The Island That Builds AI: Inside NVIDIA GTC Taipei and How Taiwan Became the World’s Most Critical Technology Ecosystem

Before NVIDIA Founder and CEO Jensen Huang took the stage for his keynote at Computex 2026, a two-hour pregame show at the Taipei Pop Music Center set the strategic context. Hosted by Goldman Sachs Taiwan semiconductor analyst Bruce Lu and Gartner VP Analyst Tracy Tsai, the event brought together Taiwan’s top technology leaders — presented in Mandarin with English subtitles — to map out how the island sits at the intersection of every critical layer of the global AI supply chain, from semiconductor fabrication to physical robots on factory floors.

A panel of Taiwan’s top technology leaders mapped out global AI supply chains during a dynamic two-hour GTC Taipei pregame warm-up show hosted by analyst Bruce Lu and VP Analyst Tracy Tsai at the Taipei Pop Music Center. Image: NVIDIA

Taiwan’s Supply Chain Is Not Incidental — It Is Structural

The event opened with a pointed argument: the AI revolution does not run on software alone. It runs on physical infrastructure, and Taiwan builds most of it.

TSMC Executive Vice President Dr. Yu outlined the trust relationship that underpins AI hardware globally. Jensen Huang, he noted, once said TSMC’s most important asset is trust — and that trust has extended across Taiwan’s entire supply chain to sustain the AI buildout. “The most important thing is that we are expanding our production capacity vigorously,” Dr. Yu said, describing simultaneous expansion across Taiwan, the United States, Japan, and Europe. He acknowledged the tension plainly: explosive AI demand means capacity remains seriously short, forcing TSMC to make difficult allocation decisions every week.

This is not a minor operational headache. Advanced chip packaging — which TSMC has invested in for 15 to 17 years — has become one of the most critical bottlenecks in AI system production. Photonics integration, chip-stacking (logic-on-logic and memory-on-logic), and silicon photonics for data transmission are all in active development for near-term mass production. The reason matters: as AI shifts from training to inference, low-latency, power-efficient interconnects become the constraint, not raw compute.

Quanta Computer Chairman Barry Lam described a parallel transformation in manufacturing. Quanta began making PCs in the 1980s, responsible for final assembly — the “last mile.” That same discipline now applies to AI server racks, except the complexity has multiplied. The industry has moved from shipping individual GPUs to shipping full rack-level systems — units containing thousands of interconnected processors — that must integrate power management, liquid cooling, and high-speed networking simultaneously.

The Shift From Components to Systems: Why It Changes Everything

The manufacturing conversation surfaced a critical insight that has direct implications for investment and competition: the era of standardised component production is over.

Simon Lin, Chief Strategy Officer of Wistron, described the shift as a “DNA change” — one that required rebuilding talent pipelines from scratch, particularly as geopolitical pressures have made mainland China-based manufacturing talent unavailable for transfer. “Talent is fundamental,” he said. Taiwan’s manufacturers have responded by training new engineers domestically and deploying AI inside their own factories to compensate for reduced headcount.

This is the practical meaning of “AI-enabled manufacturing”: companies like Quanta and Wistron are not just building AI hardware — they are using AI to build it. The loop is self-reinforcing.

MediaTek CEO Rick Tsai reinforced why Taiwan’s position is unusual globally. MediaTek can simultaneously act as a chip designer, a system integrator, and a collaborative design partner — working with NVIDIA on compute chips while also supplying CPU solutions to major cloud service providers. “A complex chip made by two different companies together is almost unheard of in the history of semiconductors,” Tsai said. That ability to operate across the stack — from mobile SoCs to cloud inference chips to edge AI — gives Taiwan’s semiconductor ecosystem a flexibility that is difficult to replicate elsewhere.

Agentic AI and the Physical World: What Factories Are Already Doing

The keynote’s second half shifted from infrastructure to application, and the message was direct: agentic AI — systems that can reason, plan, and act — is already operating in real manufacturing environments, not in pilot programmes.

Hon Hai (Foxconn) Head of AI Software Keysight Yang reported three measurable outcomes from deploying AI agents across its production lines: a 50 percent increase in row capacity efficiency, a 50 percent reduction in defect rates, and 90 percent accuracy in abnormal cause analysis. The key preconditions were not sophisticated algorithms — they were disciplined data collection and correctly identifying which pain points to target first. “Data is the most important thing,” Yang said. “You have to let AI in where your wall goes.”

Pegatron’s leadership described using NVIDIA’s Omniverse simulation environment to train robots in virtual settings before physical deployment, achieving over 60 percent improvement in deployment efficiency. Robots in Pegatron’s US facilities already perform pick-and-pack and simple assembly tasks. More significantly, the same robots are being deployed inside hospitals to handle material transport for paramedics — cutting the time clinical staff spend on logistics.

Delta Electronics Chairman Ping Cheng explained how the company is transitioning its power systems business from traditional AC infrastructure to Solid State Transformer (SST) technology. SST replaces conventional transformers with digitally controllable units that eliminate unnecessary conversion stages, reduce energy loss, and respond dynamically to the variable power demands of GPU clusters. Delta’s internal “SmartDesign” programme uses AI to compress the product design cycle by an estimated 50 percent — running virtual simulation and validation before physical prototypes are built.

The practical implication: as AI data centres consume more power (with single GPU racks now drawing over 100 kilowatts), power delivery and thermal management are becoming strategic differentiators, not commodity services.

Sovereign AI and the Computing Access Problem

A third panel addressed a less visible but increasingly urgent issue: most enterprises and academic institutions cannot actually access the GPU capacity they need to develop AI applications.

Taiwan’s first NVIDIA GP300-equipped supercomputing centre — described as the first in the world — represents a direct policy response to this problem. The founder of TAIPEI (the AI cloud provider building the facility) framed the issue in terms familiar to investors: GPUs are currently used only by a handful of AI-native startups because the operational complexity is prohibitive. The goal is to virtualise that capacity into model services — making compute accessible as a token-based API, in the same way cloud computing commoditised server access in the 2000s.

NTU Professor Hung-yi Lee illustrated the stakes from the academic side. His lab trained a speech recognition system using 16 H100 GPUs — borrowed from TAIPEI — and achieved recognition accuracy superior to OpenAI’s Whisper on mixed Mandarin-English academic lectures, running five times faster. The model was trained on just 5,000 hours of data, less than comparable open-source models from Alibaba, by developing more efficient algorithms under resource constraints. The result demonstrates what Professor Lee called the importance of knowing what to optimise, not just how to train — a skill set he argued will matter more than traditional ML engineering as AI systems increasingly design and train themselves.

Why This Moment Is Different From Previous Technology Waves

Several speakers drew comparisons to the PC era, but the consensus was that the current transition is structurally distinct. PC manufacturing standardised over decades. AI hardware is iterating faster than any previous technology cycle, with new chip architectures, cooling systems, and networking requirements arriving before the previous generation is fully deployed.

Barry Lam made the broader societal point directly, drawing on the historical pattern of automation: “When you have all automation, then you can do the dishes without doing them.” The concern that AI will displace workers mirrors fears raised by the PC, factory automation, and the washing machine — each of which ultimately expanded the economy and created new categories of work. The difference this time is speed, and Taiwan’s ability to keep pace with that speed is precisely what the event was designed to demonstrate.

The takeaway for the global technology landscape is specific: Taiwan is not merely a manufacturing subcontractor for the AI era. It is the integrated ecosystem where chips are designed, fabricated, packaged, assembled into rack systems, powered, cooled, and increasingly operated by the same AI they produce. No comparable concentration of that capability exists elsewhere at scale — which is why decisions made in Taipei about capacity allocation, talent development, and sovereign computing infrastructure carry consequences well beyond the island’s borders.

Author

Bernice Lottering