The Lyceum: Semiconductor Weekly — Jul 07, 2026

Samsung just posted the most profitable quarter in the history of semiconductors — a 19-fold profit surge — and the stock closed down 7% on the session. That single fact tells you everything about where this week's tension lives: not in whether the AI memory boom is real, but in whether it's already priced in, and what happens when China starts designing its way around the constraints. SK Hynix is days from one of the largest first-time listings ever, DeepSeek is quietly building its own inference chip, and Korea's market is wobbling on "peak-out" fear even as its trade surplus hits a record. The data and the sentiment are looking at the same numbers and drawing opposite conclusions.

The numbers are almost hard to believe — and the market didn't care. Samsung guided to a second-quarter operating profit of ₩89.4 trillion on roughly ₩171 trillion in sales in its July 7 preliminary release — a 19-fold jump year-over-year, amid explosive demand for AI memory. Price increases hit not just high-bandwidth memory (HBM, the stacked DRAM that sits on every AI accelerator) but conventional DRAM and NAND flash too. According to Samsung's Global Newsroom, HBM4 is already in mass production, with first HBM4E samples slated for this quarter. (Samsung Electronics Announces Earnings Guidance for Second Quarter 2026 – Samsun)

After raising commodity DRAM prices roughly 90% in Q1 over Q4 reference levels, then another 50–60% in Q2, Samsung is now negotiating a further hike of up to 20% for Q3, per Wccftech's reporting.

So why the sell-off? The market is no longer pricing current earnings — it's pricing the top of the cycle. Albert Yong of Petra Capital Management told Reuters that the strong quarter was "widely expected and had largely been priced in," with investors worried about the sustainability of hyperscaler AI spending. A Morgan Stanley note the day before warned the chip pullback wasn't over, citing expectations that cloud operators will soon tighten capex.

The signal to watch is the July 30 divisional breakdown — specifically, how much profit came from HBM versus commodity DRAM, and whether the foundry business is actually recovering. Broad-based profit means the peak-out crowd is early. Profit concentrated in HBM means they may be right.

The world's dominant HBM supplier is about to start trading on Nasdaq — and this is no routine dual-listing. SK Hynix is selling American depositary receipts (ADRs) representing about 17.79 million common shares, an offering valued near $28 billion on Friday's Seoul close. Per data cited by Foreign Policy Journal via Bloomberg, that rivals Saudi Aramco's $29.4 billion 2019 debut, making it potentially the largest-ever U.S. listing by a foreign company.

The proceeds aren't for buybacks. They're funding the Yongin Cluster — a massive new fab campus in South Korea coming online from 2027 — plus a $4 billion packaging plant in Indiana, per CNBC. This is capacity, not financial engineering.

The thesis is simple: SK Hynix controls 57% of the HBM market by revenue as of Q4 2025, per Counterpoint Research. Every Nvidia Blackwell GPU that ships needs HBM stacked on top of it. Korean outlet coverage, echoing Bloomberg, reports the book was oversubscribed several times, with at least three U.S. firms signaling intent to underwrite roughly ₩10 trillion.

Watch the final pricing when trading opens. Price at the top of the range and open strong, and U.S. institutions are treating AI memory as a structural trade. A weak open says even the world's HBM leader can't outrun peak-cycle fear.

The most consequential supply-chain story of the week came from a Chinese AI lab, not a fab. Reuters, citing three people familiar with the matter, reported that DeepSeek is developing its own AI chip to reduce reliance on Nvidia and Huawei. Critically, it's designed for inference — the stage where a trained model generates responses — not for training.

That choice is deliberate and smart. Training is where Nvidia's software and interconnect lead is widest and where China's lithography constraints bite hardest under U.S. export controls. Inference is more forgiving on process node and more sensitive to per-query cost, which is exactly the workload DeepSeek runs at scale. The project has reportedly been underway about a year, with the company holding discussions with chip-design, foundry, and memory partners, per The Tech Portal. (DeepSeek begins developing its own AI chips to reduce reliance on Nvidia and Hua)

But the foundry question is where this becomes real or stalls. The U.S. bars Chinese designers from the most advanced overseas foundries, and separate curbs have cut China's access to HBM. There's no named foundry, no prototype, no benchmark — just hiring and discussions, per Let's Data Science.

Watch who gets the tapeout. SMIC at 7nm is the likely domestic path, but that node carries real performance penalties for inference at scale. If a non-Chinese foundry surfaces, it's an immediate export-control enforcement story.

The headline story has been conventional DRAM's historic surge — up roughly 93–98% QoQ in Q1, driving industry revenue up 81% to $97 billion, per TrendForce. The quieter, nastier signal is buried in the NAND data: the tightest spot in all of memory is now eMMC/UFS, the embedded storage that runs infotainment systems, ADAS controllers, and instrument clusters in cars.

The mechanic is brutal and simple. eMMC/UFS process capacity overlaps with enterprise SSD production but carries far lower margins, per Tom's Hardware. When Samsung, SK Hynix, or Kioxia choose between a hyperscaler's SSD order and an automotive Tier 1's eMMC order, the hyperscaler wins every time. TrendForce separately notes that older parts like NOR Flash and SLC NAND remain structurally short for the same reason — everything is being pulled toward AI-linked demand. (Structural Shortages to Keep NOR Flash and SLC NAND Prices ...)

For automotive and industrial procurement, this is a hidden bill already in the mail. Meaningful relief isn't expected until late 2027 at the earliest. The observable signal: watch whether any Tier 1 supplier publicly flags an eMMC allocation shortfall in H2 guidance. That's when "tight" becomes "line-down risk."

Korea's May current account surplus hit a record $38.61 billion — its 37th straight month in the black — driven primarily by semiconductor exports. Samsung and SK Hynix are together planning roughly ₩3,200 trillion (about $2.1 trillion) in domestic investment, per BigGo Finance. On paper, the demand side has never looked stronger. (Samsung Electronics Q2 Operating Profit Soars 19-Fold to ₩89.4 Trillion, Beating)

Here's the catch: Samsung and SK Hynix together make up more than 40% of the benchmark KOSPI, per CNBC. That concentration means "semiconductor peak-out" stops being a stock-picker's worry and becomes a sovereign macro exposure — the entire index rides on data-center capex staying hot.

This week the KOSPI fell more than 2% on peak-out fears even as the trade data screamed strength. That divergence is the whole story. If the volatility is sentiment, it fades. If it's an early read that hyperscaler capex is softening, it shows up in Korea's export numbers within two quarters — and a country's economy, not just a stock, takes the hit.

The signal to watch: whether Korea's monthly export data starts decelerating even as memory prices stay high. That's the tell that the boom is inventory-building, not consumption.

Somewhere in Seoul a stock chart is falling while the profit line behind it goes vertical; somewhere in Hangzhou an AI lab is cold-calling foundries that may not be legally allowed to pick up; and somewhere on an automotive assembly line, a procurement manager still doesn't know the cheapest chip in the car is about to become the one they can't get. The most expensive memory on earth is now cheaper per bit than the stuff in a $400 laptop — which is either the market telling you something profound or the market having a nervous breakdown, and this week nobody could agree which.

Stay skeptical of the ticker; the trade data lies less.

Forward this to the procurement lead who thinks eMMC is a solved problem — they'll want the next two quarters back.