Can TSMC bring a new shape to AI?

Hello everyone! This is Cheng Ting-Fangsending you greetings from Taipei, where we are experiencing the heat and afternoon thundershowers of the subtropical summer.

I recently received a call from a real estate agent in Nangang, the neighbourhood in eastern Taipei where I live. “Ms Cheng”, she began, “are you interested in buying an apartment? We have some great listings.” But what truly grabbed my attention was what came next: “Did you know Nvidia’s CEO Jensen Huang is building a new R&D centre nearby? This is bound to make the area even more popular.”

Even though Huang has already returned to the US following his two-week visit to Taiwan for the Computex tech fair, he and his company continue to exert a surprisingly widespread impact on the island. Influencing the housing market in a specific district of Taipei is just one example. Night market stalls and traditional Taiwanese cuisine dining places where the CEO visited have been swamped with visitors, with some restaurant booking systems crashing due to the surge in reservations.

The AI frenzy has permeated Taiwanese society to a surprising level. At a local café, I overheard an elderly couple next to me discussing not just the record-breaking stock prices of Nvidia and TSMC, but also the subject of CoWoS, or chip-on-wafer-on-substrate. This advanced chip packaging technique, developed by TSMC, helps enhance the performance of AI chips for clients like Nvidia and AMD.

The fact that even complex industry jargon like this has become a topic of everyday conversation only underscores Taiwan’s position as Asia’s leading chip economy.

“Taiwan is probably the only place in the world where everybody knows what CoWoS is without explanation,” AMD CEO Lisa Su said.

Shape shifting

When many people think of chipmaking materials, one of the first images that comes to mind is the round shape of wafers, the substrate on which chips are built. But booming AI demand could soon change thatwrite Nikkei Asia’s Cheng Ting-Fang and Lauly Li.

Advanced AI chipsets demand not only cutting-edge chip production but also advanced packaging techniques to stack more of those chips together and enhance performance. And the 12-inch round wafers that are currently the industry’s most advanced option could soon be too small.

That is because AI computing chips keep getting bigger. Only 16 of the latest Nvidia B200 AI computing chipset, for example, can fit on one wafer.

That’s why TSMC is studying rectangular panel-like substrates that offer more than three times the usable area as round wafers. Intel and Samsung are also investigating rectangular substrates, but commercialising such a radical new approach likely remains years away.

Centres of activity

US tech giants are spending billions of dollars on cloud computing infrastructure in India as they bet the world’s most populous country will become a leading artificial intelligence hub, write of Benjamin Park and Camilla Hodgson for the Financial Times.

Amazon plans to invest about $12.7bn in cloud infrastructure in India by 2030, while Microsoft has committed about $3.7bn to India’s southern state of Telangana, local officials have said.

This would give Microsoft a data centre capacity equivalent to the annual electricity consumption of about half a million European households, according to data from Structure Research, and make India Microsoft’s largest market for its self-built data centres outside the US.

Big Tech companies are set on boosting their cloud computing capacities as they vie to dominate generative AI. India has emerged as a particular focus area thanks to a fast-growing domestic technology market and vast pool of skilled workers that authorities hope will transform the country into a main consumer and exporter of AI.

“India today is one of the most exciting markets in the world for tech,” Puneet Chandok, Microsoft’s president for India and South Asia, told the FT.

Worried mines

Seven years since the last Chinese premier’s visit, Australia continues to walk a tightrope: it wants to diversify its resource-reliant economy and open up new markets, but it remains heavily dependent on Asia’s largest economy.

This conundrum was in the spotlight this week when Premier Li Qiang visited a Chinese-backed lithium refinery in Western Australia, Nikkei Asia’s Shaun Turton writes. Critical minerals like lithium are an area Australia is trying hard to break into, forging deals with the EU and US and boosting domestic processing capabilities via government subsidies.

But the challenge is substantial. While Australia boasts the world’s largest lithium reserves, China remains its top customer. In 2022-2023, 98 per cent of Australian lithium exports went to China, according to government data. Overall trade data further underscores the dependence, with exports to China hitting a 10-year high last year.

Making memories

In the AI era, it is not only processing chips that matter. Memory chips, more precisely high-bandwidth memory chips, also play a significant role, and only three companies in the world can make them.

Micron, which trails its two South Korean counterparts, is pledging to more than triple its market share in this area to around 25 per cent by 2025.

To achieve this ambitious goal, the US company is expanding its research and development facilities for next-generation HBM in the US and is considering setting up its first-ever HBM production facility in Malaysia, Cheng Ting-Fang and Lauly Li write.

Currently all of Micron’s HBM is made in Taichung, Taiwan, where the company is expanding its facilities.