast, bank credit is constrained by the rigidity of deposit liabilities, regulatory rules, and risk-control frameworks; credit assessment relies more on physical hard collateral and short-term verifiable cash flow, and its risk-control model is highly standardized, making it difficult to adapt to the capital-intensive investment and highly uncertain operating characteristics of small and medium-sized innovative AI companies.

In recent years, China's industrial guidance funds and science and technology innovation guidance funds have, to a certain extent, filled the funding gap caused by the relatively smaller scale of venture capital compared with the United States. But in practice, although this type of funding takes the form of equity investment in transactions, some projects use arrangements such as repurchase clauses, blurring the boundary between equity and debt. Some guidance funds have the characteristics of nominal equity but real debt, and their cash flow returns are similar to those of U.S. private credit, but they are not classified as credit in accounting and regulatory terms. Looking back, many businesses in China's shadow banking system were also nominal equity but real debt, performing a financing function similar to credit or private credit. This reflects a structural mismatch in China's science and technology investment and financing tools: there are funding arrangements similar to debt claims, yet there is a lack of a standardized and compliant private credit market.

In recent years, one innovation in China's science and technology finance sector has been science and technology innovation bonds, but issuers are mainly large enterprises. Banks are also expanding investment-loan linkage business, but investment-loan linkage is more suitable for growth-stage science and technology innovation entities with controllable risks and preliminary cash flow. A question worth considering is how, in the face of the iterative evolution of AI large-model technology and its scenario applications, to innovate debt-type tools to adapt to high-risk, high-return early-stage/high-growth innovation projects.

Patient capital is mainly about risk appetite

The third dimension: when discussing technological innovation and science and technology finance, we often mention "patient capital." But what exactly is patient capital? If one buys 50-year long-term government bonds, does that count as patient capital? It certainly has a long-term attribute, but the two are not equivalent. The core connotation of patient capital is a higher willingness to bear risk. As long as one is willing to bear risk, even if the investment horizon is relatively short, it can make up for the constraints that insufficient duration places on innovation activities. Therefore, patient capital is not just about long duration; more critically, it is about having a higher risk appetite.

Comparing Chinese and U.S. data, we can see that in terms of the number of investment projects, angel-round and seed-stage projects in the United States account for 41% of all venture capital deals, while in China the figure is only 9%. From this perspective, U.S. venture capital pays more attention to "investing early" and has a higher risk appetite, which is precisely a manifestation of patient capital.

We can also find evidence within China. Taking the sources of venture capital in Zhejiang Province as an example, social private capital accounts for the highest share of local venture capital funds. Although the total investment scale of venture capital in Zhejiang is less than that in Beijing and Jiangsu, the number of investment projects leads the country. The reason lies in the characteristics of social capital: it has a high risk appetite and is more inclined to invest early and in small projects, so with the same amount of capital it can cover more startup projects.

Here there is a macro paradox worth noting. Currently, the risk-free interest rate level in the United States is higher than that in China. According to economic principles, the interest rate is a macro indicator of the degree of patience: when interest rates are low, investors need a longer time to wait for returns; when interest rates are high, it means investors obtain higher short-term returns. In reality, the United States, under a high-interest-rate environment, still has strong risk appetite. For the S&P 500, the excess return investors require relative to 10-year Treasury bonds is almost zero. That is to say, when choosing between 10-year Treasury bonds and the S&P 500, investors do not require additional risk compensation, reflecting optimistic market expectations.

China's risk-free interest rate level is lower than that of the United States, but market risk appetite is lower and the risk premium is higher. In the current A-share market, investors require an excess return of about 3 percentage points relative to 10-year Treasury bonds, implying relatively cautious expectations for the future. Given a certain risk appetite, a decline in the risk-free interest rate prompts investors to reduce allocations to safe assets and increase allocations to risky assets. But risk appetite itself can also be a cause: when risk appetite rises, investors increase allocations to risky assets and reduce allocations to safe assets, which is reflected in prices as a decline in the risk premium and a rise in the risk-free interest rate. Therefore, a very important focus for cultivating patient capital is to shape investors' optimistic expectations for the future.

Of course, people may also raise doubts: a high risk appetite, when combined with finance, can easily lead to asset bubbles and financial risks, such as the large expansion of U.S. private credit mentioned earlier. It is worth noting that the asset-heavy model of large models is fundamentally different from the asset-heavy model of real estate. Simply put, AI-related physical assets such as chips have the characteristics of economies of scale: the greater the demand, the lower the unit cost and price in the long run. Land, by contrast, has the characteristics of diseconomies of scale: an increase in demand can only be balanced through higher prices (costs for other sectors), which also gives land an extraordinary collateral attribute and strengthens the procyclicality of finance. Therefore, the degree and harmfulness of the AI large-model bubble, including related financial risks, should be smaller than the bursting of the real estate bubble in 2007-2008 and the financial crisis it triggered.

Returning to the core issue of innovation externalities, this round of AI technological breakthroughs shows a noteworthy feature: the implementation and development of commercial large models has not been backed by government support. During the R&D stage of ChatGPT, the U.S. government did not provide funding support for the project; in China, large-model teams such as DeepSeek also did not rely on government investment in their initial stages. The development of leading commercial large models in both China and the United States was initially driven by market forces. This raises the question: how can market forces support this type of innovation with strong positive externalities in the absence of direct public subsidies? One possible explanation is that high valuations in the capital market significantly lowered the threshold for innovation financing. Looking ahead, if the development of the AI industry brings profit growth in line with expectations, high valuations will be digested; if not, then after the bubble recedes, some investors will bear losses, but such privately borne losses have objectively helped promote leapfrog technological development.

The role of government under the new geopolitical situation

This brings us to the final dimension. This round of AI development appears not to require government funding support, so where does the government's role lie? This time, the government's role differs from its past support for innovation and is roughly reflected in two aspects. The first is direct administrative intervention, delineating what can and cannot be done, such as the United States restricting exports of advanced semiconductors to China. The second is the government directly taking equity stakes in companies. A typical example is the U.S. government's investment in Intel, holding a 9.9% stake and becoming its largest single shareholder. The U.S. government's intervention aims, from the perspective of geopolitical competition and security, to bring advanced chip production capacity back to the United States. There are also relevant practices domestically; you may have noticed that the government previously took a stake in DeepSeek and holds voting rights. Now countries around the world are emphasizing sovereign AI. So in the AI era, the government's role is not only to support innovation but has also added a geopolitical dimension.

Finally, back to China. At the outset we mentioned that our country's AI capital expenditure is lower than that of the United States. The United States relies on optimistic expectations in the stock market, even bubbles, to support innovation, while our capital market does not have such strong optimistic expectations, so what can we do? How can public policy play a role in promoting technological innovation, including AI? Three points are worth considering.

First is fiscal investment. In 2020, a large portion of the revenue of leading AI vendors in the image recognition (computer vision) field came from smart city and security procurement by governments at all levels. Government procurement can cultivate early market demand and play an important guiding role in technological innovation.

Second, optimize the industrial guidance fund mechanism. As mentioned earlier, guidance funds have played a significant role in filling the funding gap for scientific and technological innovation. But government funds as LPs still have several problems: unstable policy expectations, insufficient market-oriented professional investment research teams, relatively short investment horizons, relatively weak risk tolerance, and some guidance fund projects have alienated arrangements resembling credit lending, among other issues. In response to these problems, relevant mechanisms need to be continuously optimized.

Finally, I think the most important point is that vigorous consumer demand is the most fundamental driving force for innovation. Therefore, fiscal and monetary policies need to work in coordination, introducing effective measures to address the current insufficient aggregate demand, boost consumption, and promote economic growth. This is not only needed for short-term growth stabilization but is also the fundamental logic for planning innovative development and long-term economic growth.

(The author is vice chairman of the China Chief Economist Forum; this article is his speech at the Tianjin Wudadao Financial Forum; the content has been edited)

(Source: Huaxia Times website) Middle of the main text Embedded advertisement