Zero One Daily Intelligence Brief — July 1, 2026

Taiwan's memory darlings tumble as China's GigaDevice warns the price boom will cool

Top Market Signals

China's GigaDevice warns on prices and Taiwan's memory leaders tumble

Taiwan's best-performing chip trade of 2026 went into reverse on 1 July, as Winbond (2344 TT), Nanya Technology (2408 TT) and Macronix (2337 TT) all fell more than 5% after China's GigaDevice warned that memory prices had climbed too far.

GigaDevice, a Chinese NOR flash supplier, issued a formal stock-trading risk warning flagging four risks, the central one being that memory chip prices sit at historic highs, that the sharp rally is unsustainable, and that supply and demand will move back toward balance with prices likely to fall substantially. Because GigaDevice's products overlap heavily with Macronix and Winbond, and partly with Nanya, the warning read straight across to the Taiwan names.

The falls were steep. Winbond closed down 8.4% at NT$190, Macronix down 7.7% at NT$143.5, and Nanya down 7.0% at NT$421, each on heavy volume even as the wider index hit a record. This group had led the whole market this year, with Macronix up 294% in the first half, Winbond up 151%, Nanya up 134% and Powerchip (6770 TT) up 102%.

Foreign funds had already been stepping back. They net sold Winbond for nine straight sessions (286,700 lots), Nanya for six (75,000 lots), Macronix for two (60,000 lots) and Powerchip for four (145,000 lots). Our 25 June 2026 Blade carried Micron's view that the memory shortage runs to 2028; GigaDevice's note is the first public suggestion from a supplier that the price spike itself could reverse. Whether this is a pause or a top now rests on whether contract prices actually roll over.

Linked stocks: 2337 TT, 2344 TT, 2408 TT, 6770 TT

Sources: 台股衝4萬7記憶體股卻慘綠 原來是...中國記憶體大廠示警四大風險

Meta rewires retired DDR4 into new servers to dodge the DRAM squeeze

Meta (META US) has built a custom chip that lets it plug the DDR4 memory from decommissioned servers into new machines that only take DDR5, a way to add capacity without buying scarce, expensive new DRAM.

The chip, Vistara, is Meta's first in-house CXL memory-expansion ASIC, presented at ISCA 2026 on 29 June. It implements CXL 2.0 over a PCIe 5.0 x16 interface and bridges standard DDR4 modules to the host processor. Meta is deploying it at 128GB per chip, built from recycled 32GB DDR4 modules pulled from retired servers.

In Meta's MemServer system, two Vistara chips connect over PCIe to a 158-core AMD (AMD US) EPYC processor, pairing 768GB of local DDR5-6400 with 256GB of CXL-attached DDR4-2400 for 1TB in total. Software presents the older memory as a separate tier, so Linux keeps hot data on fast DDR5 and pushes colder data to the cheaper DDR4 layer; idle round-trip latency is held to about 50 nanoseconds.

The point is cost. With DRAM in short supply and prices climbing, reusing working DDR4 rather than scrapping it with the server gives hyperscalers a lever to hold down memory spend, and Korean startup Panmnesia is chasing a similar CXL approach. It is a reminder that the largest buyers can engineer around a shortage, which cuts against the demand the memory makers are counting on.

Linked stocks: META US, AMD US, MU US, 2408 TT

Sources: TechNews

Applied Materials launches tools aimed at DRAM and 3D chip stacking

Applied Materials (AMAT US) rolled out a set of new manufacturing systems built for the DRAM and advanced-packaging steps that AI chips increasingly depend on, targeting the memory-bandwidth and 3D-integration bottlenecks now shaping performance.

On the DRAM side, the company introduced an upgraded epitaxy system that it says raises drive current and transistor efficiency to improve DRAM performance and power, supporting next-generation high-bandwidth memory and DDR. The new system also shrinks tool footprint by about 20%, letting fabs pack more capacity into the same floor space.

For advanced packaging, Applied added three process tools spanning chemical-mechanical planarization, electrochemical deposition and plasma-enhanced CVD, plus two electron-beam systems, the VeritySEM metrology tool and the SEMVision defect-review platform, to bring fab-grade measurement and defect analysis to packaging lines. Both are aimed at catching sub-optical defects in HBM and 3D heterogeneous integration.

The launch tracks where the money is going. As AI models grow, memory bandwidth and packaging complexity have become the gating constraints rather than transistor density alone, and the equipment makers are repositioning toward the DRAM and back-end steps where capacity is being added fastest.

Linked stocks: AMAT US, 2408 TT, 3711 TT

Sources: 應材推一系列全新系統，搶攻 DRAM、先進封裝商機

Samsung maps a foundry roadmap to 1.4nm and builds HBM4 bases in-house

Samsung Electronics (005930 KS) laid out a three-part plan to remake its contract chip business around AI, putting firm dates on its most advanced nodes and leaning on its own memory arm to differentiate.

On process, Samsung said its 1.4nm node (SF1.4) is on track for 2029 mass production, with an improved SF1.4 Plus following in 2030. Its 2nm line will move to an enhanced SF2P Plus across 2027 to 2028 and then to SF2X. The company credited design-technology co-optimization for more than half of a 26% power reduction at 2nm, and said it had shrunk SRAM to what it calls the world's smallest cell.

The sharper differentiator is memory. Samsung is making the base die under sixth-generation HBM (HBM4) on its own 4nm logic process (SF4X), and says it has verified clean signaling at 10 Gbps with headroom to 11.7 Gbps. It also replaced a slow, largely manual chip-to-chip verification flow with a digital tool it calls 3D DRAM PHY to cut customers' design and simulation time.

The pitch is that Samsung can offer logic and memory together, connected through packaging, as a single foundry. That is the ground TSMC and the Taiwan back-end chain are also fighting over, and Samsung is betting its in-house HBM is the edge that wins AI logic customers back.

Linked stocks: 005930 KS, 2330 TT, 2454 TT

Sources: TechNews

Cloud AI crowds TSMC's 2nm line and pushes phone chips up a node

The scramble for TSMC (2330 TT) 2nm capacity is reshaping the phone-chip roadmap, with MediaTek (2454 TT) and Qualcomm (QCOM US) moving flagship processors up to 2nm partly to avoid an increasingly crowded 3nm node.

Handset system-on-chip makers are collectively lifting their 2026 flagship platforms to 2nm. Beyond the usual spec upgrade, the bigger motive is to sidestep 3nm, where demand is strongest and capacity tightest. From 2027, nearly all new cloud AI chips are set to accelerate onto 2nm as well, so the crowding now seen at 3nm is expected to move up a node.

The demand signal from the fab side is consistent. Analysts previewing TSMC's 16 July call describe 2nm demand as extremely strong, driven by AI CPUs, GPUs, ASICs and networking chips, the same leading-edge pull that is squeezing capacity for everyone else.

For MediaTek, moving to 2nm keeps it competitive at the high end but puts it in direct capacity competition with the AI-chip giants for the same wafers. The node that decides phone performance is now the node the entire AI industry wants.

Linked stocks: 2330 TT, 2454 TT, QCOM US

Sources: 雲端AI擠壓2奈米產能版圖　聯發科、高通更多手機SoC讓位, 法說行情暖身？外資終止連六賣 回頭買超台積電

Intel and Qualcomm are reported to have circled Jim Keller's Tenstorrent

AI chip startup Tenstorrent, led by veteran chip architect Jim Keller, has drawn takeover interest from both Intel (INTC US) and Qualcomm (QCOM US), according to reports that Keller has partly confirmed.

Keller said he had met with Intel and Qualcomm, and argued that Tenstorrent would surpass rival AI-chip startup Cerebras. The reporting frames the interest against a wider scramble for AI-silicon talent and intellectual property, after NVIDIA (NVDA US) hired the core team and licensed IP from inference startup Groq in a deal valued at around US$20bn.

The interest is telling for two would-be challengers. Intel is trying to rebuild a credible AI-accelerator roadmap, and Qualcomm has only just moved into the data center; both lack the in-house AI-training silicon that Tenstorrent, with its RISC-V-based designs, has been building.

Whether or not a deal happens, the pattern is clear enough: the largest chip firms increasingly see buying a startup team as the fastest way onto the AI-accelerator map, rather than building from scratch against NVIDIA's lead.

Linked stocks: INTC US, QCOM US, NVDA US

Sources: DigiTimes

先聲 First Word — Exclusives from Chinese-Language Sources

UMC is thrown into a trading timeout after a blistering run

UMC (2303 TT) will be placed under Taiwan Stock Exchange disposition measures from 2 July to 15 July after landing on the exchange's attention list six times in ten sessions, a penalty that lengthens its order-matching intervals. The stock had run from NT$141 to an intraday high of NT$185.5 on 24 June before pulling back, and still closed up 2.74% at NT$169 on the day the penalty was announced. It was one of 58 names flagged, a measure of how hot Taiwan's mature-node and hardware rally has become. (01 Jul 2026) Source: 聯電遭證交所列處置股 明起「關禁閉」至7/15

AUO sets up an innovation institute to speed its move beyond panels

AUO (2409 TT) announced a reorganization and senior management changes on 1 July, creating a new Innovation Research Institute to pull together its AI and forward-technology R&D, including Micro LED, CPO optical interconnect, AR glasses and low-earth-orbit satellite antennas. The panel maker is regrouping around three pillars, display technology, smart mobility and vertical-field solutions, as it tries to shift its identity from a cyclical panel supplier to a solutions company and blunt the swings of the display cycle. (01 Jul 2026) Source: 友達宣布組織調整 新設創新研究院深化AI布局

Pan Jit pushes a car-and-AI-server strategy at Munich Shanghai

Pan Jit International (2481 TT), a Taiwanese power-discrete maker, used the 2026 Electronica Munich Shanghai show to present a dual-core strategy built around automotive electronics and AI-server power, showcasing higher-efficiency, higher-reliability parts aimed at lifting its share of the high-end market. It is another Taiwanese component supplier steering its product mix toward the two fastest-growing demand pools, electrified vehicles and AI infrastructure. (01 Jul 2026) Source: 強茂參加2026慕尼黑上海電子展 展車電及AI Server雙核強化市場占有率

Wistron and NVIDIA build an AI platform to run smart factories

Wistron (3231 TT) said it is working with NVIDIA (NVDA US) to build a smart-factory AI platform centered on a Factory Agent Manager, combining a Factory Operations blueprint on the software side with NVIDIA industrial edge-AI hardware for low-latency, on-site inference. The ODM is applying multi-agent orchestration to equipment monitoring, anomaly analysis and process optimization, extending its AI-server push into using AI to run its own plants. (01 Jul 2026) Source: 導入 AI 代理系統 緯創、輝達聯手打造智慧工廠 AI 平台