The Lyceum: Semiconductor Weekly — May 13, 2026

The HBM shortage stopped being a memory story this week and started being a labor story, a financing story, and a regulatory story all at once. Samsung's union walked out of government mediation Wednesday morning and is eight days from a strike that could pull 3–4% of global DRAM off the market; SK Hynix touched an all-time high intraday as Microsoft, Google, and Amazon were reported to be preparing to fund its next fab; and a US–China tariff truce signed Monday quietly restored rare earth licenses that fab operators had been hoarding against. The AI demand thesis is intact. The plumbing underneath it is getting weird.

• Q1 2026 Global Semiconductor Sales Report (SIA): Worldwide chip sales reached $791.7 billion in 2025, up 25.6% year-over-year — the baseline number the next cycle gets measured against.
• 2025 Global Materials Market Report (SEMI): Record $73.2 billion in materials revenue in 2025, with packaging materials growing 9.3% — nearly double wafer-fab materials growth at 5.4%.
• NAND Flash Market Bulletin – May 6 (TrendForce): Spot prices down 30–40% over the past month while contract prices are still forecast to rise 70–75% quarter-over-quarter — the cleanest signal of bifurcation between enterprise and commodity flash this cycle.
• Microchip FY Q4 Guide (Microchip Technology): Quarterly revenue guide above estimates on stronger industrial and automotive microcontroller demand — the first clean mature-node recovery print.

Government-mediated talks between Samsung Electronics and its largest union collapsed early Wednesday after nearly 17 hours of negotiations, putting the company eight days out from an 18-day strike involving more than 40,000 employees, mostly from its chipmaking division, beginning May 21. The Korea Herald and Korea Times both reported that the two sides could not narrow the gap on performance-based bonuses.

The fight is over how to split the AI windfall. The union wants performance bonuses pegged to 15% of operating profit, the bonus cap abolished, and the entire scheme formally institutionalized — explicitly benchmarked to SK Hynix's uncapped profit-share model. Samsung proposed 10% of chip-division operating profit, with a pledge to raise the ratio if the unit overtakes SK Hynix in revenue and profit. Workers rejected a $340,000 one-time bonus offer, pointing to SK Hynix employees who reportedly received roughly $900,000 in profit-linked payouts during the boom, per Tom's Hardware's reporting.

The supply-chain math is alarming. KB Securities estimates a prolonged walkout could disrupt 3–4% of global DRAM and 2–3% of global NAND supply, with line restoration adding another two to three weeks beyond the 18 days. An April one-day action gave a preview: memory fab output fell 18% on the affected night shift, and Samsung's contract foundry output dropped 58%. The Korea Times puts daily losses near 1 trillion won ($671 million); Tom's Hardware cites estimates as high as $11.7 billion in total damage if the full strike runs.

What changes if it happens: procurement teams at every hyperscaler rebalance allocations toward SK Hynix and Micron, contract prices for Q3 get renegotiated against a meaningfully tighter supply backdrop, and the HBM4 qualification race accelerates amid a tougher environment for Samsung's June supply-commencement timeline during a fab-floor work stoppage. What failure looks like: Samsung's Suwon District Court injunction is granted before May 21 and the walkout is blocked or narrowed. The signal: that ruling is the last circuit breaker, and the Blue House has signaled it prefers dialogue over emergency adjustment — meaning government intervention to compel arbitration is unlikely.

One legal subplot worth tracking: Chosun reports that recent Korean court precedent on whether performance bonuses count as "regular wages" — with consequences for overtime and severance — is shaping the union's leverage. Institutionalizing a 15% profit share doesn't just cost Samsung that 15%. It potentially repositions the entire wage base.

When your customers offer to build your factory for you, the shortage is real.

SK Hynix shares surged more than 15% intraday on May 11, touching an all-time high of 1.949 million won and pushing market cap past $900 billion, with the KOSPI closing at a record 7,844. The move occurred amid reports of structurally tight HBM supply. TradingKey reported that Microsoft, Google, and Amazon are preparing investment proposals to SK Hynix to fund capacity expansion in exchange for forward HBM allocations. Hyperscalers signing long-term supply agreements is the old model. Hyperscalers offering capex to accelerate fab construction is a new one.

SK Hynix's Q1 2026 numbers help explain the urgency: revenue of 52.58 trillion won (up 198% year-over-year) and operating profit of 37.61 trillion won (up 405%), with a 72% operating margin — higher than Nvidia's roughly 65% operating margin in its most recent quarter. A memory company with better margins than the GPU company it sells to is not a normal market state.

What changes if hyperscaler co-investment formalizes: AI infrastructure financing reorganizes around capacity pre-buys rather than purchase contracts, and the memory oligopoly gets a third layer of customer concentration risk baked in. The signal that says it's happening: any of the three names confirming a structured deal — not an MOU but an actual capital commitment with named milestones. Goldman Sachs raised its 2026 DRAM supply-demand gap forecast from 3.3% to 4.9% this week, calling it the worst shortage in 15 years. The top three memory makers have essentially booked their entire year, and with fab construction running four to five years, no meaningful new capacity arrives in 2026.

While everyone watches the fab drama, the materials layer is printing money quarter after quarter without the volatility.

Resonac — the Japanese specialty chemicals company formed from Showa Denko and Hitachi Chemical — revised its interim earnings forecast upward this week, citing strong demand for materials used in advanced semiconductor manufacturing. Resonac supplies CMP slurries (chemical mechanical planarization — the slurries that polish each layer of a wafer to atomic flatness), semiconductor-grade gases, and advanced packaging materials. The consumables every leading-edge fab burns through with every wafer. Merck's Q1 results echoed the pattern.

The structural number underneath the corporate prints: SEMI reported global semiconductor materials revenue hit a record $73.2 billion in 2025, up 6.8%. Wafer-fabrication materials rose 5.4% to $45.8 billion. Packaging materials grew 9.3% to $27.4 billion — nearly twice as fast. Taiwan remained the largest consumer at $21.7 billion, China second at $15.6 billion, South Korea third at $11.2 billion.

What changes if the materials trend holds: chemical suppliers absorb a structurally larger share of fab operating costs as advanced packaging scales, and the supplier list — Resonac, Shin-Etsu, SUMCO, JSR, Merck, Entegris — becomes a cleaner real-time read on fab utilization than the foundries themselves. What failure looks like: materials guidance flattens or drops while fab capex stays high, signaling utilization is softening even as customers keep buying tools. Watch Shin-Etsu Chemical, SUMCO, and JSR for the next print. Materials don't have 12-month equipment lead times. They get paid every quarter the fab runs hot.

The headline tariff numbers from Monday's US–China agreement — US duties cut to 30%, Chinese duties cut to 10%, both reassessed in 90 days — are not the part of the deal chip executives are studying.

Under the framework, China agreed to suspend global implementation of its expansive new rare earth export controls and issue general licenses for exports of rare earths, gallium, germanium, antimony, and graphite to US end users and their suppliers worldwide, according to the official fact sheet. China also terminated several investigations into US chip-sector firms. Gallium and germanium feed compound semiconductors and certain deposition processes; the Chinese export control suspension removes a near-term materials chokepoint that procurement teams had been actively hedging against since late 2024.

What changes if the truce holds through mid-August: compound semiconductor supply chains stabilize, fab-floor materials procurement teams can de-risk inventories built during the panic, and the Section 232 semiconductor tariff investigation becomes the dominant remaining trade risk for the back half of 2026. What failure looks like: the 90-day window expires without an extension and license suspensions snap back, restoring a procurement scramble that hits Korean and Japanese fabs hardest. The signal that says which way it's going: whether Commerce publishes Section 232 findings before mid-August — a final rule before truce expiry would suggest Washington is preparing to escalate on a different axis.

Last week we flagged the Validated End-User program closure as the less-visible part of the Hua Hong tool restrictions. It's now in its operational phase.

Per BIS, no foreign-owned fab now holds VEU status, and former participants have 120 days from Federal Register publication to apply for and obtain export licenses. BIS has stated it intends to grant licenses to operate existing fabs in China but does not intend to grant licenses to expand capacity or upgrade technology. Operating and expansion licenses are now two different things.

The practical effect: Samsung's Xi'an NAND facility, SK Hynix's Wuxi DRAM fab, and TSMC's Nanjing mature-node plant can keep running but cannot grow. Meanwhile, per a Congressional Research Service report, Changxin Memory Technologies — the Chinese firm leading domestic HBM efforts — remains off the Entity List and can still purchase US semiconductor equipment. The asymmetry is the second-order effect procurement teams at Samsung and SK Hynix are now modeling: foreign-owned fabs frozen on Chinese expansion, Chinese domestic players still buying.

The observable signal: how many former VEU participants successfully receive operating licenses inside the 120-day window. A high approval rate keeps fabs running; a slow or selective approval process turns the compliance clock into a production clock. A separate BIS settlement document published in February lays out an enforcement view of "dual-build" routing through Korean assembly sites — meaning BIS is now policing logistics flows, not just tools. The compliance perimeter just expanded.

Applied Materials reports fiscal Q2 2026 after the bell on Thursday, May 14, and the setup is more complicated than the AI-equipment-boom narrative suggests.

Analyst Robert Castellano's review of served available markets found that Applied Materials and Tokyo Electron both lost share within their respective markets in 2025 while Lam Research, KLA, and ASML expanded. AMAT has the broadest portfolio in equipment — if it's losing ground in deposition, etch, and CMP simultaneously, that's structural rather than cyclical. Per 24/7 Wall St., Chinese domestic equipment vendors — NAURA, AMEC, ACM Research, Hwatsing, Piotech, Kingsemi — collectively held 6.5% of the global wafer fab equipment market in 2025, up from 5.6% in 2024 and just 1.2% in 2021. That incremental share is coming from somewhere.

What to listen for Thursday: China revenue concentration, commentary on the BIS settlement's routing implications, and any acknowledgment of share dynamics in CMP and etch. What failure looks like: AMAT beats the print but guides down on China, validating the share-loss thesis. What success looks like: strong Korea and Taiwan order flow that offsets China softness — meaning the HBM and N2 capex wave is doing the heavy lifting that earlier-stage Chinese mature-node spend used to. Note the unusual quiet on major equipment-order announcements between late April and now. It might mean orders already moved into supply chains. It might mean customers are reassessing. Thursday's call resolves which.

The easy reading is "another Sony–TSMC deal in Japan." The useful reading is that Japan is trying to move from being a very good supplier to the chip industry to being a place where finished, high-value silicon gets made.

Per Reuters reporting carried by Investing.com, Sony Semiconductor Solutions and TSMC plan a new joint venture in Japan to develop and manufacture next-generation image sensors. The venture would be majority-owned by Sony, with development and production lines at Sony's new fab in Koshi City, Kumamoto. The agreement is explicitly non-binding, investment phased to demand, and government support assumed rather than finalized.

Image sensors feed automotive and robotics "physical AI" applications, and a Japanese sensor line is more than a capacity note — it's an ecosystem play that pulls equipment, materials, and packaging demand into a specific region. What changes if it converts: Japan's leading-edge supplier base gets a second anchor customer beyond Rapidus, and the Kumamoto cluster becomes a credible third pole for advanced sensor and analog manufacturing. What failure looks like: the MOU sits as an MOU for 18 months and quietly gets restructured as a supply agreement. The signal: whether a definitive capex commitment with named subsidy backing is announced before year-end. Non-binding agreements are industrial-policy courtship. Definitive capex is when the ecosystem starts becoming real.

This week the AI memory boom produced a company with margins higher than Nvidia, a Japanese MOU dressed up as an industrial-policy victory, and 40,000 Korean fab workers preparing to walk because their bonus is only $340,000. Somewhere in Wuxi, a compliance lawyer is staring at a 120-day calendar and explaining to a fab manager that the difference between an operating license and an expansion license is the difference between a paycheck and a press release. See you next week.

Forward this to the friend who keeps asking why their GPU order keeps slipping a quarter.