AI infrastructure demand has triggered a memory semiconductor super-cycle that is rewriting Asia's economic geography. Two companies control the chokepoints. Every prior cycle ended in oversupply. This one might not.
SK Hynix has roughly tripled in value this year. Samsung hit a record. The Korean Kospi touched an all-time high above 7,800. UBS called the semiconductor rally "unseen in nearly thirty years."
The numbers underneath the rally tell a more specific story than a broad tech boom. South Korea exported $173.4 billion in semiconductors in 2025, up twenty-two percent year over year. Saudi Arabia exported $187 billion in oil. The two figures are approaching parity for the first time — but the structural implications are opposite. Oil represents roughly seventy percent of Saudi total exports and about a fifth of GDP. Semiconductors represent twenty-four percent of Korean exports and about nine percent of GDP. Korea has the revenue concentration of a petro-state without the economic fragility of one.
The driver is high-bandwidth memory. HBM is the specialized chip that sits next to every AI training GPU, feeding it data fast enough to keep the processor from idling. SK Hynix holds roughly sixty percent of the global HBM market and is the dominant HBM supplier to NVIDIA. The company's entire 2026 output is already sold. Goldman Sachs projects undersupply extending into 2027. New fabrication plants take eighteen to thirty-six months to build, so the supply response lags the demand signal by at least two years.
The Chokepoint
TSMC reported $35.9 billion in revenue for Q1 2026, up over forty percent year over year, with sixty-one percent coming from high-performance computing and AI. The company holds approximately seventy percent of the global foundry market. Samsung Foundry is a distant second at roughly seven percent. For the most advanced chips — three nanometers and below — TSMC's share exceeds ninety percent. No other company on earth can manufacture them at scale.
The geographic concentration is stark. TSMC's primary fabrication capacity sits in Taiwan, within range of Chinese military exercises that have increased in frequency since 2022. A six-month blockade of Taiwan could disrupt over two trillion dollars in global economic activity, according to the Rhodium Group. TSMC is building fabs in Arizona — $165 billion committed across six fabrication plants and supporting facilities — and in Kumamoto, Japan. But the Arizona timeline stretches to 2027 for three-nanometer mass production, and the Japanese expansion to 2028. For the next two years, the world's most advanced semiconductor supply chain runs through a single island.
Why This Cycle Looks Different
The memory semiconductor industry has crashed five times in thirty years. Every crash followed the same script: demand surges, companies overbuild capacity, supply floods the market, prices collapse.
Starting in 1996, DRAM prices fell more than seventy-five percent over two years after fifty new fabrication plants were announced during the boom. The oversupply helped trigger the Asian Financial Crisis. In 2008, DRAM spot prices fell twenty-one percent in a single week after Lehman Brothers collapsed. Qimonda went bankrupt. The industry consolidated. From the 2018 peak to the 2019 trough, memory revenue declined over thirty percent and gross margins compressed from fifty-nine to twenty-seven percent. In 2023, SK Hynix posted a negative twenty-eight percent net margin as pandemic over-ordering met weakening demand.
The 2025-2026 cycle has not yet produced the signature oversupply signal. HBM capacity is fully committed. SK Hynix and Samsung have signaled high-teens to twenty percent price increases for 2026 — the inverse of every prior cycle's late-stage behavior. AI consumes roughly twenty percent of global DRAM wafer capacity, creating a structural demand floor that did not exist in previous booms. The HBM total addressable market is projected to reach $100 billion by 2028, roughly tripling from 2025.
The bears will point out that this is what the peak of every cycle looks like from the inside. Sold-out capacity, rising prices, structural demand stories — all of these were present in 1995, 2000, and 2017 before the correction arrived. The difference between a super-cycle and a bubble is visible only in retrospect.
The Geographic Bet
What makes this cycle structurally distinct is not the demand story but the geographic consequence. Oil concentrated wealth in the Persian Gulf for fifty years. Semiconductors are concentrating it in East Asia now, but in a narrower corridor: South Korea for memory, Taiwan for logic. Two countries. A handful of companies. The AI boom's manufacturing heartland is smaller than the oil boom's ever was.
The winners are legible: SK Hynix, Samsung, TSMC, and the economies they anchor. The losers are harder to see. Southeast Asian economies that competed on labor costs are watching value migrate to capital-intensive fabrication they cannot replicate. Countries without advanced semiconductor manufacturing capacity are becoming permanent importers of the most strategic commodity of the twenty-first century.
The risk is equally concentrated. A Taiwan Strait conflict would be the most destructive possible disruption to the global technology supply chain — not because alternatives don't exist, but because they are years away from operational scale. The super-cycle's wealth is real. So is the fragility that comes with it.
If AMD and custom silicon designs capture more than twenty-five percent of AI training compute by the end of 2027, the concentration thesis breaks. The chokepoint widens. Until then, the semiconductor super-cycle is a geographic story as much as a financial one — and the geography has never been this narrow.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)