ADVERTISEMENT

Technology

AI’s $1.3 Trillion Future Increasingly Hinges on Taiwan

(Bloomberg)

(Bloomberg) -- When Jung Yoonseok was looking for an assembly partner for his AI chip startup, he had his pick of almost any country in Asia, including his native South Korea. Instead, the Rebellions strategy chief opted for Taiwan because of what he sees as an unparalleled combination of talent, cost and speed.

“Taiwan is small, and Taipei is small, and in that small area everything moves super fast,” the 35-year-old Harvard graduate said after one of his trips to secure production.

Jung reached the same conclusion as thousands of businesses, executives and entrepreneurs who rely on the island to turn their AI visions into reality. From Nvidia Corp. and Microsoft Corp. to OpenAI, the world’s AI frontrunners are increasingly turning to Taiwanese companies to fabricate their chips, build their servers and cool their devices. That in turn has made the island’s stock market the hottest major bourse in Asia over the past year, led by Taiwan Semiconductor Manufacturing Co. and Hon Hai Precision Industry Co.

Some investors think the $400 billion-plus rally is just the beginning. To the bulls, the world is witnessing the establishment of a ChatGPT-era manufacturing base concentrated in Taiwan. That would make the island a key beneficiary of the AI boom — and a critical determinant of its pace and direction.

“Taiwan is really the engine that’s driving AI,” said Sean King, a senior vice president at Park Strategies and a former US Department of Commerce official. 

There are risks for Taiwan. For the first time in decades, an entire technology production ecosystem will be centered not in China but its tiny neighbor. Growing tensions between the US and China may have dissuaded some AI companies from producing hardware in the mainland. Yet the rising importance of Taiwan makes it all the more alluring for Beijing, which has long described the island as a breakaway province it will eventually reclaim.

TSMC is the foundation of this success. As rivals Intel Corp. and Samsung Electronics Co. struggle, the Taiwanese company is extending its leadership in the chip industry, manufacturing virtually all of the world’s most advanced semiconductors. It’s the only place where Nvidia Chief Executive Officer Jensen Huang can get his AI accelerators made. 

The island is now chock-full of lesser-known firms that are just as essential for global AI development. These linchpins include server maker Quanta Computer Inc., power leader Delta Electronics Inc. and Asia Vital Components Co., a pioneer in creating computer cooling systems. Collectively, Taiwanese firms are poised to play an outsized role in an AI market that’s projected to reach $1.3 trillion by 2032.

“This time around the optimism for Taiwan names will be stronger and last longer than in the past,” said Edward Chen, chairman of First Capital Management, citing TSMC’s central role in selecting partners for the likes of Nvidia. “This is elevating Taiwan tech to an entirely different level.”

The benchmark Taiex has climbed more than 40% over the past year, well ahead of comparable indexes for China, Hong Kong, India and Japan.

Taiwan’s rise as a tech hub began, in many ways, in the 1980s. That’s when Japanese companies started outsourcing the production of low-end manufacturing to an island known for cheap plastic toys.

As its economy grew, these same companies became more sophisticated manufacturers and began to open factories in mainland China. But they always kept their most advanced techniques at home. 

In recent years, the increasingly aggressive US trade sanctions on China have forced companies to scout out alternative production locations, knocking the country out of many supply chains.

In less than two years, for example, those curbs have effectively sidelined China’s AI hardware industry. Taiwan’s exports of servers and graphics cards — the building blocks of data centers for training AI models — in the first nine months of 2024 were more than double China’s output, according to data collected by Bloomberg. That’s a sharp reversal from previous years.

Today, the biggest cloud service operators — Microsoft, Amazon.com Inc., Meta Platforms Inc. and Alphabet Inc.’s Google — all use Taiwanese assemblers to fill their server farms as they seek to outdo OpenAI’s ChatGPT. Global spending on AI systems and services is expected to more than double to $632 billion by 2028, according to the research firm IDC.

“Taiwan is a one-stop shop for AI-related hardware,” said Liu Pei Chen, a researcher at the Taiwan Institute of Economic Research. 

AVC Chairman Shen Ching Hang has had a front-row seat to that historic shift. His company, which started decades ago by making heatsinks — basic aluminum rails designed to cool computers — is now developing liquid cooling systems that can keep up with Nvidia’s far-hotter Blackwell generation of AI chips and systems. 

He’s succeeded by going to extreme lengths for his customers. He set up an office across the street from Amazon’s headquarters in Seattle in 2014, years before the company began working with AVC. Shen had discovered that Amazon was installing new servers for its cloud computing business at a faster pace than rivals, and he was determined to win the business for their cooling systems. 

“We went to them every day and asked if they needed our services,” the 61-year-old Shen said. “We could help you do the entire simulation for free. We have new technology. Do you want to take a look?”

Finally, three years later, Amazon said yes. The two companies are still working together today.

One of Taiwan’s strengths is this service-focused strategy, which keeps partners like Amazon, Nvidia and Apple Inc. coming back. When the pandemic triggered a global chip supply crunch, TSMC kept price hikes modest to secure customer goodwill, calling its approach ”strategic, not opportunistic.”

They’ve also shown a readiness to pivot and go where the greatest need lies. Hon Hai and Quanta, best known as assemblers of iPhones and MacBooks, now focus on competing for AI server orders. AVC, Delta and Quanta commit roughly half or more of their operating expenses each year to research and development. They’ve even heeded calls to diversify geographically, with AVC recently setting up a new factory in Vietnam, where its investments will total $450 million in a year or two.

Delta, whose headquarters are just a five-minute walk from Nvidia’s Neihu offices, says that’s borne out of necessity.

“Nvidia’s GPUs needed heat dissipation, so we made heatsinks. They needed magnetic inductors, so we made inductors,” said Ares Chen, the executive in charge of Delta’s power business. “For heat dissipation you need fans, so we made fans. When they boosted the GPU levels to be used in the data center, they needed solutions for bigger power and more heat dissipation.”

These challenges are accelerating as AI demand booms. Nvidia has stepped up its pace of product introductions and it’s now packaging dozens of its new Blackwell chips into a single unit that will help train a new generation of more powerful AI models. This NVL72 monster server, which may run to more than $3 million apiece, will generate more heat than any of its predecessors, forcing the likes of Delta and AVC to get creative about new capabilities for power distribution and cooling.

That kind of innovation is what draws AI entrepreneurs like Rodrigo Liang, CEO of the Silicon Valley chip startup SambaNova Systems. A founder looking to shop around for Taiwanese partners could talk to Delta, AVC, and Quanta in the same afternoon. The entire island can be traversed via high-speed rail from northern Taipei to southern Kaohsiung in an hour and a half. 

That proximity makes for intense competition. Suppliers large and small fight to pitch for almost any new business, eager to capitalize on the AI boom. 

“The world is at a time when AI is front and center,” Liang said. “You’re going to have a lot of companies continuing to come to Taiwan to look for those technologies.”

--With assistance from Shadab Nazmi, Debby Wu, Betty Hou, Spe Chen and Andy Lin.

©2024 Bloomberg L.P.