Zero One Daily Intelligence Brief — May 24, 2026

Vera Rubin is Taiwan supply chain's biggest product as Samsung quietly targets MediaTek

Top Market Signals

Jensen Huang: Vera Rubin will be Taiwan's biggest supply chain ramp

NVIDIA (NVDA US) CEO Jensen Huang arrived in Taiwan four days ahead of schedule on May 23, declaring that the Vera Rubin platform will involve 100-150 Taiwan supply chain partners and calling it the largest product in the history of the island's semiconductor and electronics supply chain.

Huang landed at Songshan Airport originally expected May 27, and proceeded directly to a developer event in Nangang. Speaking to media at the airport, he said Taiwan's supply chain will be "very busy" in the second half of 2026 as Vera Rubin ramps. He will meet TSMC (2330 TT) chairman C.C. Wei and founder Chang Shang-yi this week, deliver a keynote at Taipei Music Center on June 1, and attend COMPUTEX June 2.

Vera Rubin combines a new Vera CPU architecture with the Rubin GPU, and Huang described it as the most successful NVIDIA product yet. The platform is expected to be supply-constrained from launch. Huang noted that NVIDIA's full AI stack now has the company partnering with "virtually all frontier AI model companies, cloud operators, and computer companies" from Hopper through Grace Blackwell to Vera Rubin.

Huang also addressed memory pricing. He acknowledged rising memory costs will push up consumer electronics prices, including PCs and graphics cards, and urged suppliers to expand capacity. On co-packaged optics, he said NVIDIA is already one of the world's largest Ethernet networking suppliers, with copper and silicon photonics pathways remaining relevant in parallel.

Linked stocks: NVDA US, 2330 TT, 2317 TT, 2382 TT

Sources: 黃仁勳：Vera 將是台灣供應商史上最大規模產品 台鏈下半年會很忙, 黃仁勳預告：台灣供應鏈下半年會很忙 近日將拜訪張忠謀, 黃仁勳抵台 將拜訪客戶並與魏哲家見面

Samsung chairman secretly visits MediaTek to trade memory for foundry orders

Samsung Electronics (005930 KS) chairman Lee Jae-yong made a quiet visit to MediaTek (2454 TT) headquarters in Taiwan on May 21, accompanied by multiple senior executives, with the primary agenda of winning MediaTek as a Samsung Foundry customer in exchange for memory supply commitments.

Industry contacts told Commercial Times and TechNode that Lee's visit followed Samsung winning Tesla's AI6 chip as a foundry customer and is part of a broader campaign to erode TSMC's (2330 TT) hold on leading-edge design wins. Samsung is simultaneously pitching AMD (AMD US) on its 2nm process, and industry sources name MediaTek as the next major prize.

MediaTek currently places all its leading-edge silicon at TSMC. Winning MediaTek would materially shift Samsung Foundry's capacity utilization and revenue mix, which has trailed TSMC on yield and customer retention in advanced nodes. The pitch bundles Samsung's memory supply with foundry services under one roof, an integrated offering TSMC cannot match.

TSMC has historically benefited from customers keeping foundry and memory relationships separate. Lee Jae-yong's offer to bundle HBM and logic production as a combined commercial arrangement signals a willingness to compress margin on memory to gain foundry market share, a trade-off Samsung can sustain at scale.

Linked stocks: 005930 KS, 2454 TT, 2330 TT, AMD US

Sources: 李在鎔訪聯發科 挖台積牆腳, Technode

Samsung HBM packaging stalls as bonus rift spreads beyond memory division

A bonus payout gap of nearly 100x between Samsung's memory division (~US$400,000 per worker) and its device and packaging units (~US$4,000 per worker) has triggered work slowdowns in Samsung's Test and Package (TSP) operations, putting HBM4 production for NVIDIA's Rubin AI accelerators at delivery risk.

TSP handles back-end packaging and testing for HBM, running through Samsung's integrated turnkey system that routes chips through both foundry and packaging lines. The Seoul Economic Daily reported that meetings are being canceled across Samsung's non-memory and shared business units and that "decision-making on major projects has come to a complete halt" in foundry and TSP divisions as inter-departmental resentment spreads. A source warned that continued negligence on production and verification lines risks damaging Samsung's customer relationships and delivery commitments.

The timing is consequential. All three major DRAM producers are racing to fill hyperscaler HBM orders for next-generation AI infrastructure. Any slowdown in Samsung's TSP constrains HBM4 output at the precise point where NVIDIA is ramping Rubin-generation systems. This is a new development: the 21 May Blade covered Samsung averting an 18-day strike with the stock-based bonus deal; the conflict has now re-emerged inside the packaging divisions that were not party to the original labor agreement.

For NVIDIA, the delivery risk concentrates supply dependence on SK Hynix (000660 KS), which is already the preferred HBM3E supplier and is tracking its own record-margin cycle ahead of HBM4 qualifications.

Linked stocks: 005930 KS, 000660 KS, NVDA US, 2330 TT

Sources: Tom's Hardware, Google News

SK Hynix at 72% operating margins, stock nears all-time high

SK Hynix (000660 KS) is trading toward its all-time high, with Korean media citing analyst targets near the 4 million won threshold, backed by 72% operating margins that reflect the company's position as the primary HBM supplier to hyperscalers at a point of sustained memory tightness.

The 72% operating margin is materially above any historical precedent for a DRAM company and reflects SK Hynix's pricing power on HBM3E, where hyperscaler allocation commitments are multi-quarter and not repriced on spot. Multiple analyst reports this week framed the margin level as reflecting SK Hynix's "AI infrastructure status" rather than a standard memory recovery, citing long-term supply agreements and priority qualification status with NVIDIA (NVDA US) for HBM4 on the Rubin platform.

SK Hynix's position is reinforced by Samsung's packaging disruption. If Samsung's TSP slowdown delays HBM4 ramp, SK Hynix becomes the default fill-in supplier, sharpening its negotiating position on both pricing and allocation for next-generation systems.

Micron (MU US) is a third dynamic in the memory market. Micron this week announced that its Manassas, Virginia facility has begun producing 1-alpha DRAM, quadrupling DDR4 wafer output with more than US$2bn invested and US$275m in CHIPS Act funding. The Manassas fab targets automotive, defense, and industrial customers facing DDR4 shortages as mainstream fabs redirect toward DDR5 and HBM. Automotive DRAM contract prices are projected to rise 70-100% in 2026 by S&P Global Mobility.

Linked stocks: 000660 KS, MU US, NVDA US, 005930 KS

Sources: Google News, Tom's Hardware

Arm Holdings hits all-time high on NVIDIA's $200bn CPU market call

ARM Holdings (ARM US) surged to an all-time high of US$306.51 this week, up 46.5% over the trailing five sessions, after NVIDIA CEO Jensen Huang projected a US$200bn CPU total addressable market that includes China, directly expanding the royalty base for ARM-architecture silicon underlying NVIDIA's Vera CPU.

Huang's explicit inclusion of China in the US$200bn CPU TAM was notable given ongoing US export restrictions on advanced GPU silicon. The statement implies that Vera CPU can address Chinese markets through channels unavailable to Rubin GPU exports. ARM earns royalties on all shipped ARM-architecture chips regardless of which company manufactures them, so any CPU TAM expansion translates directly to ARM's royalty revenue ceiling.

NVIDIA's Vera platform separates the CPU (Vera) and GPU (Rubin) into distinct product lines for the first time, allowing Vera CPU to be sold independently into markets where Rubin GPU is restricted. Prior NVIDIA compute platforms bundled CPU and GPU into a single module, which constrained ARM's royalty exposure to bundled system sales. The bifurcation is new and structurally increases the addressable market for ARM's ISA.

ARM also announced an expanded collaboration with Red Hat to build an Agentic AI stack, adding enterprise software ecosystem momentum to a week when AI inference demand dominated technical discussion, with agentic workloads consuming orders of magnitude more tokens than standard AI interactions driving a corporate cost reckoning at Microsoft, Meta, and Amazon.

Linked stocks: ARM US, NVDA US, TSM US, 2330 TT

Sources: Google News, Google News, Google News

DeepSeek V4-Pro gets permanent 75% price cut, anchoring China's AI pricing floor

DeepSeek has made permanent a 75% price reduction on its flagship V4-Pro AI model, keeping developer access at 25% of the original price indefinitely and establishing a pricing reference point for AI inference in China that US frontier labs cannot cost-match without accepting negative margin on inference.

The announcement, confirmed by Bloomberg and Reuters, applies to DeepSeek's V4-Pro model, which handles agentic and complex reasoning tasks. At 25% of original pricing on a permanent basis, the move signals that aggressive inference pricing in the Chinese AI market is a deliberate strategy, not a promotional window. The model operates at a fraction of what OpenAI, Anthropic, and Google charge for comparable inference capability.

For US AI infrastructure, the competitive pressure operates at two levels. First, hyperscaler-funded AI inference economics in China become harder to justify at high GPU compute costs if DeepSeek is compressing the per-token revenue pool. Second, price-sensitive enterprise and developer segments globally face a widening price gap between Chinese open or semi-open models and US frontier offerings, which narrows the revenue premium that US labs depend on for R&D reinvestment.

The timing coincides with a broader agentic cost crisis: Tom's Hardware reported this week that Microsoft (MSFT US), Meta (META US), and Amazon are managing "tokenmaxxing" by employees, where agentic AI workflows consume up to 1,000 times more tokens than standard AI interactions and are draining corporate AI budgets ahead of plan.

Linked stocks: 700 HK, 9988 HK, MSFT US, META US

Sources: Google News, Tom's Hardware

先聲 First Word — Exclusives from Chinese-Language Sources

TSMC 2nm secrets case creates corporate liability for equipment suppliers

Taiwan's Intellectual Property and Commercial Court convicted former TSMC engineer and Tokyo Electron (TEL) employee Chen Li-ming to 10 years for stealing 2nm process secrets under the National Security Act -- the first such conviction in Taiwan's history under that law. TEL was fined NT$150m (US$4.7m) as a corporate entity, setting precedent that equipment suppliers can face criminal liability for benefiting from stolen process information, not just the individual thieves. (24 May 2026) Source: 供應商搶單遭重罰 1.5 億元！台積電 2 奈米案立下新準則

IC distributors WPG and WT Micro guide record Q2 profits on AI demand

WPG Holdings (3702 TT) and WT Microelectronics (3036 TT), Taiwan's two largest semiconductor distributors, both posted record Q1 profits and upgraded Q2 guidance, citing sustained AI chip demand from data center customers. Analysts are modeling both companies for consecutive quarterly records through 2026, with AI distributor volumes compressing the historical seasonal slowdown in Q2. (24 May 2026) Source: 熱門族群》AI需求延燒到半導體通路 法人看好成長動能

ASUS server division guided to double-digit Q2 revenue growth on data center demand

ASUSTeK Computer (2357 TT) server division is projected by institutional analysts to deliver double-digit Q2 revenue growth, driven by deferred customer orders converting and rising AI inference demand from enterprise clients. Server is now tracked as ASUS's fastest-growing division by sell-side coverage in Taiwan. (24 May 2026) Source: 伺服器強勁需求帶動 華碩Q2營運可望走強