Secondary market investors discuss the next round of AI opportunities at WAIC: domestic computing power chain, commercialization of model companies and valuation space

During the 2026 World Artificial Intelligence Conference (WAIC), at the "AI Insight, Co-Create the Future" AI Investment and Financing Forum hosted by Guotai Haitong, the roundtable discussion "Next Round of Investment Opportunities in the AI Era" gathered numerous investors from leading domestic and international financial institutions.

The guests participating in the roundtable include: WENG Qisen, Deputy General Manager and Chief Investment Officer of HuaAn Fund; JIA Jian, General Manager of the Research Department and Fund Manager of E Fund Management; CAO Jin, Senior Fund Manager of Fullgoal Fund; WANG Xiaojing, Chief Investment Officer of Equity, Quantitative and Multi-Asset at BlackRock Fund; and HE Xin, Chief Executive Officer of Societe Generale (China). The roundtable was hosted by CHEN Jie, General Manager of Equity Research Department of Fullgoal Fund.

The core issues of this discussion cover multiple aspects, including the continuous evolution of large model capabilities, the domestic computing power chain, the continuous expansion of global AI capital expenditure, and the next round of investment opportunities in the AI industry against the background of continuous geopolitical disturbances.

The following are the refined viewpoints of the roundtable discussion by "Suchbright":

01

Evolution of AI Technology: From Productivity Tools to End-to-End Solutions

Focusing on the most significant changes in the AI field over the past year, WENG Qisen first mentioned Anthropic.

In his view, the biggest change in the AI market over the past year is that Anthropic's coding code has officially become a productivity tool for programmers. "It has genuinely delivered a practical productivity tool," WENG Qisen said.

He mentioned that from his recent communications with contacts in Silicon Valley, he learned that Anthropic's large model coding code has entered the self-recursive stage, which means the model is in a self-iteration phase. Although the parameters of the models recently announced by Anthropic have reached a very high level, the more critical point is that relying on self-iteration capabilities, the scale of model parameters has the opportunity to continue expanding in the future. "The legendary 'stepping one's own left foot with the right foot' that we once thought only existed in martial arts stories has now truly happened in production tools, and it has happened completely."

JIA Jian's observations are close to WENG Qisen's, but he specifically regards the release of Claude Opus 4.6 as a landmark node. He believes that this model is of epoch-making significance because it has profoundly expanded the scope of AI applications.

On the one hand, it has truly turned the work of junior programmers, such as web pages and demos, into productivity tools; on the other hand, it allows people from non-computer science and non-professional backgrounds to access programming capabilities. "Code may be the reinforced concrete of the digital world. Once emergence occurs on the code side, it can actually empower a large number of industries," JIA Jian said.

He believes that in the future, we will see more and more people who have never written code begin to use AI to develop their own small tools. And this self-training model may further promote model self-training as capabilities emerge on the code side.

CAO Jin summarized the biggest change in AI this year as: shifting from selling computing power and Tokens to selling results and solutions.

In the past, the market's understanding of large models mostly focused on whether they could solve math problems, chat, or provide emotional value. But this year, the emergence of Agents has allowed large models to become a complete set of solutions: from perceiving the environment, to planning tasks, and then to specific execution.

"I think the real fission of large models is the gradual transition from the process of selling Tokens or computing power to the process of truly outputting results or solutions," CAO Jin said.

This also means that the business models and valuation logic of AI companies are changing. If a large model is just a Token consumable, investors will pay more attention to call volume, unit cost and gross profit margin; but if a large model becomes an Agent and part of enterprise processes and industrial systems, its valuation logic will be closer to that of software, platforms and infrastructure.

02

Collaboration Between Domestic Models and Domestic Hardware: Resolving Valuation Confusion in the Secondary Market

Different from domestic public fund managers who focus more on cutting-edge overseas models, WANG Xiaojing from BlackRock Fund has some different perspectives, placing the most important change over the past year on the domestic industrial chain.

WANG Xiaojing believes that the release of DeepSeek V4 has connected domestic large models and domestic hardware at the practical business level, which is of great significance for secondary market investment.

In the past, many domestic hardware manufacturers claimed that their valuations could be benchmarked against similar manufacturers in the US stock market, but investors faced a confusion: without solving the underlying business support, what justifies directly benchmarking against overseas valuations?

WANG Xiaojing believes that after domestic large models and domestic hardware are connected at the business level, they provide a solid underlying business support for the domestic hardware industrial chain. "Your valuation can now be linked to the domestic AI computing power demand, and at this point, the valuations of these underlying stocks become meaningful."

From this perspective, the investment logic of the domestic computing power chain is gradually shifting from "benchmarking against NVIDIA" or "domestic substitution" to "model business verification + hardware performance improvement + ecological collaboration".

HE Xin observed the AI industry from a more macro capital market perspective. He believes that the most important feeling over the past year is that the entire AI industry has been further capitalized in the secondary market.

In China, from Cambricon breaking through an important market cap threshold, to the full listing of domestic GPU enterprises, to Zhipin's impact on the "first large model stock" and DeepSeek's new financing, all show that the AI industry is being re-priced by the capital market. Overseas, platforms like SpaceX, which is a "computing power and model infrastructure platform", and OpenAI's preparation for listing, also reflect the same trend.

HE Xin mentioned that Societe Generale, as the only French-owned bank, participated in the IPO underwriting of SpaceX. "All these indicate that AI is beginning to be systematically priced by the capital market, which also means that it is not just a concept, but has truly developed into a long-term growing industry."

03

Evaluating Large Model Companies: Focus on Technical Parameters or ARR

When discussing how to evaluate large model companies, CHEN Jie raised a question: Should investors place more emphasis on the model's technical parameters, or on commercial data represented by ARR?

WENG Qisen believes that the two are two sides of the same coin.

In his view, the only product of a large model company is the "intelligence capability" of the large model, and this capability is also the only thing that can be monetized. Therefore, whether the model capability can be truly evaluated is a very critical factor. But as large models enter the Agentic stage, various Benchmarks are becoming more diverse. Investors should not only look at a single indicator, but also pay attention to the effectiveness of the model in different tasks.

More importantly, we cannot judge a company's capabilities only from a single point in time. "Judging a company's overall capability level and evaluating its performance in a single period only based on a single timing point may be one-sided," WENG Qisen said. Investors should focus more on a company's comprehensive capabilities, including talent, data, optimization capabilities and continuous iteration capabilities.

Commercial data is equally important. The ARR data of OpenAI and Anthropic has attracted market attention, which essentially reflects the monetization level of model capabilities. The stronger the monetization capability, the more it can strengthen the self-sufficiency of large model companies and further support team and technical iteration.

Returning to the domestic market, WENG Qisen mentioned that from DeepSeek's leadership in post-training reinforcement learning, to Tongyi Qianwen's enhancement of long-task capabilities, to Zhipin's catch-up with international giants in coding code, and Kimi's launch of a large model close to the North American SOTA level, investors need to continuously evaluate the comprehensive strength of different manufacturers at each stage to identify the true leaders and laggards.

WANG Xiaojing put forward another perspective. He believes that AI provides different services at different cost points.

For example, if a user asks AI to help find the most cost-effective combination of air tickets and hotels in the next three months, a top-tier AI may complete the task in 0.1 seconds but cost 10 yuan; a less capable AI may take 10 seconds but only cost 1 yuan or even 10 cents. For non-critical tasks, users do not necessarily choose the most powerful model.

Therefore, AI is not only competing in capability and speed, but also in cost. WANG Xiaojing believes that in the future, it is not certain that one or two top models will take over all industries. Different winners may emerge for different applications, different cost points, and different demands. "There is no need to prematurely judge which large model will eventually win. Maintaining a relatively inclusive view and focusing on the segmented winners of AI large models in each industry may be a more universal perspective for secondary market investment."

04

Export Controls and Sovereign AI: Chinese Open-Source Models Gain Global Opportunities

Since the beginning of this year, the release of the most cutting-edge large models in the United States has been increasingly affected by regulatory factors. CHEN Jie mentioned that the latest model releases of Anthropic and OpenAI have both encountered US export controls or government regulatory factors. Will such changes affect the risk pricing of American AI companies? Will it benefit Chinese open-source model companies?

JIA Jian believes that this will definitely have an impact, mainly in two aspects.

On the one hand, export controls will trigger a trust crisis on the client side, especially when doing business in non-US and non-European countries, which will damage the risk pricing and valuation of American AI companies. On the other hand, government regulation is equivalent to using government credit endorsement to tell the market which models are truly at the frontier, which to some extent has a positive impact on the valuation of these models.

But more importantly, this will trigger the issue of "sovereign AI" and the independent controllability of large models. JIA Jian mentioned that they have investigated many non-US countries in the past few years, and every time both government officials and technology leaders will mention that even if there is a gap between the capabilities of their own models and the top North American manufacturers, they will definitely insist on developing their own models.

Especially this time, after the United States has imposed similar restrictions on its allies, technology manufacturers in countries like South Korea will be more determined to promote the independent controllability of large models.

"During our investigation, we found that many non-US countries are drawing on Chinese open-source models when training their own large models." JIA Jian believes that this will obviously give China's top local open-source model manufacturers better development opportunities on the global stage.

CAO Jin believes that policy risks will definitely affect the risk premium, but it is more like a phased friction coefficient rather than a permanent risk.

He mentioned that some policies have recently reversed, for example, access to some models has been restored. This shows that the competition pattern of the large model industry is very fierce, and no enterprise can guarantee to always lead. If a large model enterprise is labeled as "closed" while another enterprise quickly releases new products, the former will also be greatly affected and may even actively lobby the government to restore access.

CAO Jin believes that for China's large models, this is definitely a phased benefit. Many AI users around the world prefer to use open-source, independently deployable, more secure and reassuring models. But he also reminds that this is more of a benefit from the perspective of access rate, which does not mean that China's model capabilities have surpassed or taken the global lead. Long-term competitiveness still depends on continuous capital expenditure investment, as well as the accumulation of computing power and data.

05

Hardware Landscape: NVIDIA Remains Strong in Training, While Decoupling Trend Is Obvious in Inference

The capabilities of large models are inseparable from the support of advanced AI hardware, but currently high-end AI hardware is still mainly monopolized by overseas manufacturers. Focusing on whether the monopoly positions of companies like NVIDIA and SK Hynix will be broken in the next three years, JIA Jian and CAO Jin gave hierarchical judgments.

JIA Jian believes that the chip side should distinguish between training and inference.

On the training side, NVIDIA has accumulated deep barriers since the release of the CUDA ecosystem. Its advantages come not only from chip performance, but also from the software ecosystem, developer community and tool chain. Therefore, in the next three years, external changes may not necessarily shake NVIDIA's advantages on the training side.

But on the inference side, JIA Jian believes that the loosening of barriers is only a matter of time. There are many new hardware optimization ideas on the inference side, including PD separation and more segmented AF separation. These software and hardware optimization opportunities provide a catch-up window for second-tier overseas manufacturers and domestic GPU manufacturers.

The storage side should also be viewed separately. DRAM still relies on planar scaling, while China is subject to great restrictions in the EUV field. Therefore, in three years, it is more likely that the market share will increase steadily, and it is difficult to achieve a huge leap. However, NAND relies on vertical stacking and does not involve key equipment restrictions such as EUV, so domestic top manufacturers have better opportunities to catch up with global giants.

CAO Jin also believes that in the next 2 to 3 years, whether in chips or storage, the advantages of overseas giants will still be very strong. But he observed an obvious trend: the entire industry is decoupling these leading giants.

Taking NVIDIA as an example, the GPU is far ahead on the training side, and there is a large gap between ASIC chips and GPUs. But whether it is AF separation or PD separation, the essence is to decouple a GPU that processes multiple tasks into different functional modules. In this process, more enterprises have the opportunity to enter the supply chain, including GPU enterprises and ASIC enterprises.

He believes that cloud manufacturers have been reducing their dependence on NVIDIA, so they will design many software solutions to differentiate functions. "We don't need a large, all-in-one super chip. What we hope is that more players will come in, reduce costs, achieve refinement, and improve efficiency."

Similar decoupling may also occur on the storage side in the future. In the past, when people talked about storage, they thought of the HBM production capacity bottleneck and insufficient CoWoS packaging production capacity, which also formed an oligopoly in the storage field, a pattern dominated by a few giants. But with the continuous decoupling of packaging and storage businesses, in the future, the businesses undertaken by giants such as SK Hynix and Samsung may become simpler and simpler, and other businesses will be more differentiated.

06

Domestic Chips: Only When the Entire Domestic Computing Power Chain Is Fully Connected Will There Be Greater Valuation Space

Focusing on whether domestic AI chips should prioritize inference or training, WANG Xiaojing believes that inference and training are equally important.

He mentioned that DeepSeek V4 has already brought incremental valuations to upstream semiconductor manufacturers on the inference side, because the market realizes that the combination of domestic models and domestic hardware can be independent of the overseas model system, forming its own computing power demand and business support.

At the same time, he also noticed that some domestic manufacturers have claimed to have fully realized training and inference on domestic hardware. For example, Meituan's LongCat 2.0 is said to have been completed on the domestic hardware chain. If this is verified, it will bring greater valuation enhancement to the entire domestic hardware industrial chain.

WANG Xiaojing believes that some general hardware modules, such as optical communication, have previously mainly benefited from the demand of the global computing power industrial chain. If they can also be applied in the domestic computing power industrial chain in the future, they may benefit from both the global computing power and domestic computing power demand curves at the same time.

Therefore, fully connecting the entire domestic computing power chain is very important for long-term industrial trends.

07

Can the Trillion-Dollar Capex Be Recovered

Global cloud manufacturers continue to increase AI capital expenditure. CHEN Jie asked: The combined capital expenditure of several of the largest American cloud manufacturers next year may exceed 1 trillion US dollars. Can such a huge investment be earned back in the future?