应用材料CEO Dickerson在EPIC中心揭幕式演讲：泛论AI算力的技术与产能竞赛，无订单/指引等新增事实

Applied Materials CEO Gary Dickerson. Credit: Monica Chen

The contest for leadership in AI computing is being fought on two fronts—technology and manufacturing capacity—and chipmakers must advance chips, memory, advanced packaging and system interconnects simultaneously to keep pace with demand, Applied Materials President and CEO Gary Dickerson said.

Speaking at the unveiling of the company's Equipment and Process Innovation and Commercialization (EPIC) Center on October 12, Dickerson said AI computing demand remains in its early stages, with substantial room for growth. However, rapidly evolving workloads and computing architectures are forcing the semiconductor industry to accelerate R&D, production qualification and capacity expansion.
Applied Materials is focusing on three areas to address these challenges: materials innovation, integration innovation, and yield innovation. The company aims to advance them through EPIC's high-velocity co-innovation model, bringing companies together earlier in the development process to improve efficiency and accelerate commercialization.
Dickerson divided the competition for AI computing leadership into two areas. The technology race centers on device and chip innovation, along with system and packaging innovation, with the goal of increasing the number of tokens processed per second per watt of power consumed.
The capacity race involves both data center infrastructure and semiconductor manufacturing capacity, with challenges ranging from improving fab yields and wafer output to accelerating the ramp-up of new facilities.
Single-chip performance alone is no longer enough to determine competitiveness. Even as chip designs deliver greater computing power, actual supply can remain constrained if chip and packaging yields or wafer output fail to keep pace. New fabs must also ramp up production quickly to translate capital investment into usable capacity.
Data centers reshape wafer demand
AI data centers are changing the semiconductor industry's demand structure, accounting for a growing share of wafer starts in advanced logic and DRAM as traditional applications such as smartphones and PCs lose relative share.
In advanced logic, data center demand is increasingly overtaking mobile devices and PCs as a primary growth driver. In DRAM, the share of wafer production driven by data centers continues to rise, while the relative contribution from mobile devices and PCs declines.
The shift is forcing semiconductor manufacturers to reconsider their innovation priorities. Data center chips are generally larger and more complex, placing greater demands on chip design, process integration, advanced packaging, and multi-die interconnects.
High-bandwidth memory (HBM) adds further pressure on wafer capacity. According to Dickerson, HBM requires four times as many wafer starts as conventional memory to produce the same number of bits.
As AI servers consume more HBM, changes in the memory product mix are also affecting how fab capacity is allocated. With demand for HBM, advanced logic and advanced packaging rising simultaneously, wafer output, manufacturing yields and capacity allocation are becoming increasingly important across the supply chain.
AI workloads themselves are also diversifying. Dickerson identified four major categories driving growth: physical AI, agentic AI, generative AI and perception AI.
Physical AI encompasses robotics, autonomous vehicles, smart environments and digital twins, using technologies such as multimodal models, world models, reinforcement learning and imitation learning.
Agentic AI covers applications including software development, workflow automation and customer support. Its underlying technologies include large language model (LLM) reasoning, tool calling, reinforcement learning and test-time scaling.
Generative AI supports chatbots, content creation and code generation, primarily through transformer-based LLMs and diffusion models. Perception AI encompasses computer vision, speech recognition and advanced driver-assistance system (ADAS) perception, using convolutional neural networks (CNNs) and vision transformers (ViTs).
The rise of agentic AI has already changed computing workloads and system configurations, including increasing demand for CPU processing, Dickerson said. How physical AI and other emerging computing technologies will develop remains uncertain.
That uncertainty reinforces the need for system-level co-innovation. As different AI applications place varying demands on CPUs, GPUs, memory and interconnects, chip designers and manufacturers must determine technology combinations and integration approaches earlier in development to reduce the risk of encountering bottlenecks later.
Semiconductor growth drives manufacturing complexity and R&D spending
Dickerson also outlined how semiconductor industry growth has been accompanied by increasing manufacturing complexity over the past five years.
Global semiconductor revenue is projected to reach approximately US$1.5 trillion in 2026, roughly three times the US$500 billion recorded in 2021, according to figures he presented.
Manufacturing complexity has increased alongside market growth. As leading-edge processes advance from 5nm to 2nm, the number of steps in advanced semiconductor manufacturing has risen to approximately 1.5 times its previous level, according to Dickerson's presentation.
According to Dickerson's presentation, R&D investment by the top five semiconductor companies and wafer fabrication equipment suppliers is estimated to reach US$67 billion in 2026, up from approximately US$40 billion in 2021, an increase of roughly 1.7 times.
Over the past year, the industry's focus on manufacturing yields and output efficiency has become particularly pronounced. Semiconductor competition increasingly depends not only on technological advances and performance metrics but also on the ability to translate innovation into products that can be manufactured reliably, at high yields and in large volumes.
EPIC aims to accelerate semiconductor commercialization
As manufacturing challenges become more complex, Dickerson said individual companies can no longer solve every problem independently. Semiconductor R&D must shift from developing individual components toward collaboration across the entire system technology stack.
Traditional semiconductor development follows a sequential process, beginning with materials exploration and progressing through equipment development, process development, and module integration before moving into pilot production and high-volume manufacturing.
Taking a technology from initial concept to commercial chip production can require more than a decade.
Applied Materials' EPIC Center is designed to shorten that cycle by allowing research, equipment development, process engineering and module integration to proceed in parallel within a shared environment, reducing delays between development stages.
The company aims to accelerate technology commercialization by up to twofold through this collaborative approach.
Dickerson emphasized that successful cross-company collaboration requires more than exchanging data and technical information. A culture of cooperation and long-standing relationships built on trust are equally essential to making the model work.