Four Ledgers, One Number — The Week Korea's Customs, Japan's ICT Print, and Broadcom's Guide All Cleared +180%

Four Ledgers, One Number

On August 14, three completely separate accounting systems reported the same growth rate.

Korea's Ministry of Trade published July semiconductor exports at $32.7B, up 166.29% YoY — the third consecutive month above +150% (May +154.3%, Jun +173.87%, Jul +166.29%). Japan's MOF confirmed July semiconductor/computer exports surged 178.8% YoY, part of a record $53.36B ICT print. And in San Jose, Broadcom guided fiscal 2026 AI revenue at $56 billion, up 180% from prior year.

Three ledgers. Three currencies. Three different reporting cadences. One number: roughly +180%.

That kind of convergence is not supposed to happen. National trade data lags corporate earnings by design — customs settles physical shipments after they clear ports, while forward guidance prices intent. When both agree, it means the AI capex → supply-chain revenue conversion has fully synchronized. There is no longer a "lead" indicator.

Why the Convergence Matters

For the last 24 months, the semi cycle has been read through fragmented signals: TSMC monthly revenue, Samsung memory prints, Nvidia's data-center segment, ASML bookings. Each measured a different slice of the same wave, and each had its own lag structure. Bulls argued the wave was still building because guidance kept beating; bears argued it had already peaked because trade data was slower to react.

This week resolved that debate — in favor of the bulls, but with a twist. Korea's export surge, Japan's customs print, and Broadcom's forward guide all clustered at +180%. That doesn't mean growth is decelerating. It means the entire supply chain — from GPU tape-out in Santa Clara through advanced packaging in Hsinchu, HBM stacks in Icheon, and Advantest testers in Gunma — is now moving at the same effective velocity. Every link is fully utilized and pricing to that utilization.

The corollary is uncomfortable. If four independent measurement systems now travel in lockstep, they will also decelerate in lockstep. There is no diversification between the KR trade balance, JP customs, TSMC capex, and Broadcom guidance the moment big-tech capex intent changes.

The Korean Anchor

Korea's $32.7B July semi export is essentially an HBM tape-out invoice. SK Hynix has become the physical throat of the AI supercycle: iHBM shipments to Nvidia and to Broadcom's custom silicon program register as customs-cleared exports the following week. SK Square posted Q2 operating profit of ₩19.2 trillion — a number driven overwhelmingly by SK Hynix equity income, per the disclosure.

Yet the retail-vs-foreign split turned strange this week. Foreign investors net-sold SK Hynix and Samsung Electronics even as retail accumulated on the way down. Two mutually contradictory reports circulated on August 14 — one claiming foreign inflows into semis, another documenting net-outs. Reconciled: foreigners rotated inside the sector, out of the two megacaps and into second-tier substrate and equipment names (Solbrain Q2 op-profit +128% YoY, Semitech +11× YoY, TLB +86% YoY, Agicland H1 revenue +273% YoY).

The rotation is not bearish on Korea. It is bearish on the pure-play HBM trade and bullish on the Korean below-the-die stack that had been trading at half the HBM multiple.

The Japan Mirror

Japan's July ICT exports at $53.36B (semis +178.8%) is the same wave measured with different physics. What Japan ships is upstream of what Korea ships: silicon wafers, photoresist, test equipment. Advantest closed at record highs for a fourth session as Nomura raised its target; DIC posted record H1 operating profit on chip materials and lifted full-year guidance; Japan's advanced packaging market is now forecast at $7.4B by 2035 (9.7% CAGR).

The Japanese story has a hidden constraint the market has not yet priced. On August 15, Agora Web ran a piece flagging power shortages as Japan's semi Achilles' heel. TSMC's Kumamoto ramp, Rapidus in Chitose, and any Advantest capacity expansion all draw on grids that were not designed for AI-era loads. Japan can measure the AI supercycle at +178.8% on customs data today; it cannot deliver the physical throughput to grow that further without new base-load generation.

The Taiwan Convergence

TSMC's board approved a $29.4 billion capex plan and signed a binding JV with Sony Semiconductor to co-manufacture smartphone image sensors — a deal that formalizes what Kumamoto was always going to become. Meanwhile Taiwan's contract manufacturers are the fastest-moving nodes in the Rubin supply chain: ASUS entered Nvidia's Vera Rubin supplier list this week and traded toward NT$1,000; Compal +18% with a target of 30–40% server mix by 2027; Gigabyte doubled Q2 profit YoY and raised capex to NT$2.8–3.0B; Tripod Technology posted H1 EPS +42.6% on server PCBs. Foreign investors net-bought NT$206B on Taiwan for five straight sessions.

The Taiwan tape reads exactly what Korean and Japanese trade data measure — but from the box-build angle rather than the input-materials angle. Same wave, third-order visibility.

The US Anchor

Broadcom's $56B / +180% AI guide is the number that reconciles the three trade-data prints. It is also this week's most-cited figure in Micron's re-rating (stock +2.1% on tightening DRAM/NAND, valuation debated as undervalued given the HBM ramp). x86 CPUs are now forecast to dominate a $200B AI CPU TAM through 2030 — a curious counter-narrative to the ARM-in-server thesis that has powered Nvidia Grace.

The one dissenter is Pat Gelsinger, who this week called HBM "lousy." SK Hynix's response was diplomatic: HBM is "not the final answer." The subtext: HBF (High Bandwidth Fabric) is accelerating toward commercialization, and if it displaces HBM in the 2028+ generation, the current +180% convergence flips sign quickly.

What to Do About It

The convergence framework produces three actionable positions:

1. Long the second tier, not the memory megacaps. If the +180% number is now fully priced across KR/JP customs, TSMC capex, and AVGO guidance, further upside for the megacaps requires above-consensus growth — increasingly hard. The delta lives in the below-die names (Solbrain, Semitech, TLB, DIC, Advantest) that were still trading at reset multiples entering the summer.

1. Watch Broadcom's next quarterly guide as the leading break. Because AVGO is the only one of the four ledgers that is forward-looking, its next revision telegraphs whether the convergence is holding or cracking. A cut from +180% to +150% would trigger a KR August-September print decel roughly two months later.

1. Price HBF as a 2028 tail risk, not a 2026 story. Gelsinger's "lousy" and the HBF news are noise this cycle, signal next cycle. If SK Hynix pursues a US listing now, it does so precisely to price HBM optionality before HBF displaces it.

The AI supercycle is now measured identically in four accounting systems. That is the bull case and the bear case in one number.

Key Sources: - SK Square Q2 Operating Profit Hit 19.2 Trillion Won on SK Hynix Equity Gain (자본시장뉴스, 2026-08-15) - Broadcom guides $56B AI revenue for 2026, up 180% from prior year (eciks.org, 2026-08-14) - July ICT Exports Hit Record High at $53.36B; Semiconductors Surge 178.8% (디지털투데이, 2026-08-14) - Taiwan Weekly: TSMC Sony JV & $29.4B Capex, MSCI Weight Hike, AI Server Surge (cnyes, 2026-08-15) - Gelsinger calls HBM 'lousy', SK Hynix says it's not memory's final answer (Digitimes, 2026-08-14) - plus 15 more