The AI memory shortage has escalated from a commercial supply squeeze into a matter of "economic security," according to SK Group Chairman Chey Tae-won, who spoke to reporters on 19 July 2026 at the 49th KCCI Jeju Forum. Chey said foreign governments now treat access to high-bandwidth memory (HBM) the same way they treat energy or rare-earth supply, and that direct pressure between capitals is imminent. It is the first time the leader of the world's largest HBM producer has framed the shortfall as a geopolitical problem rather than a market cycle.
TL;DR
- Chey said customers want 60 to 100 percent more AI memory in 2027 than this year.
- Governments are treating HBM access as "economic security" — government-to-government pressure is imminent.
- SK Hynix holds 58 percent of HBM revenue (Q1 2026); Samsung and Micron trail at 21 percent each.
- No major producer has meaningful new HBM capacity arriving in 2027.
- Chey warned "abnormal" prices will invite competitors and said prices "have to normalise".
What did SK Hynix's chair actually say?
Speaking at the KCCI Jeju Forum, Chey said the shortage is no longer something companies can quietly absorb. Customers have requested 60 to 100 percent more AI memory in 2027 than this year, against AI accounting for more than half of semiconductor consumption and demand growth of at least 50 to 60 percent.
His most striking line: "No company has meaningful new capacity coming online next year." He described the current environment as "near-chaotic lobbying" from both corporate buyers and foreign officials, and said: "Right now, companies absorb the pressure. Governments will start pressuring other governments soon."
Why is HBM the chokepoint?
HBM is the stacked DRAM paired with Nvidia and AMD GPUs. Without it, an AI accelerator cannot feed compute units fast enough to train or serve large models, so HBM allocation rations who can build frontier AI and at what scale. Only three suppliers matter at volume, and one ships more than half of what the world uses.
That concentration is what makes Chey's warning politically loaded. SK Hynix held 58 percent of the HBM market by revenue in Q1 2026 (Counterpoint Research), with Samsung and Micron at 21 percent each. When a single Korean company controls the majority of a critical AI input, governments have an obvious incentive to intervene — as seen in the sovereign AI push HCLTech is backing in India.
How is SK Hynix responding on capacity?
Chey stressed that SK Hynix is building as fast as possible. The company pulled forward the first Yongin clean room to February 2027 from May and committed an additional 21.6 trillion won ($14.52 billion) in March 2026. The Cheongju M15X plant is being converted to dedicated HBM DRAM production.
A longer-term expansion into southwestern Korea is jointly planned with Samsung at the Gwangju military airfield. But Chey did not promise relief for 2027 buyers — fabs take years, HBM stacking is yield-constrained, and pulling schedules forward by months does not close a 60 to 100 percent demand gap.
Financially, SK Hynix has room to spend. It posted 79 trillion won in revenue and 47 trillion won in operating profit in 2025. Morgan Stanley projected 2026 operating profit could reach 179 trillion won ($123 billion). Its Nasdaq ADR debut on 10 July 2026 raised $26.5 billion.
What about US fabs and Washington's pressure?
Chey was notably cool on large US HBM fabs. He acknowledged US sites are being examined but framed it as an internal capital decision. SK Hynix's confirmed US footprint is a $3.87 billion packaging and R&D facility in Indiana. He was openly dismissive of Commerce Secretary Lutnick's push for Korean chipmakers to reshore production.
The subtext: Korea sees HBM as a strategic asset it does not want to hollow out. The contrast with logic manufacturing is stark — TSMC on 16 July 2026 announced a further $100 billion Arizona investment, taking its total US commitment to $265 billion. Memory suppliers are not moving on the same schedule.
Why does Chey want prices to fall?
Counterintuitively, Chey warned against sustained high HBM prices. He called current levels "abnormal" and said, "Prices have to normalise. Otherwise, the market shrinks and competitors flood in." He cited Tesla CEO Elon Musk's interest in chip manufacturing as an early sign of the kind of entrant that outsized margins attract.
The logic: incumbents make more per wafer at today's prices, but persistent shortages push customers to fund alternatives. Apple's M7 Ultra with 1.5 TB unified memory for AI servers is one signal that big buyers are already exploring routes that reduce exposure to discrete HBM.
What does this mean for AI buildouts and deal flow?
For hyperscalers and model labs, Chey's comments confirm that HBM — not GPU wafer capacity — is the binding constraint through 2027. That has consequences for infrastructure commitments like the $10 billion Meta-Anthropic compute deal and Meta's $50 billion Hyperion data centre. Announced compute is only as real as its memory allocation.
Longer-horizon bets face the same physics. SoftBank's $5 trillion AI programme through 2040 depends on HBM output growing several-fold over a decade. The Financial Times reported on 19 July 2026 that memory shares fell as investors priced in the opposite risk: that capacity expansion overshoots and produces oversupply by 2028. Both risks can be true in sequence.
FAQ
Q: What is HBM and why does the AI memory shortage matter? A: HBM is stacked DRAM paired with AI accelerators like Nvidia GPUs. Without enough of it, AI chips cannot ship in usable systems, so HBM supply directly caps how much AI compute can be deployed.
Q: How large is SK Hynix's share of the HBM market? A: According to Counterpoint Research, SK Hynix held 58 percent of the HBM market by revenue in Q1 2026, with Samsung Electronics and Micron each holding about 21 percent.
Q: When will new HBM capacity come online? A: SK Hynix pulled its first Yongin clean room forward to February 2027 and is converting Cheongju M15X to HBM DRAM. Chey said no producer has meaningful new capacity arriving in 2027.
Q: Is SK Hynix building HBM fabs in the United States? A: Not at scale. Its confirmed US footprint is a $3.87 billion packaging and R&D site in Indiana. US fab sites are under review, evaluated on economics, not political pressure.
Q: Why does SK Hynix's chair want HBM prices to come down? A: Chey said prolonged "abnormal" pricing would attract new entrants and provoke retaliation. Normalising prices protects long-term market structure even at the cost of near-term margins.
Q: How does this affect big AI infrastructure deals? A: Announced GPU and data-centre commitments only become deployed compute if HBM is available. With supply constrained through 2027, hyperscalers may see delivery timelines slip regardless of spend.
Discussion
0 comments