Zero One Investment Research Daily Intelligence Brief, September 1, 2026

NVIDIA pays US$3.5bn to put its interconnect inside chips that compete with its GPUs

Top Market Signals

NVIDIA puts its interconnect inside MediaTek's custom chips

MediaTek (2454 TT) priced US$3.9bn of overseas convertible bonds on 31 Aug 2026, about NT$125bn and the largest such issue in Taiwan's capital market. NVIDIA (NVDA US) took US$3.5bn of it, close to 90%, with Alphabet (GOOGL US) and overseas institutions taking the remaining US$400m. It is the first time NVIDIA has invested in a Taiwan-listed company.

The money buys a position in the part of the market that competes with NVIDIA's own accelerators. MediaTek designs custom AI chips, called XPUs, for cloud customers who want an alternative to buying GPUs. Under the expanded agreement MediaTek will build those chips on NVIDIA's NVLink Fusion interconnect and use NVHBM, NVIDIA's custom memory product, so a customer's own silicon plugs into NVIDIA racks, switches and networking rather than replacing them. Jensen Huang described the two lines as complementary, with the GPU as the general accelerator and the XPU as the customer-specific one.

Huang rejected the description of the investment as circular financing, pointing to a ten-year joint technology roadmap and a partnership that has run since 2023 through automotive platforms, the GB10 superchip in early 2025, custom XPU work in mid-2025 and RTX Spark PC processors this year. He also said NVHBM will go to MediaTek and to first partners including Amazon (AMZN US), which requires the memory and the XPU to sit together inside a CoWoS package.

Rick Tsai said supply of high-bandwidth memory and substrates is genuinely difficult right now, while arguing that sitting at the center of the AI infrastructure supply chain is what makes Taiwan useful to a customer with that problem. Economic Daily News reported that TSMC (2330 TT) is the other obvious beneficiary, since NVIDIA GPUs, Google TPUs and MediaTek's main chips are almost all made there, with ASE Technology (3711 TT) and Sigurd Microelectronics (6257 TT) picking up back-end work as the customers bind themselves together.

Linked stocks: 2454 TT, NVDA US, 2330 TT, 3711 TT

TSMC puts AI accelerator utilization below 40%

TSMC told the SEMICON Taiwan advanced process forum on 31 Aug 2026 that moving data between chips can consume up to 50 times the energy of moving it inside a chip, that data movement may account for 60% of system activity, and that this holds the utilization of expensive AI accelerators below 40%. The point of the presentation was that the industry is now spending most of its money on parts that sit idle most of the time.

Li Hsiang, who runs TSMC's high performance computing and global AI business development, put new AI infrastructure additions at 30 to 40GW a year in the coming years, against roughly 5 to 6GW a year for data centers historically. She said global AI inference token volume has risen close to 500 times since 2022, and that inference has overtaken training as what drives system expansion.

The packaging roadmap is the answer TSMC is offering. It is mass producing the largest CoWoS package to date this year at 5.5 times reticle size with yields above 98%, moves to a 14 times reticle version carrying 20 stacks of high-bandwidth memory in 2028, and adds a larger version carrying 24 stacks in 2029. On logic, A14 enters mass production in 2028 alongside N2X and N2U, with A13 and A12 following in 2029.

Li put total system compute in 2029 at 50 times the 2024 level, with advanced packaging rather than transistor scaling doing most of that work. Read against the utilization number, the claim is that most of the next several years of AI performance comes from moving data more cheaply between chips, while faster compute dies add less.

Linked stocks: 2330 TT, NVDA US, 3711 TT

Samsung locks 70% of memory output to 2031

Samsung Electronics (005930 KS) has reportedly committed up to 70% of its memory division capacity to long-term contracts running through 2031, and the spot market has emptied out behind it. MegaGrid Supply data reported by Seoul Economic Daily puts 36GB HBM3E at US$2,100 on spot against contract prices of roughly KRW500,000 to KRW700,000, which is four to five times. A 16-layer HBM4 part fetches US$3,500 on spot.

The second cause is yield. HBM4 is early in its ramp and yields below HBM3E, so each shipped part consumes more DRAM wafers than the previous generation did. Ahn Ki-hyun of the Korea Semiconductor Industry Association said HBM4 sells for close to twice the prior generation partly because ramping it absorbs DRAM capacity that would otherwise have gone to other products.

The split shows up cleanly in Korea's trade data. July DRAM export volume fell 13.2% while export value rose 18.5%, so the entire increase came from price. For buyers outside the long-term contracts, including consumer electronics makers and the AI labs now buying desktop hardware, there is no volume left to bid for at any price.

Linked stocks: 005930 KS, 000660 KS, MU US, NVDA US

Unimicron's US customers carry the tariff risk

Taoyuan prosecutors questioned 14 Unimicron (3037 TT) employees as suspects on 28 Aug 2026, including the general manager and deputy general manager of its printed circuit board business, over allegations the company shipped China-made boards back to Taiwan and relabeled them as Taiwanese. Nikkei Asia reported the raids on the Guishan headquarters and a Zhongli plant, with four more people questioned as witnesses. Bail ran to NT$15m for the PCB general manager and NT$12m for his deputy.

The exposure does not sit mainly with Unimicron. Since August last year US Customs and Border Protection has applied an additional 40% tariff on goods it determines were transshipped through a third country to avoid duties, and the agency has no authority to waive it. If relabeled boards reached the United States, that penalty lands on the importer of record, which means Unimicron's American customers. The company supplies NVIDIA, Intel (INTC US), Google and Amazon.

Unimicron told the Taiwan Stock Exchange it is cooperating fully and does not expect a material impact on operations, and said separately that the disputed products are boards rather than substrates. Liberty Times reported a second round of searches at the Guishan plant on the morning of 31 Aug.

The speed of the disclosure is itself informative. Prosecutors published names and bail amounts within 24 hours of raiding one of Taiwan's most important listed technology companies, in a week when the White House has Taiwan on its list of 40 territories at elevated risk of illegal transshipment. Taipei is demonstrating that it polices the difference between made in Taiwan and made in China itself.

Linked stocks: 3037 TT, NVDA US, INTC US, 3189 TT

OpenAI bought tens of thousands of Mac minis

OpenAI has bought tens of thousands of Mac mini and Mac Studio desktops and wants more, according to The Information, using them for reinforcement learning and to train agents that operate a computer directly. Anthropic rents Mac minis through Amazon Web Services for similar work. Neither company is buying them as office equipment.

Apple Silicon's unified memory pool is what makes the machines useful, since the processor and graphics units draw on the same memory rather than each holding their own, and a high-end Mac Studio can be configured with far more of it than a comparable card. Desktop thermals also suit sustained load better than a laptop. Apple has started promoting clusters of Mac Studios as a way to run large frontier models.

Apple was not set up for this. The report describes a company with no engineering team dedicated to enterprise customers, no developer relations staff for them and no complete enterprise AI strategy, which held an unusual Business at the Park event in June for executives from Disney, Ford and Anthropic. Mac revenue rose about 29% YoY last quarter to roughly US$10.4bn, making it Apple's fastest-growing line.

The memory shortage then arrived on top of the demand. Lead times on some high-end configurations stretched to months, Apple conceded supply and demand would take months to rebalance, and some corporate buyers began looking at NVIDIA's DGX Spark instead. Apple moved its Mac refresh forward this year rather than waiting for the usual October window.

Linked stocks: AAPL US, NVDA US, MU US, 005930 KS

Micron's Taiwan union survey backs a strike 80%

A preliminary survey of Micron's (MU US) Taiwan unions returned 80% in favor of striking over performance bonuses, and the same dispute is now running at all three of the largest memory makers. The unions want payouts comparable to those at Samsung Electronics and SK Hynix (000660 KS), whose Korean workers have been paid on the same cycle of record memory profits.

Micron says it will pay its highest performance bonus on record in 2026. The disagreement is about the Incentive Pay Plan, the variable component employees measure against what the Korean makers have distributed rather than against Micron's own history.

The timing matters more than the amounts. Micron has just brought the former AUO plant in Tainan into operation, a site it bought for NT$7.4bn in late 2024 and converted in about six and a half months, at the point where every additional DRAM wafer is being bid for by AI customers. Labor is now one of the few inputs to memory supply that has not already been contracted years forward.

Linked stocks: MU US, 005930 KS, 000660 KS, 2408 TT

MediaTek caps panel packaging savings near 10%

MediaTek has been working on fan-out panel-level packaging for about seven years and took roughly three of them to get a first product into mass production, senior director Hsu Jun-chung told the SEMI panel-level packaging forum on 31 Aug 2026. Yields on existing products now match wafer-level equivalents. The cost saving from the whole exercise is at most around 10%.

The geometry is the appeal. A 300mm wafer is round, so the larger a package gets the more of the wafer edge is wasted, while a 510 by 515mm or 600 by 600mm square panel can be filled and then cut into sub-panels or strips. Effective capacity on a 600 by 600mm panel is several times that of a 300mm wafer. Area alone does not become cost, though. Without uniformity, process control and yield, a larger panel raises both scrap and equipment write-off risk.

AI packages are pushing several difficulties at once. Package area is moving from about 1.2 times reticle size toward more than 10 times, architectures are moving from 2.5D to multi-die and 3D, and power is going from 100W to the 1,000W class while integrating high-bandwidth memory, silicon bridges, substrates and custom memory. Molding and thermal steps make large panels expand, contract and shift dies, and stress differs between the center, the edge and the corners, so alignment has to be compensated zone by zone with feed-forward and feedback control rather than caught at final inspection.

Qualification speed is the harder limit. Advanced packaging currently takes eight to nine months to introduce a new product, and establishing a stable process window on panels is likely to take longer, against AI platforms that turn over faster than that. Hsu said no single company can deliver panel-level packaging alone, which puts chip designers in the position of coordinating foundry, assembly, materials, equipment and design software suppliers themselves.

Linked stocks: 2454 TT, 3711 TT, 6239 TT, 2330 TT

先聲 First Word: Exclusives from Chinese-Language Sources

C Sun's order book is now 92% advanced packaging and advanced boards

C Sun Mfg (2467 TT), a TSMC supply-chain equipment maker founded in 1966, said advanced packaging accounts for 54% of its unfilled orders and advanced printed circuit boards a further 38%. July revenue was NT$1.17bn, up 34.3% MoM and 132% YoY, and the January to July total of NT$6.21bn has already passed all of 2025. An equipment order book is a forward read on where packaging capital is actually being committed, ahead of any capacity announcement. (31 Aug 2026) Source: 台積電供應鏈志聖迎60周年慶　AI核心業務占上半年營收已占70%

Genius Electronic Optical is sampling co-packaged optics components

Genius Electronic Optical (3406 TT), an iPhone lens supplier, reported July net profit of NT$330m on revenue of NT$2.42bn, up 39.3% MoM, and has moved into co-packaged optics parts including V-grooves, fiber ferrules, metalenses, fiber array units, micro lens arrays and molded glass. V-grooves, prisms and ferrules are already at the sampling stage. Taiwan's consumer optics base finding a second customer set in data center interconnect widens who benefits from the shift to optical links. (31 Aug 2026) Source: 玉晶光獲利靚 7月每股賺2.92元 切入 CPO 獲市場肯定

UMC is building silicon photonics on thin-film lithium niobate

UMC (2303 TT) is moving its silicon photonics work onto thin-film lithium niobate and pushing it toward 12-inch production lines, corporate technology R&D vice president Ting Wen-chi said, as optical modules move toward 1.6T and 3.2T. The material choice is a bet that conventional silicon modulators run out of headroom before the industry reaches 400G per lane. Taiwan's second foundry taking a different path to TSMC on photonics gives the optical supply chain a second qualified source. (31 Aug 2026) Source: 聯電矽光子鎖定12吋TFLN　突破「暴力升級」400G單通道難關