Tesla's capacity cut and Korea's +203% export print mark peak demand, but NVIDIA's rising DSI and China's 343 DUV units are the quieter signals suggesting the cycle has turned.
Tesla's confirmation that it is throttling AI5 chip capacity because memory supply cannot keep pace is the single most revealing data point in the current cycle[1] — and the company simultaneously raised the AI5 memory specification from 72GB to 96GB,[2] meaning per-unit DRAM demand is rising even as unit output is being cut. Pair that with Korea's August customs HS8542 semiconductor exports hitting $38.2B, up +203% YoY,[3] and the picture is a supercycle running at redline. For SK Hynix (000660), this is the context the market is pricing — the question is whether that pricing captures what comes next.
The HBM–DRAM split is the defining feature of this cycle. Samsung Electronics (005930) is tracking toward Q3 operating profit that may breach 100 trillion won for the first time[4] — a number that only works arithmetically if HBM pricing holds. SK Hynix (000660) sits one layer deeper as the primary HBM supplier to NVIDIA (NVDA), whose inventory days of supply climbed to 119.7 in the latest print from 115.1 the prior quarter.[5] That rising DSI at the single most important customer is the first number that does not fit the "demand is infinite" narrative. If NVDA is accumulating memory faster than it ships systems, the pull-through assumption weakens.
The supply-chain neighborhood tells a parallel story. Hanmi Semiconductor (042700), the TC bonder central to HBM stacking, and Soulbrain (357780), a key etchant vendor, both ride SK Hynix capex — Korea's semi equipment-materials export index printed 274.05 in August, +82% YoY.[6] The capex is real. But across the strait, Nanya Tech (2408) selected Pingtung for a new DRAM fab[7] the same week China was reported to have secured 343 immersion DUV lithography units,[8] and Yangtze Memory's IPO filing points to NAND capacity surge from the mainland.[9] Taiwan's Quanta (2382) posted August revenue of NT$424 billion, +177% YoY[10] — AI server pull-through remains intact, but each new fab announcement shortens the pricing-dominance window. Foreign investors net-sold 20 trillion won of KOSPI equities since September, with roughly 17 trillion won concentrated in Samsung and SK Hynix.[11] Flow data is screaming what the fundamentals are only whispering.
What Cuts Against
The counter-signals could equally reflect catch-up supply discipline rather than imminent inflection. NVIDIA's elevated DSI may simply pre-stage Blackwell and Rubin ramps rather than signal glut — the GB300 assembly line hitting a 95% busbar success rate on Foxconn's demonstration run[12] implies the system-level bottleneck has moved, not disappeared. And the custom HBM programs now underway between SK Hynix, Samsung, NVDA, and AMD[13] could re-price the merchant HBM market lower while leaving the economics of specific hyperscaler contracts intact. "Peak" in merchant pricing does not always equal peak in the business.
Where This Lands
The data points toward a cycle whose demand side is at or near its narrative peak while supply responses — Chinese DUV accumulation, Nanya Pingtung, Yangtze NAND — are only now building the next wall. The 20 trillion won foreign outflow from Korean memory names is price-coherent with that reading. If NVDA DSI continues rising through the Oct 15 TSMC earnings print and Samsung's Q3 confirmation, the inflection thesis strengthens; if DSI compresses as Blackwell ships, the story is pre-ramp pacing, not peak. The resolution should arrive inside the next two earnings weeks.
AI-generated synthesis of public disclosures, prices, trade data, and supply-chain relationships. Directional commentary uses hedged language and is informational, not a solicitation to buy or sell any security.
What to watch
- TSMC Q3 earnings call Oct 15 — CoWoS/HBM-adjacent commentary and 2027 capex guide set the ceiling for Korean memory read-through
- Samsung Electronics (005930) Q3 preliminary earnings (mid-October) — whether operating profit actually breaches the 100 trillion won threshold
- NVIDIA (NVDA) next quarterly DSI print — 119.7d trajectory decides pre-ramp pacing vs genuine demand peak
Sources
- [1]조선일보 — Tesla reduces AI chip capacity due to memory shortage— Tesla is throttling AI5 chip capacity because memory supply cannot keep pace
- [2]新浪财经 — Tesla raises AI5 chip memory from 72GB to 96GB— Tesla simultaneously raised the AI5 memory specification from 72GB to 96GB
- [3]Korea customs HS8542 export, 2026-08— Korea's August customs HS8542 semiconductor exports hitting $38.2B, up +203% YoY
- [4]한겨레 — Samsung Q3 operating profit may breach 100 trillion won for first time— Samsung Electronics is tracking toward Q3 operating profit that may breach 100 trillion won for the first time
- [5]NVDA DSI filings-derived, 2027-Q1/Q2— NVIDIA inventory days of supply climbed to 119.7 in the latest print from 115.1 the prior quarter
- [6]Korea semi export index 403Y003, 2026-08: 274.05 (+82.47% YoY)— Korea's semi equipment-materials export index printed 274.05 in August, +82% YoY
- [7]technews — Nanya Tech selects Pingtung for new DRAM plant— Nanya Tech (2408) selected Pingtung for a new DRAM fab
- [8]뉴시스 — China secures 343 immersion DUV lithography tools— China was reported to have secured 343 immersion DUV lithography units
- [9]user.guancha.cn — Yangtze Memory IPO signals Chinese NAND capacity surge— Yangtze Memory's IPO filing points to NAND capacity surge from the mainland
- [10]Quanta (2382) monthly revenue 2026-08— Taiwan's Quanta (2382) posted August revenue of NT$424 billion, +177% YoY
- [11]Chosunbiz — Foreign investors sell 20 trillion won in Korean stocks; Samsung, SK Hynix lead outflows— Foreign investors net-sold 20 trillion won of KOSPI equities since September, with roughly 17 trillion won concentrated in Samsung and SK Hynix
- [12]cnyes — NVIDIA and Foxconn deploy robotics in GB300 assembly, busbar rate tops 95%— GB300 assembly line hitting a 95% busbar success rate on Foxconn's demonstration run
- [13]뉴시스 — Memory industry accelerates custom HBM development— custom HBM programs now underway between SK Hynix, Samsung, NVDA, and AMD
If this analysis was helpful · ☕ Support Us · ✈️ Telegram