SK Hynix has reportedly broken ground on a new advanced memory packaging facility in West Lafayette, Indiana, that should boost the supply of US-made high-bandwidth memory (HBM), a key component in high-end AI accelerators from the likes of Nvidia and AMD.

Unlike the system memory used in notebooks, smartphones, and general purpose servers, HBM requires specialized packaging and test services to stack multiple layers, often 8-12, of DRAM together to form a single module.

Today, most HBM memory is manufactured in Korea by Samsung and SK Hynix, with Micron being the United States' primary source of homegrown memory. That will change slightly when SK Hynix's Indiana plant, expected to cost nearly $4 billion to build, is completed in late 2027.

Late last week, South Korean newspaper the Herald Economy reported that SK had notified local officials that construction of the facility's foundation was now underway, with vertical construction slated for later this year.

Alongside the packaging plant, the facility will also house an R&D facility where SK Hynix will develop future generations of chips. 

According to the industry watchers at TrendForce, the facility is unlikely to begin full-scale operations until the second half of 2028, around the time when various brands of HBM4E memory expected to hit the market beginning with Nvidia's Rubin Ultra platform.

This week, SK Hynix also broke ground on a similar advanced packaging facility at the Cheongju Technopolis Industrial Park located in Heungdeok-gu, Chungcheongbuk-do, South Korea. 

While SK's new plant in Indiana should help to bolster supply of US-made HBM, the plant will come online too late to offer much if any relief from the current memory shortage, which is expected to peak later this year.

SK's Indiana plant will, however, address growing demand for US semiconductor manufacturing capacity. Since TSMC began producing chips at its Arizona fab early last year, Nvidia, Apple, and AMD have announced plans to produce chips in the US.

• Datacenter boom keeps dirty coal plants alive in the US
• Forget one chip to rule them all: With TPU 8, Google has an AI arms race to win
• Oil crisis? What oil crisis? IT spending de-coupled from wider war shock
• Peace President's Iran war piles more pain on already battered PC market

Ironically, gaps in the US semiconductor ecosystem have meant some chips produced at TSMC's US facilities, notably Nvidia's Blackwell GPUs, have had to make a round trip to Taiwan and back for final integration.

TSMC is working with Amkor Technology, an outsourced semiconductor assembly and test provider, to bring its chip-on-wafer-on-substrate (CoWoS) advanced packing capabilities to the US, but this won't happen overnight.

Back in October, Amkor broke ground on the $7 billion, 750,000 square foot facility, with construction expected to be completed in the middle of next year and production slated to start in early 2028. That's coincidentally just in time for SK Hynix's HBM factory to ramp production – it will need advanced packaging capabilities to turn DRAM into HBM, although no telling yet whether Amkor will be its provider.

In the meantime, Intel's Foundry division has reported growing interest in its own advanced packaging capabilities, which while different from TSMC's CoWoS, can be made compatible with silicon manufactured by the Taiwanese foundry giant.

According to Intel CFO David Zinsner, advanced packaging is likely to be one of the early revenue drivers for the company's foundry business.

Speaking at the Morgan Stanley Technology, Media, and Telecom conference last month, Zinsner suggested Intel could announce a packaging win before the end of the year. ®

More about

• AI
• High bandwidth memory
• Nvidia

More like these