Technology · AI
Integrated Systems Will Drive Next Wave of AI Growth, Says TSMC
As inference workloads surge 500-fold since 2022, the chipmaker is betting on packaging innovation and system-level integration to power industrial AI deployment

KEY TAKEAWAYS
- ·TSMC argues future AI leaders will build integrated computing systems spanning chips to data centers, not just better individual components, as the industry industrializes.
- ·Global AI inference token volume has surged roughly 500-fold since 2022, making inference the dominant driver of system expansion as agents run continuously in enterprises.
- ·TSMC is developing 3DFabric packaging, photonic interconnects, and advanced memory technologies to address bottlenecks where data movement consumes 60 percent of system activity.
The Integration Imperative
The artificial intelligence industry is entering what Taiwan Semiconductor Manufacturing Co. calls "true industrialization," and the winners will be those who master system-level design rather than component-level performance. That was the message from April Li, TSMC's director of AI and high-performance computing business development, speaking at the Semicon Taiwan IC forum in Taipei.
According to Li, AI agents are now embedding themselves into enterprise operations while investment accelerates in specialized infrastructure and physical AI applications. The shift marks a fundamental change in how the industry must think about compute architecture.
"Market leaders will not be the ones who simply make better models, but those who build the most integrated systems," Li told attendees. The statement reflects growing recognition across Asia's semiconductor ecosystem that isolated advances in chip design or software no longer suffice as AI workloads become more demanding and diverse.
Inference Demand Reshapes the Stack
Behind the push for integration lies a dramatic expansion in inference computing. Global inference token volume has climbed roughly 500-fold from 2022 levels, according to TSMC. Unlike the training phase that dominated early AI investment, inference involves deploying trained models to handle live requests and execute tasks continuously.
The economics have flipped. Where inference was once considered a lightweight afterthought, it now drives the bulk of system expansion as AI agents run persistently in enterprise environments. More sophisticated reasoning models and agentic systems consume significantly more tokens than single-query interactions, compounding the load.
Li identified four pressure points emerging from these workloads: logic scaling, interconnect efficiency, memory performance, and power delivery with cooling. Data movement alone can represent up to 60 percent of system activity in typical scenarios, leaving accelerators running below 40 percent utilization. That inefficiency underscores why packaging and interconnect technology have become as strategic as transistor count.
Multi-Die Architecture and Photonics
By 2030, AI packages may house more than 1 trillion transistors, according to TSMC's projections. That scale makes multi-die architectures and heterogeneous integration unavoidable. The company is responding with its 3DFabric platform, which combines system-on-integrated-chip 3D stacking with chip-on-wafer-on-substrate advanced packaging techniques.
TSMC is also developing a compact universal photonic engine platform designed for high-speed optical data transmission. The move into photonics reflects broader industry momentum toward optical interconnects as electrical signaling hits bandwidth and energy limits. Copackaged optics and silicon photonics were highlighted themes across this year's Semicon Taiwan event, which drew more than 1,300 exhibitors and an expected 100,000 attendees.
On the memory front, TSMC is working to boost high-bandwidth memory performance by applying advanced logic technology to base dies. The approach aims to reduce bottlenecks that throttle accelerator utilization when data cannot move fast enough between compute and storage.
From Wafer to Data Center
Li emphasized that TSMC's role has evolved beyond manufacturing discrete chips. The company now seeks an end-to-end understanding of the supply chain "from silicon to the data center to the token," working with ecosystem partners to validate technologies early across that entire span.
"Gone are the days when we can just ship wafers across the fence," she said during a question session. The comment captures how tightly coupled design, packaging, and deployment have become in AI infrastructure. Chipmakers that once handed off finished silicon to systems integrators now participate in rack design, thermal management, and even software optimization to ensure their products perform in real-world configurations.
The shift has strategic implications for Asia's semiconductor industry. Foundries like TSMC, packaging specialists such as Powertech Technology, and optical component suppliers are being pulled deeper into customers' architectures. Collaboration extends beyond traditional fab relationships into joint engineering efforts with hyperscalers, AI labs, and enterprise customers.
Regional Context
Taiwan's semiconductor ecosystem is positioning itself at the center of this integration wave. Powertech recently announced its advanced fan-out panel-level packaging line is fully booked through 2030, a signal of how critical advanced packaging has become. Meanwhile, India's AM Intelligence ordered 9,000 Nvidia Vera Rubin systems for deployment next year, illustrating the regional race to build out AI infrastructure.
China is pursuing a different path, concentrating data center construction in less populated western provinces like Guizhou to manage power and cooling costs. The geographic distribution of AI infrastructure across Asia reflects varied strategies around energy access, regulatory environments, and proximity to end users.
For TSMC, the integration thesis aligns with its broader ambition to remain indispensable as compute architectures fragment. The company is betting that system-level complexity will favor partners who can coordinate across logic, memory, packaging, and interconnects rather than specialists optimizing individual layers. Whether that bet pays off will depend on how quickly the industry can standardize interfaces and overcome the coordination costs of deeper collaboration.
RELATED STORIES
Spot something wrong? Email editor@briefasia.com. We log every correction publicly.



