The US biotech ETF (XBI) is up +88.8% in a year, and money is rotating from semiconductors into biotech. Three engines drive it: an M&A supercycle ($123B+ in H1), rate-cut hopes, and a reopened IPO window. Then two events in mid-August redrew the theme map: the first-ever Phase 3 win for an mRNA cancer therapy, and lab-validated protein design by an AI (Claude). This post explains all 8 themes in beginner language: next-gen obesity, China licensing, M&A, radioligand therapy, CNS, autoimmune cell therapy, AI drug discovery, and the mRNA renaissance. It then maps the Korean names actually connected by contracts (Alteogen, HanAll, LigaChem, ABL Bio, ST Pharm, OliX and more) on a 'contract substance × price momentum' grid. Includes 8 visuals. Data as of the 2026-08-19 close.
While the KOSPI rose 11.49% from August 10 to 14, retail investors net sold around 7 trillion won and foreign investors net bought in the mid-6 trillion won range. But when the full period since the July 30 low is split into phases, foreign buying was not continuous. It came in two bursts. Foreigners absorbed capitulation supply in a single day on July 31, sold back into the stall period in early August, and bought again from August 12 once the inflation-driven rally began. The methodological lesson of this breakdown is that cutting the window from August 1 instead produces the opposite conclusion: retail net buying and foreign net selling. And the rebound's buying returned to the same spot as the crash's selling. Half of net buying was concentrated in Samsung Electronics and SK Hynix, where foreign and institutional investors absorbed the supply that retail sold. The remaining caveat is futures and program trading. With cash equities bought but futures shorts not fully unwound, the structure is mixed, and confirming a full return to trend depends on four follow-up checks.
This piece verifies the claim that AI's bottleneck has moved from the GPU to HBM. Morgan Stanley's report frames the bottleneck as moving along the supply chain, from chips to power, memory, networking, and cooling, but the figure of roughly 50 billion Gb in 2027 HBM demand that circulates alongside it does not come from that document. It comes from Korean brokerages and research aggregation, and three separate estimates converge on the same range. The core of the claim holds up: compute grows three to four times per generation while bandwidth grows more slowly, agent inference occupies memory through the KV cache, and HBM consumes three times the wafers per bit while yield at 16 layers falls below half, so 2027 supply is effectively locked in by physics. But the same arithmetic is moving buyers too. NVIDIA cut module capacity in half once memory cost reached 29% of system cost, reports say the next flagship's HBM specification will be set below the prior generation, and prefill-only chips skip HBM entirely. It lays out the bottleneck's reality alongside the workarounds it is provoking, and where that leaves the profits of Korean memory makers in between.
Two things became official on a single day, August 10. The SEC excluded data center securitization bonds from asset-backed securities regulation, and NVIDIA announced a $500bn third-party capital platform with six partners ranging from Apollo to KKR. The shift of AI capital expenditure financing from corporate balance sheets to capital markets was confirmed on both the regulatory and structural fronts in one day. The same week, CoreWeave and Nebius demonstrated demand through earnings, SanDisk disclosed that multi-year contracts now cover two-thirds of its NAND shipments, and SK Hynix Chairman Chey Tae-won pulled out, on camera, a supply request that Jensen Huang had handwritten on a wafer. US inflation data broke the odds of a rate hike, and the KOSPI rose for four straight days. This piece reviews a week in which all three observation items left by the previous post were filled in, and lays out what remains unresolved.
Last week the KOSPI rose 3.76%, fell 4.58%, and fell another 0.60% over three days. The trigger for the plunge on the 6th was not macro but guidance from AMD and SanDisk, and oil prices and interest rates were actually favorable that day. This piece scores five candidate axes for a rebound (macro, NVIDIA earnings, shareholder returns, export data, and product qualification) for impact and timing. The conclusion is that a single trigger is not enough. NVIDIA beat estimates for four straight quarters without its stock rising, and Samsung Electronics closed lower on the day it reported its highest-ever earnings. Then on August 7, SK Hynix announced a $38bn capacity expansion, making shareholder returns and supply discipline two sides of the same decision. The earliest leading signal is not a trigger but the first instance of two consecutive up days on good news.
Microsoft and Amazon have, for the first time, shown convincing evidence that AI capital spending is converting into real revenue and profit. Azure grew 43% and topped $100bn in annual revenue. AWS grew 36.7%, with an operating margin of 39.4%. All of the $51bn net increase in Microsoft's commercial backlog came from customers outside the frontier model companies. Yet over the same period, foreign investors sold more than 11tn won of Samsung Electronics and SK Hynix alone while net buying the rest of the Korean market. Demand was proven, so why did memory get sold? This piece lays out three hypotheses, a timing lag, a pricing regime shift, and supply arrival, and scores each one. It then examines how NVIDIA's move to cut Vera Rubin's module capacity in half over memory costs changes that scoring.
A fact-checked analysis of the joint U.S.-Japan yen intervention, how the Fed's FIMA repo facility actually works, the oil and Treasury rally after Trump paused new Iran strikes, and the implications for Korean memory stocks and the August macro regime.
On August 3, the KOSPI fell 5.12% while Samsung Electronics and SK Hynix dropped 8.8%, and the KOSDAQ actually rose 2.44%. This asymmetry, large-cap memory giving back twice as much as the index, is the real information of the day. The third-quarter contract price deceleration data cited as the cause is actually a month old; what is new is not the data but the market's room to look at it. As oil fell from $116 to $79 and the first joint US-Japan intervention since 1998 reversed the yen, the market shifted its question from the discount rate to how long the profits last. The deceleration stems not from oversupply but from a limit on consumers' willingness to pay, and the contract structure has actually shifted in suppliers' favor. The 2019 data point shows that even when bits grew 19%, a 47% price decline cut revenue by 37.5%. So the question for this cycle is shifting from whether prices rise further to whether they hold.
Micro evidence improved at the end of July, but oil, U.S. long yields, and the yen carry trade remain the three macro gates for the August rebound. This report separates confirmed facts from overstatements, maps the August 3-7 calendar, and sets scenarios for big tech and memory stocks.
An independent report closing out the July earnings season. It synthesizes ten earnings reports and calls, from TSMC to Samsung Electronics, SK Hynix, Microsoft, Meta, Amazon, Apple, and IBM, two record-breaking listings (SK Hynix's Nasdaq ADR and CXMT's Shanghai debut), and the Kimi K3 shock into a single picture. The data is consistent. Cloud growth accelerated across the board and everyone talked about a supply shortfall, yet the market has started grading capital expenditure against revenue proof in the very same quarter. The least-noticed variable is memory pricing. It became Amazon's stated reason for raising capital expenditure, Apple called it a hundred-year flood, and the recipients on the other side of that flow are Korea's two memory makers. Building on this evidence, the piece sorts out what earnings actually answered for the market's four worries: token demand, frontier-lab profitability, capex ROI paired with supply expansion, and semiconductor profitability beyond 2028.