How a Taiwanese Chipmaker Became the Linchpin of the AI Revolution

In the span of a few short years, artificial intelligence (AI) has gone from a niche research topic to one of the most transformative and talked-about technologies on the planet. Companies and governments around the world are racing to develop and deploy ever more sophisticated AI systems for everything from voice assistants and autonomous vehicles to drug discovery and financial forecasting.

But behind the breathtaking pace of progress in AI lies an often overlooked but absolutely critical enabling technology: semiconductors. The advanced processors that power today‘s leading AI applications require the latest and greatest in chip manufacturing technology. And one company stands head and shoulders above the rest in its ability to produce those chips: Taiwan Semiconductor Manufacturing Company (TSMC).

TSMC‘s Unrivaled Manufacturing Prowess

Founded in 1987, TSMC is the world‘s largest dedicated independent semiconductor foundry, with over 50% market share in 2021.^1 The company specializes in the manufacture of logic chips, which include CPUs, GPUs, and other processors used in a wide range of electronic devices.

But it is in the field of AI that TSMC has truly become indispensable. The company‘s leading-edge manufacturing processes, known as "nodes," offer unmatched transistor density and energy efficiency that are critical for the demanding workloads of machine learning and deep learning.

TSMC‘s 7nm node, which first entered mass production in 2018, packs in over 96 million transistors per square millimeter—more than a 50% increase over its previous 10nm node.^2 The company followed that up with the launch of its 5nm node in 2020, which further increased transistor density to nearly 172 million per square millimeter.^3

These advanced nodes have enabled a new generation of AI chips with unprecedented capabilities. NVIDIA‘s A100 GPU, which is manufactured on TSMC‘s 7nm node, delivers up to 624 TFLOPS of AI performance in a single chip.^4 That‘s nearly a 20x increase over the company‘s previous generation V100 GPU.

And TSMC is not resting on its laurels. The company is already mass producing chips using its 4nm node, with plans to start volume production of 3nm later this year. Looking further ahead, TSMC has outlined a roadmap that includes 2nm and 1.4nm nodes in 2025 and 2027 respectively.^5

To put those numbers in context, Intel, TSMC‘s largest rival in advanced chip manufacturing, is not expected to launch its own 7nm node until late 2022 or early 2023—more than four years after TSMC.^6 Samsung, another leading foundry, has also struggled to keep pace with TSMC‘s relentless innovation.

The AI Chip Arms Race

TSMC‘s manufacturing leadership has made it the partner of choice for the world‘s top designers of AI chips. In addition to NVIDIA, the company counts AMD, Google, Amazon, Microsoft, and dozens of other leading tech firms among its customers.

These companies are engaged in an intense arms race to develop ever more powerful and efficient AI accelerators to train and run massive machine learning models. And they are increasingly turning to custom chip designs tailored specifically for AI workloads.

Google‘s Tensor Processing Units (TPUs) are perhaps the most well-known example of a custom AI chip. First introduced in 2016, TPUs are designed from the ground up for machine learning and offer significant performance and cost advantages over traditional GPUs. Google uses TPUs to power everything from search ranking and translation to AlphaGo, the first computer program to beat a world champion at the ancient game of Go.^7

But Google is far from alone in its pursuit of custom AI silicon. Amazon Web Services has its own Inferentia chips for deep learning inference, while Microsoft has developed the Brainwave FPGA-based accelerator. Chinese tech giants like Alibaba and Baidu are also heavily investing in AI chip development.

All of these custom AI chips have one thing in common: they are manufactured by TSMC. The Taiwanese foundry‘s ability to turn cutting-edge designs into working silicon has made it the linchpin of the AI revolution.

Geopolitical Risks and the Future of AI

TSMC‘s dominance in advanced chip manufacturing has not gone unnoticed by policymakers and national security experts. As AI becomes ever more central to economic competitiveness and military superiority, access to leading-edge semiconductor technology has taken on new strategic importance.

This has put TSMC squarely in the middle of rising tensions between the United States and China. The vast majority of TSMC‘s manufacturing capacity is located in Taiwan, which China claims as its own territory and has vowed to reunify with the mainland, by force if necessary.

Any conflict between China and Taiwan would have devastating consequences for the global semiconductor industry. TSMC‘s fabs are clustered on the island‘s western coast, well within range of Chinese missiles. Even a temporary disruption to TSMC‘s operations could cause widespread shortages of critical components and bring the advancement of AI to a screeching halt.

Recognizing this risk, the United States has been pushing TSMC to build new fabs on American soil. In 2020, the company announced plans for a $12 billion fab in Arizona that will produce 5nm chips starting in 2024.^8 But that facility will represent only a small fraction of TSMC‘s total capacity, and will be generation behind its most advanced nodes.

Truly diversifying leading-edge chip production away from Taiwan will require massive investments and a concerted effort to build up domestic semiconductor ecosystems. The United States, European Union, Japan, and other countries have all announced ambitious plans to boost their own chipmaking capabilities, but it will take years if not decades for them to catch up to TSMC.

In the meantime, the world‘s continued reliance on TSMC as the single source for the most advanced AI chips creates a significant risk factor for the future development of artificial intelligence. As AI systems become ever more complex and power-hungry, they will require a steady supply of next-generation processors that only TSMC can provide.

Conclusion

The rise of artificial intelligence is one of the defining technological trends of our time, with the potential to transform virtually every aspect of the global economy and society. But the AI revolution would not be possible without the advanced semiconductor manufacturing capabilities of TSMC.

The Taiwanese foundry‘s unrivaled ability to turn cutting-edge designs into working silicon has made it the linchpin of the AI ecosystem. From NVIDIA‘s GPUs to Google‘s TPUs, virtually every major AI chip in use today is manufactured by TSMC.

This concentration of critical technology in one company and one geopolitical hotspot creates significant risks for the future of AI. Any disruption to TSMC‘s operations, whether from natural disasters or military conflicts, could derail the rapid pace of progress in machine learning and set back the development of AI by years.

Diversifying leading-edge chip production and building up domestic semiconductor supply chains must be a top priority for countries that want to maintain their competitiveness in the age of AI. But in the short to medium term, the world will continue to depend on TSMC as the backbone of the AI revolution.

As TSMC goes, so goes the future of artificial intelligence.