Zero One Investment Research Daily Intelligence Brief, August 14, 2026

Phison sees years of NAND shortage as Quanta guides notebooks down over 20%

Top Market Signals

Phison says NAND stays short for years

Phison Electronics (8299 TT), a leading maker of the controller chips that manage flash memory, posted the biggest quarter in the history of Taiwan's memory industry and then told investors the shortage behind it runs well past the point most of the market has been modeling. Chief executive Pan Chien-Cheng said supply is tight "not just until next year, but many years," and that 2027 will very likely be shorter than 2026.

Second-quarter revenue was NT$67.9bn, up 65.7% QoQ and 279.5% YoY. Gross margin reached 65.31%, up 3.49 percentage points QoQ and 36.25 points YoY, and net profit rose 72.8% QoQ and 3,419% YoY to give earnings per share of NT$118.57. First-half profit of NT$41.4bn exceeded the previous 23 quarters combined. The prior Taiwan memory record, set by module maker Innodisk on 6 August, lasted one week.

Pan's argument for the duration is a build-time argument. A new NAND plant takes as long as four years to go from investment to usable output, so supply cannot answer a demand shift quickly. He said the binding question has moved from what memory costs to whether a buyer can get any, and that 2027 and 2028 will still not balance. He expects prices to keep rising but by shrinking increments, on the view that pushing memory makers toward 90% gross margins would damage the industry that has to keep buying.

The company was direct about what produced the margin. Cheap NAND bought before the run-up flatters the current cost line, and that benefit fades as inventory turns, so Pan said margin cannot keep climbing in a straight line. Phison is holding roughly eight months of supply, funded partly by an US$800m convertible bond issued during the quarter, and spent NT$15.1bn on research, 22.2% of revenue. Enterprise and embedded products now carry the business: AI ecosystem solutions reached 38% of revenue with 68% QoQ growth, embedded reached 32% with revenue more than doubling QoQ, and retail fell below 5%. The board proposed a NT$60 dividend, a 32% payout against the roughly 55% Phison has paid historically, retaining the rest for research and capacity.

Linked stocks: 8299 TT, 2408 TT, 000660 KS

Quanta's AI servers set records as notebooks fall

Quanta Computer (2382 TT) reported its best quarter on record and, in the same session, guided third-quarter notebook shipments down more than 20% QoQ. The two halves come from the same cause: the components its AI server business is winning on are the components its notebook business can no longer afford or obtain.

Second-quarter revenue was NT$1.04tn, up 28.1% QoQ and 105.6% YoY. Gross margin returned above 5% at 5.02%, operating margin reached 3.3%, and net profit of NT$28.65bn was a single-quarter record, up 35.2% QoQ and 69.9% YoY for earnings per share of NT$7.43. AI servers ran at 75% to 80% of total server revenue in the first half, with management targeting 80% for the year.

On notebooks, Quanta shipped 11.5m units in the quarter, ahead of its own plan, because brands pulled orders forward from the second half once component prices started moving. Management now expects third-quarter notebook shipments to fall more than 20% QoQ and full-year shipments to decline by a double-digit percentage, and said it sees no near-term improvement in CPU and memory availability. Consumer brands are taking the heavier hit.

Chief financial officer Yang Chun-lieh said customer AI roadmaps and order visibility now reach into 2028, and the company raised 2026 capital spending from NT$30bn to NT$40bn, a rise of more than 33%. AI server capacity at the end of this year should be double the end-2025 level, and 2028 capacity double that again. The board approved a US$973m injection into the US subsidiary for California assembly capacity, and the Hwaya Technology Park plant bought from AUO is due for handover in 4Q27 and production in 2028, chosen partly because a former panel fab already has the high-voltage supply that finished AI servers need for testing. From this quarter Quanta also moves some high-value components on selected projects from buying them itself to taking them on consignment, which reduces the working capital that expensive AI servers tie up.

Linked stocks: 2382 TT, 2317 TT, 2324 TT

Neoclouds now approach Taiwan's server assemblers

A new customer line ran through the first-half results calls of Foxconn (2317 TT), Quanta (2382 TT), Wistron (3231 TT), Pegatron (4938 TT) and Compal (2324 TT): the emerging cloud operators the industry calls neoclouds. The detail worth having is the direction of travel. Quanta told investors it had not gone looking for these customers. They came to Quanta, because the chips they need are short and assembly capacity that can deliver at scale is shorter.

The economics differ from the hyperscaler business in both directions. Neocloud customers want heavily customized machines and are generally not able to supply the components themselves, so the assembler funds the expensive parts and carries the working capital. In exchange, the assemblers said gross margins on those orders are better than on comparable orders from the large cloud service providers, whose spending this year passes US$800bn.

Each of the five is at a different point. Compal said neocloud customers are the main driver behind server revenue that matched all of last year in the first quarter and grew a further 70% to 80% in the second. Pegatron, whose server business is up tenfold this year, started with neoclouds deliberately to build engineering and factory capability before approaching the largest cloud operators, and now splits the work so that a subsidiary handles smaller neocloud and enterprise accounts while the parent pursues the largest ones. Foxconn expects neocloud revenue to contribute before the end of this year, alongside new model developer, sovereign and enterprise customers. Rotating chief executive Chiang Chi-heng said the barrier to entry keeps rising, because an AI system now requires high-power design, liquid cooling, system integration and validation together.

The customer set is real rather than promotional. By the end of June, nine neocloud operators sat inside the global top 40 cloud providers, against five inside the top 30 at the end of the first quarter, according to Synergy Research Group data cited in the reporting. CoreWeave, the largest of them, lifted installed compute from 1GW at the end of March to 1.5GW at the end of June across 51 data centers. Synergy puts the neocloud market at US$400bn by 2031.

Linked stocks: 2382 TT, 2324 TT, 4938 TT, 2317 TT

Nanya plans two new DRAM fabs

Nanya Technology (2408 TT) is planning two new 12-inch wafer fabs, one at the Yunlin Industrial Park and one at the Pingtung Science Park, in what Formosa Plastics Group describes as the largest investment in its history. Group administrative center president Wu Chia-chao confirmed the plan and said the total will exceed the NT$300bn committed to the Linkou 5A fab due to reach volume production in 2028, putting the program close to NT$400bn.

Where the fabs go was decided by infrastructure rather than by convenience. TSMC's domestic expansion has left power supply in northern Taiwan tight enough that a northern site was not available to Nanya, so the company asked the government for land and was routed through a cross-ministry process coordinated by Presidential Office secretary-general Pan Meng-an. Nanya assessed both southern sites and is leaning toward starting at Yunlin, because the surrounding infrastructure there is complete and the Pingtung site would not finish in time to meet the customer demand the capacity is meant to serve. Pingtung stays in the next phase.

The output is aimed at next-generation and custom DRAM, with Yunlin planned around fourth-generation 10nm class (1D) processes and below, positioned for memory demand as AI inference spreads from data centers to edge devices. Nanya has been running its existing capacity full and still cannot meet customer orders.

Customers have already paid toward this. In March, SanDisk, Cisco, Kioxia and SK Hynix's Solidigm unit together put NT$78.7bn into a Nanya private placement, taking equity in exchange for capacity. Our 7 August Blade reported the board lifting 2026 capital spending to NT$69.7bn and setting a 2026 to 2029 ceiling of NT$346.6bn that includes EUV lithography tools. Today's plan sits on top of that budget rather than inside it, which is why the group framing has moved from a Nanya capex decision to a Formosa Plastics one.

Linked stocks: 2408 TT, 2330 TT, 000660 KS

Cerebras' hardware revenue falls as cloud triples

Cerebras Systems (CBRS US) grew second-quarter revenue 74% YoY to US$180m and raised its full-year guidance, but the composition moved sharply. Hardware revenue fell 23% YoY to US$54.1m, while the business of renting out its own AI compute nearly tripled to US$126m. For a company whose case rests on selling an alternative to NVIDIA silicon, the shrinking line is the one that matters.

Chief executive Andrew Feldman gave a physical reason rather than a demand one. Delivery and revenue recognition on hardware orders is lumpy, and some customers do not yet have enough data center space to install the equipment they have ordered. That is the same constraint Wistron's chief technology officer described in our 10 August Blade, when he said GPU lead times had shortened enough that six months of notice suffices and that the binding limit on AI data centers had moved to electrical infrastructure and power distribution delivery schedules. Buyers with capital and chips are now waiting on buildings and power.

The guidance does not read as a demand problem. Cerebras guided third-quarter revenue to about US$215m against a market estimate near US$213m, with gross margin of 38% to 40%, and raised full-year revenue guidance to US$880m to US$890m from US$855m to US$865m, saying it plans to more than triple revenue in 2027. Shares fell about 15% in early trade on 13 August, with the reaction attaching to the hardware line rather than to the totals. For scale on how early the customer expansion still is, AMD and Intel grew revenue 50% and 25% respectively last quarter.

Linked stocks: CBRS US, NVDA US, 3231 TT

Lotes' margin falls as cheap materials run out

Lotes (3533 TT), which makes the sockets that hold processors onto a server board, reported second-quarter gross margin of 45.47%, down 4.08 percentage points QoQ and 5.53 points YoY. Servers were 55.36% of revenue. A supplier with that mix, in this cycle, still lost margin.

Spokesperson Tsai Ming-jui gave the reason plainly. The low-priced raw material Lotes bought last year was consumed during the first quarter, so the second quarter reflects current market prices for the first time. He said margin improves when oil prices settle, when the 2027 server platforms reach volume, and when the new products reach scale, which places the recovery outside this year. Net profit was NT$1.984bn, down 17.1% QoQ but up 164.8% YoY, for earnings per share of NT$17.67.

Two new products are moving into production against that cost backdrop. SOCAMM, the compact memory module format built for AI servers, has entered mass production. Quick disconnects, the couplings that let a liquid cooling loop be opened without draining it, began volume shipment in July at about 1% of revenue, and Lotes expects 2% to 3% by the end of the year if capacity expansion holds. Desktop revenue is falling on component shortages, the same input squeeze seen from the other side at Quanta.

The 2027 case rests on platform timing. Tsai said the new Intel and AMD server platforms are shipping in small validation quantities now, should reach limited production in the fourth quarter, and reach real volume early next year. Server mix rises from here as the two new product lines ramp. What would change the timing is those platform dates slipping.

Linked stocks: 3533 TT, INTC US, AMD US

先聲 First Word: Exclusives from Chinese-Language Sources

Samsung moves commodity memory packaging to Vietnam

Samsung Electronics is reported to be shifting part of the back-end memory capacity handled at its Cheonan and Onyang sites in South Chungcheong province to Thai Nguyen in Vietnam, according to industry sources cited by Korean outlet DealSite. Freeing Korean back-end lines for high bandwidth memory means the commodity DRAM and NAND that PC and phone makers buy gets packaged further from home and, on the margin, stays scarcer. (13 Aug 2026) Source: 三星傳調整記憶體產能　天安溫陽聚焦HBM、一般製程移至越南

Delta and Lite-On both lift capex to about 9% of revenue

Delta Electronics (2308 TT) and Lite-On Technology (2301 TT) have each raised capital spending to roughly 9% of revenue to chase AI data center power and thermal business. Power suppliers used to follow system makers' specifications and deliver; now that NVIDIA and AMD change platforms annually, a supplier that cannot fund research and build capacity near the customer at the same cadence risks being designed out of the next generation. (13 Aug 2026) Source: 台達電、光寶科擴大資本支出攻AIDC　佔營收比重均約9%

Taiwan's chip designers start collecting on AI vision sensing

Novatek (3034 TT), Himax Technologies (HIMX US), Realtek (2379 TT), Pixart Imaging (3227 TT) and Elan Microelectronics (2458 TT) all say revenue from AI vision sensing is now rising quickly and will become a meaningful part of their revenue mix. Novatek, the display driver chip leader, called it the key driver of continued growth in its system-on-chip business. Several of these are mid-sized Taiwanese designers better known for display drivers, touch controllers and mouse sensors, and edge AI is giving them a second growth line. (13 Aug 2026) Source: AI視覺感測市場需求加溫　聯詠、原相、義隆與瑞昱迎收割