Kao Yung-chung (高永中) has a phrase he keeps returning to — and every time he says it, he laughs a little.
"Tomorrow's Material."
It has been the compound semiconductor industry's self-description for more than thirty years. Every major technology wave — fiber-optic communications, 4G, 5G, electric vehicles — seemed to announce that the sector's moment had finally arrived. And each time, large-scale commercialization moved slower than the market expected.
"We've always been talking about Tomorrow's Material," said Kao, chairman of IntelliEPI Technology, a Taiwan-listed epitaxial wafer maker headquartered in Allen, Texas. "But that tomorrow never seemed to come."
Generative AI, he now believes, may be the event that changes the accounting.
Silicon Handles Computation, but AI Data Centers Need More Than That
For the past half-century, semiconductor progress was largely a silicon story. CPU and GPU performance advanced along process nodes; market expectations of technological progress were shaped almost entirely by transistor counts and clock speeds.
That framework is now incomplete. When an AI data center deploys tens of thousands of GPUs in dense clusters, the binding constraint is no longer how fast individual chips compute. It is how quickly data can move between chips, across racks, and between machines — while keeping power consumption within manageable bounds.
Kao observed that high-speed optical communications has accelerated from 100G to 400G to 800G, and is now approaching 1.6T, with generational transitions coming faster than before. Copper interconnects increasingly fail under these speeds and densities, forcing the industry to push optical components closer to the compute core. The progression from pluggable optical modules to silicon photonics to co-packaged optics (CPO) represents different architectural approaches to the same underlying problem: move data faster, over shorter distances, at lower energy cost.
Silicon is excellent at computation but cannot complete the full optical interconnect stack on its own. High-speed lasers, photodetectors, and modulators — the components that handle actual light transmission and detection — still depend on compound semiconductors, particularly indium phosphide (InP) and gallium arsenide (GaAs). As AI systems grow larger and more interconnected, Kao argued, the dependency on these materials deepens rather than fades.
"A lot of people now think we're doing AI," he said. "But we've always been doing compound semiconductors."
Four Materials, Four Distinct Battlefields
One of the points Kao most wants international readers to understand is that compound semiconductors are not a single material competing to replace silicon. They are a family of distinct platforms, each solving a different set of problems.
Indium phosphide is the foundation for high-speed optical communications — lasers, photodetectors, and modulators used in data center interconnects. Gallium arsenide (GaAs) has long underpinned mobile radio-frequency components and is now also the substrate for quantum dot lasers, an emerging technology with strong potential in optical communications. Gallium nitride (GaN) excels in high-frequency, high-power, and high-efficiency power applications. Gallium antimonide (GaSb), with its infrared characteristics, holds a specialized position in defense, aerospace, and thermal imaging.
These materials are not in competition with each other, and none of them is positioned to displace silicon entirely. They each occupy a specific role in the AI infrastructure stack. IntelliEPI's long-term commitment to all four — InP, GaAs, GaN, and GaSb — reflects Kao's conviction that future systems will require more material cooperation, not less, and that single-platform approaches will hit limits that only compound semiconductor integration can solve.
The Smartphone Era Was a Preview; AI Is Rewriting the Whole Data Center
Looking back at the industry's development, Kao identified the mass adoption of smartphones as the moment compound semiconductors first reached true consumer scale. The rollout of 4G and 5G drove demand for radio-frequency front-end components, bringing GaAs power amplifiers into hundreds of millions of handsets and demonstrating, for the first time, that non-silicon materials could ship in consumer device volumes.
But the smartphone transformed a single terminal device. AI is transforming entire data centers.
As GPUs move from standalone machines into large-scale clusters, the requirements for chip-to-chip interconnection, power conversion, thermal management, and sensing all expand simultaneously. Materials are no longer just a line item in a component spec sheet — they are beginning to directly constrain whether a system can scale.
This, in Kao's view, is the decisive difference between AI and the technology waves that preceded it. Previous cycles positioned compound semiconductors as a niche complement to silicon. In the AI era, when system efficiency rather than raw compute becomes the critical variable, material selection is being evaluated at the same level of importance as chip design and advanced packaging.
Thirty Years of Patience, Now Visible to the World
Kao was careful not to declare that Tomorrow's Material has fully arrived. The compound semiconductor industry has been here before — at the threshold of its moment — and pulled back. He maintained his characteristically measured posture on timelines for technologies such as CPO and quantum dot lasers: the direction is clear, but manufacturing readiness must be demonstrated, not assumed.
What has changed, in his reading, is the structural position of the materials themselves. InP, GaN, GaSb, and related compounds — long treated as niche markets — have now been drawn into national security frameworks, supply chain policy debates, and defense procurement reviews in ways that were unimaginable a decade ago. The competition in semiconductors is no longer only about who achieves the next process node. It is also about who can secure a reliable chain from materials and equipment through to customer-qualified production.
Kao has spent more than thirty years building exactly that chain — through the fiber-optic era, through mobile, through 5G, and now into AI. He did not change direction when each new wave arrived. He stayed with the same materials, the same equipment domain, the same manufacturing discipline.
"AI hasn't changed our direction," he said. "It just made that direction visible." （Related： Exclusive | He Turned His PhD Into a Working Machine—and Built IntelliEPI the Same Way ｜ Latest ）
