SK Hynix’s $60 Billion Buildout Does Not End the AI Memory Squeeze

What SK Hynix Actually Announced

SK Hynix’s new Cheongju plan is large, but the timing matters more than the headline number. On July 2, the company announced KRW 100 trillion in investment for Cheongju: KRW 80 trillion for M17, a NAND production fab, and KRW 20 trillion for P&T7, an advanced packaging facility and related infrastructure. SK Hynix says M17 construction starts in 2027, with operations targeted for the first half of 2029. P&T7 is expected to be completed by the end of 2027.

Reuters put the package at about $64.4 billion. PC Gamer framed it as more than $60 billion while noting the practical consumer angle: the memory shortage is not ending soon. That is the right caution, even if the consumer-PC framing is too narrow.

This is not one interchangeable “AI chip factory.” The Cheongju package is specifically NAND plus packaging. NAND matters for storage in AI systems and data centers. Packaging matters because high-bandwidth memory is not just a wafer-output problem; stacking, testing, substrates, and integration can become binding constraints. But the most direct DRAM/HBM capacity expansion is tied to broader Korean plans, including SK Hynix’s Yongin strategy.

The Bottleneck Is Becoming Capex

The useful read is not “AI shortage solved” or “AI bubble gamble.” It is that a shortage has turned into dated industrial capex. Scarce memory raises margins. Higher margins justify fabs. Fabs require years. The resulting capacity affects late-decade AI costs, not next quarter’s server availability.

SK Hynix’s Cheongju spending sits inside a much larger domestic plan. The company has described a KRW 1,100 trillion mid-to-long-term South Korean investment strategy: KRW 600 trillion for Yongin, KRW 100 trillion for Cheongju, and KRW 400 trillion for a southwestern semiconductor cluster. South Korea’s government has separately promoted mega-projects totaling KRW 1,558 trillion, including KRW 800 trillion for four memory fabs in the southwest.

That does not make execution automatic. Official industrial plans are not capacity. Power, water, permits, equipment delivery, skilled labor, yields, customer qualification, and the durability of AI capex all remain hard constraints. The relevant fact is that Seoul and the memory majors are trying to convert AI demand into physical industrial capacity, not that they have already done so.

Why It Matters for AI Costs

Memory is no longer a secondary component in AI economics. HBM bandwidth affects accelerator utilization and training throughput. NAND affects storage-heavy data-center architecture. Advanced packaging affects whether memory can be integrated at the scale and performance AI accelerators require.

TrendForce projected SK Hynix would remain the largest HBM supplier in 2026, with roughly 50% of global HBM bit output, down from 59% in 2025, while Samsung rises from about 20% to 28%. TrendForce also said Samsung, SK Hynix, and Micron were in final HBM4 validation stages in early 2026 and expected Nvidia to use all three because no single supplier could fully meet Vera Rubin requirements.

That detail matters. AI supply chains are not waiting on one magic fab. They are constrained across memory wafers, packaging, validation, and customer allocation. SEMI’s forecast points in the same direction: global 300mm fab equipment spending is projected to rise 18% to $133 billion in 2026 and another 14% to $151 billion in 2027, with AI-driven HBM, NAND flash, memory capacity additions, and policy-backed localization named as drivers.

What Remains Uncertain

The dollar value shifts with exchange rates, so “$60 billion” and “$64 billion” are both approximations of the KRW 100 trillion plan. It is also not clear how much of P&T7 will be dedicated to HBM-related work versus other advanced packaging and testing.

The bearish risk is overbuilding into a later downturn if AI infrastructure spending slows. The bullish risk is that demand keeps outrunning supply despite these projects. For now, the practical conclusion is narrower: SK Hynix’s buildout shows the AI memory bottleneck is being industrialized, but the relief is measured in years, not months.