Semiconductor Watch | Octobeer 1, 2026
Samsung closes the HBM gap on SK Hynix, TSMC's 3nm/2nm/CoWoS stays too tight to satisfy everyone, and an extended US-China trade truce starts pulling orders back toward the mainland — the bottleneck just moved to geopolitics.
The semiconductor bottleneck keeps moving down the value chain — and this week it moved somewhere new: geopolitics.
Five threads this week — Samsung and SK Hynix's share war over Nvidia and OpenAI's HBM business, TSMC's tight 3nm/2nm/CoWoS capacity, ASML aligning Intel, TSMC and Samsung behind a new photomask standard, an extended US-China trade truce already redirecting orders toward the mainland, and fresh capital and legal moves around HBM — all point to the same shift: the constraint on AI chips is no longer just what can be built, but where and under what trade terms.
1. Nvidia Leans on Samsung and SK Hynix — and Now So Does OpenAI
Source: SamMobile
Nvidia CEO Jensen Huang told reporters in New York this week that Nvidia is "happy to cooperate with Samsung for the next 33 years," pointing to a three-decade partnership history with Samsung, SK Hynix, and LG — and noting that Nvidia's own technology runs inside Samsung's fabs. The comment underscores how dependent Nvidia's accelerator roadmap remains on the two Korean memory makers that together control 83% of the global memory market, even as the two compete fiercely with each other for Nvidia's business.
That dependency now cuts more than one way: OpenAI announced this month it is working directly with Samsung on next-generation chips, the clearest sign yet that AI labs, not just Nvidia, are negotiating their own memory supply relationships.
What It Means: When the world's most valuable chip company is publicly reaffirming a 33-year relationship, it's managing perception as much as supply — a signal to Samsung and SK Hynix that Nvidia's allocation decisions remain the deciding factor in who wins HBM share, even as AI labs start building their own direct lines to Korea.
The CODEW Lens: Nvidia doesn't need to own a fab when it can decide, with a single allocation call, who wins the memory war underneath it.
2. Samsung Closes the HBM Gap on SK Hynix
Source: Rest of World
Samsung has more than doubled its share of the high-bandwidth memory market over the past year, clawing business away from longtime leader SK Hynix and forging new pacts with both Nvidia and OpenAI. SK Hynix still holds the technical lead — it says its custom HBM now delivers up to 5.15x the inference performance of standard parts — and the deeper existing Nvidia relationship, but analysts say Nvidia's own allocation decisions remain the single biggest factor determining who wins share.
All three major DRAM makers, including Micron, are racing to turn HBM from a commodity into a customized, contractually locked-in product. In China, CXMT's disclosed $5.2 billion equipment buildout signals Beijing's parallel push toward domestic DRAM equipment independence — a memory race running almost entirely outside the Samsung-SK Hynix-Micron competition for Nvidia's business.
What It Means: HBM is moving from a market where technical leadership decided share to one where contractual lock-in does. Samsung's gains show that catching up on customization and securing a marquee customer relationship can matter as much as holding the raw performance lead.
The CODEW Lens: HBM stopped being a spec sheet competition. It's now a contract competition, and Samsung just proved contracts can be won from behind.
3. TSMC's Capacity Stays Tight as ASML Resets the Mask Standard
Source: Digitimes
TSMC's 3nm, 2nm, and CoWoS advanced-packaging capacity stayed tight through the third quarter as customer demand kept shifting: Nvidia and Apple continue to dominate advanced process and packaging resources, AWS has expanded AI chip tape-outs to win more 3nm capacity for its Annapurna Labs division, and MediaTek is drawing on Google TPU-related orders for its own share of 2nm, 3nm, and CoWoS allocation. TSMC is reportedly now evaluating a second US manufacturing hub in Texas — potentially housing up to six advanced wafer fabs — on top of its existing 12-fab Arizona cluster.
Separately, ASML issued joint statements with Intel Foundry, TSMC, and Samsung Electronics converging all three leading-edge chipmakers on a shift from the industry's decades-old 6-inch photomask to a larger 12-inch format for High-NA EUV lithography. Only TSMC attached a concrete timeline to the transition. In Singapore, VSMC — the Vanguard-NXP joint venture — opened its new 12-inch wafer fab on September 28, betting on value-added R&D and IP rather than competing on cost alone.
What It Means: TSMC being the only signatory with a concrete 12-inch mask timeline is an early tell on who actually ships High-NA EUV capability first — and its Texas evaluation shows US capacity competition is no longer just an Arizona story.
The CODEW Lens: A joint statement with no shared timeline isn't alignment. It's TSMC setting the pace and daring Intel and Samsung to keep up.
4. The Trade Truce Extension Starts Pulling Orders Back to China
Source: Digitimes / CNBC
Treasury Secretary Scott Bessent confirmed this week that the US and China have extended their trade truce — the "Busan Agreement" reached last October — from its original November 10 expiration to January 10, following Xi Jinping's state visit to Washington. Digitimes reports the extension is already redirecting orders for non-sensitive goods back toward mainland China, a development that could weigh directly on order volumes for Taiwanese manufacturers who had picked up business during the earlier standoff. Bessent said it remains unclear whether a broader deal or another rolling extension follows by January.
In a related development, Nvidia's Jensen Huang and AMD's Lisa Su both joined the advisory board of Tsinghua University's School of Economics and Management this week — placing two of the AI chip industry's most prominent executives in a formal advisory relationship with a Beijing-based institution at the same moment Washington and Beijing are negotiating chip-related trade access.
What It Means: An eight-week truce extension is now a bigger swing factor for Taiwanese manufacturers' booking volumes than any single TSMC capacity decision. Trade policy doesn't sort itself out as cleanly as a tight fab market does — and it can move order books faster than any factory can.
The CODEW Lens: Two chip CEOs joining a Beijing university board in the same week Washington extends a trade truce isn't a coincidence worth ignoring — it's a hedge placed in plain sight.
5. AMD Deepens Its Own Korea Ties as HBM's Legal Fights Begin
Source: Digitimes
AMD CEO Lisa Su is reportedly planning to revisit South Korea in October to deepen ties with Samsung and SK Hynix around HBM4 and NPU development — a trip that follows her new Tsinghua advisory role and underscores how central Korean memory partnerships have become to AMD's roadmap, not just Nvidia's. Samsung Electro-Mechanics disclosed 2026 supply agreements totaling $2.6 billion as server demand surges, a smaller but telling sign of how far AI-driven capital commitments now reach into the broader component supply base.
On the legal front, memory-technology firm Netlist is challenging Micron, Nvidia, Broadcom, and Google over alleged use of its HBM-related patents — a reminder that as HBM shifts from commodity to specialized, contractually-locked product, the intellectual property underneath it is becoming a competitive battleground of its own.
What It Means: AMD following Nvidia's exact playbook — deepen Korean ties, hedge with Beijing in the same stretch of weeks — shows the entire industry is converging on the same dual-track strategy regardless of which chip company is in the lead.
The CODEW Lens: Once HBM patents are worth suing over, the category has officially graduated from commodity to crown jewel.
The Semiconductor Signal
The bottleneck in AI chips has moved past manufacturing and into geography.
Every layer inside the fab this week — memory, process node, packaging, photomasks — is tight but legible: demand exceeds supply, and pricing and allocation sort out who gets what. Trade policy doesn't sort itself out that cleanly, and an eight-week truce extension is now a bigger swing factor for Taiwanese order books than any single capacity decision TSMC makes. Meanwhile, every major executive named this week — Huang, Su, and the Samsung and SK Hynix leadership racing for Nvidia and OpenAI's business — is hedging across both the US and Chinese markets at once.
| Development | Layer of the Stack |
|---|---|
| Nvidia / Samsung / SK Hynix | AI chip allocation & supplier leverage |
| Samsung vs. SK Hynix HBM share | Memory & packaging |
| TSMC capacity / ASML masks | Foundry & manufacturing |
| US-China trade truce extension | Supply chain & geopolitics |
| AMD Korea ties / Netlist patent fight | Capital & competition |
The CODEW Lens: Watching only the fabs tells you what chips can be built. Watching the trade truce and the advisory-board appointments tells you whether they can actually ship where the demand is.
Sources
→ Rest of World — Samsung and SK Hynix battle for U.S. AI memory dominance
→ SamMobile — Nvidia CEO Jensen Huang happy to cooperate with Samsung for the next 33 years
→ Digitimes — TSMC 3nm, 2nm and CoWoS stay tight as AWS, MediaTek, TPU shift capacity
→ Digitimes — TSMC reportedly evaluates Texas fabs as US regional competition intensifies
→ Digitimes — ASML lines up Intel, TSMC and Samsung behind 12-inch masks
→ Digitimes — VSMC bets on value over cost as Singapore fab builds next-generation R&D and IP capabilities
→ Digitimes — US-China trade truce extension draws orders back to China, threatening Taiwanese booking volumes
→ Digitimes — CXMT's US$5.2 billion buildout deepens China's push for domestic DRAM equipment
→ Digitimes — Jensen Huang, Lisa Su join Tsinghua University, advisory board
→ Digitimes — AMD's Lisa Su reportedly revisiting South Korea in October to deepen HBM4, NPU ties
→ Digitimes — Samsung Electro-Mechanics' disclosed 2026 supply agreements total US$2.6 billion
→ Digitimes — Netlist challenges Micron, Nvidia, Broadcom, Google over alleged HBM patent use
→ NBC News / CNBC — U.S. and China agree to extend trade truce through Jan. 10
The CODEW Stat
5 threads · 83% of global memory held by Samsung + SK Hynix · $5.2B CXMT equipment buildout · Jan. 10 trade-truce deadline This week's semiconductor developments all converged on the same test: not just who can build the most advanced chip, but who can get it where AI demand actually is. Samsung closed the HBM gap on SK Hynix while both compete for Nvidia and OpenAI's business; TSMC stayed the capacity chokepoint for Nvidia, Apple, AWS, and MediaTek at once, with ASML now aligning the industry behind a new photomask standard it alone has a timeline for. Underneath all of it, an eight-week trade-truce extension started pulling orders back toward China, and two of the industry's most prominent CEOs hedged by joining a Beijing university board in the same stretch of days. The bottleneck isn't just what can be built anymore — it's where, and under whose trade terms.
Editorial Note
Semiconductor Watch tracks the developments reshaping the global chip industry — wafer pricing, process-node competition, memory and packaging, foundry capacity, supply-chain geopolitics, and the capital flowing into AI-era manufacturing.
Educational content only. Not investment or business advice. Analysis is based on company announcements, official disclosures, and reporting from Rest of World, SamMobile, Digitimes, NBC News, and CNBC cited above. Metrics referenced are labeled as reported, calculated, or CODEW-derived. Some products referenced may be affiliate partners — see our Affiliate Disclosure for full details. Platform coverage, data sources, and methodologies can change as the intelligence platform evolves.
Reviewed by Erwin Castro
on
Thursday, October 01, 2026
Rating:
No comments: