Over the past two years, the global semiconductor industry's focus has been almost entirely on NVIDIA's GPUs, TSMC's (Taiwan Semiconductor Manufacturing Company) advanced process nodes, and the data center capital expenditures of tech giants, which often reach tens of billions of US dollars. From the continuation of Moore's Law to breakthroughs in shrinking process nodes, the entire industry seems to be cheering for the geometric growth in the number of transistors.
Data Source: MIR DATABANK
From the perspective of the domestic industrial landscape, China's semiconductor wafer manufacturing materials market has shown an overall trend of steady expansion and counter-cyclical growth, returning to a high-speed growth track after a brief adjustment in the industry in 2023. Given the continuous release of capacity by domestic wafer fabs and the rapid rise of the AI computing power industry, market demand for high-end semiconductor materials continues to surge. The focus of industry growth has comprehensively shifted from traditional basic materials to high-end functional materials adapted for AI data centers, advanced packaging, and high-speed optical communications, ushering in structural growth dividends for the domestic materials industry.
However, as tens of thousands or even hundreds of millions of high-performance chips are densely packed into super data centers, the physical limits of traditional semiconductor materials are being rapidly exhausted. Physical bottlenecks such as severe attenuation in electrical signal transmission, soaring chip heat generation, and structural deformation under high integration levels are intertwined, forming insurmountable physical walls.
The evolution of computing power is shifting from "process node dividends" to "material dividends." The era of relying solely on transistor scaling has slowed down, and the stability of the entire AI computing power pillar is tilting towards downstream advanced chemical material systems. Under the current market landscape, the core criterion for evaluating the importance of a semiconductor material lies in whether it can solve the four most critical bottlenecks in data centers: the memory wall, the power wall, the interconnect wall, and the physical limits of packaging.
The main materials currently constraining the release of AI computing power include High-K high-dielectric precursors and GMC (Granular Epoxy Molding Compound), Silicon Carbide (SiC) and Gallium Nitride (GaN), Indium Phosphide and Thin Film Lithium Niobate (TFLN), and special glass substrates. We will break them down one by one below.
01. Breaking the "Memory Wall": The Rigid Demand of HBM for High-K Precursors and GMC
The physical essence of the memory bottleneck. Against the backdrop of generative AI large models with tens of billions or even trillions of parameters, the memory bandwidth bottleneck caused by "computing power running faster than memory read/write" has become the absolute shortboard constraining the overall performance of AI servers. To break this limitation, High Bandwidth Memory (HBM), which vertically stacks multiple layers of DRAM chips, emerged as the times require and has quickly become the standard configuration for AI chips.
However, this "vertical building" model of HBM, while greatly expanding the width of data throughput, also brings two extremely harsh physical defects: severe microscopic leakage and serious structural thermal stress.
High-K (High Dielectric Constant) precursors are the leakage locks at the microscopic scale. As DRAM process nodes approach 10nm (1α/1β/1γ) and more advanced nodes, the gate capacitance dielectric layer of transistors becomes extremely thin. Traditional materials such as silicon dioxide can no longer effectively prevent the quantum tunneling effect of electrons in the microscopic world, leading to soaring leakage rates and power consumption. To maintain the charge storage capacity of capacitors and lock electrons, the industry must comprehensively shift to advanced High-K precursor materials such as hafnium-based (Hf) and zirconium-based (Zr) materials (gaseous sources).
Currently, the high-end market in the field of High-K precursor materials is mainly dominated by France's Air Liquide, Germany's Merck, and some local specialty gas companies in South Korea. The stability of their supply chains directly determines the HBM production line yield of the global top three memory giants (SK Hynix, Samsung, Micron).
GMC (Granular Epoxy Molding Compound) is the protective shell that solves thermal stress warpage. During the manufacturing process of HBM, 12 or even 16 layers of bare dies are connected together through TSVs (Through-Silicon Vias) and microbumps, with the overall thickness compressed to the micrometer level. In the working state, the temperature difference between the center and the edge of the chip is extremely large. The mismatch of the Coefficient of Thermal Expansion (CTE) will cause severe warpage, cracking, or pin breakage of the entire stack. At this time, Granular Epoxy Molding Compound (GMC) must be introduced for high thermal conductivity and void-free encapsulation.
Currently, the global high-end GMC market for HBM is almost monopolized by a few veteran material manufacturers such as Japan's Sumitomo Chemical. The capacity and quotas of such underlying materials actually constitute an invisible limitation on the release of HBM capacity.
02. Conquering the "Power Wall": How Gallium Nitride and Silicon Carbide Achieve a Power Efficiency Revolution
Current AI data centers are transforming from traditional "computing centers" into "power consumption black holes." The power consumption of a single standard AI rack has soared from the past 10kW to 100kW or even higher. The high-voltage AC power introduced from outside the data center needs to undergo multiple processes such as voltage transformation and rectification, and is finally converted into the low-voltage DC power of 1V or even a fraction of a volt required by the GPU chip cores. In this multi-stage conversion power transmission chain, power devices based on traditional silicon (Si) materials (such as MOSFETs and IGBTs), due to the inherent physical limitations of their materials, have huge conduction losses and switching losses, resulting in about 10% of electrical energy being wasted in the form of waste heat.
To conquer this "power wall," third-generation (wide bandgap) semiconductor materials represented by Gallium Nitride (GaN) and Silicon Carbide (SiC) are setting off a thorough substitution wave in AI server Power Supply Units (PSUs).
Data Source: MIR DATABANK
In the past, the global 6-inch SiC substrate market was long monopolized by US companies Wolfspeed and Coherent (formerly II-VI), and Japan's Rohm, with prices remaining high. However, with the concentrated release of capacity by domestic substrate giants such as Tianyue Advanced, Tianke Heda, and Sanan Optoelectronics, global SiC substrate prices have experienced a cliff-like drop.
Currently, Chinese companies have caught up in market share in the field of silicon carbide. Among the GaN devices used in data center power supplies, the domestic supply chain for consumer-grade GaN (such as mobile phone fast charging) is already highly mature. However, for the high-voltage, high-power automotive/industrial-grade GaN epitaxial wafers required by data centers, the core technologies (such as GaN-on-Si epitaxial growth and lattice defect control) still highly rely on European and American giants such as Infineon (which acquired GaN Systems), EPC, and Navitas. Domestic companies such as Innoscience and Juneng Chuangxin are going all out to accelerate penetration into high-voltage, high-current data center power-level products.
03. Breaking Through the "Interconnect Wall": The Trend of Multi-Material Heterogeneous Integration in the 1.6T Era
The essence of AI clusters is the collaborative computing of tens of thousands of GPUs. The interconnect bandwidth between chips and between servers directly determines the effective output rate of overall computing power. "Optics advancing, copper retreating" — using optical fibers instead of copper wires for signal transmission between racks and even at the board level — has become an irreversible iron law in the industry. At the node where 800G optical modules are evolving towards 1.6T and 3.2T, the limitations of traditional single "Silicon Photonics" materials under ultra-high-frequency modulation, due to the lack of a direct bandgap and weak electro-optic effects, are completely exposed. Heterogeneous integration of multi-material fusion has become the most cutting-edge market reality today.
As a second-generation compound semiconductor, Indium Phosphide, with its excellent direct bandgap physical characteristics, is the core substrate for manufacturing high-speed optoelectronic devices (such as laser chips and detector chips).
The upstream of Indium Phosphide is the rare scattered metal "indium", with over 70% of the global indium reserves and production located in China. With the implementation of relevant export controls, the raw material supply chains of overseas optical module giants are extremely tight. However, in the core link of pulling metal indium into crystals to produce high-purity single-crystal substrates, Japan's JX Metals and Japan's Sumitomo Electric still occupy the major global shares. Although domestic companies such as XianDao Technology and Youyan New Material have achieved absolute independence in the resource end and low-to-mid-end substrates, they are still in the critical stage of high-performance sample validation and capacity ramp-up for 6-inch/8-inch ultra-high-purity indium phosphide single-crystal substrates facing 1.6T optical modules.
If Indium Phosphide is the "engine" that generates optical signals, then how to efficiently and distortion-free modulate the electrical signals output by the GPU onto optical waves is another core technical bottleneck. Thin Film Lithium Niobate (TFLN), with its extremely high electro-optic coefficient, ultra-wide modulation bandwidth, and extremely low insertion loss, is becoming the definitive route for 1.6T and above ultra-high-speed optical modulators.
The supply chain structure of Thin Film Lithium Niobate is very special. As the source of lithium niobate wafers (Bulk Crystals), Japanese optoelectronic giants Fujitsu and Sumitomo possess almost absolute dominance. However, Chinese companies are at the forefront of the world in downstream "thin-film (Smart Cut)" processes, lithography etching, and chip design. Represented by domestic manufacturers such as Advanced Fiber Resources (AFR), they are joining forces with upstream domestic equipment manufacturers to accelerate the breaking of Japan's monopoly in the field of modulator chips, gradually achieving full industry chain independence from wafer imports.
04. Subverting "Advanced Packaging": The Substitution of Traditional Substrates by Special Glass Substrates
As chip process nodes approach the physical limit of 1nm, relying on advanced packaging technologies represented by TSMC's CoWoS to "piece" multiple logic computing chips (Logic) and multiple HBM memory chips together on the same interposer has become the only path to continue the benefits of Moore's Law.
However, as the AI chip package size evolves from the early 50mm×50mm all the way to 100mm×100mm or even larger sizes, organic substrates (such as ABF substrates), which have been the cornerstone of the industry for the past few decades, are beginning to frequently encounter challenges from physical limits: organic materials are highly prone to thermal expansion deformation and in-plane warpage under multiple rounds of high-temperature curing processes in advanced packaging, directly leading to alignment failures and solder joint fractures of tiny pitches.
To thoroughly subvert this traditional route, global semiconductor and packaging giants such as Intel, TSMC, and Samsung are going all out to turn to the ultimate answer for next-generation packaging substrates: Special Glass Substrates (Glass Core Substrates).
The formulations and overflow fusion processes of electronic-grade and semiconductor-grade special glass have extremely high patent barriers. Currently, those who can provide ultra-flat, defect-free special glass substrates that meet the requirements of Intel and TSMC are still the global top three traditional display glass giants: US-based Corning, Japan's AGC, and Japan's NEG.
Conclusion: Material Evolution Determines the Future Form of Data Centers
Looking at the current semiconductor industry transformation driven by data centers, a clear underlying industrial logic can be drawn: All algorithm optimizations and all system-level architectural innovations must ultimately rely on the physical characteristics of underlying materials to be implemented.
From the High-K precursors used inside memory chips to lock electrons, to the third-generation semiconductors used at the power end to improve conversion efficiency; from the Indium Phosphide and Thin Film Lithium Niobate responsible for efficient light emission and light control in the optical interconnect network, to the special glass substrates that are about to thoroughly subvert the physical form of advanced process chips. AI's extremely greedy pursuit of computing power is forcing the entire semiconductor material system to undergo drastic iteration and reconstruction at an unprecedented speed.
In this material transformation forced by physical limits, whoever can take the lead in the R&D, engineered mass production, and secure and controllable supply chains of these critical, high-end materials will firmly grasp the most unshakeable underlying discourse power in the next stage of the global computing power general competition.