I titoli col trend più forte

Micron Technology (NASDAQ: MU, headquartered in Boise, Idaho, founded 1978) is the only US-based volume producer of memory semiconductors and one of just three companies in the world with meaningful scale in both DRAM and NAND flash — the other two being SK hynix (000660.KS) and Samsung Electronics (005930.KS). Memory is the part of a computer that holds data while it is being worked on (DRAM) or stored permanently (NAND); it is a commodity in the sense that chips from different vendors are largely interchangeable, which historically made the industry brutally cyclical: prices collapse when capacity outruns demand and spike when it does not. Micron's revenue is organised around four business units: Compute & Networking (data center and PC DRAM, including HBM), Mobile (smartphone memory), Embedded (automotive, industrial, IoT) and Storage (SSDs and NAND). The AI-era pillar of the franchise is HBM — High Bandwidth Memory — vertically stacked DRAM dies connected by through-silicon vias and mounted next to an AI accelerator; every NVIDIA or AMD data center GPU ships with several HBM stacks, so HBM demand is directly geared to AI accelerator volumes and carries structurally higher margins and ASPs than commodity DRAM. Micron entered HBM as the third supplier behind SK hynix and Samsung and has been closing the gap generation by generation (HBM3E 8-high and 12-high, now HBM4). Manufacturing is capital-intensive and geographically concentrated: leading-edge DRAM fabs in Taiwan (Taichung), Japan (Hiroshima) and Singapore, NAND in Singapore, plus large US greenfield projects in Boise (Idaho) and Clay (New York) supported by CHIPS Act funding, and a multi-billion-dollar HBM advanced-packaging and DRAM expansion in Singapore. Structural risks: (1) the memory cycle itself — three-player oligopolies have historically broken down when everyone adds capacity at the same time, and Micron's earnings swing from large losses to record profits within 18 months; (2) customer concentration, with a growing share of HBM revenue tied to a handful of AI accelerator and hyperscaler buyers; (3) technology and qualification risk on each HBM generation, where a yield stumble can cost an entire product cycle of share; (4) enormous and largely fixed capex commitments that must be funded through the trough as well as the peak; (5) geopolitical exposure — China revenue restrictions, export controls, tariffs and fab concentration in Taiwan and East Asia.

Micron is in the strongest phase of any memory up-cycle on record. Fiscal Q3 2026 (reported 24 June 2026) delivered record revenue of $41.5B, up 74% sequentially and roughly 346% year over year, with gross margin at ~84.9% and EPS of $25.11 — all above guidance. Management guided fiscal Q4 to approximately $50B in revenue ($49.0-51.0B) with gross margin near 86% and non-GAAP EPS around $31. The driver is HBM plus an extremely tight commodity DRAM market: HBM capacity is sold out under contract, HBM4 shipments to the lead accelerator customer are ramping roughly twice as fast as HBM3E 12-high did, more than $1B of HBM4 revenue has already been booked, and HBM4 pricing is reported to carry a step-up of at least 50% versus HBM3E. Micron has also signed multi-year Strategic Customer Agreements (SCAs) that lock in volume and pricing further out than the traditional quarterly contract cycle — the key argument bulls make that this cycle is structurally different. On HBM share, Micron sits third: on NVIDIA's HBM4 allocation the split is roughly mid-50% SK hynix, mid-20% Samsung, ~20% Micron, though Micron has overtaken Samsung in overall HBM market share. Capacity is being expanded aggressively, including a ~$24B Singapore fab alongside the Idaho and New York US projects. The stock has re-rated violently: shares closed around $1,010-1,017 in early September 2026, up roughly 256% year to date, giving Micron a mega-cap market capitalisation. Valuation remains the crux of the debate — the forward P/E is somewhere between ~7x and ~14x depending on whose estimates are used, which looks cheap on peak-cycle earnings and expensive if those earnings are the peak. Sell-side is heavily positive (roughly 44 buys vs 4 holds, average target above $1,500), while prediction markets price much more modest outcomes. Next catalyst: fiscal Q4 2026 results on 30 September 2026, where the swing factors are September-quarter memory ASPs (any sequential softness would be the first crack), fiscal 2027 capex guidance, and early commentary on calendar-2027 HBM contract pricing.

Taiwan Semiconductor Manufacturing Company (TSMC, 2330.TW on the Taiwan Stock Exchange, ADR TSM in New York) is the world's largest dedicated semiconductor foundry: it does not design or sell its own chips, it manufactures them on behalf of others. The 'pure-play foundry' model it invented in 1987 split the industry in two — 'fabless' designers (NVIDIA, Apple, AMD, Qualcomm, Broadcom, MediaTek) own the blueprint, TSMC owns the factory. The customer therefore never becomes a competitor, which is precisely why almost every major chip designer entrusts TSMC with its most valuable designs. Revenue comes from wafers: silicon discs processed inside 'fabs' (fabrication plants) and priced by technology 'node' — the smaller the node (5nm, 3nm, 2nm), the more transistors per unit area, and the higher the price per wafer. The leading edge is the profit engine: advanced nodes command premium pricing and carry the highest margins, while mature nodes amortise already-depreciated equipment. Alongside wafer manufacturing, TSMC has become the bottleneck in advanced packaging — CoWoS (Chip-on-Wafer-on-Substrate), the technology that stitches a logic die to stacks of high-bandwidth memory, is effectively mandatory for AI accelerators, and TSMC controls the bulk of world capacity. The competitive moat rests on three reinforcing pillars. First, capital: a leading-edge fab costs $20bn+ and TSMC spends $40-65bn a year in capex, a level only two other firms on earth (Samsung Foundry and Intel Foundry) even attempt. Second, yield: manufacturing know-how accumulated over decades translates into a higher share of working chips per wafer, which is the real currency of the business — a node that works at 70% yield versus a rival's 50% is dramatically more profitable for both foundry and customer. Third, switching costs: porting a complex design to another foundry takes years and hundreds of millions of dollars, so customers are locked in for a product generation. The result is a structural near-monopoly at the leading edge, with TSMC holding roughly two-thirds of global foundry revenue and an even larger share of sub-5nm capacity. Financially the model is unusual for a manufacturer: gross margins in the 50-70% range, high returns on capital, and a large, lumpy capex cycle that makes depreciation a key swing factor for margins. The structural risks are equally distinctive. Geography is the dominant one: the majority of leading-edge capacity sits on the island of Taiwan, inside a geopolitical flashpoint between Beijing and Washington — a factor sometimes described as the 'silicon shield' (the world's dependence deters conflict) and sometimes as single-point-of-failure risk. TSMC's answer has been a multi-year geographic diversification into Arizona, Japan, Germany and beyond, but overseas fabs cost more to run than Taiwanese ones and structurally dilute gross margin. The business is also cyclical and concentrated: a handful of customers account for a large slice of revenue, and demand tracks end markets (smartphones, PCs, and since 2023 above all AI datacentre) that can swing hard. Finally, it operates inside export-control and tariff policy — Washington restricts what can be sold to China, and semiconductor trade measures directly shape where TSMC must build. Headquarters are in Hsinchu, Taiwan; the company is led by CEO C.C. Wei.

TSMC is the world's largest contract chipmaker and the only viable supplier of leading-edge logic at scale: Nvidia, Apple, AMD, Broadcom and Qualcomm all design chips and hand them to TSMC to be built. It sells manufacturing capacity rather than products, so its pricing power comes from being one to two nodes ahead of anyone else, not from a consumer brand. Its second moat is CoWoS advanced packaging - the technique that stacks logic and high-bandwidth memory into one AI accelerator package - which has been the industry's hard bottleneck since 2024. In 2026 the company raised capex by $8bn to a $62bn midpoint and guided full-year US-dollar revenue growth above 40%. This line is the Taiwan-listed ordinary share; slot 5 (TSM) is the same company's US ADR.

Asia Vital Components (AVC), founded in 1991 and headquartered in New Taipei City, Taiwan, is a vertically integrated thermal-management manufacturer: it designs and builds the parts that keep electronics from overheating. It started as a maker of cooling fans and heat sinks for PCs and notebooks and has evolved into one of the main thermal suppliers for data-centre and AI servers. Its product range spans air cooling (fans, heat sinks, vapour chambers, heat pipes) and, increasingly, liquid cooling — cold plates that sit directly on the GPU/ASIC package (direct-to-chip), manifolds that distribute coolant inside the rack, CDUs (coolant distribution units) and full liquid-cooled chassis. The economics of the business are driven by content per server: every generation of AI accelerator dissipates more watts, so each rack needs more and more expensive cooling hardware, which lifts both revenue per unit and mix-driven gross margin. AVC sells mainly to server ODMs/OEMs and hyperscaler supply chains, which makes customer concentration high and ties its cycle tightly to AI capex. Manufacturing is spread across China, Taiwan and newer capacity in Vietnam, a footprint that also serves as tariff/geopolitical diversification. It is listed on the Taiwan Stock Exchange under ticker 3017.

The story right now is a full-blown AI-cooling upcycle, with the stock behaving as a high-beta derivative of AI server capex. 2025 revenue was about TWD 139.6 billion, up roughly 95% year on year, and growth accelerated into 2026: Q1 2026 revenue of about TWD 49 billion (+110% YoY) with a 29.8% gross margin, followed by Q2 2026 (reported 12 August) revenue of about TWD 49.1 billion (+66% YoY), a record gross margin of 32.57% (+8.16 points YoY) and EPS of TWD 24.37 (+137% YoY). The margin expansion — not just the revenue growth — is the key point: it comes from mix, as lower-margin air cooling gives way to liquid cooling content (cold plates, manifolds, chassis), which management has framed as the start of the "liquid cooling era", with liquid-cooling penetration expected to pass 50% during 2026. Design-win breadth is unusually wide for a component supplier: AVC is qualified on Nvidia GB300 and the next-generation VR200 platform as well as on custom ASIC programmes (Google TPU, Groq LPU), and management has pointed to ASIC revenue overtaking GPU revenue next year — relevant because it diversifies away from a single accelerator roadmap. New Vietnam capacity has entered mass production on GB200 products and starts contributing revenue from Q3 2026. Management claims order visibility stretching to 2029 and sell-side estimates cluster around roughly 60% revenue growth for 2026; analyst positioning is overwhelmingly positive (consensus close to Strong Buy). The other side of the ledger is what this has already done to the price: the shares trade around TWD 2,700-2,800 after a multi-fold re-rating, so the multiple already embeds continued hyperscaler capex growth, sustained liquid-cooling mix gains and no pricing pressure from competitors (Auras, Cooler Master, Vertiv and the broader Taiwanese thermal supply chain) all expanding capacity into the same demand.

Dell Technologies is one of the world's largest makers of computing hardware. Its business splits into two halves. ISG (Infrastructure Solutions Group) sells servers, storage and networking gear to data centres — this is where the AI servers live: Dell buys Nvidia GPUs, integrates them into racks with power, cooling and cabling, and ships a working system to the customer. CSG (Client Solutions Group) sells PCs, laptops and workstations to companies and consumers. Dell is essentially an integrator and a logistics machine rather than a chip designer: its edge is supply-chain scale, direct enterprise sales relationships and the ability to deliver very large deployments fast. The economic consequence is structurally thin margins — the expensive part of an AI server (the GPU) is bought from Nvidia, so Dell captures a small slice of a very large revenue number. Company-level gross margin runs around 20%, versus 70%+ for a chip designer. Dell was taken private in 2013 and returned to the public market in 2018; founder Michael Dell remains CEO and a controlling shareholder through a dual-class structure.

Dell is in the middle of an AI-server boom that has transformed the size of the company. In Q2 FY2027 (reported early September 2026) revenue hit a record $47.0bn, up 58% year over year, with non-GAAP EPS of $7.04 — more than $2 above consensus. AI server revenue doubled to $16.4bn, the company booked $60.9bn of AI orders in the quarter alone, and exited with a record $95bn AI backlog, with management saying the pipeline is now 'multiples of backlog'. Full-year guidance was raised to roughly $192bn of revenue with about $74bn of AI server revenue, close to three times the prior year. Critically, the margin story improved: ISG operating income rose 225% to $4.8bn with a 15% operating margin, up 620bp year over year — a reversal of the FY2026 pattern, when the AI mix shift pushed ISG margin down to 11.7% and group gross margin from 23.7% to 20.2%. The open question now is memory cost: server DRAM prices were guided up 13-18% sequentially into Q3 as DDR5 supply is diverted toward HBM for GPU packaging, and Dell buys memory as an input. The stock trades around 19-22x forward earnings, roughly double its five-year average of about 11x, so the market has already re-rated it from a low-multiple hardware name toward an AI infrastructure name.

Lenovo Group is the world's largest PC maker by unit share and, since acquiring IBM's ThinkPad business in 2005 and IBM's x86 server line in 2014, a full-stack hardware company. Listed in Hong Kong and controlled in part by China's state-linked Legend Holdings, it runs three segments: IDG (Intelligent Devices — PCs, tablets, Motorola smartphones), ISG (Infrastructure Solutions — servers and storage, including AI servers) and SSG (Solutions & Services — support contracts, managed and as-a-service offerings). The economics differ sharply by segment: IDG is a high-volume, low-margin assembly business whose profitability depends on component costs and PC replacement cycles, ISG is a scale game where Lenovo assembles other people's silicon (mainly Nvidia GPUs) into racks and earns a thin slice of a very large revenue number, and SSG is the small but high-margin services annuity. In plain terms: Lenovo buys expensive chips, puts them in boxes, and sells the boxes — so its gross margin is structurally thin (high single digits to low teens at group level) and its earnings are highly sensitive to memory and GPU input prices it does not control.

The dominant theme is Lenovo's transformation from a PC maker into an AI-infrastructure supplier, and the extraordinary re-rating that has come with it: the shares are up roughly 200% in 2026 and trade at ~HK$32 against a 52-week high near HK$34.9, which is why the momentum screen picks the name up. The trigger was the Q1 FY2026/27 report (quarter ended June 2026, published 13 August 2026), the strongest quarter in group history: revenue of US$26.9bn, +43% year-on-year and an all-time quarterly record, with adjusted net income up 176% to US$1.1bn, above the US$1bn mark for the first time. The mix shift is the story. The Infrastructure Solutions Group (servers and data-centre hardware) nearly doubled, +98% to US$8.5bn, and management put the AI-server pipeline at US$54bn, +157% versus the prior period; ISG had turned profitable for the first time in the preceding fiscal year (US$19.2bn revenue, a US$140m year-on-year operating-profit improvement) and has since printed a record ~9.1% operating margin quarter, above the group's ~5.7% adjusted operating margin, so the mix is currently accretive rather than dilutive. AI-related revenue reached US$9.3bn, +60%, or 35% of the group. The legacy businesses are also running hot: Intelligent Devices (PCs and smartphones) +27% to US$17.1bn and Solutions & Services +28% to US$2.9bn at a 24.2% operating margin. The counterweight is cost: DRAM and NAND are in a violent up-cycle (DDR5 up ~70% year-on-year, some parts up to +170%, contract prices +90-95% quarter-on-quarter in early 2026) and Lenovo has called elevated memory pricing the 'new normal' into the 2030s, passing it through with list-price increases in January and again in July 2026. Memory inflation squeezes the PC bill of materials while helping the server mix — the two effects run in opposite directions inside the same P&L. Sell-side positioning is crowded on the bull side (roughly 20 buy ratings, no sells, average target near HK$45), and the valuation sends a mixed signal: trailing P/E around 29x versus a forward P/E near 10x, i.e. the cheapness is entirely a function of forward estimates that extrapolate the current AI-server and pricing cycle. Next scheduled catalyst is the Q2 FY2026/27 interim report, due around mid-November 2026.

NVIDIA designs the graphics processors (GPUs) and the full-rack systems that train and run artificial intelligence. Think of an AI model as a factory: the GPU is the machine on the assembly line, and NVIDIA is by far the dominant machine maker. Founded in 1993, it started by selling 3D graphics cards to video-game players, then discovered in the late 2000s that the same chips — thousands of small cores working in parallel instead of a handful of fast ones like a CPU — were ideal for the matrix arithmetic behind neural networks. From 2023 the generative-AI wave turned that side business into the core one: today the Data Center segment (accelerators, networking, complete server racks) generates the overwhelming majority of revenue, with gaming, professional visualization and automotive reduced to minority lines. The real moat is not only silicon: it is CUDA, the software layer released in 2006 that lets developers program GPUs, plus the libraries and frameworks built on top of it over almost twenty years. Migrating that stack to a competitor means rewriting and re-validating code, which is why NVIDIA has kept a share of the AI accelerator market above 80% despite credible rivals. It is fabless — it designs the chips and outsources manufacturing, mainly to TSMC — and sells primarily to hyperscalers (Microsoft, Amazon, Google, Meta, Oracle), to AI labs and to so-called neoclouds. The business is structurally cyclical and concentrated: a handful of customers accounts for a large slice of revenue, and demand follows the capital-expenditure cycle of a few very large technology buyers.

NVIDIA is the epicentre of the AI trade and, at roughly $5.3 trillion of market value, one of the largest companies ever listed. Fiscal Q2 2027, reported on 26 August 2026, was another beat-and-raise: revenue of about $96.2 billion, roughly double year over year, with the Data Center segment at about $89.0 billion (+117% y/y) and EPS of $2.22 against $2.09 expected. The guide for the current quarter, roughly $108 billion, came in above consensus, and management flagged fiscal-2028 revenue growth of around 70% — an unusually long forward commitment, framed explicitly as supply-constrained rather than demand-constrained, with the CFO noting that signalled demand would support something closer to 140% growth if supply allowed. The Vera Rubin platform is the ramp now driving the cycle. Two things have changed the texture of the story in recent weeks. First, memory: DRAM and HBM prices are spiking, memory now represents around 62% of Vera Rubin's bill of materials, and NVIDIA has guided gross margin to trough at roughly 71-72% in fiscal Q4 2027 — down from the ~75% investors got used to. The company is responding by pushing more than 15% price increases through to customers on Grace Blackwell and Vera Rubin systems, which protects the margin but shifts the cost squeeze onto the buyers whose capex budgets sustain the entire cycle. Second, China: the guidance is built assuming essentially zero China revenue, so recent H200 export approvals are optionality rather than something already in the numbers — while Jensen Huang has publicly described a Chinese AI-chip market of roughly $50 billion that export controls may keep permanently out of reach. The debate has therefore moved off "is demand real" and onto three narrower questions: whether hyperscaler capex digests after this build-out, whether custom ASICs (OpenAI's Broadcom project, AMD's and Qualcomm's data-center pushes) erode share at the inference end, and whether the margin trough is one cycle of memory inflation or a structural reset.

Rexel is a French B2B distributor of electrical products and services: it buys cables, switchgear, lighting, HVAC, solar, EV-charging and industrial automation equipment from manufacturers (Schneider, Legrand, ABB, Siemens and others) and resells it to electrical contractors, installers, industrial plants and utilities. It is one of the two or three largest players in a fragmented global market, operating in roughly 17 countries through a network of around 1,900 branches plus digital channels, with Europe as the largest region, North America second and Asia-Pacific the smallest but fastest-growing. The business model is classic distribution: thin gross margins on the goods themselves, with profit coming from scale in purchasing, logistics density, own-brand products, services and working-capital discipline. Adjusted EBITA margin runs in the 6% area, so results are highly sensitive to volumes, price/mix and the copper cycle (cable prices flow straight through revenue). Demand is driven by construction and renovation activity, industrial capex and the electrification of the economy - grid upgrades, data centres, solar, heat pumps and EV infrastructure - which makes Rexel a cyclical company with a structural growth tailwind attached to it. Growth is also bought: Rexel regularly makes bolt-on acquisitions of regional distributors to consolidate local markets.

2026 has been a re-rating year for Rexel. H1 2026 sales came close to EUR 10bn for the first time in a half-year, with Q2 at about EUR 5.25bn (+6.1% year-on-year), Europe back to volume growth (+4.4% same-day, led by HVAC, solar and EV charging) and Asia-Pacific up around 17% on industrial automation in China and India and Australian solar. Management raised full-year guidance in July to roughly 5% same-day sales growth (from a 3-5% range) while keeping the adjusted EBITA margin target at 6.2% or better, and lifted the data-centre growth expectation from above 20% to above 50% for the year - AI-driven electrical demand has become the visible swing factor in the story. The Axelerate 28 plan (productivity, fixing underperforming countries, accretive M&A) still points to a mid-term EBITA margin above 7%. The stock has responded: shares around the mid-EUR 30s after a roughly 50% one-year total return, trading near 18x earnings, in line with European trade distributors and close to the average analyst target in the high EUR 30s. In other words, the operational upgrade is real, but the easy discount is gone and the market now needs the data-centre pipeline and the margin path toward 7% to keep delivering.

Wistron Corporation (3231.TW) is one of Taiwan's largest contract design-and-manufacturing groups (ODM/EMS), spun out of Acer's manufacturing and R&D arm in 2001 and listed in Taipei. Its historical core was designing and building notebooks and PCs for global brands, but over the last decade the centre of gravity has moved decisively to enterprise and data-centre hardware: server motherboards, GPU baseboards, compute trays, liquid-cooled AI server nodes and full rack-level assembly, sold mainly to chip vendors (Nvidia above all), hyperscalers and brand OEMs such as HP, Dell and Lenovo. Alongside that, the group keeps smaller lines in displays, automotive and medical electronics, IT services and recycling, and it spun off Wiwynn, its hyperscaler-focused subsidiary, in 2016. The economics are those of a scale contract manufacturer, not of a chip designer: revenue is enormous (well above NT$1 trillion a year) because the bill of materials — GPUs, HBM and other memory, SSDs — flows through the income statement, while gross margins sit in the mid single digits and profit depends on volume, mix, engineering content and manufacturing yield rather than on pricing power. Wistron builds where its customers need it (Taiwan, China, Mexico, the Czech Republic, the US and Southeast Asia), exited iPhone assembly by selling its Indian plant to Tata in 2023, and today is best understood as a leveraged, low-margin operating proxy on data-centre capital expenditure: when AI rack demand accelerates, revenue and profit compound very fast; when orders pause, the same operating leverage works in reverse.

Wistron is in the middle of an extraordinary AI-server upcycle. Q2 2026 set records on every line: consolidated revenue of about NT$895 billion (roughly +64% year on year), net profit of NT$14.83 billion — up around 128% year on year — and EPS of NT$4.72, taking first-half EPS to NT$7.78, about +94% year on year. Momentum did not slow after the quarter: August 2026 monthly revenue came in at roughly NT$460 billion, +166% year on year and +49% month on month, pointing to another record Q3. Management said AI accounts for more than 95% of server revenue, that both the VR and B2 platforms are already in mass production with no transition gap in Q3, and that second-half AI server shipments should exceed the first half with momentum carrying into the first half of 2027; Nvidia is reported to have locked up capacity at the Zhubei server plant through 2026. A second, higher-margin leg is emerging in networking: switch shipments have grown roughly tenfold this year at gross margins above the corporate average. The company is reinvesting aggressively, committing over NT$10 billion to capacity expansion and an AI computing centre. The valuation still reflects the ODM label rather than the growth rate — the stock trades in the mid-teens on trailing earnings (recent price around NT$186 against trailing twelve-month EPS near NT$12.75) — but that discount exists because the business model is thin-margin pass-through: AI server gross margins across the top Taiwanese ODMs run only about 5.3% to 8.3%, memory and GPU cost inflation is a live 2026 headwind, and Wistron holds only an estimated 8-10% of AI rack assembly versus roughly 40% for Foxconn and 25-30% for Quanta. The next checkpoint is the Q3 2026 investor conference, expected in late October.

GIGABYTE Technology (2376.TW) is a Taiwanese hardware manufacturer listed in Taipei, historically known as one of the world's largest makers of PC motherboards and graphics cards (consumer brands AORUS and GIGABYTE Gaming). Over the past decade the company has pivoted its centre of gravity towards enterprise computing: through its Giga Computing subsidiary it designs and assembles rack servers, GPU servers and full AI/HPC systems built around NVIDIA, AMD and Intel silicon, including HGX and GB-series rack-scale platforms, plus liquid-cooling and data-centre infrastructure. The business model is that of a Taiwanese ODM/OBM: high-volume board and system assembly with relatively thin gross margins, where profitability depends on scale, component allocation from chip vendors, and product mix — consumer boards and graphics cards carry structurally higher margins than AI server racks, while AI servers carry far higher revenue per unit. Customers span PC retail channels, enterprises, cloud service providers and the newer wave of AI-focused data-centre operators. Manufacturing is concentrated in Taiwan and mainland China, with capacity additions in Taiwan, Vietnam and the US to serve AI demand and to reduce tariff/geopolitical exposure. Revenue is reported monthly, as is standard for Taiwan-listed hardware names.

The dominant theme is a violent trade-off between explosive AI-server growth and collapsing gross margin. Q2 2026 (reported 14 August 2026) was a record quarter on the top and bottom lines: revenue of NT$144.21bn (+37.3% QoQ, +41.0% YoY), operating profit of NT$8.12bn (+9.5% QoQ, +53.4% YoY) and net profit attributable to the parent of NT$6.61bn (+116.6% YoY), giving record quarterly EPS of NT$9.86. Yet gross margin fell from 12.01% in Q1 to 9.62% in Q2, a 2.39pp sequential drop, because the mix shifted hard: lower-margin AI servers jumped from 72% to 82% of revenue while the motherboard and graphics-card business was in its seasonal trough. The market punished the mix rather than rewarding the profit record — the stock closed limit down (−9.94%, −NT$39.5) on the session after the release, with foreign institutions selling roughly 7,888 lots. Management's response was to lean further into the AI cycle: full-year capex guidance was raised from NT$2.3–2.5bn to NT$2.8–3.0bn to expand AI server capacity, following the Q1 2026 call where AI server revenue was reported up about 110% YoY and capacity was said to be doubling ahead of the GB300 generation. Management also argues H2 will be stronger, with the Q3 motherboard and graphics-card peak season lifting blended margins. The open question for the next two quarters is therefore whether gross margin stabilises around 10% or keeps eroding as AI racks take an ever larger share of the mix, and whether the added capacity is absorbed by orders rather than by inventory.

ASUSTeK Computer (2357.TW) is a Taiwanese hardware maker, historically known worldwide for consumer PCs: laptops (Zenbook, Vivobook), the ROG gaming brand, motherboards and graphics cards, where it has long been one of the top-two motherboard vendors globally. Its business model is that of a branded assembler: it designs and sells finished machines but outsources most manufacturing, so profitability depends on the gap between component costs (CPUs, GPUs, memory) and the price it can charge — a structurally thin margin, typically low single-digit operating margin, very different from a chip designer. Since the AI cycle began, ASUS has added a second leg: the Infrastructure Solutions business, which builds AI servers and HPC racks around NVIDIA platforms and sells them to cloud providers, sovereign-AI projects and enterprises. The company is organised in three groups — System (consumer and commercial PCs), Infrastructure Solutions (servers/AI) and Open Platform (motherboards, components) — and also holds a large stake in industrial-computing group Advantech, which is a meaningful part of its balance-sheet value.

ASUSTeK has been re-rated during 2026 from a cyclical PC brand into an AI-server play. Q2 2026 (reported 12 Aug 2026) was a record: revenue NT$241.1bn, +39% YoY and the first quarter ever above NT$200bn, with operating profit +230% YoY and net profit +94% YoY. Server revenue roughly doubled YoY and commercial PCs grew more than 50% YoY for the sixth consecutive quarter. Management raised full-year server growth guidance to at least +150% YoY and guided Q3 servers +10-15% QoQ (>150% YoY) with PC revenue +15-20% QoQ. The stock trades around NT$1,010 (early September 2026) after a large run, and the debate is no longer whether the AI-server ramp is real but how much of it is already in the price: Morgan Stanley cut the ADR/GDR line to Underweight over valuation and margin risk, while the broader sell-side consensus stays on Buy. The main live variable into H2 is the memory shortage: conventional DRAM contract prices are seen up 13-18% QoQ in Q3 2026 and NAND up 10-15%, so ASUS is simultaneously a beneficiary of the AI build-out and a payer of its input-cost bill on the consumer PC side.