SK Hynix announced a $4 billion expansion of its U.S. HBM manufacturing capacity to capitalize on sustained AI chip demand.

SK Hynix announced Thursday that it plans to begin volume production of its next-generation HBM4E memory chips at its Indiana facility in the third quarter of 2029. The South Korean chipmaker is positioning itself for what Chief Executive Kwak Noh-Jung expects to be a prolonged shortage of advanced memory driven by the expansion of artificial intelligence.

Kwak stated that the company anticipates the memory shortage will persist through the end of 2030, signaling a bullish outlook for a market that has recovered from a severe downturn and emerged as one of the primary beneficiaries of the AI infrastructure buildout. “We see no clear signs of a memory downturn,” Kwak said, adding that any eventual slowdown was more likely to involve a moderation in demand than a sharp decline.

These comments underscore how fundamentally AI has reshaped the memory-chip market. High-bandwidth memory (HBM) has become an essential component of advanced AI accelerators, enabling processors to access vast amounts of data at extremely high speeds while consuming less power than conventional memory configurations. For SK Hynix, this demand has transformed HBM from a relatively specialized product into a major profit driver. According to Counterpoint Research, SK Hynix captured 58% of the global HBM market by revenue in the first quarter of 2026, compared with 21% each for Samsung Electronics and Micron Technology. This leading position provides a significant advantage as AI chipmakers compete for increasingly scarce advanced memory.

To capitalize on this shift, SK Hynix is investing more than $4 billion in its Indiana project, marking a major expansion of its U.S. manufacturing footprint. The facility, located at Purdue Research Park in West Lafayette, is scheduled to begin cleanroom operations in the second half of 2028, followed by volume production of HBM4E in the third quarter of 2029. By 2030, SK Hynix expects Indiana to serve as a critical U.S. production base for HBM as the company builds what it describes as an advanced domestic memory supply chain. The site could eventually reach an annual production capacity of hundreds of thousands of wafers.

The Indiana facility will house a next-generation HBM packaging line and an AI semiconductor research and development center. This combination is strategically vital because HBM production extends beyond conventional wafer manufacturing. Advanced packaging has become increasingly critical to AI-chip performance, as memory must be closely integrated with processors to meet the bandwidth and power requirements of large-scale AI systems. Geographically diversifying production also mitigates risk. While much of the world’s semiconductor manufacturing remains concentrated in Asia, U.S. technology companies are rapidly expanding domestic data-center capacity. Locating advanced memory packaging and research in the United States reduces logistical constraints and improves supply-chain coordination with major customers, including Nvidia, Microsoft, and Alphabet’s Google.

The timing aligns with extraordinary growth in AI computing demand. On Wednesday, Nvidia forecast that its revenue would increase 70% in the next fiscal year, reinforcing expectations that spending on AI data centers will remain elevated. Structurally, HBM differs significantly from traditional memory products. The chips are vertically stacked and tightly integrated with AI processors, requiring sophisticated manufacturing and packaging capabilities. These technical barriers limit new market entrants and have allowed leading suppliers to command stronger pricing than typically seen in commoditized memory markets.

This dynamic helps explain Kwak’s unusually long-range outlook. Historically, memory manufacturers have endured sharp boom-and-bust cycles, with oversupply causing prices and profits to collapse. However, the rapid expansion of AI infrastructure has introduced a new demand source less dependent on traditional PCs and smartphones. SK Hynix is effectively betting that AI accelerators, data centers, and increasingly sophisticated computing systems will continue absorbing HBM output even if other semiconductor segments weaken. Beyond Indiana, the company approved approximately 54.3 trillion won ($38.3 billion) in investments through 2031, including 35.2 trillion won for the second phase of its semiconductor manufacturing facilities in South Korea.

The U.S. project is bolstered by Washington’s push to build domestic semiconductor capacity. In December 2024, the U.S. government finalized $458 million in CHIPS Act grants and up to $500 million in loans for SK Hynix. The Indiana facility is projected to generate approximately 7,000 direct and indirect jobs locally while strengthening the U.S. semiconductor supply chain.

SK Hynix is actively positioning itself at the center of the AI hardware ecosystem, recognizing that access to advanced memory has become a key constraint on scaling AI systems. Its HBM leadership places it in direct competition with Samsung and Micron, both of which are investing heavily to expand their market share. The three memory giants are now competing not only on production volumes but on their ability to deliver more advanced HBM generations featuring higher bandwidth, lower power consumption, and improved integration with AI processors. SK Hynix’s decision to commit to HBM4E production years in advance signals its expectation that technological leadership and manufacturing capacity will remain critical as newer chip generations replace current ones.

Kwak’s forecast that shortages could persist until the end of 2030 represents more than a near-term pricing call; it is a bet that AI will fundamentally reshape the memory cycle and sustain demand for advanced products well beyond the current wave of data-center construction. As always in semiconductors, the risk lies in companies collectively building capacity faster than demand materializes. Yet SK Hynix’s investment strategy suggests it currently views the greater danger as being unable to supply the next generation of AI systems rather than possessing excess advanced memory capacity.