The throat of AI computing power is being choked by materials.
Intel CEO Lip-Bu Tan has named diamond, silicon carbide (SiC), indium phosphide (InP), gallium nitride (GaN), and glass substrates as five key materials critical to breaking the bottleneck. Meanwhile, industry insiders have bluntly warned: "Insufficient InP laser production capacity is the core choke point for co-packaged optics (CPO)."
This means that no matter how powerful Nvidia's GB200/300 chips are, without the support of advanced materials, data simply cannot flow at high speed.
This is not just a capacity issue – it is a challenge pushing the limits of materials science. From SiC/GaN power delivery and glass-substrate packaging (as highlighted by Tan), to lithium niobate and SOI for optical interconnect solutions, and diamond and ceramic substrates for heat dissipation under extreme power loads, materials have become the invisible battleground in the second half of AI. Whoever breaks through first will hold the key to computing dominance.
Diamond
As AI computing power accelerates at breakneck speed, heat dissipation in high-end chips is becoming increasingly severe. Nvidia's Rubin Ultra, for instance, is expected to exceed 2,300W per chip, with local heat flux densities surpassing 1,000 W/cm² – an almost unimaginable level of heat. Traditional copper and aluminum heat spreaders are hitting their physical limits. Diamond, with its ultra-high thermal conductivity, has emerged as a leading candidate among new semiconductor materials.
Widely known for its extreme hardness – the hardest naturally occurring substance, with a Mohs hardness of 10 – diamond is also one of nature's best thermal conductors, with a thermal conductivity of up to 2,200 W/(m·K), 5.5 times that of copper and 11 times that of aluminum. It is also an electrical insulator, and its coefficient of thermal expansion closely matches that of silicon, SiC, and GaN. Industry observers have noted: when single-chip power exceeds 1,400W, diamond is no longer an "option" – it becomes a "necessity."
Nvidia's next-generation Vera Rubin architecture GPUs have fully adopted a "diamond–copper composite material + 45°C direct-contact liquid cooling" solution, and the industry has even marked 2026 as the "first year of large-scale diamond heat dissipation adoption."
Silicon Carbide (SiC)
AI data center rack power is surging from tens of kilowatts to hundreds of kilowatts, with megawatt-scale racks on the horizon. As current and voltage levels rise, traditional silicon-based power conversion suffers from staggering losses. Moreover, AI demands materials that dissipate heat quickly, consume less power, occupy minimal space, and withstand high temperatures.
SiC offers high breakdown voltage, stable performance at high temperatures, and switching losses far lower than silicon. Its compact device footprint and excellent reliability make it the ideal choice for 800V high-voltage DC (HVDC) architectures – achieving efficiencies above 97% in high-voltage input rectification and power factor correction (PFC) stages, and reducing transmission losses by over 30% in 800V HVDC systems. New data centers built by Nvidia, Microsoft, and other tech giants have already adopted SiC-based power architectures as standard. Domestic Chinese manufacturers such as TankeBlue Semiconductor are accelerating 8-inch substrate and epitaxial mass production to reduce wafer costs.
Gallium Nitride (GaN)
GaN and SiC form a clear "high-voltage vs. low-voltage" division: SiC stays in the server room, while GaN moves into the racks.
GaN features high electron mobility and exceptional power density, excelling in high-frequency, low-voltage, high-density power conversion. In GPU power modules, high-frequency DC-DC converters, and server power supplies – where space and response speed are critical – GaN can shrink power supply size while boosting power density. Samsung's 8-inch GaN process has already achieved a 30% cost reduction and is accelerating its adaptation for large-scale data center deployment.
Indium Phosphide (InP)
Indium phosphide is irreplaceable – it is the only mass-production substrate material for laser and photodetector chips, perfectly matching the 1310nm and 1550nm low-loss windows of fiber-optic communication. In the CPO domain, the industry's mainstream approach integrates optical engines and electrical chips on the same substrate, and the high-speed light-emitting devices at the core of CPO rely 100% on InP lasers.
Returning to that industry executive's warning: behind those words lies a stark set of numbers. In 2025, global InP substrate demand is estimated at 2.0–2.1 million pieces, while effective supply is only 600,000–700,000 pieces – a supply-demand gap exceeding 70%, expected to persist until 2030. Goldman Sachs has issued an even more direct forecast: "Light source supply will remain tight through 2027, and may only begin to balance in the second half of 2028 as supply chain capacity expansions come online."
Thin-Film Lithium Niobate (TFLN)
Leveraging the strong linear electro-optic effect (Pockels effect) of lithium niobate, thin-film lithium niobate enables high-bandwidth, high-linearity (high-fidelity) optical signal modulation at low driving voltages. This makes it well-suited for medium- to long-distance, high-bandwidth optical transmission applications such as backbone networks and metro networks. In addition, its high refractive index contrast and excellent optical confinement allow for higher integration, improving device size and power consumption.
In optical modulators, thin-film lithium niobate offers an electro-optic coefficient three times that of InP and over a hundred times that of silicon photonics. With just 1.8V, it can achieve ultra-high-speed modulation beyond 100GHz, cutting power consumption by 30–50%, while a single channel can easily run at 400G.
As AI computing power explodes, data center optical interconnects are racing from 400G to 800G, 1.6T, and 3.2T. The performance limitations of traditional materials are becoming increasingly apparent, and thin-film lithium niobate is poised to become a critical material for next-generation high-speed optical transmission modulation. Currently, only four companies worldwide have mastered mass-production capabilities for 8-inch high-end wafers, with a supply-demand gap exceeding 70%.
SOI (Silicon-on-Insulator)
If silicon is the "flesh" of a chip, SOI is the "skeleton" of silicon photonics. This "top silicon–buried oxide–substrate" sandwich structure, by inserting a silicon dioxide insulating layer beneath the silicon layer, completely eliminates the crosstalk between electrons and photons. It reduces parasitic capacitance by over 60%, allowing electrical signals to run faster and more efficiently. More importantly, the large refractive index difference between the top silicon and the oxide layer naturally forms the "walls" of an optical waveguide, enabling optical signals to bend and transmit with minimal loss on the chip.
In the AI era, SOI is the irreplaceable core substrate for manufacturing optical modules, CPO engines, and quantum chips. Global production capacity is heavily concentrated at France's Soitec, making domestic localization an urgent priority.
Glass Substrates
As GPUs, HBM, and chiplet stacking drive packaging toward large form factors, high density, and high-speed interconnects, traditional organic substrates and silicon interposers are increasingly facing limitations in size, cost, and performance.
Glass substrates, through glass vias (TGV) for vertical electrical interconnection, precisely fill the market gap between organic substrates and silicon interposers. They offer a tunable coefficient of thermal expansion (matched to silicon chips), extremely low dielectric loss, and ultra-high surface flatness. Intel tests have shown they can reduce pattern distortion by 50% and increase interconnect density by up to 10 times.
Global semiconductor leaders are rapidly accelerating their investments – Intel, TSMC, Samsung Electro-Mechanics, SK Group, and LG Group have all entered the arena, kicking off a commercial race around glass substrates.
Ceramic Substrates
In high-density integrated packaging, high-power devices operating simultaneously generate massive amounts of concentrated heat, and the packaging substrate is the primary heat-dissipating component. As AI chip power crosses the 2,000W threshold, traditional FR-4 organic substrates, with their poor thermal conductivity (only 0.3 W/(m·K)) and thermal expansion mismatch, face severe warpage and delamination risks. Ceramic substrates, with their inherent advantages of high thermal conductivity, high electrical insulation, and excellent thermal matching, have become a necessity for high-power scenarios.
Among them, aluminum nitride (AlN), with a thermal conductivity of 170–220 W/(m·K), is the top choice for heat dissipation in 800G/1.6T high-speed optical modules. Silicon nitride (Si₃N₄), with its excellent flexural strength and thermal shock resistance, serves as the packaging foundation for SiC/GaN high-voltage power modules.