Zero One Investment Research Daily Intelligence Brief, September 7, 2026

DeepSeek picks 160,000 Huawei chips over NVIDIA; TSMC's 3nm overtakes 5nm

Top Market Signals

DeepSeek buys 160,000 Huawei chips over NVIDIA

DeepSeek is reported to have ordered 160,000 Huawei Ascend 950DT accelerators for a data center in Inner Mongolia, per a single Chinese-language report, worth about US$2.56bn at the chip's roughly RMB111,000 market price, and chose them over NVIDIA's (NVDA US) China-market H20.

The price gap is not what decided it. A 950DT runs about US$16,000 against US$15,000 to US$25,000 for an H20. Founder Liang Wenfeng told an investor meeting in July that Huawei's rack-scale systems can do everything NVIDIA's GB300 can at the same latency, and that the cost of that is four Huawei chips per NVIDIA GPU and a two-year technology lag. DeepSeek is accepting both.

The 950DT is built on Huawei's Da Vinci v5 architecture and is thought to be made on SMIC's (981 HK) N+2 process or an improved N+3 version. Each chip carries 144GB of Huawei's own HiZQ 2.0 high-bandwidth memory at 4.0 TB/s, with 2.0 TB/s of interconnect and native support for the FP8 and FP4 low-precision formats. Huawei links up to 8,192 of them in its Atlas 950 SuperPoD, rated at one exaflop of FP8 compute.

Delivery is the limit. Huawei builds only a few hundred thousand 950DT units a year, so a single order of this size could take more than a year to fill, and the chips are going to inference work rather than training. Inner Mongolia is drawing this build for its power economics, with Alibaba's data center at Ulanqab in the same region running at about RMB0.32 to RMB0.35 per kilowatt hour, and ByteDance is in talks for a further 5GW to 6GW of capacity in the province.

Linked stocks: NVDA US, 981 HK, AMD US

Analysts see TSMC 3nm passing 5nm in 2H26

Analysts cited by Economic Daily News estimate TSMC (2330 TT) will book more than NT$400bn of 3nm output value in 3Q26, a record for the node and over 30% of company revenue, with the 3nm family passing 5nm for the first time in 2H26.

The capacity came from a decision management had already flagged. Chairman C.C. Wei said TSMC historically did not add capacity once a process reached its target output, but expanded 3nm anyway on AI demand. A new fab takes two to three years to build, so the company is converting 5nm equipment across to 3nm and running 7nm, 5nm and 3nm as a flexible pool. Capacity is still tight.

TSMC expects 3nm gross margin to exceed the company average in 2H26, so the node about to become its largest is running above the blended margin. First-year 2nm output is also scaling faster than the company expected.

A16 is due to be production-ready in 2H26, A14 enters mass production in 2028, and both A13 and A12, the second-generation backside power rail node, are planned for 2029. Industry reports suggest the first A16 customers include Apple (AAPL US) and NVIDIA, with OpenAI's Broadcom-designed accelerator also placing orders, though none of that is confirmed by TSMC.

Linked stocks: 2330 TT, AAPL US, NVDA US, AVGO US

Taiwan fab-construction backlog reaches a record NT$881bn

Taiwan's five largest semiconductor facility-engineering contractors are holding more than NT$880bn of signed orders between them, a record, and the work is booked out to 2028 and beyond. The five are 6139 TT, 2404 TT, 6196 TT, 5536 TT and 6691 TT.

The book is heavily concentrated in one contractor. 6139 TT alone holds NT$440.1bn, half the group total, and its chairman said the company has taken as much as NT$600bn of turnkey contracts in Singapore over the past four years, with more expected. 2404 TT holds NT$193.9bn, also a record, on continued fab building by TSMC and Micron (MU US).

6196 TT's backlog reached NT$135.1bn with visibility to at least 2028, and its chairman said customer projects for 2029 and 2030 are already in planning. She added that advanced packaging equipment for CoWoS (chip-on-wafer-on-substrate) is the busiest line this year, and that the company has moved into CoPoS, the panel-level successor, alongside it. 5536 TT holds over NT$60bn, 63% of it semiconductor work, and 6691 TT NT$51.8bn with visibility to the end of 2027.

The demand behind the backlog is TSMC's own statement at SEMICON Taiwan that it has 20 fabs under construction and still cannot meet customer demand, plus a renewed White House threat of tariffs of up to 200% on chipmakers that do not build in the United States. Analysts cited by Economic Daily News expect the combined book to pass NT$1tn.

Linked stocks: 6139 TT, 2404 TT, 6196 TT, 2330 TT

Amazon lifts 2026 capex to US$220bn

Amazon (AMZN US) has raised its 2026 capital spending plan to US$220bn from US$200bn, an increase of a tenth, and its Taiwanese rack suppliers are shipping into that raise now.

AWS's AI and in-house silicon business has reached US$25bn of revenue and is still growing at triple-digit percentage rates. Chief executive Andy Jassy said the shortage of AI compute runs into 2027 and that he can already see 2028 demand, which he described as staggering. Annapurna Labs, the AWS chip design unit, has most of its custom accelerators, or ASICs (application-specific integrated circuits), built on TSMC 3nm.

Alchip (3661 TT) said 3nm AI accelerator demand is running better than it had assessed and that it may raise its 2027 shipment and revenue estimates. It expects record quarterly revenue and profit this quarter and a further increase in 4Q26, with tape-out on current projects due by year end and momentum running at least three to four years on work already in hand.

Wiwynn (6669 TT) assembles the AWS accelerator racks and supplies the switch trays, while Foxconn (2317 TT) supplies CPU trays. Analysts expect Wiwynn's 2026 revenue to grow about 40% to a record. Foxconn's rotating chief executive said the group made new progress on custom accelerator projects this year, including new customers, and targets more than 40% share of that server market.

Linked stocks: 3661 TT, 6669 TT, 2317 TT, AMZN US

Samsung gains 12 points of AI memory share

Samsung Electronics (005930 KS) lifted its share of the high-bandwidth memory (HBM) market by 12 percentage points in a single quarter to 33%, closing on SK Hynix (000660 KS), which still leads with 50%.

HBM is the memory stacked next to AI accelerators, and it is the part of the memory market SK Hynix has led through this cycle. DigiTimes, citing Korean outlets ET News and Seoul Economic Daily and a market research firm, reported the share move alongside the rise of China's CXMT.

A 12 point share gain in one quarter is large for a product where customer qualification runs long, and it would leave AI memory supply less concentrated in a single vendor. The report gives quarterly share only and does not break out the volumes or pricing underneath it, so how much of the move is new qualification and how much is shipment timing is not visible from this data.

Linked stocks: 005930 KS, 000660 KS, MU US

Phison buys cheap flash as it dismisses glut

Phison Electronics (8299 TT) is buying up low-priced NAND flash inventory from Chinese module makers while its chief executive argues the market is misreading what comes next in supply.

Pua Khein-Seng said he recently met two NAND makers and both told him they do not know how large the 2027 shortfall will be. He said the assumption that capacity expansion by the makers will produce oversupply is too naive. Phison expects AI-related products to reach 40% of its mix by the end of this year.

Phison is a controller and module company, so accumulating cheap flash puts its own balance sheet behind the view that supply stays tight into 2027. The makers' new capacity arriving ahead of the demand they are describing is what would change that.

Linked stocks: 8299 TT, MU US, 000660 KS

先聲 First Word: Exclusives from Chinese-Language Sources

Taipower opens a sandbox for TSMC's surplus green power

Taiwan's state utility launches its third regulatory sandbox in October to address green power matching, where the renewable electricity a chipmaker buys does not line up with when it actually uses it. TSMC has committed to 60% renewable power by 2030 and 100% by 2040, so that target now depends on matching as well as on supply. (7 Sep 2026) Source: （獨家）台積電「餘電」問題有解？　台電沙盒3.0計畫10月上路

Anthropic's Claude enters Samsung's chip design flow

Samsung Electronics is testing Anthropic's Claude models inside its semiconductor design business, according to Korean media reports. Generative AI is extending into chip design, among the most specialized engineering workflows in the industry, and Anthropic is moving on Korean enterprise adoption while it does. (6 Sep 2026) Source: DigiTimes

SK Hynix denies an eSSD controller deal with LG

SK Hynix has denied discussing handing part of its next-generation enterprise SSD controller backend physical design to LG Electronics, talks Korean media reported had been running since July 2026. The denial sets back LG's ambition to build a Korean equivalent of Broadcom in custom silicon design. (7 Sep 2026) Source: 樂金「韓版博通」起步受挫　SK海力士否認討論eSSD控制器委託