Hello! I'm your financial analysis partner. Today, we're not talking about the complicated numbers related to chip manufacturing processes, but about the "invisible champions" that lie behind the chips and determine the upper limit of AI computing power—semiconductor materials.
Over the past two years, everyone has been focused on NVIDIA's GPUs and TSMC's manufacturing processes, thinking that the smaller the transistors and the stronger the computing power, the more dominant that company would be. But the reality is that we're almost hitting a wall just by making transistors smaller alone. The real bottleneck has shifted to the materials.
The core point of this article is quite sharp: The competition in AI computing power is shifting from the "process dividend" to the "material dividend." Whoever can develop high-end materials that overcome these physical limitations will have the upper hand in the future computing power race.
Below, I'll break down this complex news into four sections that are easy for everyone to understand, so you can see what exactly this material revolution entails.
---
1. Not enough memory? Building taller chips while preventing "power leaks" and "deformation"
[Core issue]
Today's large AI models have trillions of parameters and calculate extremely fast, but the speed of memory read and write can't keep up. It's like having a wide highway (computing power) with only two lanes at the tollbooth (memory), causing a traffic jam. This is what's known as the "memory wall."
[Solutions and material breakdown]
To solve this problem, the industry has developed HBM (High Bandwidth Memory). Simply put, it involves stacking multiple layers of memory chips vertically to increase data throughput. However, this creates two major issues that require specific materials to overcome:
- Issue 1: Microscopic power leaks (electrons wandering around)
- Phenomenon: The higher the chips are stacked and the thinner the layers, the easier it is for electrons to escape through the material (quantum tunneling effect), leading to power leaks and heat generation.
- Saving grace material: High-K dielectrics. Think of it as a special type of "insulating glue" or "lock." Traditional silicon can no longer hold electrons securely; materials like hafnium (Hf) and zirconium (Zr) are needed to confine them within the capacitors.
- Who supplies these? Companies like Air Liquide in France, Merck in Germany, and Korean specialty gas companies. Their products directly affect the yield of HBM produced by giants like Samsung, SK Hynix, and Micron.
- Issue 2: Deformation due to thermal expansion and contraction (the building tilts)
- Phenomenon: The center of the chip gets hot while the edges stay cold, causing different materials to expand at different rates, which can cause the entire stack to warp or crack, and the pins to disconnect.
- Saving grace material: GMC (Granular Molded Compound). This acts as a heat-conducting and strong "protective layer" around the chip, preventing deformation due to temperature changes.
- Who supplies this? Currently, it's almost monopolized by a few companies like Sumitomo Chemical in Japan. The production capacity of this material is essentially the hidden ceiling for HBM production.
[Plain language summary]
HBM is essential for AI, but its production requires High-K dielectrics and GMC. These technologies are mainly in the hands of companies in Europe, America, and Japan, representing key points in the supply chain.
---
2. High electricity costs? Using "new materials" to make power supplies more efficient
[Core issue]
AI data centers are true power guzzlers—a single rack's power consumption has increased from 10 kilowatts to over 100 kilowatts. Electricity needs to be transformed from high voltage to low voltage suitable for chips. Traditional silicon-based materials are inefficient, with up to 10% of the energy wasted as heat. This is the "power consumption wall."
[Solutions and material breakdown]
To reduce power consumption, we need to replace traditional silicon-based power devices with third-generation semiconductor materials such as silicon carbide (SiC) and gallium nitride (GaN):
- Silicon carbide (SiC): Ideal for high-voltage, high-power applications with very low losses.
- Shift in the market: Previously dominated by American companies like Wolfspeed and Japanese companies like Rohm, the price was extremely high. Now, Chinese companies like Tianyue Advanced, Tianke Heda, and Sanan Optoelectronics have increased production, leading to a sharp drop in SiC prices. China has made significant progress and is offering better value for money.
- Gallium nitride (GaN): Great for high-frequency, fast-switching applications.
- Current situation: The technology for high-voltage, high-power, and automotive-grade GaN used in data centers is still dominated by European and American companies like Infineon and EPC. Chinese companies like Innosilicon and Jueneng Chuangxin are working hard to break this monopoly.
**[Plain language summary]
AI consumes a lot of power, and traditional silicon materials are insufficient. Replacing them with SiC and GaN can save a lot of energy. China has made significant progress in SiC, but Europe and America still lead in high-end GaN technologies.
---
3. Slow network speeds? Replacing "copper" with "light," relying on "Chinese resources and Japanese technology"
[Core issue]
AI clusters use thousands of GPUs working together, requiring fast communication. Copper wires are no longer sufficient for high-speed signals; we need to switch to fiber optics. This is the "interconnection wall."
[Solutions and material breakdown]
To upgrade from 800G to 1.6T or even 3.2T, traditional silicon-based optical technology is inadequate, so we need a mix of materials for heterogeneous integration:
- Indium phosphide (InP): The "engine" of optical signals, used to convert electrical signals into light signals.
- Resource advantage: China has over 70% of the world's indium reserves and production. With export restrictions, foreign companies like Coherent face tight supply issues, giving China a significant resource advantage.
- Technological challenge: Although China has the minerals, the core process of refining indium into high-purity single-crystal substrates (especially for large sizes) is dominated by Japanese companies like JX Metal and Sumitomo Electric. Chinese companies like Guangdian Technology and Youyan New Materials are making progress in this area.
- Thin-film lithium niobate (TFLN): Used to modulate electrical signals into light waves efficiently and with minimal loss. It's the key for ultra-high-speed optical modulators.
- Supply chain: The raw lithium niobate crystals are dominated by Japanese companies like Fujitsu and Sumitomo Electric. However, Chinese companies are leading in the downstream processes of thin-film fabrication and chip design.
**[Plain language summary]
We need to use light for AI communication. Producing optical components requires InP and TFLN. China has a strong resource advantage in InP, while Japan excels in manufacturing; China is making progress in TFLN technology.
---
4. Chips are getting too large? Traditional substrates aren't enough; special glass is the new solution
[Core issue]
As chip manufacturing processes approach 1nm, the performance of individual chips is reaching its limit, so multiple computing chips and HBM memories need to be combined in advanced packaging. As packaging sizes increase (from 50mm to over 100mm), traditional organic substrates deform easily under high-temperature treatments, causing issues with pin alignment and soldering. This is the "packaging physical limit."
[Solutions and material breakdown]
Companies like Intel, TSMC, and Samsung are turning to special glass substrates.
- Why glass? Glass is much flatter than organic materials, has a lower thermal expansion coefficient, and is more resistant to high temperatures, making it the perfect solution for large-scale packaging.
- Who supplies these? The formulas and melting processes for electronic and semiconductor-grade glass are highly patented. Currently, the three major display glass companies—Corning in the US, AGC in Japan, and NEG in Japan—supply the ultra-flat, defect-free glass required by Intel and TSMC.
**[Plain language summary]
As chips grow larger, traditional plastic/organic substrates become unreliable. Glass substrates are the future trend due to their flatness and heat resistance. However, this market is still dominated by US, Japanese, and Chinese companies are working to catch up.
---
In conclusion: Materials determine the upper limit of computing power
Putting these four points together, we see a clear chain of logic:
1. Memory bottleneck → Requires High-K dielectrics and GMC (dominated by Europe, America, and Japan).
2. Power consumption bottleneck → Requires SiC and GaN (China has made progress with SiC, and GaN is catching up).
3. Interconnection bottleneck → Requires InP and TFLN (China has a resource advantage, and Japan has strong manufacturing processes; China is working to improve its processes).
4. Packaging bottleneck → Requires special glass substrates (dominated by US and Japanese companies with high technical barriers).
Implications for everyone:
- Don't just focus on chip design: Half of the competition in future AI computing power will be determined by materials.
- Pay attention to bottlenecks and breakthroughs: Chinese companies are showing strong competitiveness and resource advantages in SiC, InP, and TFLN, representing investment and industrial opportunities.
- Supply chain security: Those who control the mass production and supply chain of these high-end materials will have the upper hand in the next generation of computing power.
This material revolution, driven by physical limitations, is just beginning.