An Investor's 17 Judgments on Embodiment, Models, and Computing Power

Author: Xiaoyan, Tencent Technology

Editor: Xu Qingyang

In recent years, the hottest keywords in China's tech investment circle have revolved around AI, robotics, and embodied intelligence.

In the large model arena, Zhiyu is one of the first companies discussed in the context of "China's version of OpenAI." Jieyue Xingchen and Shengshu Technology are also at the center of popular directions such as foundational models and video generation. In the robotics sector, Ubtech has already entered the capital market, while companies like Galaxy Universal Robotics, Yinshi Robotics, and Tashizhi Hang represent different explorations of robotics from the body, models to scene implementation.

Behind these star enterprises is a common investment institution—Qiming Venture Partners. Founded in 2006, it manages 11 US dollar funds and 7 RMB funds, with a total managed asset amounting to 9.5 billion USD.

Having invested multiple times in technological waves, what is Qiming Venture Partners' methodology?

Recently, Qiming's managing partner Zhou Zhifeng shared insights on Qiming's investment methodology, discussing the underlying development trends in cutting-edge fields such as large models, embodied intelligence, and computing chips, and breaking down the core standards that differentiate concept hype from real industry implementation.

As an investor, Zhou Zhifeng is a true deep experiencer of AI. During discussions, he often approaches from everyday perspectives, painting a real picture of the era of AI technology implementation. He mentioned that his elders at home are already adept at using short video platforms and can smoothly use AI for daily information queries; in his own work environment, AI is also deeply integrated, as he relies on AI tools for data analysis and material organization, and even with a tight schedule, he continues to pay attention to the content iteration of AI-generated short videos.

These fragmented, real-life snippets intuitively validate a core trend: AI is no longer just a professional concept confined to laboratory theories or journal papers, nor merely a capital story lingering in financing press releases in the primary market, but has truly sunk into the mobile terminals of ordinary people, completing the leap from cutting-edge technology to a daily tool for everyone.

The investor's responsibility is to identify which technological directions, product forms, and companies are most likely to bring these changes to reality before they occur on a large scale.

Zhou Zhifeng summarizes this thinking as "half a step faster" investment—not necessarily the earliest, nor waiting for market consensus to chase in, but entering after the technological breakthrough and before the commercial explosion point.

Zhou Zhifeng, Managing Partner of Qiming Venture Partners

Taking Zhiyu as an example, in May 2020, the release of GPT-3 made Qiming Venture Partners realize that the Scaling Law was being validated, and large models had crossed an important technological node. Based on this judgment, Qiming invested in Zhiyu in December 2021. At that time, ChatGPT had not yet been released, and "All in AI" was not yet a collective slogan in the investment circle. It wasn't until November 2022 that ChatGPT burst onto the scene, and generative AI was truly pushed in front of the public and the venture capital market.

So, where will the next "half a step" opportunity arise in the hot directions of AI, robotics, computing power, and chips? During the on-site discussion, Zhou Zhifeng did not provide a direct answer but instead broke down the changes occurring in these tracks.

Faced with these hot tracks and the influx of hot money, Zhou Zhifeng believes that the closer one gets to the eye of the storm, the more one must return to the most fundamental question: Is there real value being created, and can it withstand the verification of commercial logic? He repeatedly mentioned that the market is likely to enter a "show me the money" phase. The newer the thing, the more one must be wary of the gap between short-term attention and long-term value; the noisier the market, the more companies must prove they can turn technology into revenue and imagination into reality.

Next, for reading efficiency, Tencent Technology has organized Zhou Zhifeng's core judgments in a first-person narrative.

01 "Embodied intelligence is the field that kills my brain cells the most, no exceptions"

1. After communicating with many embodied intelligence companies, I mainly have three feelings.

First, both the primary and secondary markets are particularly optimistic about this track. The core reason is that this may be the first industry in history that combines the "shipment scale of smartphones" and the "unit price of passenger cars." If this industry matures, there will be an annual shipment of 1 billion units, with an average price of about 30,000 USD, equivalent to 200,000 RMB. This is a top-tier track in the history of human commercial development over the past two to three hundred years, without exception.

Second, everyone is now scrambling for IPOs, essentially competing for scarcity dividends. The secondary market itself has this characteristic: when the first one or two companies in a large track go public, due to the scarcity of targets, they can enjoy super capital dividends, which is intuitively reflected in stock prices and market values that soar beyond conventional logic, so everyone wants to be the first to go public.

Third, many companies are becoming increasingly difficult to distinguish. We have monitored that there are over 370 companies related to embodied intelligence in China, and we basically receive two or three new projects every week. Their team backgrounds, technical routes, and implementation scenarios are becoming increasingly similar: they are mostly professors, little geniuses, executives from autonomous driving giants, or AI model backgrounds; they all talk about VLA and world models; and their implementations mostly revolve around industrial manufacturing, logistics, and commercial services, with recent appearances of bionic robot scenarios. But the problem is that there is currently no objective standard for evaluating technical levels or model capabilities that can truly determine who has stronger technology and implementation capabilities.

Therefore, after the first one or two companies go public, they may soar to a high market value due to scarcity. However, six months to a year later, the market is likely to enter a "show me the money" phase, focusing only on whether it can be implemented and whether it can be converted into sales revenue and profit margins.

If by the end of this year or mid-next year, the implementation does not meet expectations, even if successfully listed, the market value may drop to a few hundred billion range, and the primary and secondary markets will also experience valuation inversion. The original valuation support in the primary market will not hold, and the difficulty of subsequent financing for companies will significantly increase.

The key is still technology. If there are no key breakthroughs in the robotics track, especially if the technical route cannot converge, the industry will find it difficult to implement on a large scale. Currently, many companies exploring robotics are still using proprietary scenario models, rather than the highly publicized general models. If technical convergence cannot be achieved, large-scale implementation scenarios cannot be opened up, and in the end, they can only do some optional demo projects, making it difficult to achieve commercial scale.

So what I can do now is to continuously and actively look at all the new projects that come out, ensuring comprehensive information collection and continuous tracking of the industry landscape.

Currently, the primary market is still desperately raising funds, and companies with relatively stable scales are desperately preparing for IPOs. But ultimately, regardless of whether they go public or not, the market is still looking at commercialization. If the industry cannot produce real implementation results, the market is likely to face a deep adjustment.

2. I have never believed that world models are an entirely new track; the probability of VLA and world models merging in the future exceeds 50%.

It is more like a technical path that has been hyped into a popular concept in the primary market. Recently, about 30 new startups in world models have emerged, and compared to previous companies with VLA technical routes, there is no essential difference in commercialization implementation.

3. Currently, embodied intelligence lacks objective evaluation standards.

Language models have many benchmarks, but embodied intelligence faces the labor force of the physical world, making evaluation much more difficult. Currently, there are about three to five embodied intelligence benchmarks in the global market, but these lists have recently faced much skepticism. Some companies achieve high valuations by gaming the rankings, which industry insiders can see through as meaningless.

Before standards converge, we are currently more focused on bottom-up logic: first, whether the algorithm and model architecture routes align with our deductions; second, whether the team has rich engineering experience; third, data strategy. Data may be the most critical variable moving forward. Language models achieved the Scaling Law with 10 billion tokens; video models achieved it with tens of millions of clips. Currently, the top embodied model companies in China and the US have only about hundreds of thousands of hours of data, which is an order of magnitude away from sufficient scale. However, it is likely that the leading embodied intelligence companies in China and the US will reach this data scale this year, so breakthroughs may occur.

Once technological breakthroughs happen, evaluation will become simpler. For example, in an industrial scenario, if a large factory has 25 processes, without any post-training or only very simple post-training, we can look at the success rate of robots completing these tasks. If the success rate exceeds 50%, the large factory will genuinely pay to purchase the robots; if it is only 5%, that indicates it is still not feasible.

4. Data is a technical bottleneck that embodied models need to overcome, but it may change rapidly in the next year or two.

To create embodied models, it may require 1 to 2 million hours of training data, and at this stage, the quantity of data is more important than the quality of individual data points. Regarding various data pairing schemes, the industry professional term is data strategy, and this area has formed a consensus in recent months. Previously, the industry was divided into several data routes, such as the data collected by Tesla's self-developed real machines, which has the highest authenticity, making it easier for models to learn and adapt. When the models are deployed to hardware for execution, the match with the training data is entirely consistent.

However, the threshold for collecting this type of real machine data is extremely high. Even leading companies like Tesla find it challenging to advance, with a very limited annual output; it requires deploying 1,000 robots, with dedicated personnel, and each device can only effectively collect data for one or two hours a day, making it extremely inefficient. It may take ten years to accumulate 1 million hours of data, and if Tesla finds it difficult, other companies will find it even harder. This type of data is of top quality but is scarce in total. Earlier, Google and OpenAI tended to use video data; Google itself has deep expertise in video models, but the vast amounts of general video data are disconnected from practical robotic scenarios. For example, videos of conference room scenes are unlikely to teach robots practical skills and instead introduce a lot of low-quality noise data into the models.

Between the two extremes is the UMI data that has emerged in the past year, where workers' operations are recorded in real operational scenarios using wearable devices, making it easier to align with model training needs. Currently, leading companies in China and the US plan to procure a total of 1 million hours of training data this year, with real machine data accounting for only 1% to 3%, UMI data about 70%, and video data about 20%. Nuo Yiteng has split its motion capture business to develop independently, and motion capture technology can optimize UMI data and real machine data collection, now able to supply various training data across all categories.

In addition to scale, tactile data will also become important. For example, when a robot picks up a seemingly ordinary but actually heavier liquid, humans immediately perceive the change in weight and adjust their grip; however, current real machine data, texture data, and video data mostly lack this tactile information.

Therefore, there are now a number of companies attempting to create tactile fabric solutions to develop machines with tactile perception, collecting tactile data. This direction is hot for investment, but currently, there is no company in the world whose technology has reached complete maturity.

5. In terms of embodied intelligence models, China's advantages mainly lie in three areas: data, implementation scenarios, and hardware support.

It is currently difficult to quantify the technical differences between models in China and the US because the gap essentially lies in computing power. Technology has not fully converged, and technological exploration and research are particularly like sailing in the dark sea at night in search of Treasure Island.

The US does not have computing power limitations, and leading companies can send out 30 ships simultaneously every night. Each round of exploration feedback is crucial for finding direction, and each team reports back on their routes. For example, if they sail 5 nautical miles at a 30° angle today without finding the target, they won't need to repeat that route in the future.

China is currently limited by chip restrictions, allowing it to deploy only one ship per night, which highlights a core gap. The overall trajectory of large language models is already clear, so the perceived gap does not seem significant. However, if the industry experiences the next technological leap, statistically speaking, exploring 30 routes simultaneously compared to just one route significantly increases the likelihood of the U.S. achieving a technological breakthrough first. While the current model differences appear small, the long-term overall gap is not negligible.

However, China has clear advantages in data, industrial application scenarios, and hardware support.

Several leading companies in the U.S. are sourcing data from Chinese enterprises, indicating their own data reserves are insufficient.

Secondly, in terms of industrial application scenarios, China has massive manufacturing enterprises like CATL and BYD, which have ample physical factories for collaborative R&D.

Thirdly, regarding hardware support, humanoid robots consist of about 1,200 components, with over 90% of the supply chain concentrated in China's Yangtze River Delta and Pearl River Delta. Chinese companies can quickly iterate both the body and the model. Once a mismatch between the model algorithm and hardware execution is identified, suppliers can be adjusted and optimized within two weeks.

In summary, China has significant advantages in hardware and data, while the U.S. has strengths in model development, but the gap between the two is not vast.

Regarding the debate over whether humanoid robots are for "show" or "practical use," many discussions fail to clarify concepts.

Embodied intelligence-related algorithms can generally be divided into three main directions: Manipulation, Navigation, and Locomotion.

First, Manipulation involves physical task control, and both embodied intelligence and world models fall under this direction. Currently, the industry has not formed a unified and mature route. Second, Navigation technology has matured and has been applied to autonomous driving. Third, Locomotion refers to performance actions like running and martial arts, which are more about display.

All three belong to the realm of robotic AI algorithms, but the core that determines whether robots can create large-scale commercial value is still control technology. Locomotion has developed more maturely, with Yushun being a global leader in this area, and UBTECH also has a solid foundation. Therefore, it is normal for people to think they are merely showcasing their capabilities, as this is indeed their strong suit. Recently, over 360 new robotics companies have been established, all focusing on control; Yushun and UBTECH are well-funded and have built relevant R&D teams, so their capabilities in this area are also strong.

Looking solely at performance control scenarios, the global market ceiling is only about $1 billion. In contrast, the global scale of industrial manufacturing-related robotic application scenarios is much larger, with the two not being on the same level. Simply put, control technology matured earlier; in the past, robots could only perform functions like dancing and performing. It will take until this year or next for control-related technologies to converge, allowing robots to land in truly practical large-scale scenarios.

02 "Two Changes in the AI Field Exceed Expectations, One Change Falls Short"

In the next year or two, the valuation of AI companies will ultimately return to revenue and delivery capabilities. Traditional enterprise software companies may have a price-to-sales (P/S) ratio of 5-15 times, while those in hot sectors with leading technologies can reach 20 to 100 times. For companies like Zhipu, whether they can maintain their valuation depends on whether they can achieve significant revenue growth. If they can reach a revenue scale of 10 billion, a 100 times P/S would correspond to a market value of 1 trillion; however, if revenue is only 1.5 billion, the market valuation will face a correction, and the same logic applies to the robotics sector.

A company's ARR (Annual Recurring Revenue) represents its growth potential, while recognized revenue reflects cash flow. The robotics industry is no different; ultimately, it still depends on overall revenue, and these financial indicators are the fairest measure.

Therefore, the most important things for AI companies are two: whether model capabilities can continue to improve and whether real usage and revenue can be generated from the customer side. These two factors determine whether a company has long-term value.

In the past year, there have been two changes in the AI field that exceeded expectations, while one change fell short.

The first exceeding expectation is AI computing power. The total computing power and the speed of transition from training to inference, as well as the shift in computing paradigms or demand paradigms, have all exceeded expectations. For example, a major domestic tech company had a computing power budget of about 50 billion yuan last year, and this year's budget is more than six times that of last year.

Thus, whether it is the emergence of a large number of new-generation AI chip companies in the primary market or the speculation in the secondary market around HBM memory and optical communication sectors, the various hot trends in the industry are fundamentally driven by huge computing power demand, and the underlying logic is coherent. As for whether the short-term surge in individual stocks is reasonable, I cannot judge, but the overall heat and growth of the computing power market have indeed far exceeded my expectations.

The second exceeding expectation is the speed of development of model technology itself and the rapid formation of consensus around models in the market. For instance, in January of this year, with the emergence of intelligent agents represented by crayfish, coding capabilities became a core competitive advantage of large language models, which I did not anticipate when we discussed the top ten outlooks at WAIC last year, where we only mentioned that coding capability was important.

Because coding capability has brought about the capabilities of intelligent agents, I believe the industrial value of intelligent agents is exponentially higher than AI products dominated by chatbots in the past two to three years. Additionally, a positive feedback loop has formed: the computing power consumption generated by intelligent agents is thousands of times that of purely conversational products, which also explains why the growth in the computing power sector has exceeded expectations; these two are interrelated.

The development of model technology and the capital market's enthusiasm for model companies have also exceeded expectations. The speed at which the market has formed a consensus around high-quality model companies is extremely fast, with leading companies' market values able to reach trillions; a large number of neo-labs and new model startups have emerged, with founders mostly being post-95s and post-00s, and these projects can achieve angel round valuations of 2 billion to 3 billion, which I have never seen in such a hot market in all my years of experience.

Falling short of expectations is AI applications, especially 2C applications. Last year, I judged that 2025 would be the inaugural year of the AI application era. Looking at it now, the overall market for AI applications is still exceeding expectations, but the way it opens up is somewhat different from what I thought last year. Today, AI applications are mainly related to AI coding, including the development of intelligent agents, which I did not anticipate. I thought we might see AI truly empowering various industries this year, perhaps producing a few 2C applications with a bit of hope of becoming the next Tencent, ByteDance, or Alibaba, but it seems that no new generation of 2C applications has emerged that has particularly excited the entire market.

The first generation of AI applications established in 2022 and 2023, mostly represented by conversational tools and emotional companion products like CharacterAI, have now largely stagnated, and the industry is caught in a competition of product homogeneity. User growth has also slowed compared to the rapid growth of the previous two years. Our internal review concluded that the core issue lies in the fact that the user growth and traffic logic of the internet and mobile internet do not work for 2C products in the AI era.

AI toys and AI short dramas are examples. Some AI toy companies have sold hundreds of thousands of units, but 90% of users do not activate the AI interaction function long-term. Companies admit this is actually a good thing because if hundreds of thousands of users were all high-frequency dialogues, continuously consuming tokens, the company would not be able to bear the costs. In AI short dramas, the proportion of AI-generated content has increased rapidly, but it is difficult to produce true blockbusters.

This indicates that the short drama industry relies on a large-scale foundation, but core monetization is highly dependent on blockbuster works, and at this stage, AI is still unable to produce many blockbusters. This also illustrates that in artistic creation, human artistic expression and conception play a significant role; it cannot simply rely on AI to generate exquisite character visuals to support high-quality content.

In the past year, video model technology has achieved leapfrog growth.

The new generation of video models, such as the globally explosive Seedance 2.0, adopts the MoE architecture, significantly enhancing intelligent capabilities. It now supports 4K resolution. Because of this, many Hollywood films and advertisements from major brands like Coca-Cola and McDonald's have segments that are entirely or partially generated by AI, relying on the model's high-definition generation capabilities.

In this round of world models, they can empower video generation, achieve object movement and collision effects, and restore real physical laws, which was completely unpredictable a year ago. In the past year, related companies have seen rapid business growth, with leading players falling into two categories: three major global companies, ByteDance's Seedance, Kuaishou's Keling, and Google's Veo; and startups like our invested Shenshu Technology, Aishi Technology, and Video Rebirth, all of which have achieved tenfold growth in business and revenue.

Now, Hollywood, the advertising industry, wedding, and conference companies are all using these technologies. Various application scenarios have suddenly opened up, and I predict that the overall commercialization scale of the industry will see a significant increase this year.

Seedance, Keling, and Google's Veo 3 have core advantages in computing power and data.

Seedance, Keling, and Google can be considered a type, and even if Keling splits, it can still rely on Kuaishou's computing power and data support; the core advantage of these three companies is their own computing power scale, which gives them an edge over startups like Shenshu Technology. After the video model upgrade, both training and inference scales need to keep up, and these companies have tens of thousands to hundreds of thousands of cards, giving them a clear advantage.

However, I believe there are still opportunities for startups: the technology has not yet fully converged, and startups are not lagging behind large companies in terms of talent and technological exploration iteration speed. I believe Keling's decision to split will also help retain top talent. The underlying logic of the relationship between VCs and startups is that although startups are smaller, their equity incentive mechanisms and ability to concentrate all resources to exert effort give them advantages over large companies.

The market scale is rapidly expanding, and after scaling, the division of labor will become more refined, with clear differentiation in the commercialization efforts of various companies. First, regarding language models, the three leading companies in the U.S. have different user experiences; some feel that Gemini offers a better chat experience, but from a technical and industrial consensus perspective, OpenAI's ChatGPT has the largest user base and was the first to launch a conversational chatbot, with many optimizations focused on conversational scenarios.

For English conversational scenarios, ChatGPT leads globally in fluency; Gemini, backed by Google, has access to vast online data, with advantages in information retrieval and organization; Anthropic, starting from first principles, has advantages in coding and intelligent agent capabilities from the beginning, and the three have formed a clear differentiation.

Each video generation company also follows different routes: ByteDance focuses on the C-end, Keling targets B-end business, and our invested Shenshu Technology also focuses on certain B-end scenarios, with a very clear trend of industry differentiation. The requirements for model characteristics in B-end scenarios and C-end scenarios also differ significantly.

Now there is a hidden danger: after the formation of AI consensus, a large amount of hot money is flooding in.

After the consensus is formed, a large amount of hot money is entering the secondary market first. Currently, there are not many pure hardcore AI listed companies, and the pool of funds that can be absorbed in the secondary market is not large. It is now evident that hot money in the secondary market is starting to flow back to the primary and primary-and-a-half markets. Many companies have just completed financing and do not lack funds, yet institutions are still willing to raise valuations by 50% to 100%, immediately adding another round of investment. This influx of hot money has a significant impact on the industry; companies receiving funds beyond their needs may disrupt their strategic judgments and daily operations. However, I also understand entrepreneurs; when someone actively offers a higher valuation and a large amount of funding, choosing to refuse is inherently counterintuitive and difficult to do.

In the short term, this is more beneficial for us, but in the long term, the market will become very chaotic. As I mentioned earlier, there are now nearly ten embodied intelligence companies with valuations exceeding 20 billion, and more than ten companies with valuations of over 10 billion, all established only two or three years ago, which is itself quite abnormal.

Many companies are pouring large amounts of capital into the AI sector, which is likely to lead to chaotic competition: first, the cost of computing power is skyrocketing; a server from NVIDIA that originally cost 3 million yuan is now being traded for over 10 million, raising the overall cost of computing power across the industry. Second, there is a vicious competition for talent, leading to skyrocketing salaries in the industry. Third, on the client side, there is disordered competition, as companies lack mature commercial scenarios and can only cluster together to compete for major clients, merely comparing revenue scales.

These phenomena will harm the industry's development in the long run, and the current market is filled with irrational fervor.

12. Currently, the capital market has a kind of "mysterious infatuation" with young AI entrepreneurs.

First, most institutions missed the investment opportunities in large model companies two to three years ago. Many institutions did not invest back then due to a lack of conviction and determination regarding AI. By early this year, there was a consensus on large models, prompting many institutions to rush to catch up, which will certainly provide special capital dividends to newly emerging model companies.

Second, after DeepSeek emerged, many were shocked. Numerous media reports stated that the core team consists of doctoral students from Peking University and Tsinghua University, rather than veterans in the AI field. This has led to the impression that the younger the team, the smarter they are, and the less historical baggage they carry, the more likely they are to succeed. Many investors in the market now tend to trust young teams. This is not to say that young teams are bad; we have also invested in very young teams and seen many projects. I just believe that using the youth of entrepreneurs as the main criterion for investment is very subjective and unsustainable.

Third, some new cutting-edge model laboratories have emerged in the U.S. The core personnel of the top three overseas companies are now earning annual salaries of over ten million dollars. Some outstanding young researchers, due to their generous incomes and lack of financial worries, choose to independently establish new model companies, which also encourages many young people in China to try their hand at it.

The logic behind this wave is understandable, but when we evaluate individual projects, we will not simply invest because the founder is young. We must focus on whether their technical route is disruptive or capable of achieving a tenfold improvement. Even if the founder is just a doctoral student or a recent graduate without relevant industry experience, we will gather corroborative information from multiple sources to comprehensively verify the team's real hard power and the choices of their R&D direction, conducting a complete assessment and analysis.

13. In times of madness, we must heed the lessons of history. In noisy times, philosophical thinking may become more important; we need to think through and understand these matters.

For young companies that have been established for only two or three years, my advice is to look more at history. The bigger the wave, the more we should learn from historical lessons.

In fact, similar situations have occurred in the past during major waves—such as the internet and mobile internet eras—only the scale has become increasingly exaggerated. In the late 1990s, many individuals could raise funds and complete IPOs within two years. But ultimately, dust returns to dust; without truly creating value, even if they are favored by the capital market, they may still fail.

Thus, in times of madness, we must heed the lessons of history. In noisy times, philosophical thinking may become more important; we need to think through and understand these matters.

The essence of investment is that we invest in a company that can achieve scalability in the future, allowing us to realize returns through IPOs and other means. For example, we would invest in McDonald's but would never invest in a Michelin three-star restaurant. No matter how profitable a Michelin restaurant is, if it cannot achieve capitalization and lacks an IPO amplifier, it cannot provide an exit.

03 "If intelligent agents do not emerge in a year, the demand for computing power will be reassessed"

14. This year, several major tech giants in the U.S. have raised their annual AI computing power budget from over $700 billion to over $800 billion; China's budget is around $100 billion.

This statistical data may not be entirely accurate, but the direction is clear: this is currently the largest demand in human commercial society that is more certain than the robotics field and has a shorter landing cycle, with very clear demand.

To give some specific examples, a few years ago, everyone was focused on model training, but as of this year, we know that ByteDance's demand for computing power and inference has gradually shifted from a 1:1 ratio to more inference. Including Zhipu, the usage of tokens has increased; if it can maintain the growth of the past few months, its inference computing power will definitely increase rapidly. Essentially, major tech companies like ByteDance will see their annual inference computing power demand reach the level of millions of cards in two to three years. So this is a huge demand, very grounded and real.

Therefore, the market demand for GPU companies is very large. Even if GPU companies only capture 1% of this vast market, it means they can achieve annual revenues of tens of billions, and their corresponding market value can easily reach hundreds of billions. However, the market will first disperse and then converge. China can currently accommodate many GPU companies, but how many will remain in five or ten years? I believe it will definitely converge.

15. Currently, there are roughly three types of domestic GPU routes:

The AI cloud chip companies in the market correspond to three different technical routes, with the latter two having emerged only in the past year and being particularly active in the current primary market. The first type is domestic GPU manufacturers, with companies like Biren Technology, Muxi, Moore Threads, Kunlun Core, Cambricon, and Huawei as benchmarks. The current core factor is who has secured production capacity from the domestic supply chain. Currently, only a few companies have a promising outlook for stable and continuous supply.

Due to supply chain issues, the supply of advanced process capacity is limited, leading to the emergence of two new routes that can better meet future AI inference demands and avoid supply chain problems. The two routes are 3D stacked DRAM and DDR. Currently, there are nearly ten companies pursuing the 3D stacking route and DDR route, many of which are leading enterprises, with current valuations generally in the range of 10 billion to 20 billion yuan, and their financing scales are quite large.

The logic behind the market's optimism for these companies is straightforward: as long as the products can be mass-produced, they can immediately solve the supply capacity problem; furthermore, the future incremental space in the inference market is enormous, and these companies will always capture corresponding market shares.

16. The market's expectations for computing power demand are still wavering, which is something to be cautious about.

What the entire market needs to be most vigilant about is that just a few days ago, there was already a shock when reports emerged that Meta planned to sell off some redundant computing power. This news was not officially released by Meta, and the market immediately began to doubt all previous industry forecasts. That night, the South Korean stock market directly triggered a circuit breaker, with stocks of Samsung and SK Hynix plummeting, while related Hong Kong stocks and A-shares also fell in sync.

If subsequent AI applications fail to maintain sustained explosive growth, for example, if the development of agents does not show improvement after a year, or if commercialization does not advance further, or if the capabilities of models do not improve further and can only maintain the current level, the overall application scale will not be able to increase, and expectations for computing power growth will be shattered, leading to concentrated risks in the primary and secondary markets.

17. The gap between domestic high-end AI chips and NVIDIA lies primarily in the software ecosystem.

Currently, the chips used for model training are basically NVIDIA's, and all model training systems are built on the CUDA ecosystem. To achieve high efficiency and low cost in inference, compatibility with CUDA is essential. This is not just a challenge for Chinese companies; AMD has also struggled with this issue for over a decade.

Some changes are now emerging. First, compared to the AI 1.0 era, the convergence of large model algorithms is higher, making operator optimization relatively easier and reducing the adaptation barriers brought by CUDA. Second, in the past six months, the ability to write large model code has improved; now all non-NVIDIA chip manufacturers are using large models to automatically adapt operators. However, regardless of the developments, the CUDA ecosystem remains NVIDIA's biggest competitive barrier.

On the hardware side, it has been said for many years that our advanced power supply processes lag behind TSMC by a generation, and the number of transistors in chips is fewer. To achieve the same computing power, we must increase the chip area, which raises costs and increases heat dissipation pressure, leading to a series of chain problems. In simple terms, our current high-end chips lag behind NVIDIA by at least one generation.

Disclaimer: This content is provided for general branding and informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online events, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets or to use any services. Crypto assets are highly volatile and may result in loss. WEEX services and online events may not be available in all regions and are subject to applicable laws, regulations, and eligibility requirements. You are responsible for ensuring that your use of WEEX services complies with local laws and for carefully assessing the risks before participating in any crypto-related activities.