Human-written opinions, AI deep dives, single-ticker research, and factor notes — all the analysis output in one place.
SLC NAND's 120–170% H2 forecast, Kioxia's Japan-cap rally, and Quanta's doubled June revenue point to an AI-server storage bottleneck whose primary read-through sits at SK Hynix.
SLC NAND contract prices are forecast to surge 120–170% in H2 2026 on structural shortage. On the same day Kioxia hit Japan's top market cap in a semiconductor-led rotation, and Taiwan AI-server ODM Quanta (2382) posted June revenue of NT$385.2B — MoM +23.7%, YoY +102.9%. Three data points, three lenses, one story: AI-server storage tightness has escalated from a spot-pricing anomaly into a hyperscaler-level bottleneck, and the primary beneficiaries sit in Korea.
Samsung's 19x profit, Korea's +154% HS8542 exports, and Quanta's +94% revenue confirm the AI memory boom is real — but a plunging tape and rising NVIDIA inventory suggest the market is already pricing the top.
Samsung Electronics (005930) Q2 operating profit surged ~19x YoY, briefly topping NVIDIA (NVDA)'s quarterly record. Korea's May HS8542 semi exports hit $29.43B, +154% YoY. Quanta (2382) May revenue was NT$311.5B, +94% YoY. Every fundamental confirms the AI memory build — yet memory equities plunged this week and NVIDIA's DSI climbed 105.9d → 115.1d in three quarters. The tape and the tape's biggest customer are disagreeing with the fundamentals.
The market sold the companies that rent out GPUs as though the rent had already collapsed. In the numbers I actually track, it hasn't — the newest chip hit a new high the day after the crash. This is a column about the gap between what the stocks did and what the price did, written by someone who trusts the price sheet over the panic.
Flat DDR5 prints hide an allocation regime already tightening around Samsung and SK Hynix — the customer roster is the tell, not the tape.
Two 36氪 datapoints from this week collided at the same node: the FT reported Apple lobbying the White House to greenlight DRAM sourcing from Pentagon-blacklisted CXMT, while Bloomberg's Gurman said Apple will skip the M6 Pro/Max entirely as a memory squeeze reshapes its 2026 silicon roadmap. Both point at the same constraint — Apple can no longer plan its product cadence around commercially available DRAM alone. DDR5 RDIMM 32GB spot held flat at $1,275 on 2026-06-28, and DDR5 16Gb 4800/5600 sat at $46.73, but the spot tape is the wrong tape to watch. The scarcity is showing up in customer behavior, not the trailing spot print.
China's WF6 weaponization and Japan equipment China-sales contraction are converging to make Korea's specialty gas and chemicals layer — not the equipment vendors above it — the strategic bottleneck in the AI memory chain.
Two reports landing this week rewire the etch and CVD gas supply chain in ways the consensus PM has not yet priced. China's tungsten export controls escalating into WF6 — the metal precursor for tungsten plug and barrier deposition — were flagged as a structural pressure point for global memory and foundry. The same week, Japanese chip equipment makers reported a 10% YoY drop in China sales. Korean semiconductor exports simultaneously hit $32.04B in May, with HS8542 customs exports up 154.29% YoY.
Last column I asked why supply can't catch up. This one asks the scarier question on the other side: build all of it, and does anyone actually show up? Part of the answer I can put on a map — 28 gigawatts under construction in the US alone. Part of it I'm guessing at. A column about a buyer list that got longer, written by someone still sorting measurement from hope.
The textbook says the memory cycle fixes itself: shortage lifts prices, profits fund new fabs, supply floods back, prices fall. The first three steps are happening right now — capex just hit a record. The fourth isn't. This is a column about why the loop's last link looks broken, written by someone who isn't sure yet.
For two years this work read the chip supply chain as a sequence — Japan leading Taiwan, Taiwan leading Korea, a lag you could trade on. Testing it killed that idea. Then a wider look brought a different version of it back, with one catch: the cycle is still there, but right now it can't be read.