Zero One Investment Research Daily Intelligence Brief, July 24, 2026

Intel's data center revenue jumps 59%; NVIDIA funds Amkor to move AI packaging onshore

Top Market Signals

Intel's AI CPU demand drives a 15-year high

Intel (INTC US) reported 2Q26 revenue of US$16.1bn, up 25% YoY, with its data center and AI unit growing 59% YoY, the fastest expansion for the group in 15 years. Shares rose about 10% after hours, and the move was driven by a 3Q26 revenue and profit guide above consensus rather than by the quarter itself.

The forward signal is that AI server buildouts are now pulling through general-purpose server CPUs alongside GPUs, a demand stream that had lagged the GPU cycle for two years. Intel described a tight supply environment for its server parts and raised its planned capex to more than US$20bn to add capacity over the next two years, a commitment it would not make on a one-quarter blip.

The print does not settle the foundry question that has hung over the stock. The data center recovery is a product story, and it says little about whether external foundry customers are ramping. The Arizona fabs remain the swing factor, and management left the timeline for new US capacity unchanged.

Linked stocks: INTC US, NVDA US, AMD US

Sources: Fierce-network, AI晶片需求旺　英特爾第3季財測優於預期

NVIDIA's US packaging deal pressures Taiwan test houses

NVIDIA (NVDA US) signed a US$1.5bn strategic agreement with Amkor Technology (AMKR US) on 23 July to fund the expansion of Amkor's advanced packaging and test capacity in Arizona, and Taiwan's packaging and test names sold off the next session on the read-across. Amkor rose about 16% on the news.

The funding takes the form of a prepayment, and the two sides will co-develop packaging and test methods for next-generation AI platforms, with a focus on heterogeneous chip integration. Amkor is already a NVIDIA packaging partner and signed a 10-year agreement with TSMC in June, so the deal reads as NVIDIA pre-committing cash to lock down US capacity while advanced-packaging supply stays tight.

Taiwan's test houses fell across the board on 24 July. ASE Technology (3711 TT) dropped more than 6% intraday to NT$609, and ChipMOS (8150 TT) and Tong Hsing (6271 TT) fell 6% to 7%, with King Yuan Electronics (2449 TT) and second-tier testers down 4% to 5%. The selloff reflects positioning, not a change in order books. ASE, Powertech (6239 TT) and other testers hold results calls next week, where the question is whether onshore US capacity signals eventual share loss or simply meets demand that Taiwan cannot absorb alone.

Linked stocks: AMKR US, NVDA US, 3711 TT, 8150 TT

Sources: 艾克爾獲輝達大單！封測卻走低 日月光、南茂重挫逾5％, Amkor獲輝達15億美元大單擴產先進封裝 盤後噴

AMD readies its NVIDIA rival on TSMC 2nm

AMD (AMD US) used its Advancing AI 2026 event to detail the Instinct MI455X accelerator and the Helios rack that ties 72 of the GPUs into one system, its first design built to compete head-on with NVIDIA's 72-GPU rack. The MI455X is a 320 billion transistor chip whose compute dies are fabricated on TSMC's 2nm gate-all-around process, with the cache and I/O dies on N3P and the whole assembly joined through TSMC CoWoS-L packaging and six HBM4 stacks.

AMD claims the MI455X delivers up to 34 times the token throughput of the prior generation and up to 18 times lower token cost, the metrics that matter for inference economics. Its sixth-generation EPYC Venice server CPU, already in production, is the industry's first high-performance computing product to reach mass production on TSMC 2nm.

The Taiwan read-through is a broad server wave rather than a single winner. Gigabyte's Giga Computing, MSI (2377 TT), ASUS (2357 TT), Advantech (2395 TT), Pegatron (4938 TT) and MiTAC all launched sixth-generation EPYC systems this week, with the first boxes shipping in November. TSMC (2330 TT) captures the process and packaging value on both the GPU and the CPU regardless of how the AMD versus NVIDIA share contest resolves.

Linked stocks: AMD US, 2330 TT, NVDA US, 2377 TT

Sources: 超微 AI 晶片進擊，最新 GPU 與 CPU 產品採 2 奈米製程，台積助攻, Tom's Hardware

Transcend's quarterly profit tops any prior full year

Transcend Information (2451 TT) reported 2Q26 net profit of NT$11.9bn, up 46.5% QoQ and a record, a single-quarter figure that exceeds the company's profit for any full year in its history. Earnings per share reached NT$27.7, and the stock jumped 9.4% to NT$267.5, touching its NT$268.5 limit.

The driver is the memory price surge working directly into a module maker's margin. Transcend buys DRAM and NAND, packages it into branded modules and cards, and sells into industrial, embedded and edge markets where supply is tight and pricing has run hard. First-half net profit reached NT$20bn, up more than 24 times YoY, with first-half EPS of NT$46.63.

The forward question is durability. Management pointed to robotics, electric vehicles and AI PCs as sustaining demand for high-capacity, high-reliability storage, and flagged new products on BiCS8 218-layer 3D NAND. A module maker's earnings are geared to the spread between component cost and module price, so the same cycle that lifted this quarter turns against Transcend whenever memory prices roll over.

Linked stocks: 2451 TT, 000660 KS, MU US

Sources: 焦點股》創見：Q2賺贏歷年全年 亮燈漲停

Alphabet lifts capex toward US$205bn on compute shortage

Alphabet (GOOGL US) raised its 2026 capex guidance to US$195bn to US$205bn from US$180bn to US$190bn, said 2027 spending will rise significantly again, and stated plainly that its compute supply still falls short of demand. It plans to lease third-party capacity from 3Q26 as a bridge while it builds its own.

The spend is backed by contracted demand. Alphabet cited a remaining performance obligation of US$514bn, roughly half of which converts to revenue within 24 months, and framed the shortfall as serving a small number of very large cloud customers on multi-year deals. Google Cloud began delivering TPU systems to third parties this quarter but told investors external TPU revenue stays small through 2026 and ramps in 2027.

For Taiwan, the clearest beneficiary is Foxconn (2317 TT), which supplies both NVIDIA GPU racks and Google TPU trays and puts its combined GPU-plus-custom-silicon share above 40%. Quanta (2382 TT) carries AI order visibility into 2027 and 2028 and is adding three US sites. The capex raise extends order visibility for the rack assemblers by another year, which is what lowers the risk that this year's build pace fades in 2027.

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

Sources: Google上修資本支出並直言算力供不應求　鴻海「雙平台」通吃, NVIDIA 訂單動能強勁 廣達、鴻海 AI 伺服器商機續旺

DeepSeek's founder says Huawei can replace NVIDIA's GPUs

DeepSeek founder Liang Wenfeng told an investor meeting that Huawei's Atlas 950 SuperPoD can fully substitute for NVIDIA's GB200 and GB300 rack systems on performance and price, even though Chinese silicon remains about two years behind on a per-chip basis. His framing, that four Huawei chips roughly equal one NVIDIA chip, is a claim about system-level parity through brute-force scaling rather than chip-level parity.

The numbers he gave show the gap and the workaround at once. Training an 800 billion active-parameter model, he estimated, would need about 50,000 NVIDIA GB300 units or about 200,000 Huawei 950 chips. DeepSeek holds roughly 20,000 NVIDIA H100-equivalent GPUs and has taken about 16,000 Huawei Ascend 950 cards, equal to roughly 4,000 NVIDIA GPUs, to validate its models on Chinese hardware.

The signal for investors is that China's largest AI labs are now willing to trade energy and capital for domestic supply, and Liang argued NVIDIA's CUDA software advantage is eroding. For NVIDIA the risk is gradual demand diversion at the China margin rather than a hit to the current quarter, because the four-to-one silicon penalty means Huawei-based training burns far more power and floor space for the same result, which caps how fast Chinese buyers can switch.

Linked stocks: NVDA US, 2330 TT, AMD US

Sources: DeepSeek 創辦人梁文鋒放話華為一年就可擺脫輝達，美中差距只在算力, 華為950超節點「落後2年」梁文鋒：能平替NVIDIA、貴200％都行

先聲 First Word: Exclusives from Chinese-Language Sources

Zhen Ding sees the global PCB market topping US$100bn in 2026

Zhen Ding Technology (4958 TT) chairman Shen Ching-fang broke ground on a third Shenzhen smart-manufacturing campus, an investment above RMB 10bn aimed at AI server high-end HDI and substrate-like boards plus flexible boards for AI phones and glasses, and said the global PCB market should exceed US$100bn in 2026. The world's largest PCB maker is committing capital to move from a consumer-electronics base toward AI compute infrastructure as its second growth curve. (24 Jul 2026) Source: AI 掀 PCB 黃金時代 沈慶芳：今年全球市場規模有望突破千億美元

Samsung Electro-Mechanics lands a second large AI-server MLCC order

Samsung Electro-Mechanics signed a supply contract worth about KRW 300bn for AI server multilayer ceramic capacitors, following a roughly KRW 450bn order in June, as AI server expansion pulls demand for high-performance passive components. The back-to-back wins point to a tightening AI-server passives market, a read-across for Taiwan's MLCC makers even as Yageo trades weak on profit-taking. (24 Jul 2026) Source: 三星電機AI伺服器MLCC訂單爆發　繼6月大單後再拿3,000億韓元合約

LG Innotek's glass-substrate advance narrows the gap with Intel and Samsung

LG Innotek has improved its through-glass-via process and is in talks to apply it to LG's glass substrate business, a technology industry watchers say is now near commercial readiness. The step positions LG to chase Intel and Samsung Electro-Mechanics in glass substrates, the packaging platform that could succeed today's organic substrates for high-end AI chips. (24 Jul 2026) Source: 樂金玻璃基板TGV技術大突破　助樂金Innotek加速追趕英特爾、三星電機