Zero One Investment Research Daily Intelligence Brief, July 21, 2026

Microsoft backs AMD's first rack rival to NVIDIA as SK warns of memory chaos

Top Market Signals

Microsoft commits to AMD's Helios, NVIDIA's first rack rival

Advanced Micro Devices (AMD US) has secured Microsoft (MSFT US) as a launch customer for Helios, its first rack-scale AI system and the first credible challenger to NVIDIA's (NVDA US) top-end GB300 and Vera Rubin racks. Microsoft said on July 20 it will deploy Helios in volume across its own data centers and offer it to Azure customers, giving AMD a hyperscaler anchor for a product that ships in the second half of this year.

Helios packages 72 next-generation Instinct MI455X GPUs with 31.1TB of HBM4 memory, and AMD targets 260 TB/s of internal scale-up bandwidth, roughly on par with NVIDIA's Vera Rubin NVL72. According to Tom's Hardware, Microsoft will also add two Azure instance families built on AMD's sixth-generation Epyc Venice CPUs, one for agentic AI and data pipelines and one for chip-design workflows. AMD says 8 of the 10 largest AI companies now run workloads on its Instinct GPUs, including OpenAI, Cohere and Musk's xAI.

The competitive read is that AMD is converting one-off model-lab deals into recurring hyperscaler capacity. CNBC, reported by the Liberty Times, put NVIDIA's share of data-center GPUs above 95% against AMD's roughly 4.5%, with Futurum estimating a Helios rack at US$5.0m to US$5.5m versus US$3.5m to US$4.0m for a Vera Rubin rack. Futurum's Daniel Newman argues AMD could take 20% to 25% of the market over time, a shift worth hundreds of billions in revenue if it holds.

For Taiwan, the read-through runs to the rack builders. Wiwynn (6669 TT) and Wistron (3231 TT) assemble AI racks for the same cloud customers now qualifying Helios, and a second high-volume GPU platform broadens their order base beyond NVIDIA.

Linked stocks: AMD US, MSFT US, NVDA US, 6669 TT

Sources: Tom's Hardware, 對決輝達Vera Rubin！超微推出首款AI機櫃Helios 微軟率先採用

SK chairman sees 2027 memory demand up to doubling

SK Group chairman Chey Tae-won told a Korea Chamber of Commerce forum that AI semiconductor demand could grow 60% to 100% in 2027, with total memory demand up 50% to 60%, while the major suppliers have almost no capacity additions planned. He described current memory prices as abnormally high and said the scramble for supply must be called chaos, a rare public warning from the parent of the world's largest HBM maker.

Chey's framing is that the shortage widens before it eases. He argued that AI companies can absorb higher memory costs through investment, but PC and smartphone makers can only pass them into finished-goods prices, raising the risk of what he called chipflation. To avoid that, he said SK Hynix (000660 KS) should build not only in Korea's Honam region but also in the United States, tying capacity expansion to trade pressure as well as pricing.

The Taiwan read-through is that the memory tightening is broad, not vendor-specific. Nanya Technology (2408 TT) president Lee Pei-ying said DRAM supply stays tight for several more quarters, with 3Q26 contract prices above 2Q26 and 4Q26 operations improving further. Local brokers flagged Nanya, Winbond and ADATA as beneficiaries alongside TSMC and ASE. Micron (MU US) rose on the same commentary in US trading, and the picture squares with TSMC chairman C.C. Wei's own call last week that AI demand runs strong into 2029 and 2030.

The caution is Chey's own: he flagged that persistent price spikes could turn into a geopolitical problem, and that memory makers who only raise prices without adding supply will invite new entrants.

Linked stocks: 000660 KS, 2408 TT, MU US, 2330 TT

Sources: SK 集團：記憶體搶貨混亂 明年價格漲勢不會趨緩 南亞科、華邦等利多, 有人施壓？SK 崔泰源：記憶體太貴、考慮赴美設廠, 崔泰源喊「能蓋廠的地方都要蓋」　AI晶片2027需求最少增5成

Google designs its own chip to run Gemini cheaper

Alphabet (GOOGL US) is designing a new server chip, internally called Frozen v2, to run its Gemini models more efficiently, according to a report in The Information. The chip is slated for 2028 and could deliver 6 to 10 times more tokens per unit of power than Google's current AI silicon. Google did not confirm or deny the report, and its shares rose about 3% on July 20 ahead of earnings later this week.

The move fits a widening effort by the largest AI buyers to design their own inference silicon and lean less on NVIDIA. OpenAI unveiled its first custom inference chip, Jalapeño, in June, and Anthropic is reported to be discussing a chipmaking partnership with Samsung. Efficiency has become the selling point as investors question AI spending: Google has guided to US$180bn to US$190bn of capital spending this year and needs to show the outlay converts into cheaper serving.

The read-through is twofold. Every hyperscaler ASIC that comes online caps the incremental GPU each buyer needs from NVIDIA, even as total compute keeps growing. It also lifts the merchant design houses that co-develop this silicon, with Broadcom (AVGO US) and Marvell (MRVL US) the two names positioned to capture custom-accelerator work as more of it moves in-house.

Linked stocks: GOOGL US, AVGO US, NVDA US

Sources: TechCrunch

Kimi K3 rout looks overdone for chip suppliers

Chip-design and AI-infrastructure stocks sold off on July 17 after China's Moonshot released Kimi K3, an open model it says leads on frontend-code benchmarks, and several analysts now argue the reaction went too far. Cadence (CDNS US) fell 9.47% to US$330.11 and Synopsys (SNPS US) dropped 7.85% to US$384.28, both to multi-month lows, after a blog post claimed the model could design and verify a chip in 48 hours using open-source EDA (electronic design automation) tools.

BNP Paribas told clients the selloff was unwarranted. Its analysts noted Kimi K3's cited result came on a 45nm process, which does not compare with the sub-5nm nodes where leading chips are designed, and that top chipmakers still pay for Cadence and Synopsys tools despite decades of free open-source alternatives. Moonshot itself paused new subscriptions over the weekend, citing compute strain that ran far beyond its expectations.

The more important read is on who benefits if model prices keep falling. Kimi K3 carries a 2.8 trillion parameter count, the largest of any open model, but data from Artificial Analysis shows it costs 50% to 70% more per completed task than GPT-5.6 because of weaker token efficiency. Atreides Management's Gavin Baker argues that anything squeezing the model layer is a net positive for power, semiconductors, hyperscalers and software, because it stops two or three labs from becoming a buyer's monopoly over compute. That logic, he notes, is why NVIDIA's Jensen Huang publicly backs open models.

For Taiwan, the near-term winner is the cloud supply chain: Inventec (2356 TT) is flagged as a beneficiary of the mainland compute demand that models like Kimi K3 pull forward.

Linked stocks: CDNS US, SNPS US, NVDA US, 2356 TT

Sources: Kimi K3 模型嚇跌 EDA 雙雄　專家喊買、資安股為鑑, 中國 AI 模型嚇壞華爾街　Baker：反益晶片／AI 基建, The Verge

Samsung pay gap drives foundry staff to leave

Samsung Electronics (005930 KS) faces a growing exodus from its contract-chip unit, with 81.5% of 8,297 foundry employees surveyed by the company union saying they intend to leave, a level the union classifies as very high above its 62% threshold. The trigger is a new bonus structure that rewards the memory division far more richly than the rest of the company, opening a morale gap just as Samsung tries to win advanced-node share from TSMC.

The numbers explain the anger. Under the union agreement, memory staff can earn a special bonus worth 10.5% of annual operating profit when Samsung clears KRW 200tn of operating profit between 2026 and 2028. With 2026 operating profit estimated near KRW 300tn, each memory employee could receive around US$400,000 this year, against roughly US$4,000 for a mobile-division worker, a gap of about 100 times. Discontent was highest in foundry, followed by System LSI at 75.4%, while memory sat at just 32.7%.

The timing matters for the foundry race. DigiTimes reports Samsung is evaluating doubling the initial capacity of its Taylor, Texas fab as 2nm yields improve and large customers engage, and that roughly 30% of the customers queued for TSMC's 2nm node are said to be weighing Samsung as a second source. A foundry trying to close a technology and trust gap with TSMC cannot afford to lose its most experienced process engineers at the same time, which is the tension this survey exposes.

Linked stocks: 005930 KS, 2330 TT

Sources: 三星獎金制度掀內亂！記憶體員工獎金多，81% 晶圓代工部門想離職, 三星泰勒廠產能擬翻倍　台積電2奈米排隊客戶傳3成考慮三星

Samsung and SK Hynix add a new AI memory tier

Samsung Electronics (005930 KS) and SK Hynix (000660 KS) are pushing CXL memory (Compute Express Link, a standard that lets servers pool extra memory outside the CPU's own channels) as a new tier to get around the memory ceilings that constrain AI servers. Reports this week say the two, along with Micron (MU US), are stepping back from their own in-house CXL controller chips in favor of standard modules, treating server customers who need more capacity as the priority.

The strategic read is that the memory makers want to expand capacity sold per server without waiting on HBM alone. CXL modules sit between fast HBM and slower storage, adding addressable memory for large AI models, and standardizing them speeds adoption. Samsung has been showing its CMM-D module line as the vehicle for this tier.

The read-through lands on the merchant controller designers. If Samsung, SK Hynix and Micron pull back from proprietary controllers, the silicon that manages these CXL modules moves to third parties, and reports name Marvell (MRVL US) among the likely beneficiaries. The theme reinforces a pattern across this cycle: as memory content per AI server rises, value accrues not only to the DRAM makers but to the controller and interface chips around them.

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

Sources: 三星、SK海力士搶進CXL記憶體　布局AI時代記憶體階層架構, 科技1分鐘：三星CMM-D 2.0

先聲 First Word: Exclusives from Chinese-Language Sources

Advantech readies edge-AI platform for NVIDIA's next Jetson modules

Advantech (2395 TT) has prepared its next-generation edge-AI platform to support NVIDIA's new Jetson T2000 and T3000 modules, with launch planned for 1Q27 as it targets humanoid robots, autonomous machines and smart edge computing. The move slots Advantech into the robotics and edge-AI supply chain as Jetson Thor-class modules move toward volume. (20 Jul 2026) Source: 研華新一代AI平台 明年首季上市

Gold Circuit rides three AI PCB demand streams into a 2H peak

Gold Circuit Electronics (2368 TT) is capturing three AI printed-circuit-board demand streams at once, spanning ASIC AI, CPU servers and telecom switches, with its third-generation ASIC AI platform ramping from 2Q26 toward a second-half shipment peak and revenue growing quarter by quarter. It is a clean read on how the ASIC build-out feeds Taiwan's high-end PCB makers. (20 Jul 2026) Source: 最牛一輪／金像電俏 兆豐64叫好

Ennoconn's NCR order lifts its 2026 earnings outlook

Ennoconn (6414 TT) is guided by brokers toward 2026 earnings of about NT$26.34 per share as a multi-billion NT dollar order from NCR ramps hardware shipments and average selling prices, while its shift toward subscription-based ESaaS lifts gross margin. The item shows how a large anchor customer plus a software-mix shift can reset an industrial-PC maker's earnings power. (20 Jul 2026) Source: 樺漢2026年 EPS 拚26.34元 法人看好毛利率同步看增