US missing pieces of AI chip puzzle despite TSMC’s $65bn bet

Taiwan Semiconductor Manufacturing Company’s decision to bring its latest technology to America is a big step forward for US President Joe Biden’s quest for security in the vital tech supply chain — but still leaves Washington short of being able to completely produce the most complex chips in the US.

The world’s biggest chipmaker by sales must also pull off an intricate balancing act as it steps up its US presence, satisfying customers such as Nvidia without damaging its highly profitable business model, which has underpinned the development of the global semiconductor industry for more than 30 years.

TSMC’s planned $65bn of investments in Arizona are part of a construction race in the US that involves other global chipmakers such as Samsung and Intel, which are also taking big subsidies from Washington.

But producing chips for purposes such as AI is still likely to involve plants in Asia, a reflection of the complexity involved in packaging various types of chip together to boost their performance and efficiency.

“It’s really not that simple to onshore everything. Having the logic [chip] foundry in the US and then a bit of the packaging there is not enough,” said Myron Xie, an analyst at boutique consultancy SemiAnalysis.

TSMC — which makes chips under contract at hugely complex and expensive fabrication plants, or fabs — plans to start manufacturing 2-nanometre chips in the US in 2028. This is an upgrade from the company’s previous plans. At that time 2nm technology is expected to be the latest in mass production worldwide, whereas previously the company had intended each new US fab to start operating with process technology one generation behind Taiwan.

TSMC has also committed to offer a third plant using 2nm or even newer technology by 2030.

Washington is paying a hefty price for the upgrade, with US$6.6bn in grants and up to $5bn in loans for TSMC. The money comes from the 2022 Chips and Science Act, which aims to onshore advanced chipmaking for the US. Commerce secretary Gina Raimondo has said the US will be on track to make about 20 per cent of the world’s most advanced chips by the end of the decade.

But while Washington’s money offers some incentive, TSMC’s most important motive for stepping up its commitment to the US was to bring its own US strategy in line with the needs of Nvidia and other vendors of the AI chips that have become the most potent driver of global semiconductor demand.

While TSMC will kick off 2nm volume production in Taiwan next year, its original plans would have offered less powerful 3nm chips only from 2028 in the US, putting it years behind the AI chip cycle, analysts said.

The new plan will allow Nvidia and other sellers of AI chips to shift some of their orders from Taiwan to Arizona.

US chipmaker AMD, one of Nvidia’s biggest rivals in the AI chip market, plans to be one of the first customers of the Arizona plant with high-end graphics processing units and central processing units, according to a person familiar with the company. 

But giving clients the right to choose where their chips are made departs from TSMC’s established practice. It would reduce the company’s flexibility in allocating capacity, which has helped it to generate gross profit margins of more than 50 per cent.

People familiar with discussions between TSMC and its clients said any access to a specific plant will have to be laid down in separate agreements with individual customers, possibly in exchange for a price premium or an upfront deposit.

Industry executives and analysts said despite enhancing options for AI chipmakers, TSMC’s stepped-up investment leaves gaps in bringing the manufacturing of entire chip products onshore.

“We think Nvidia will start adopting 2nm in 2026. So with the concrete plans for 2nm from 2028 at TSMC’s second Arizona fab, that will still be behind,” Xie said.

For TSMC to move more quickly would disrupt one of its key advantages as a chipmaker — its ability to outperform rivals in achieving high yields for new process technology, keeping defective chips to a minimum. Its research and development engineers in Taiwan, managing the start of the production stage, are considered vital.

“It’s not that we don’t want to bring new process [technology] to the US even earlier,” said a person familiar with TSMC’s considerations. “But we need the vicinity to our global research and development centre whenever we ramp up a new node. That means we have to ramp in Taiwan first.”

Mobile chips made for Apple, TSMC’s largest client, may bear the heaviest impact from the resulting gap between capacity in US and Taiwan. TSMC usually makes smartphone chips first with its latest processing technology, serving high-performance computing products a year or two later.

“Apple has always been the first adopter of a node. So if the Arizona fabs are a bit behind, then maybe they could only meet Apple’s needs for older models,” Xie said.

Furthermore, TSMC by itself cannot ensure AI chips will be made in the US, as commerce secretary Raimondo has claimed.

To make the complex devices, various logic and memory components made in South Korea and Taiwan need to be assembled and integrated, a process known as advanced packaging. 

This month, South Korean memory-chip maker SK Hynix announced it would build an advanced packaging facility in the US state of Indiana to produce “high-bandwidth memory” chips (HBM) — made by stacking memory chips on top of a logic base die made by TSMC — to be used in Nvidia’s most powerful GPUs.

But the memory chips themselves, known as Dram, will continue to be produced by SK Hynix’s facilities in South Korea. “Samsung and SK Hynix are making the minimum possible investment in the US, and only then because of geopolitical pressure, so they are unlikely to build separate plants in the US for advanced memory chips,” said CW Chung, an analyst at Nomura.

An alternative source will open up when US memory-chip maker Micron’s leading-edge fab in Idaho starts production in 2027. “Nvidia could get the Dram components out of there if they really want to bring the entire chain onshore,” said one industry expert who declined to be named.

However, an onshore location would still be needed for a different advanced packaging process in which the HBM is installed on the GPU module, which is made by TSMC. The Taiwanese contract chipmaker has shown no appetite for building an advanced packaging facility in the US, in part because its Arizona capacity is too small to make such a plant viable.

A senior US administration official said Amkor, a US-based semiconductor assembly and test provider, could help fill the gap: its US site comes online a year after TSMC’s. But it lacks the ability to make a key packaging component used to connect logic and memory parts.

TSMC’s Korean rival Samsung Electronics, which makes both logic and memory chips, is set to offer an alternative by building an advanced package facility as part of more than $20bn in new investments in Texas due to be announced next week.

A person familiar with the plans said the facility would offer Samsung’s equivalent of the advanced packaging technique that TSMC uses to produce Nvidia’s AI chips. Samsung declined to comment on the planned investment, which was first reported by the Wall Street Journal.

Still, analysts doubt this will be enough to convince TSMC customers to jump ship.

“Samsung manufactures all the memory in Korea, so shipping memory to the US to do HBM packaging might not be ideal. But it would make sense if they’d also make the second stage of packaging in the US,” Xie said.

But Xie said Samsung was a laggard in HBM technology. “Customers would rather go for best in class rather than an all-in-one facility,” he said.