TSMC boosts Biden’s AI chip ambitions with $11.6bn US production deal

Unlock the Editor’s Digest for free

The world’s biggest chipmaker, Taiwan Semiconductor Manufacturing Co, has agreed to make its most advanced products in Arizona from 2028, in a boost to White House efforts to bring semiconductor production on to home soil.

TSMC will make the latest cutting-edge 2-nanometre chips in a fabrication plant, or fab, it is building in Phoenix, Arizona, marking an upgrade from its previous plans.

That facility will be the company’s second in the US. The first, which is also in Arizona and was announced in 2020 under the Trump administration, will begin production next year.

TSMC also said on Monday that it will increase its total investment in the US from $40bn to $65bn to build a third fab, with 2nm or even more advanced technology, which will be operational by 2030.

The Taiwanese company and the US commerce department said on Monday that Washington would provide it with support worth $6.6bn in grants and up to $5bn in loans.

The subsidies fall under the Chips Act, which was passed in 2022 to boost US chipmaking. Last month, the Biden administration unveiled a deal for $8.5bn in grants and up to $11bn in loans for Silicon Valley’s Intel, which has pledged $100bn in new investment.

TSMC’s commitment helps the White House move towards its goal of bringing 20 per cent of the world’s advanced semiconductor manufacturing onshore by 2030.

Growing fears of a Chinese invasion of Taiwan, where 90 per cent of cutting-edge chips are currently made, have prompted the US to step up efforts to boost its domestic semiconductor production.

“TSMC is expanding its manufacturing capabilities in Arizona such that for the first time ever we will be making, at scale, the most advanced semiconductor chips on the planet here in the United States of America,” said US commerce secretary Gina Raimondo. “[We are] massively strengthening our national security position.”

“Our US operations allow us to better support our US customers, which include several of the world’s leading technology companies,” said Mark Liu, chair of the leading contract chipmaker.

The latest plans will bring semiconductor production in the US closer to the state of the art, as artificial intelligence drives demand for ever more computing power.

TSMC previously planned to run its US fabs on older manufacturing technology than its most advanced mass production in Taiwan.

By 2026, most AI chips will run on 3nm, meaning that the production capabilities of TSMC’s first Arizona plant will fall short.

By the time TSMC’s second Arizona fab opens in 2028, Nvidia and other AI chip vendors are likely to have migrated to 2nm, said an engineer familiar with the process. Therefore, TSMC’s original plan to have that plant run on 3nm “didn’t make sense”, a company executive said.

The US hopes the TSMC deal will mean that some of the most advanced chips used in AI could be partly made in the US by the end of the decade, reducing the reliance of chipmakers such as Nvidia and AMD on Asian production.

“The chips that TSMC makes . . . underpin all AI. Tens of thousands of leading-edge chips are required to train a single frontier AI model [such as OpenAI’s GPT4],” Raimondo said. “And now, because of this announcement, these chips will be made in the United States of America.”

But industry executives and analysts said that claim went too far.

“Having 2nm in the second fab doesn’t mean Nvidia will not be buying chips made in Taiwan any more,” said a person familiar with TSMC’s plans, “it just means that they will have an option to issue a special requirement that a certain amount of their chips come from that [Arizona] fab.”

TSMC’s leading-edge fab investment at home continues to far outpace that in the US. It will start 2nm mass production next year and plans to build “multiple” more fabs operating on that technology in three locations in Taiwan, Liu told investors in January.

Video: The race for semiconductor supremacy | FT Film

Via

Leave a Comment

ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT ArT