How niche technology has become a choke point for AI

Subramanian Iyer, an electrical engineer and educator, has long specialized in a sleepy niche in the semiconductor industry that has now become a major chokepoint in the global race for artificial intelligence leadership.

That gap is a technology called advanced chip packaging, which combines up to dozens of components in palm-sized modules. As computing profits from the traditional practice of shrinking transistors so that there are more of them on each chip have dwindled, Nvidia and other chip giants have turned to packaging as a fundamental way to deliver semiconductors capable of performing more complex tasks for artificial intelligence.

Dr. Iyer, 72, a former IBM technologist and now a professor at the University of California, Los Angeles, has helped guide packaging advances for decades. But he and chip industry executives have watched with dismay as America’s leadership in the industry has slipped to the same company that dominates advanced chip manufacturing.

Taiwan Semiconductor Manufacturing Company, which makes high-end chips for Nvidia and other AI leaders, also packages almost all of them. Its key suppliers and partners are also largely in Taiwan, where it faces the same threat from China that has caused U.S. policymakers to invest billions of dollars in supporting domestic chip production.

Packaging bottlenecks have become a hot topic in Silicon Valley as TSMC struggles to keep up with demand. Dr. Iyer tried to help by drawing up plans for a packaging research and development center, funded with $1.1 billion from the Biden administration and slated for construction in Arizona, but the Trump administration effectively killed the effort last year.

“The bottom line is that they threw the baby out with the bath water,” said Dr. Iyer. “We ended up in a place where we’re even more dependent on TSMC.”

The bottleneck underscores how the U.S.’s dependence on Taiwan has not eased despite efforts by the Biden and Trump administrations. The Biden administration has earmarked more than $50 billion to jump-start domestic chip manufacturing under the Chip and Science Act of 2022. President Trump, who has protested grants to chipmakers, has instead pushed for deals with U.S. companies that include equity stakes and threatened tariffs on foreign companies to essentially accomplish the same thing.

After the chips are made, “the packaging is the most important thing,” said Patrick Gelsinger, a former Intel CEO who lobbied for the Chip Act. “And our packaging supply chains may be even more uncertain.”

A Trump administration official disputed that packaging was not a top priority. Nine packaging projects have received CHIP funding, and the Commerce Department is evaluating some important research projects in the area, the official said on condition of anonymity.

Some big American companies are trying to break this logjam. Intel, a longtime leader in the technology, has chosen packaging customers and said last week that it hired a new manager lead advanced packaging. Applied Materials, the largest maker of chip-making tools, is building a $5 billion research facility in Silicon Valley with partners, with packaging as a key spinoff.

Amkor Technology, a packaging specialist, is building its first US plant in Arizona and is expected to take over some packaging work for TSMC as part of a 10-year deal. Amkor, which was awarded a $407 million grant by the Biden administration, increased its own potential investment in the site to $7 billion after talks with Trump administration officials and signs of buying interest from Nvidia and Apple.

Such customers show “very strong demand, they want to build the American ecosystem,” said Kevin Engel, president and CEO of Amkor.

Packaging has long been seen as an afterthought that U.S. chipmakers headed to low-wage countries in Asia. The US share of chip packaging is around 3 percent, according to the Global Electronics Association.

Chips do not work without packaging. The process, also called assembly, typically wraps bare pieces of silicon in a protective plastic with connectors that pass signals to other chips on the circuit board.

Nowadays, chips are often placed on an intermediate layer called a substrate, which is usually made of plastic and fiberglass and embedded copper cables.

Dr. Iyer, who earned his doctorate from UCLA before joining IBM in 1981, made another key advance. At IBM, he developed one of the first “interposers,” a layer made of silicon that holds several chips side by side and sends signals between them at higher speeds.

For the larger packages that some AI processors need, TSMC and Intel put small pieces of silicon in the substrates to act as communication bridges between the chips.

“You can’t make them without advanced packaging,” said Mark Gardner, Intel’s vice president and general manager of packaging and testing. “Without it, we would be in a very different place in the world of artificial intelligence.”

TSMC offers an advanced packaging called CoWoS, which stands for chip on wafer on substrate. For example, Nvidia’s new Rubin processor uses CoWoS to combine two large AI chips with eight high-speed memory stacks, each containing 12 chips, cramming 336 billion transistors into a single package. By 2029, TSMC predicts a 48-fold increase in the number of computing transistors per package compared to 2024.

But the company, which received CHIPS Act funding to build large chip factories in Arizona, doesn’t plan to use CoWoS in the state until 2028 or 2029. Any chips it makes there now must be shipped to Taiwan for packaging.

TSMC is already trying to catch up with AI-driven orders. Its CoWoS production is about 30 percent below demand, estimates International Business Strategies analyst Handel Jones, who says TSMC makes about 95 percent of all advanced packaging.

“I just see demand going higher and higher,” said Kevin Chang, TSMC’s senior vice president. “It will definitely cause a lot of restrictions.

And costs remain high. An advanced chip package can cost $500, said Jan Vardaman, president of research firm TechSearch International. Simpler packages can cost closer to $40, industry officials said.

Some chip startups deliberately avoid designs relying on CoWoS, which can take months to design the desired interposer.

“I didn’t want to touch it,” said Sha Rabii, co-founder and president of Majestic Labs, which designs an AI processor that can use simpler packaging and cheaper memory chips. “It takes a key part of the supply chain challenge off the table.”

In the long run, Dr. Iyer and industry executives, said new packaging concepts are needed. Syenta, an Australian start-up, has designed one that uses electrochemical techniques to create extra-large packages with fewer manufacturing steps and up to a 20-fold increase in communication bandwidth, said Jekaterina Viktorová, the company’s CEO.

Others are packaging research partners. Resonac, a Japanese chemical manufacturer, recently announced a consortium of 12 packaging companies and a pilot production line near Silicon Valley.

But government support remains uncertain. The Biden administration has designated the nonprofit Natcast to oversee $7.4 billion in research funding from the CHIPS Act, with packaging as the top candidate.

In 2023, Dr. Iyer joined the Commerce Department’s chip program and proposed a plan for a facility where researchers could develop and test new semiconductor materials, chips and packaging. He returned to UCLA in the fall of 2024.

The Commerce Department did not announce the proposed funding and location of the facility at Arizona State University’s research park until January 2025. Under Mr. Trump, nothing happened until Commerce Secretary Howard Lutnick declared in August that Natcast was an illegally created organization and won back its funding. The group disputed the claim, but disbanded last fall.

Not all industry executives mourn the potential packaging facility in Arizona. Mr. Gelsinger, now a general partner at venture capital firm Playground Ventures, says it should be in an existing research facility, rather than starting from scratch.

Dr. Iyer, who was not involved in the site selection, agreed. But he added, “I thought I put together a good program. I didn’t expect it to fall apart.”