Zero One Daily Intelligence Brief — July 8, 2026

Samsung's record quarter triggers a peak-earnings selloff; DeepSeek enters the AI chip race

Top Market Signals

DeepSeek designs its own AI inference chip

Chinese AI developer DeepSeek is designing its own inference chip to cut its reliance on NVIDIA (NVDA US) and Huawei, Reuters reported, citing three people familiar with the plan.

The chip is aimed at inference, the stage where a trained model generates answers for users, rather than training. Reuters says the effort began about a year ago and is still early, with DeepSeek quietly recruiting chip-design engineers and holding talks with design, foundry and memory firms. Inference silicon is the fastest-growing pocket of AI compute because purpose-built parts are cheaper to run and draw less power than general-purpose GPUs, and a domestic chip also insulates the company from US export controls that founder Liang Wenfeng has called a major constraint.

The read-across points more at Huawei than at NVIDIA. Analysts cited by TechNews argued the news barely dents NVIDIA, whose China business has already gone to zero under US restrictions, and NVIDIA shares closed up 0.71% at US$196.93 after an intraday dip of 2.3%. The pressure lands on Huawei instead: it holds roughly half of China's US$50bn AI-chip market, and with Alibaba, Baidu and now DeepSeek all designing their own inference parts, that lead is narrowing. DeepSeek still faces years of work and heavy capital, plus US limits on advanced foreign foundries and on access to high-bandwidth memory (HBM, the stacked memory AI chips depend on). It is raising a first external round of about US$7bn at a US$52bn to US$59bn valuation to fund the push.

Linked stocks: NVDA US, 2330 TT

Sources: DeepSeek 進行自研晶片計畫，預估衝擊華為比輝達更嚴重, 工商時報

Samsung's record quarter triggers a chip selloff

Samsung Electronics (005930 KS) reported second-quarter operating profit of about KRW 89.4tn, roughly 19 times the KRW 4.7tn it earned a year earlier and above the LSEG consensus of KRW 87.6tn, yet the stock fell 6.9% and dragged Asian tech down with it.

Our 6 July Blade previewed the roughly 18-fold jump; the print landed even bigger, and the reaction showed the beat was already in the price. Samsung shares dropped as much as 10.1% intraday before closing down 6.9%, wiping more than US$80bn (NT$2.57tn) of market value, while SK Hynix (000660 KS) fell 6%. Korea's Kospi briefly plunged 8.2%, tripping a circuit breaker, and closed down 4.9%. The selloff spread across the region: Japan's Nikkei 225 fell about 2.1% with Kioxia down 11.26%, and Taiwan's TAIEX shed 1,077 points, or 2.31%, its eighth-largest point drop on record.

The worry is durability, not the quarter itself. Strategists quoted by Reuters said the record results triggered fear that the AI-chip boom cannot keep climbing, and pointed to the risk that US hyperscalers slow their AI infrastructure spending. Foreign investors sold NT$54.7bn of Taiwan stocks in a day and the New Taiwan dollar fell to a 14-month low of 32.143. In the US the Philadelphia Semiconductor Index dropped 4.65%, TSMC's ADR fell 4.25% and Micron (MU US) fell 4.71%.

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

Sources: 憂AI晶片榮景持續性 三星市值一口氣蒸發2.6兆元, 外資股匯大逃殺 多頭迎來終局？法人：緊盯三關鍵別急認輸

NVIDIA steps in as AI's central bank

NVIDIA (NVDA US) is using its own investment-grade credit to backstop AI compute projects, a model research firm SemiAnalysis likens to a central bank as it forecasts global AI-related debt will pass US$7tn by 2029.

The problem SemiAnalysis describes is that the GPU rental market has no public pricing and most financing is tied to five-year contracts, which forces startups that need short-term compute into large prepayments or long leases. NVIDIA's answer is a six-year minimum-revenue guarantee to smaller cloud operators, the so-called neoclouds. NVIDIA commits to buy unsold compute at a pre-agreed price, cited at a six-year average of US$2.36 per GPU-hour; below that level the operator keeps all the revenue, and above it NVIDIA takes 40%.

The strategic aim is to widen NVIDIA's buyer base. Lenders, reassured by NVIDIA's AA rating, will extend loans that match the guarantee's term, which lets neoclouds secure the chips, capital and space they previously could not assemble and offer shorter contracts to a broader set of customers. That keeps the GPU market from being cornered by a handful of hyperscalers, and it enlarges the pool of companies that can commit to buying NVIDIA systems, the constraint that would otherwise cap NVIDIA's addressable market as the current long-contract model saturates.

Linked stocks: NVDA US

Sources: 研究報告：三年後產業債務逾7兆美元 輝達扮 AI 央行解融資痛點

AMD pledges US$10bn to Taiwan's chip packaging

AMD (AMD US) will invest more than US$10bn across Taiwan's supply chain to expand advanced-packaging capacity for next-generation AI infrastructure, and has proposed a panel-level packaging partnership with Powertech Technology (6239 TT), according to market reports relayed by the Economic Daily News.

The tie-up brings AMD's core CPU into Powertech's 2.5D panel-level process, a way of building chips on large rectangular panels rather than round wafers to lift throughput and cut cost. Analysts read the win as a Tier-1 customer validating Powertech's technology, which should help it chase further custom-silicon and AI orders. Powertech is spending heavily to meet that demand: group capital expenditure this year is set at about NT$50bn, with NT$30bn to NT$40bn going into Powertech itself and NT$20bn to NT$30bn of that earmarked for fan-out panel-level packaging (FOPLP).

Powertech is building a new FOPLP line for a large US customer, with product qualification and small-volume output this year and a meaningful revenue contribution expected from 2027. The move fits the broader pattern of AI compute pulling investment into Taiwan's packaging layer, where capacity, not wafer fabrication, is increasingly the bottleneck.

Linked stocks: AMD US, 6239 TT

Sources: 力成看旺 押長天期, 力成6月營收月減3.7%

Delta raises power prices a second time

Delta Electronics (2308 TT), the largest maker of switching power supplies, has begun a second round of price increases of about 5% to 20%, market sources said, as surging raw-material costs work through the AI hardware chain.

Our 2 July Blade flagged TXC lifting frequency-component prices 10% to 30%; the same pricing wave has now reached the power-supply majors. Alongside Delta, China's power leader Megmeet, another NVIDIA supplier, has sent customers notice of increases of roughly 10% to 20% across its product line. The pressure comes from upstream inputs: copper used in transformers, PCB copper foil and wiring is up 10% to 20% this year and makes up more than a fifth of material cost, while power semiconductors such as MOSFETs and passive components have risen 10% to 30%.

Delta declined to comment on the reports and said any increase would be negotiated product by product with no fixed timetable. Rival Lite-On (2301 TT) said it would assess pricing case by case. Chairman Cheng Ping recently noted that component shortages and fast AI growth have customers pushing to build capacity quickly, with Delta now planning expansions a year out across the US, China, Thailand and Taiwan.

Linked stocks: 2308 TT, 2301 TT, NVDA US

Sources: 台達電傳啟動第二波漲價 電源產品漲幅上看20% 力保毛利率表現

Intel targets the AI memory bottleneck with a new design

Intel (INTC US) has disclosed a patent for an ultra-high-bandwidth memory architecture it calls XBM (Cross-Batch Memory), pitched to replace HBM4 with lower cost and higher bandwidth and aimed at commercialization after 2030.

The design attacks the limits of today's high-bandwidth memory, which faces shortages, high prices and heavy power draw as AI demand climbs, to the point that some in the industry are turning to LPDDR as a stopgap. XBM's core idea is a one-transistor, one-capacitor DRAM cell built in the back-end metal layers rather than on the front-end silicon, which frees area for denser through-silicon vias (TSV, the vertical connections that carry data between stacked chips) and lifts bandwidth. Each die holds 0.5GB to 5.0GB at the same size as HBM4, moves data over a UCIe chiplet interface at up to 32 GT/s, and scales from 96 data blocks in an 8-high stack to 192 in a 16-high stack, for a claimed doubling of total bandwidth.

Intel plans to pair XBM with a companion technology called ZAM, and Qualcomm (QCOM US) is pursuing a similar approach. The caveat is Intel's record: its earlier HMC and MCDRAM memory efforts never reached the market, so XBM is a statement of intent to re-enter high-end DRAM during a shortage, not a shipping product.

Linked stocks: INTC US, QCOM US

Sources: TechNews

先聲 First Word — Exclusives from Chinese-Language Sources

AI cooling trio posts record June revenue

Taiwan's three AI thermal-management names all reported strong June revenue, led by Jentech Precision (3653 TT), whose NT$2.65bn was up 57.6% YoY and a monthly record. Asia Vital Components (3017 TT) rose 66.1% YoY to NT$17.62bn, its second-highest month, and is lifting liquid cold-plate capacity fivefold to one million units a month; Auras Technology (3324 TT) grew 62.1% YoY. It confirms AI liquid cooling as a sustained demand line rather than a one-quarter spike. (8 Jul 2026) Source: 散熱三雄6月營收靚 健策創紀錄

Quanta books its first NT$1tn quarter

Quanta Computer (2382 TT) reported record June revenue of NT$385.2bn, up 102.9% YoY, taking 2Q26 revenue to NT$1.04tn, its first quarter ever above NT$1tn and up 28.1% QoQ. Management points to AI-server demand, with Vera Rubin shipments starting in the second half and full-year capital spending set to double capacity across three new plants. It is a direct read on how fast the AI-server build-out is scaling through Taiwan's largest ODM. (8 Jul 2026) Source: 廣達月季營收雙雙創新高 法人：Vera Rubin 出貨 有望帶動全年業績攻頂

Greatek takes over onsemi's Philippines plant

Greatek Electronics (2441 TT) will acquire onsemi's (ON US) assembly plant in Tarlac, the Philippines, and has signed a long-term supply agreement to keep packaging and testing onsemi's parts. The deal, part of onsemi's Fab Right restructuring, is expected to close within three to six months pending approvals. It hands Greatek offshore capacity and deeper ties to an international IDM in automotive, industrial and AI chips, as Philippine packaging capacity draws wider investment including ASE's (3711 TT) local expansion. (8 Jul 2026) Source: 超豐電子接手安森美菲律賓廠擴大全球封測版圖