Zero One Daily Intelligence Brief — June 25, 2026

Micron and Qualcomm reignite the AI chip rally after Taiwan's record foreign selloff

Top Market Signals

Micron's quarter quadruples revenue as memory shortage runs to 2028

Micron (MU US) reported fiscal 3Q26 adjusted revenue of US$41.5bn, more than four times the US$9.3bn it booked a year earlier and well ahead of the US$35.7bn consensus, then guided next-quarter revenue to about US$50bn. The stock rose roughly 16% after hours.

Adjusted earnings reached US$25.11 per share against US$1.91 a year earlier, and adjusted gross margin expanded to 84.9% from 39%. The margin jump is the read-through that matters: AI memory is pricing-led, not just volume-led, and high-bandwidth memory (HBM, the stacked DRAM that sits beside AI accelerators) is carrying the mix.

Management told investors that demand still runs far ahead of supply and that the shortage will extend beyond 2027. According to TechNews and CNA reporting, Micron plans HBM4E mass production next year, and Daiwa flagged that the company is locking in five-year customer contracts, an unusually long commitment for a commodity-cycle business. Micron also expects early product shipments from its Taiwan fab site by mid-2027.

The cycle framing is now industry-wide rather than company-specific. Research cited by TechNews puts the global memory market on track to triple this year and break US$1tn for the first time in 2027. Our 18 June Blade flagged Winbond cracking NVIDIA's Vera Rubin memory supply; Micron's print is the same supercycle measured at the largest US vendor, and the after-hours move pulled Asian memory names up with it.

Linked stocks: MU US, 000660 KS, 2408 TT

Sources: 美光財報、財測超預期 HBM供不應求估到2028年 盤後股價飆16％, 美光：記憶體晶片需求延續至2027年後, 今年全球記憶體市場估飆 3 倍、明年首破 1 兆美元

Qualcomm puts products and customers behind its data center push

Qualcomm (QCOM US) used its investor day to turn an ambition into a roadmap, unveiling the Dragonfly data center portfolio of CPUs and AI accelerators and naming Meta as a multi-generation data center CPU customer. The shares rose more than 12% after hours.

The company projected data center chip sales of more than US$15bn a year by fiscal 2029, with "billions" of dollars already expected in fiscal 2027, and set a longer-term target of US$40bn for its combined non-handset businesses. To fill the software gap, Qualcomm agreed to buy AI startup Modular for about US$3.9bn in stock. Its accelerator line, anchored by AI200 and AI250, pairs with a near-memory design Qualcomm calls High Bandwidth Compute (HBC), which stacks logic beneath the DRAM and which the company claims delivers up to six times the bandwidth per watt of conventional HBM.

Our 16 June Blade flagged Qualcomm's move into data center AI inference; this is the execution, with named customers, a dated revenue target, and an acquisition behind it. The Meta CPU win matters most, because it gives Qualcomm a reference hyperscaler designing its compute footprint around an Arm-based alternative to x86.

The competitive frame is Qualcomm attacking both the GPU and the memory stack at once. Reuters tied the Micron and Qualcomm forecasts to a roughly US$400bn single-session rally across AI chip stocks, a sign the market is reading merchant data center silicon as additive demand rather than a share-shift story.

Linked stocks: QCOM US, META US, NVDA US

Sources: 高通看旺AI資料中心商機 非手機晶片營收目標上修至400億美元, Bloomberg, Qualcomm, Meta ink data center CPU deal

OpenAI's first custom chip routes more inference demand to TSMC

OpenAI and Broadcom (AVGO US) unveiled Jalapeño, OpenAI's first custom silicon, a purpose-built inference processor for large language models manufactured by TSMC (2330 TT). It is a reticle-sized application-specific integrated circuit (ASIC, a chip designed for one workload rather than general use) that the partners say was taken from concept to engineering samples in about nine months.

Unlike a repurposed training part, Jalapeño pairs a large compute chiplet with HBM rather than cheaper DRAM, a choice OpenAI says targets the data-movement and latency bottlenecks specific to reasoning and agentic workloads. Engineering samples are already running machine-learning jobs in the lab, including OpenAI's own GPT-5.3-Codex-Spark, at target clock speed and power.

The partners claim performance-per-watt substantially better than current state-of-the-art hardware, but disclosed no benchmarks, so the figure is unverified. By their own framing, the part is meant to beat AMD's Instinct MI350 and NVIDIA's Blackwell generation; whether it competes with NVIDIA's Rubin and AMD's MI400 is untested.

The supply-chain read-through is the durable one. Every hyperscaler-class buyer that designs its own inference ASIC still routes leading-edge production and advanced packaging through TSMC, and Broadcom adds another marquee custom-silicon program to a book that already includes Google and Meta. The chip challenges NVIDIA at the system level while deepening both companies' dependence on the same Taiwanese foundry.

Linked stocks: AVGO US, 2330 TT, NVDA US

Sources: Tom's Hardware, TechNews

SK Hynix files for a US$29bn Nasdaq listing to fund the HBM build

SK Hynix (000660 KS) filed to list American depositary receipts (ADRs) on Nasdaq next month, seeking to raise up to KRW45.45tn, about US$29.4bn, in what would rank among the five largest share sales in history. The Seoul-listed shares jumped about 12% on the plan.

The raise is capacity finance. Our 12 June Blade flagged SK Hynix guiding to triple wafer capacity by 2034, and the company has been targeting a net cash position near KRW100tn to fund HBM expansion and strategic investment. A US listing widens the funding base and, by mirroring TSMC's ADR structure, gives global index and AI-focused funds a cleaner way to own the second-largest HBM supplier.

The timing reads as opportunistic. SK Hynix has been one of 2026's strongest large-cap chip stocks, and Samsung Securities projected a valuation uplift from the listing. With Micron's print landing the same week, the company is raising into peak sentiment for the memory cycle, locking in capital while AI demand visibility is at its highest.

Linked stocks: 000660 KS, MU US, 005930 KS

Sources: SK海力士擬赴美發行ADR　募資290億美元, 向台積電ADR看齊！SK海力士赴美上市 尋求集資294億美元

Taiwan absorbs a record foreign selloff even as export orders run hot

Taiwan equities took their heaviest single-day foreign hit on record on 24 June, with overseas investors net selling NT$177.4bn and the TAIEX falling 1,057 points, or 2.24%, to 46,043. Foreign investors dumped about 38,000 lots of TSMC (2330 TT) and pushed their net short futures position to a record 83,600 contracts.

The trigger was external. A 7.87% drop in the Philadelphia Semiconductor Index the prior session, on renewed worry that the Federal Reserve could restart rate hikes, hit Taiwan's AI-heavy electronics names hardest. The selling hit high-beta exposure on a macro signal, with no company-specific news behind it.

The forward indicator points the other way. Taiwan's May export orders rose 47% year on year, a 16th consecutive month of growth, and the orders line leads shipments by roughly one to three months. Hours after the Taipei close, Micron's blowout and Qualcomm's data center roadmap reignited the AI chip complex, and Korean and Japanese indices rebounded sharply the next morning. The order book and the demand signals strengthened into the selloff, not out of it.

Linked stocks: 2330 TT, 2454 TT, 3711 TT

Sources: cnYES 鉅亨網, 外資賣超1774億史上單日最大 台股嚇跌千點, 6月25日盤前／外資雖無情 但美光財報亮眼 指數盪震中再戰新高

NVIDIA's China black market price doubles as export curbs tighten

The grey-market price of NVIDIA's (NVDA US) flagship DGX B300 server in China has more than doubled over the past six months, climbing from about RMB4m to over RMB8m, according to Financial Times reporting on Chinese chip traders. The squeeze comes as Washington tightens enforcement against smuggling and Beijing raises import scrutiny to push domestic alternatives.

The price action says the demand is real and price-inelastic. Tighter US enforcement and China's own customs friction are both meant to choke off NVIDIA hardware, yet strong demand from Chinese enterprises has overwhelmed reduced supply, doubling what buyers will pay rather than cooling interest.

The read-through is that unmet Chinese demand remains a live part of the NVIDIA story even with the company largely shut out of official channels. It also strengthens Beijing's case for domestic accelerators, since a doubling in smuggled-hardware cost is exactly the gap a local substitute is meant to close. The same dynamic raises the stakes on any future licensing thaw, where restored legal access would meet visible, unmet demand.

Linked stocks: NVDA US, 2330 TT

Sources: 輝達越禁越夯 中國黑市價翻倍, NVIDIA禁售AI晶片中國黑市價翻倍　喊價突破人民幣800萬

ASE runs 15 plants at once and signals packaging price hikes

ASE Technology (3711 TT) chief operating officer Tien Wu told the company's annual meeting that AI demand has far outrun expectations and triggered a long-term structural shift in semiconductors, with the group building or expanding about 15 plant sites simultaneously and still unable to keep pace with orders. He did not rule out raising the capital expenditure guide again.

On pricing, Wu was unusually direct, describing three layers of adjustment: reflecting raw-material cost, which he called inevitable, reflecting investment cost, and adjusting for supply and demand. For the back-end packaging and test leader to name raw-material pass-through as a certainty signals that 2026 price increases are moving from negotiation to baseline.

The capacity commentary answers the question raised by TSMC's recent 10-year Amkor deal, flagged in our 18 June Blade. Rather than read TSMC's in-house and Amkor packaging build as a threat, Wu framed ASE as capacity-constrained across its own footprint, with demand deep enough to support both. Advanced packaging remains the gating step for AI accelerators, and the firms that own it are gaining pricing power, not losing it.

Linked stocks: 3711 TT, 2330 TT, 6239 TT

Sources: 日月光擴產 15座廠齊動工 營運長吳田玉：產能嚴重供不應求, 日月光投控營運長吳田玉暗示 漲價是必然趨勢, 台積電結盟Amkor撼動全球封測版圖？　日月光15座新廠衝刺擴產「毋驚」

先聲 First Word — Exclusives from Chinese-Language Sources

AuthenX debuts with Taiwan's first in-house fiber-array packaging line

Optical interconnect firm AuthenX lists on Taiwan's emerging-stock board on 26 June at NT$120 per share, with chairman Wu Mao-jen flagging the island's first self-developed fiber array unit (FAU) packaging line set for pilot production in 4Q26. FAU packaging aligns optical fibers to silicon photonics chips, a bottleneck step as the industry moves toward co-packaged optics, so a domestic line adds Taiwan supply to a part of the chain Japan and the US currently dominate. (24 Jun 2026) Source: 合聖 6/26 每股 120 元登興櫃，首條自主 FAU 封裝產線 Q4 試產

Genius Electronic Optical emerges as the winner on Meta's cheaper AI glasses

Meta (META US) launched a self-branded Meta Glasses line starting at US$299, more than 20% below the prior Ray-Ban entry model, with Taiwanese lens maker Genius Electronic Optical (3406 TT) cited as the principal supply-chain beneficiary. A lower price point is aimed at moving AI glasses from early-adopter volumes toward the mass market, which would lift unit demand for the optical lenses Genius supplies. (24 Jun 2026) Source: Meta 新 AI 眼鏡 萬元有找較上一代便宜逾二成 玉晶光躍大贏家

UMC's May profit jumps 203% on mature-node tightness

United Microelectronics (2303 TT) reported May net profit of NT$8.45bn, up 203% year on year, for EPS of NT$0.68, driven by tight mature-process foundry supply and rising quotes. Separately, AI research firm FundaAI reported UMC is working with Intel on a 3nm process slated for production at an Arizona fab. The profit jump shows the foundry price cycle reaching beyond the leading edge into mature nodes, where capacity is also constrained. (24 Jun 2026) Source: 最牛一輪／聯電鍍金 元大61沾光