Last call: Apply for the Autumn 2026 Baiguan China Tour before this Sunday (Sep 13) to claim early-bird discounts.
Check out the last remaining available trips here.
Recently, we have invited several industry experts to talk about robotics, new energy, and supply chains in China. Today, we have invited another real industry heavyweight, who goes by “Dr. Joe,” to talk about one of the most strategically important industries of today: semiconductors.
Dr. Joe was an early co-founder of perhaps the most consequential semiconductor foundry business in China, and is still actively in the industry right now.
In this conversation, Dr. Joe will share an insider account of exactly how far China’s semiconductor industry still lags behind the Western world and what a self-sufficient Chinese semiconductor industry would mean for the rest of the world.
Baiguan: Most discussions of China’s semiconductor industry begin from the premise that its foundries still lag behind TSMC. But people disagree about the size and cause of that gap. Is the main constraint capital, talent, equipment, process know-how, or simply the time required to accumulate experience?
Dr. Joe: There is no simple answer because there are many kinds of semiconductor products. When people speak of “the gap,” they are usually referring to leading-edge CMOS technology, particularly at the five-nanometer and three-nanometer nodes. In that specific area, the single biggest and perhaps the only constraint is access to EUV lithography equipment, which is not available in China.
But most semiconductor products do not require EUV. In those areas, China is moving very quickly, and in some cases it is already leading. CMOS image sensors are one example. The major global suppliers include Sony, Samsung, and OmniVision, while Chinese companies have also become highly competitive in specialized applications. In areas such as security cameras and satellite imaging, some Chinese CIS products are already among the best in the world.
The picture is therefore very different from the common perception that China is behind across the board. In many specialized segments, the gap has already reversed. China’s semiconductor exports have also reached record levels over the past three years. The real lag is concentrated in leading-edge CMOS manufacturing, not in the semiconductor industry as a whole.
Baiguan: So your argument is that, wherever EUV is not essential, China is either already competitive or likely to become competitive relatively soon. The remaining gap is concentrated at the most advanced logic nodes.
Dr. Joe: That is correct. The clearest limitation today is advanced CMOS at five nanometers, three nanometers, and beyond. Even there, however, alternative architectures are becoming more important. Huawei’s Tau Scaling Law, 3D integration, and heterogeneous integration are all attempts to improve system performance without relying exclusively on conventional transistor scaling.
In the near term, these approaches can produce chips with performance comparable to products made with more advanced process nodes. They do not mean that EUV will never be needed. But they can reduce the importance of EUV for a period of time and create other routes to comparable system-level performance.
Baiguan: When you say China will eventually need EUV, how far away is “eventually”?
Dr. Joe: My prediction is that within five years, and possibly sooner, China will have an EUV solution or an alternative technology that serves the same purpose.
Baiguan: That’s a strong statement.
Dr. Joe: It is important to remember that the current ASML architecture is not the only possible way to generate EUV light. Elon Musk has recently signaled interest in free-electron laser, or FEL, technology for the proposed Terafab in Texas. This would still use EUV wavelengths, but the source architecture would be different. Instead of pairing an individual light source with every lithography machine, a very large accelerator-based or synchrotron-like source could potentially serve an entire fab, or even several fabs.
Both China and the United States are exploring this general direction. There is no law of nature that says advanced lithography must follow ASML’s present technical path. ASML’s system is the result of decades of successful engineering and iteration, but other engineering solutions are possible. ASML itself is a Dutch company, not an American or Chinese one, which is another reminder that this remains a global technological competition. That is why I am optimistic that EUV will become less of an absolute bottleneck within five years.
Baiguan: In other words, public debate may be too fixated on one company and one present-day architecture.
Dr. Joe: Exactly. Technology is not static. The semiconductor industry works with roadmaps that extend five to ten years into the future, and there are usually several proposed solutions to each major technical problem. The market ultimately decides which approach wins. Outsiders often take the most impressive technology of the present moment and assume it will remain the only viable path indefinitely. That is rarely how this industry develops.
AI chips, the memory wall, and 3D integration
Baiguan: Is this progress in alternative chip architectures one reason Chinese AI models such as Kimi, GLM, and DeepSeek have advanced so quickly?
Dr. Joe: That is a separate issue.
An AI model is primarily an achievement in algorithms and software. Hardware is another part of the system. Both the United States and China combine computing infrastructure with model development, but they do not necessarily have the same balance between the two.
If access to the strongest chips is limited, developers can compensate by simplifying models and reducing the amount of computing power required. That is part of what Chinese model developers have been doing, up until now.
If more advanced 3D and heterogeneous-integration chips become available, however, they will give those developers a stronger hardware platform and allow them to build more capable models.
Baiguan: So the potential of these alternative hardware technologies has not yet been fully reflected in the performance of Chinese AI systems?
Dr. Joe: Not yet. I expect the first important results to appear next year. I am directly involved in the development of a chip based on 3D integration, so I have been following this very closely.
NVIDIA’s GPUs are extremely powerful, and the memory capacity around them is very large. But the movement of data between memory and the processor remains constrained by what the industry calls the “memory wall.” A GPU may spend a large share of its time waiting for data rather than performing calculations. My own estimate is that this waiting time can reach roughly 70 percent in some workloads.
If 3D integration substantially shortens those data paths, much more of the GPU’s theoretical computing capacity can be utilized. In my view, a well-designed system based on this approach could match or exceed the performance of NVIDIA’s highest-end chips, even if the underlying logic is not manufactured on the world’s most advanced process node.
Baiguan: Jensen Huang has made the point that the rest of the industry is not standing still either. TSMC and other leading companies are also investing heavily in advanced packaging and 3D integration. Does China really have an advantage here, or is everyone moving in the same direction?
Dr. Joe: Everyone is moving. The question is not whether other companies are progressing, but who is ahead in a particular architecture and who controls the relevant intellectual property.
For instance, today’s HBM systems are generally based on 2.5D integration and chip-to-chip connections. The next step is wafer-scale 3D integration using wafer-to-wafer bonding. That is a more advanced form of integration. I believe the industry will need wafer-to-wafer technology by the HBM5 generation.
There is at least one Chinese company with important intellectual property in wafer-to-wafer stacking. I do not want to name the company here, but it is an example of a field in which a Chinese company may be ahead rather than behind. If the rest of the industry eventually adopts the same route, major international memory companies may need access to Chinese intellectual property.
Same thing is actually happening in memory. Historically, DRAM cell area moved from roughly 12F² to 9F² and then to 6F². The next generation is 4F², which is close to the theoretical lower limit for a conventional DRAM cell. Samsung, SK hynix, Micron, CXMT, and other companies are all working on this.
One company in China announced a 4F² product in 2024 and began delivering products in 2025. Samsung, by comparison, is targeting production around 2028. If that timetable holds, the Chinese company will be two or three years ahead. Combine 4F² cells with wafer-to-wafer integration, and China could become a leader in some next-generation DRAM technologies without first solving every problem in EUV lithography.
Baiguan: It doesn’t seem this type of information gets reflected in today’s market consensus at all.
Dr. Joe: Indeed. But engineers who attend the relevant technical conferences can see these developments, but many people in finance and the broader public still use an older picture of the industry. I expect that picture to change substantially over the next two years.
Baiguan: How much does China’s wider manufacturing ecosystem contribute to this progress?
Dr. Joe: China is already the world’s largest manufacturing center, and semiconductors are the most technologically advanced extension of that base. Different provinces have their own semiconductor programs, while the national Fifteenth Five-Year Plan identifies three especially important four-character concepts: 存算一体computing in memory, 三维堆叠3D integration, and 光电融合optoelectronic integration.
These are not uniquely Chinese ideas. The entire industry is moving in these directions. What matters is that they have been elevated into a national technology strategy, and Chinese companies are moving quickly in all three.
Computing in memory and 3D integration are intended to address the memory wall. HBM is extremely important today, but technically it is an awkward and transitional solution. The industry needs a more direct way to place memory and computing together. I expect samples based on a new approach next year and volume production the year after that.
Samsung, SK hynix, and Micron are also pursuing these technologies. My point is not that they will stop innovating, but that one Chinese company is currently ahead in a particular implementation. The same kind of intellectual-property dynamic has already appeared around YMTC’s Xtacking NAND architecture. In the future, some global leaders may need to license or otherwise work with technology developed in China.
Baiguan: Does the same optimism apply to semiconductor materials, including photoresists and silicon wafers?
Dr. Joe: Japan remains the leader in many semiconductor materials, and this is still an area of Chinese dependence. But there are now a very large number of Chinese companies conducting materials research. I expect another fundamental change within several years.
My broader prediction is that around 2032, the overall picture will flip. That is only six years away. In some categories China will still be catching up, but in others the direction of the gap will reverse.
In the paid sections below, Dr. Joe will continue to answer questions on:
How global talent flows have shaped and will shape the semiconductor industry
Why Dr. Joe returned to China
The triumphs and mistakes of SMIC, China’s largest domestic foundry
What a self-sufficient Chinese semiconductor supply chain would mean for the rest of the world.
Will China bring “over-capacity” for semiconductors as well?
