2Q26 Letter--China’s indigenous semiconductor ecosystem
Dear Investors and Friends,
This quarter’s special topic examines the indigenous semiconductor ecosystem taking shape in China across its four building blocks: logic foundry, memory, fabrication equipment and fabless chip design.
The international semiconductor ecosystem has been one of the most rewarding ponds an equity investor could fish in. Each of its four building blocks has produced a dominant franchise with extraordinary shareholder returns. In logic foundry there is TSMC (Taiwan Semiconductor Manufacturing Company; a foundry is a factory that manufactures the chips other companies design); in memory, SK Hynix (the South Korean memory-chip maker); in wafer-fabrication equipment, ASML (the Dutch maker of chip-manufacturing machines, including the lithography systems that print circuit patterns onto silicon); in fabless design, Nvidia (a “fabless” company designs chips but outsources their manufacture to foundries). These are not merely good companies; they are among the highest-quality businesses in the world, with deep moats, real pricing power, and decades of compounding behind them.
The natural question for us is whether China’s emerging ecosystem can produce domestic equivalents — businesses of similar quality and similar returns.
We have two starting observations. First, on the supply side, China is advanced in chip design but constrained in foundry capacity: Huawei’s LogicFolding and SMIC’s 7nm-class process show genuine design ingenuity. At the same time, China cannot purchase the most advanced lithography equipment, foundry yields are sub-optimal (below 40%), and advanced-node capacity is a small fraction of expected demand. Second, on the demand side, Beijing treats compute as public infrastructure, with at least 80% domestic content required. At Variis, we are mindful that involution and state capture have destroyed returns in other Chinese industries, so our aim is to determine whether we can find durable, well-governed franchises in China’s semiconductor ecosystem as potential investments.
Our work on China’s semi ecosystem serves a second purpose. The same ecosystem that may yield new investments is reshaping the markets of businesses we already own; studying it informs our view of those risks as much as of the opportunities.
Of the companies we examined in depth, two — AMEC, the Shanghai etch and deposition equipment maker, and Montage, the memory-interface chips designer — are progressing through final qualification for our Focus List. The others, including SMIC, Hua Hong, YMTC, CXMT, NAURA, and several fabless AI accelerator designers including Cambricon, are either less promising or downright problematic.
A reminder of how ideas earn their way into our Focus List; this letter is a worked example of that funnel. We reduce a universe of roughly 1,400 MSCI Emerging Markets constituents, plus potentially thousands more non-index names, to a Focus List of about 100 through quantitative screens (on liquidity, ownership and other factors) and a qualitative assessment against our three quality filters: moat, management and stakeholder. Focus List ideas are modelled, written up and peer reviewed before any capital is at risk. Valuation then determines portfolio entry: we hold 20–30 investments, sized on expected return across bear, base and bull scenarios weighted 20/60/20. We do not need to know everything about everything; we do need depth on everything on our Focus List. The work below is what our qualification looks like in practice—including, for most of these names, why we set them aside.
A brief primer on the supply chain
The ecosystem's four segments correspond directly to how chip production works. A foundry ("fab") is the factory that imprints electronic circuitry onto a thin silicon disc, or wafer. Each wafer yields multiple identical chips ("dies"), cut apart once manufacturing is complete; the proportion that functions correctly is the yield. Equipment makers supply the specialised machinery — etching, deposition, lithography — required to build and operate a fab. Fabless design companies own no factories; they design chips and contract foundries such as TSMC or SMIC to manufacture them. Memory is a distinct chip category (DRAM, NAND) produced in dedicated fabs.
Why this ecosystem exists
Two forces explain the China semi ecosystem’s present form. The first is the export-control regime: Huawei’s 2019 Entity List designation, the 2020 block on advanced lithography exports, and the October 2022 and December 2024 US Commerce packages asserting jurisdiction over any tool worldwide containing US technology, joined by the Dutch block on advanced immersion scanners from September 2023. The policy intent was to deny China leading-edge compute; the effect was also to hand every company inside the boundary a protected demand base and a national substitution mandate.
The second force is Chinese state direct capital: the national “Big Fund”, municipal money behind CXMT in Hefei and YMTC in Wuhan, Shanghai’s and Beijing’s state-asset arms behind Hua Hong and NAURA, and Shanghai’s STAR Market as an equity funding channel for loss-making champions. China’s State Council has a policy platform that elevates compute to public infrastructure alongside power and water, with estimates of RMB 7–8 trillion (USD 1.0–1.1 trillion) of investment over 2026–2030 and a floor of at least 80% domestic content — implying around RMB 450–610 billion (USD 62–85 billion) of domestic AI-accelerator procurement a year.
Beijing now mandates localisation on the demand side too: since November 2025, state-funded data centres may not use foreign AI chips, and Micron has been barred from critical infrastructure since 2023. The captive market is no longer a by-product of American policy; it is Chinese policy.
The USD 1.0–1.1 trillion capex plan must be measured against the capacity to build the chips it implies. At card prices of RMB 50,000–70,000 (USD 7,000–9,800), the plan's domestic-accelerator budget of roughly RMB 450–610 billion a year implies purchases of 7.5–10 million accelerators annually. These are fabricated on China's advanced-node capacity — SMIC's 7nm-class lines plus HLMC's nascent 7nm — which our work puts near 17,000 wafers per month in 2026, rising to about 42,000 by 2030. But a big part of that capacity is absorbed by Huawei's mobile SoCs and domestic CPUs. We assume this non-AI base stays flat at roughly 12,000 wafers per month, so the whole 25,000-wafer increment over 2026–2030 flows to AI, lifting accelerator allocation from about 5,000 wafers per month to about 30,000.
Converting wafers to chips takes two steps: candidate dies per wafer, and the share that work (yield). A large accelerator die carries roughly 90 gross candidates; at ~20% yields in 2026 that is about 18 good dies, rising to roughly 31 as yields approach 35% by 2030. So output runs near 1.1 million accelerators in 2026 (5,000 × 12 × 18) and about 11 million in 2030 (30,000 × 12 × 31.5). Against budgeted purchases of 7.5–10 million a year, domestic supply covers barely a tenth of the programme in 2026 and stays far short through 2027–2028, drawing level with demand only at the decade's end. The binding constraint is fab capacity, not budget or demand.
Logic foundry: SMIC and Hua Hong
SMIC and Hua Hong are China’s leading logic semi fabs. SMIC’s gap to the global leader TSMC is one of scale and economics, not only technology. TSMC produces around 500,000 advanced wafers per month against SMIC’s 12–15,000; first-quarter revenue was USD 36 billion against USD 2.5 billion; gross margin 66% against 20%; and TSMC’s expected 2027 capital spending of USD 59 billion compares with USD 8–9 billion at SMIC, even though SMIC is reinvesting over 70% of revenue. SMIC’s engineering prowess is nonetheless impressive: a June teardown of Huawei’s latest smartphone chip showed SMIC’s newest process matching the transistor density of a mature 7nm TSMC node without the advanced lithography tools TSMC uses—though SMIC suffers materially higher costs and lower yields.
Neither SMIC nor Hua Hong qualifies for our Focus List. SMIC sits on the US Treasury list barring American investment, a red flag under our stakeholder filter. Hua Hong is a state-owned enterprise (51.6% held by Shanghai’s municipal state-asset commission) earning a 13% gross margin, and in April received US letters ordering equipment makers to halt tool shipments to facilities linked to its advanced-node path. On business quality, both are leading in China only because policy isolates China; on management, both answer to industrial-policy objectives before shareholder return.
Memory: YMTC and CXMT
YMTC and CXMT are China’s leading memory semiconductor manufacturers. They are competitors of Samsung and Hynix, in which Variis has investments.
CXMT, China’s DRAM leader, filed its IPO prospectus in May. CXMT is a fast-growing business that is gaining share in the global DRAM market. Revenue was RMB 9.1 billion, RMB 24.2 billion and RMB 61.8 billion (USD 1.3, 3.4 and 8.6 billion) across 2023–2025, and first-quarter 2026 revenue of RMB 50.8 billion (USD 7.1 billion), up 719% year on year.
We have two concerns about CXMT. First, SemiAnalysis, a research firm, estimates that CXMT’s cost per bit remains more than 30% above the three global leaders. Its roughly 70% first-quarter operating margin is high because of the global DRAM shortage, not because of any intrinsic advantage. CXMT suffers the same problem that dogs SMIC and Hua Hong—critical semi fab equipment, especially ASML’s advanced lithography tools, is not available to Chinese producers, and this puts them at a cost and product quality disadvantage.
Second, CXMT is very much like a state-owned enterprise even if its share register doesn’t make that transparent. Because CXMT has been a loss-making venture for years until very recently, it lacked capital to build capacity. Given China’s policy driven desire to have a domestic DRAM champion, CXMT was able to build fabs it couldn’t afford using state capital. As a result, two of CXMT’s fab entities are only about 31% owned by the company but consolidated despite majority state ownership.
Our interest in CXMT is not only prospective; as shareholders of Samsung and SK Hynix, we also track it as a risk. SemiAnalysis expects CXMT's share of global high-bandwidth memory (HBM) capacity — the memory used in AI accelerators — to rise from almost nothing to around 12% by 2028, mostly serving domestic customers that US export controls have cut off from the global leaders. CXMT's capacity additions will weigh on global DRAM pricing once the shortage normalises, and its protected home market funds its climb up the HBM learning curve.
YMTC, the NAND champion, is not yet listed, is an SOE, and has been on the US Entity List since 2022. Both CXMT and YMTC do not pass our quality filters: sanctions exposure, state ownership above 30% sitting behind formal declarations of “no controlling shareholder”, and franchises that rest on policy isolation rather than earned advantage.
Equipment: NAURA and AMEC
NAURA is China’s broadest toolmaker and now fifth globally by sales (2025 revenue estimated at RMB 46.8–52 billion, USD 6.7–7.5 billion). It is an SOE, controlled by Beijing’s state-asset commission alongside the Big Fund, and was added to the US Entity List in December 2024 together with multiple subsidiaries.
AMEC is China’s leading etch equipment specialist. Its etch tools run at the most advanced domestic logic lines, it has gained share consistently for five years, and it is expanding into thin-film deposition, contesting a market the Western majors have dominated. It was founded in 2004 by Yin Zhiyao, formerly a senior engineer at Applied Materials and Lam Research, and the founding team still holds meaningful equity alongside state investors—a founder-led business built by returning industry veterans rather than a state platform. 2024 revenue was RMB 9.1 billion (USD 1.3 billion), up 45%, with net profit of RMB 1.6 billion (USD 220 million) despite a deliberate 94% increase in research spending to RMB 2.5 billion (USD 340 million).
One item remains open in final Focus List qualification. On moat, AMEC is investing heavily in more than twenty new tools in new categories as it tries to expand its addressable market beyond etch. If new tools fail validation, enter only low-share second-source positions, or require pricing concessions to win acceptance, the company could grow revenue while generating lower incremental returns. We are working to see if we can gain conviction that AMEC can succeed outside their historical etch stronghold.
Fabless design: five sub-categories
China’s AI-chip demand was estimated at USD 39.5 billion for 2025, and the April 2025 ban on Nvidia’s China-market chip impacted the main foreign supplier; domestic chips are expected to reach 80% of Chinese AI-chip shipments in 2026. Domestic China fabless design, especially in the context of the Nvidia ban and now China’s own policies, is a huge growth area.
We have considered listed and near-listed fabless names across five categories: AI accelerators and GPUs; captive silicon (the in-house chip arms of Huawei, Alibaba, and Baidu, which are not directly investable but are competitively relevant and will eat up most foundry capacity thereby creating a constraint for everyone else); independent CPU architecture; semiconductor IP (reusable, licensable circuit-design building blocks) and ASIC (application-specific integrated circuit, a chip custom-built for one purpose) design services; and memory-adjacent fabless semiconductor businesses.
The AI accelerator businesses we considered include Cambricon, Hygon, Moore Threads, Biren, Enflame (upcoming IPO), and Muxi. For independent CPU architecture, we considered Loongson. For semiconductor IP and ASIC design services, we have done initial work on Verisilicon. And finally, in the memory adjacent space we looked at GigaDevice and Montage. It was Montage that most clearly grabbed our attention.
Montage is the leading global supplier of memory-interface chips — the chips that manage data traffic between a server’s processors and its memory. Founded in Shanghai in 2004 by Howard Yang and Stephen Tai, both veterans of American chip companies, it holds around 45% of its core global market and is tracking roughly USD 1 billion of revenue, a 50% net margin and over 20% return on equity for 2026. Its design capability is proven in open global competition, and its manufacturing is unconstrained by export controls — the chips do not need advanced lithography and are made on mature nodes at TSMC. Demand should grow as AI servers carry more memory per processor, and export controls have closed China’s interconnect market to Western competitors, giving Montage a protected domestic franchise alongside its global one.
Conclusion
We’d be remiss in not considering a world in which the export controls are removed—their existence is so fundamental to the whole China semi ecosystem. It looks far-fetched today, but a growing number of serious people argue the controls are counterproductive: they guarantee a captive market, force-feed local innovation and accelerate exactly the indigenous ecosystem they were meant to prevent, while closing one of the world’s largest end-markets to American and allied companies. Even if export controls were removed, the market would not fully reopen: Beijing’s own procurement mandates now bar foreign chips from state-funded demand regardless of what Washington does. A return of Nvidia, TSMC and the memory leaders to unrestricted competition would not be fatal to businesses with real capability, but it would test how much of today’s price is franchise and how much is government protection — most acutely for the fabless cohort, whose multiples already discount years of uninterrupted domestic dominance.
Finally, we remain deeply grateful for your trust and support. Your partnership enables us to stay focused on what we do best: identifying exceptional businesses and investing with conviction. We look forward to updating you again next quarter.
Thank you for your continued interest and support!
Leila, Eko, Rufus and Jamie
Disclaimer
FOR PROFESSIONAL INVESTORS AND ADVISORS ONLY
The contents of this document are communicated by, and the property of, Variis Partners LLP. The information and opinions contained in this document are subject to updating and verification and may be subject to amendment. No representation, warranty, or undertaking, express or limited, is given as to the accuracy or completeness of the information or opinions contained in this document by Variis Partners LLP or its directors. No liability is accepted by such persons for the accuracy or completeness of any information or opinions. As such, no reliance may be placed for any purpose on the information and opinions contained in this document. The information contained in this document is strictly confidential and is not intended to be advice or an offer or solicitation to invest. The value of investments and any income generated may go down as well as up and is not guaranteed.