0

SK Hynix expects the AI memory shortage to last through 2030

In sum – what we know:

• Packaging, not fabrication – The Indiana site handles the 2.5D/3D stacking and interposer work on wafers shipped in from Korea, closing one of the weakest links in the US chip supply chain.
• A late-decade timeline – Cleanrooms come online in the second half of 2028 and volume HBM4E production is targeted for Q3 2029, with roughly 1,000 workers on site.
• Shortage through 2030 – CEO Kwak Noh-Jung says demand will outstrip supply even past 2030, and doubling capacity over five years may still not cover it.

SK Hynix has broken ground on a more than $4 billion advanced packaging facility at Purdue University’s Research Park in West Lafayette, Indiana — a 133-acre site the company is calling its first “Made in USA” mass-production base for next-generation high-bandwidth memory. The plant will package and test HBM4E, the memory that gets stacked alongside GPUs and AI accelerators in data centers, and it’ll house an R&D testbed for future packaging technologies. To be clear, this isn’t a wafer fab. Fabricated wafers will arrive from SK Hynix’s plants, most likely in South Korea, and Indiana will handle the 2.5D/3D stacking and advanced interposer work that turns them into usable HBM. That’s arguably the most supply-constrained step in the entire AI memory chain right now.

Cleanroom operations are scheduled for the second half of 2028, with volume HBM4E production targeted for the third quarter of 2029 — a slight pushback from an earlier NIST summary that had anticipated mass production in late 2028. CEO Kwak Noh-Jung told reporters the facility should eventually reach annual output in the hundreds of thousands of wafers, and the company expects roughly 1,000 workers on site during commercial operations, plus construction and indirect jobs around West Lafayette. 

The move puts SK Hynix in a notably different position than its rivals. Samsung and Micron are both investing heavily in advanced memory, but neither has committed to a U.S.-based HBM packaging hub at this scale. For the current market leader in HBM — the company Nvidia has leaned on most heavily for HBM3E — planting a packaging flag on American soil is an aggressive bet that proximity to U.S. customers will matter for the next decade of AI memory.

Supply chain impact

The Indiana plant slots neatly into what the CHIPS and Science Act was designed to do — pull critical semiconductor steps onto U.S. soil and reduce reliance on packaging hubs concentrated in East Asia. Advanced packaging has long been one of the weakest links in the American supply chain. Wafers fabbed in Arizona still typically travel to Taiwan or elsewhere in Asia for assembly, which undercuts much of the resilience argument. A high-volume HBM packaging line in Indiana closes at least part of that loop, and it diversifies operations away from China, Taiwan, and other Asian chokepoints that policymakers have been fretting about for years.

There’s a practical business case here too, not just a geopolitical one. Most of the customers for HBM4E are American companies. Packaging and testing memory in Indiana means shorter lead times, simpler logistics, and security assurances that shipping finished stacks solely from Asia can’t match. Nvidia and TSMC have already started producing Blackwell wafers in Arizona, and SK Hynix supplying HBM from Indiana would extend that on-shoring trend one layer deeper into the AI hardware stack.

SK Hynix gets a direct pipeline of engineers and technicians through workforce programs, internships, and research collaborations with Purdue, and Purdue gets a flagship tenant for its push to brand Indiana as a “Silicon Heartland” node in the global semiconductor ecosystem. The plant will still face the standard local scrutiny that follows any large semiconductor facility, particularly around water usage, energy consumption, and infrastructure demands, of course.

Long-term memory shortage

The groundbreaking came with a forecast that should get more attention than the ribbon-cutting. SK Hynix leadership says the current memory shortage — driven overwhelmingly by AI data center and inference build-outs — will persist through the end of 2030. Kwak has gone further, saying customer demand will likely exceed the company’s supply capacity even beyond 2030. 

The company isn’t hedging much on the near term either. Kwak has described the trajectory as leading to the “worst-ever” memory supply shortage in 2027, and SK Group chairman Chey Tae-won has warned that wafer shortages could exceed 20% in the coming years, simply because building sufficient capacity takes four to five years. SK Hynix plans to roughly double its memory capacity over the next five years, with the expansion tilted toward high-margin HBM and high-performance DRAM — which is exactly the product mix the Indiana facility is built around. Even so, the company acknowledges that doubling capacity may not fully cover projected demand.

If those forecasts hold, the implications cut both ways. Elevated prices for AI-grade memory support strong supplier profitability, which is good news for SK Hynix and its shareholders, but they raise costs for the cloud providers and AI companies buying the hardware. Sustained scarcity could also constrain the availability of cutting-edge accelerators outright, slowing deployments and pushing AI companies toward more efficient models and alternative hardware designs. We’ve already seen some of that behavior during the current crunch.