TSMC Weighs Making Advanced Memory Chips to Capture More AI Spending

Taiwan Semiconductor Manufacturing Co. is considering expanding into the advanced memory chip manufacturing sector, a move that would give the world’s largest chipmaker an even bigger share of the AI hardware supply chain, according to two people with direct knowledge of the matter.

The Taiwanese chipmaking giant recently started drawing up plans for making a specific type of advanced memory chips, 3D DRAM, at the request of Nvidia, its biggest customer, the two people said. An intense shortage of memory chips, driven by the AI hardware boom, has sent prices soaring and prompted chip designers such as Nvidia and Apple to look for sources of additional supply. Apple earlier this year hiked prices for its devices, citing rising memory costs.

DRAM, or dynamic random-access memory, holds data while a processor uses it. Unlike traditional DRAM, which puts memory cells in a single layer, 3D DRAM stacks layers of cells vertically to fit more memory. Memory chip manufacturing is now dominated by Samsung Electronics, SK Hynix and Micron Technology. The trio is also developing 3D DRAM technology, setting up a contest for the next generation of memory manufacturing.

It couldn’t be learned when TSMC plans to begin production or how the company intends to secure the manufacturing facilities required. TSMC could opt to build new factories, the two people added.

Another avenue it is exploring is acquiring an existing plant and converting it for memory production, the two people with direct knowledge said. Micron is doing something similar: in March, Micron bought a plant from Taiwan’s Powerchip Semiconductor Manufacturing Corp. for $1.8 billion, with plans to convert the site for advanced memory production.

TSMC’s choice of facilities and production scale would shape how quickly it becomes a competitor to established memory suppliers. TSMC has a history of making standard memory chips from the 1990s to the late 2000s. The company was disclosing revenue from memory until the second quarter of 2008.

“We don’t comment on market rumors or speculation,” TSMC said in a statement. “TSMC focuses on our core business at all times, which is dedicated [integrated circuit] foundry. Our mission is to be the trusted technology and capacity provider of the global logic [integrated circuit]] industry for years to come,” the statement said. Nvidia didn’t reply to requests for comment.

The move into memory chip manufacturing would leave TSMC vulnerable to a notoriously cyclical segment of the global semiconductor industry. Memory shortages in the past have typically spurred a rush to expand production, but frequently demand has declined by the time additional supply can be delivered. That drives down prices and chipmakers’ profit margins. The boom-and-bust pattern could motivate TSMC to hold off on the memory venture, especially if supply increases or demand subsides in the coming quarters.

TSMC’s memory initiative is led by Hsu Kuo-chin, a vice president who oversees advanced packaging, the technology that connects multiple chips into one tightly integrated system, the two people said. Hsu previously led Micron’s Taiwan operations, giving him experience in both memory manufacturing and chip integration, according to his company biography.

Another complication of TSMC’s memory push is that in 2024 it struck a partnership with SK Hynix to use TSMC’s manufacturing technology for the base chip beneath SK Hynix’s next-generation high-bandwidth memory. HBM stacks several memory chips to supply data quickly to AI processors; the base chip manages the connections.

TSMC has been exploring memory technology for years. It has several patents for vertical DRAM designs describing ways to pack more memory into each chip by building storage structures upward.

But turning research into reliable, large-scale production could take years, especially when it comes to 3D DRAM. Packing in additional layers increases manufacturing complexity and opportunities for defects. Samsung, SK Hynix and Micron have spent decades refining how to make memory cheaply and consistently.