Capex Locked, Stocks Slipped — The Week Four Countries Committed to 2028 While Wall Street Called the Peak

The week of August 4–8, 2026 will be remembered less for what memory stocks did than for what the physical semiconductor supply chain quietly agreed to. In a five-day window, SK Hynix's board approved 54.3 trillion won (~$38.3B) for two new fabs — Y2 in Yongin (35.2T won, HBM and next-gen DRAM) and M15X in Cheongju (19.1T won, next-gen NAND). On the same days, Taiwan's Nanya Technology (2408) raised its 2026 capex ceiling from NT$52B to NT$69.7B (+34%) to prepay equipment for its 5A fab. Japan's Daifuku, the fab-automation supplier feeding both, reported record H1 revenues of ¥355.5B with an all-time-high semiconductor equipment backlog. And in Washington, Trump signed an executive order imposing a 15% tariff on polysilicon effective December 6 — the raw feedstock all three countries depend on, with Japan negotiating a lighter carve-out.

Four countries. Four separate announcements. One coordinated commitment to build 2028 capacity — made in the same week Wall Street tried to declare 2Q27 the cycle top.

The pattern breaks the usual capex-lag story. Memory has always been a cycle where demand pulls capacity a year late, and the top is called somewhere near the second capex announcement. This time, the announcements landed simultaneously across geographies and across layers. SK Hynix (Korea, HBM) and Nanya (Taiwan, commodity DRAM) sit on opposite ends of the DRAM stack; neither would normally raise capex the same week unless both were being pulled by the same customer sheet. That sheet has one name at the top: hyperscaler AI training clusters that need HBM and inference boxes that need commodity DRAM in the same rack.

The demand-side confirmation came from three unrelated data points. First, iPhone 18 Pro is running short of DRAM weeks before its September launch — Apple, the single most disciplined memory buyer on Earth, could not lock enough LPDDR5 despite months of lead time, forcing what industry sources describe as SKU rebalancing. Second, NVIDIA is reportedly evaluating lower-spec HBM for its upcoming Rubin Ultra chips because the top-bin HBM4 volumes SK Hynix and Samsung can guarantee for 2027 are already spoken for. Third, DDR5 16Gb spot closed the week at $51.60 — a level PC OEMs have not seen since 2007 in normalized terms, per weekend commentary tracing the twenty-year price-decline reversal. Korea's June semiconductor exports came in at $33.6B, up 173.9% YoY — not a base-effect artifact. May was $29.4B (+154%) and April $25.2B (+158%). Three consecutive months of triple-digit growth against tough comps is a physical volume signal, not a price signal.

Yet the same week, equity markets punished the exact companies at the center of that physical signal. SK Hynix fell roughly 15% in two sessions on foreign selling, trading down to 1.42 million won by Thursday. Citi cut Micron's price target by 18%, publishing a note that named 2Q27 as the cycle peak. Japan's Nikkei semiconductor complex — Advantest, Tokyo Electron — dragged the index lower on the same session Daifuku reported records. The disconnect is not a mystery: fund managers are trying to sell the second-derivative deceleration (growth from 173% cannot compound) even while first-derivative volumes keep accelerating. What is unusual is that the physical-side companies are voting with their capex against that thesis. You do not commit $38B, prepay for a 5A fab, and expand automation backlog if you believe the top is nine months away.

Taiwan's memory tier is telling the same story in a different key. Winbond, Macronix, and Phison — a NOR/DRAM specialist, a code-storage NOR house, and a NAND-controller fabless — each hit all-time monthly revenue records in July, simultaneously. This is the Taiwan mid-tier that services embedded, industrial, and AI-adjacent boards the top three DRAM/NAND houses don't fully cover. When all three print records the same month, it means the commodity tail of the memory demand curve is being pulled up alongside HBM, not diverging from it. Zentel's July profit exceeding its entire paid-in capital — a Taiwan DRAM-controller fabless — is a fourth-angle confirmation.

Japan's participation deserves its own reading. Daifuku's H1 record and SMC's 96% Q2 profit jump are equipment-side confirmations that Korean and Taiwanese capex is real, not press-release capex. Dai Nippon Printing's record Q2 on semiconductor demand extends the confirmation to the materials layer (photomask, packaging substrates). Meanwhile the Nikkei is pricing Japan's structural weakness (design gap, talent gap, per a Nikkei column midweek) rather than the near-term earnings — exactly the wrong signal if the Korea-Taiwan-US equipment order book is the leading indicator. Japan's negotiation of a lighter polysilicon tariff carve-out matters here too: it protects the material inputs for the Korean and Taiwanese fabs Daifuku and DNP are equipping.

The polysilicon tariff itself is the fourth-country wildcard that ties the loop. Fifteen percent, effective December 6, 2026, on the raw feedstock for both semiconductor and solar-panel wafers. In the short term, this raises input costs for US-domestic fabs (Intel, Micron, TSMC Arizona) and pushes marginal costs into 2027 memory. In the medium term, it accelerates the domestic-polysilicon buildout the CHIPS Act envisioned but never delivered — and gives Korea and Taiwan a cost-of-goods buffer versus US-made competitors precisely as their 2028 fabs come online. Japan's carve-out preserves the JP→KR/TW materials corridor. The tariff structure implicitly rewards the four-country supply chain that just committed the capex.

For positioning, the split week says three things. First, the physical-side alignment across countries is stronger than the equity-side pessimism — if you believe the capex, the 2Q27 peak call is early by four to six quarters. Second, the marginal buyer of memory equipment is Taiwan (Nanya's +34% revision) more than Korea (SK Hynix was already committed), which raises Daifuku's, TEL's, and Advantest's Taiwan-mix exposure. Third, the polysilicon tariff creates a cost bifurcation: Korean and Taiwanese output gets a tailwind versus US-domestic memory, which is why Micron's target cut landing the same week SK Hynix committed $38B is not incidental.

The most useful frame: the physical semiconductor economy just aligned across four countries on a single 2028 bet, and the equity market decided to trade the opposite side.

Key Sources: - SK Hynix Invests 54.3T Won in New HBM and NAND Fabs (TheElec, 2026-08-07) - SK hynix approves $38B+ investment in two new memory fabs (SiliconANGLE, 2026-08-07) - Taiwan Weekly: Memory Capex Surge, AI Server Strength (Cnyes, 2026-08-08) - Trump signs EO imposing 15% tariff on polysilicon (DCD, 2026-08-07) - Sino-American Silicon Posts Record Q2 Revenue (TechNews TW, 2026-08-07) - plus 8 more