Hook
On a gray Tuesday morning in Frankfurt, I received a Bloomberg terminal alert: Macquarie had named a "top pick" in China’s AI chip sector. The stock—unnamed in the brief, but universally assumed to be SMIC or HiSilicon’s listed proxy—jumped 8% before the Shanghai composite opened. My first instinct was not to chase momentum, but to audit the narrative. After years auditing Ethereum multisigs and watching DeFi protocols implode on faulty assumptions, I’ve learned that the most dangerous investment theses are those that conflate government intent with technical reality.
Macquarie’s report reportedly cited "policy-driven demand" and "export-control resilience" as catalysts. But underneath that surface lies a seven-dimensional fault line: technology, supply chain, capacity, demand, geopolitics, competition, and valuation. Each dimension carries a score, and the sum tells a story far more complicated than a simple "top pick."
Context
China’s AI chip industry is at a paradox. It is simultaneously the most protected market on earth (government procurement mandates, national AI supercomputing centers) and the most constrained (EUV embargo, DUV licensing, EDA restrictions, ARM architecture limits). The total addressable market for domestic AI chips in China is projected at $800–$1000 billion by 2027 (including servers), but that figure depends on an aggressive assumption: that local fabs can deliver 7nm-equivalent capacity at scale, that chiplet packaging can mask the 2.5-node gap with TSMC, and that the software ecosystem (CANN, PaddlePaddle) can meaningfully erode CUDA’s lock-in.
Macquarie’s pick likely falls into one of two categories: a foundry play (SMIC, with its N+2 capacity and government-backed capex) or a design-house play (HiSilicon surrogate through Huawei’s ecosystem, or Haiguang Information via x86 licensing). Each has a fundamentally different risk profile. Foundries are capital-intensive, low-margin, and exposed to equipment denial. Design houses are IP-constrained, customer-concentrated, and vulnerable to CSP self-silicon disruption. To evaluate which—if either—deserves the “top pick” label, we must decompose the industry along seven axes.
Core: Seven-Dimensional Analysis
1. Technology & Process (Score: 4/10)
China’s leading AI chips—Huawei’s Ascend 910B, Cambricon’s Siyuan 590—are fabricated on 7nm FinFET (SMIC’s N+2 node). That places them roughly 2.5 generations behind TSMC’s 3nm GAA, a gap of about 3 years. Transistor architecture remains FinFET; GAA is not yet on SMIC’s roadmap. Yield rates for N+2 are estimated at 50–60% per third-party supply chain analysis, compared to >90% for TSMC 7nm. That yield gap translates into wafer costs 50–70% higher, compressing gross margins for both foundry and design.
Unlike TSMC, which uses EUV for 7nm+ nodes, SMIC relies exclusively on DUV with multiple patterning. That increases layer count, decreases yield, and raises cost. The next node—5nm-equivalent N+3—is under development but reliant on DUV, with estimated mass production no earlier than 2026. China is compensating through chiplet architectures (Huawei’s Ascend 910C uses 2.5D interposers) and advanced packaging (an area where domestic fabs JCET and Tongfu have reached CoWoS-S 2018-level capability). But 3D SoIC and hybrid bonding remain inaccessible due to equipment restrictions from TEL and Suss MicroTec.
IP sovereignty is a mixed picture. Huawei has a perpetual ARMv8 license (but no v9); Haiguang holds an x86 license (Zen1 architecture, no further upgrades); Cambricon and Horizon use proprietary or RISC-V architectures. The software stack—Huawei’s CANN, Baidu’s PaddlePaddle—is improving rapidly, but CUDA’s dominance means that even with parity in raw compute, migration costs remain prohibitive for most enterprise workloads.
2. Supply Chain (Score: 3/10)
China’s AI chip supply chain is hyper-concentrated and fragile. Upstream, ASML DUV lithography (NXT:1980i series) requires Dutch export licenses, and actual deliveries in 2024 fell 30% short of committed volumes. Japanese materials suppliers (TOK, JSR) have cut high-end photoresist shipments by ~40%. EDA tools from Synopsys, Cadence, and Siemens EDA are restricted for advanced nodes. Domestically, Shanghai Micro Electronics Equipment (SMEE) has not yet delivered a 28nm immersion DUV; its 90nm tool is insufficient for AI chips.
Downstream, customer concentration is extreme. For Cambricon, the top five customers account for >80% of revenue, with the single largest (a government data center operator) representing 45%. Haiguang’s top five account for ~70%. These customers are price-sensitive and politically motivated, which suppresses ASPs and extends receivable cycles to 6–12 months.
Supply chain vulnerability is rated HIGH. A full DUV ban (including the 1980i series) could halt advanced fab expansion within 6 months, given that critical etch/dep tools from Lam, TEL, and AMAT still dominate ~70% of installed capacity. Chinese domestic alternatives (AMEC for dielectric etch, Naura for thin film) cover only mature nodes.
3. Capacity & Capex (Score: 5/10)
SMIC’s overall fab utilization was 70–75% in Q4 2024, dragged by mature-node oversupply, but its N+2 lines are essentially full. Planned expansions include a $8.8 billion 12-inch fab in Lingang (28nm+, target 100k wafers/month by 2025), and a rumored Huawei-backed N+2 dedicated line ($12 billion, 30k wafers/month, equipment move-in H1 2025). SMIC’s capex-to-revenue ratio is 60–70%, well above TSMC’s 35–45%, reflecting the enormous cost of building capacity under sanctions.
Depreciation is a hammer. SMIC uses 5–7 year straight-line depreciation, meaning new fabs will depress overall gross margins by 5–8 percentage points (from ~18% to potentially sub-10%). Cash-flow break-even requires >85% utilization and stable pricing, but mature-node pricing is competitive and advanced-node pricing is politically capped. The implied risk: investors are underwriting SMIC for strategic value, not profit generation.
4. Market Demand (Score: 8/10)
Demand is the strongest pillar. Government and state-owned enterprise AI server procurement accounts for 50–60% of domestic AI chip revenue, growing at 30–40% CAGR. Internet giants (Baidu, ByteDance, Alibaba) contribute another 20–30%, though they are increasingly self-siliconing. Autonomous driving and edge computing add 10–15% at 40%+ growth.
Ascend 910B shipments reached an estimated 300,000–400,000 units in 2024, with expectations to double in 2025. However, supply is the bottleneck: SMIC’s 7nm-class capacity is fully utilized, and packaging capacity from JCET for 2.5D interposers is only ~10,000 wafers/month, causing lead times of 6 months or more.
Inventory levels are elevated due to strategic hoarding. Channel stock-to-sales ratios are 3–4 months vs. a healthy 2 months, but this is “strategic stocking” rather than demand inflation. Normalization is expected by Q3 2025 unless export controls tighten further.
5. Geopolitics & Export Controls (Score: 9/10 – higher = more risk)
This is the most critical dimension. The US “Foreign Direct Product Rule” captures any chip, equipment, or software that uses US technology, with a presumption of denial for China-linked entities. The Netherlands expanded DUV controls in September 2024 to cover all immersion lithography. Japan’s July 2023 restrictions cover 23 equipment categories, including high-end photoresist.
China’s countermeasures (gallium, germanium export controls) have limited impact on AI chip production because GaN/GaAs are not used in logic chips. The National Fund III (¥344 billion) and local matching funds totaling >¥500 billion provide capital, but they cannot substitute for blocked equipment.
The decoupling risk is HIGH. Base case (2025–2027): China remains stuck at 7nm, using chiplet and packaging to close the gap to ~NVIDIA A100 performance. Pessimistic case (full DUV ban): process regresses to 14nm, widening the gap to 5+ years. Note: Macquarie’s thesis implicitly assumes the base case and relies on continued export controls to sustain the “domestic substitution” premium.
6. Competitive Landscape (Score: 6/10)
In the global AI training market, Chinese firms hold <1% share. In China’s domestic market, NVIDIA (via H20 and L40S compliant variants) still holds ~50% share, with Huawei at ~15%. The inference market is more fragmented, with Horizon, Cambricon, and local CSPs competing.
R&D spending is asymmetrical. NVIDIA spent $19 billion in 2024; Huawei’s HiSilicon spent an estimated $3 billion; Cambricon spent $250 million. Chinese firms benefit from lower engineer salaries (1/3 of US), but the absolute gap is overwhelming.
New entrants include CSP self-silicon (Baidu Kunlun, Alibaba Yitian), which threaten to displace third-party design houses. The real moat for domestic AI chip companies is not hardware leadership but software stack lock-in with domestic frameworks (CANN + MindSpore) and “Xinchuang” certification—a regulatory barrier that could vanish overnight if import restrictions ease.
The five-force analysis reveals intense rivalry, strong buyer power (government procurement), strong supplier power (monopoly equipment/IP), high substitute threat (NVIDIA via compliance channels, CSP self-silicon), and medium threat of new entrants. The market is high-risk, low-margin, and dependent on policy continuity.
7. Valuation & Financial Health (Score: 4/10)
Gross margins are suppressed: Cambricon 30–35%, Haiguang 45–50%, SMIC 15–20%. All are significantly below NVIDIA (70%+) and TSMC (55%+). R&D is fully expensed, which depresses reported profits but accurately reflects cash spend.
Operating cash flow is weak: Haiguang’s OCF/net income ratio is 0.67, Cambricon’s is negative. Free cash flow is negative for all but Haiguang. Companies depend on equity financing.
Valuations are extreme: Haiguang trades at ~80x trailing PE; Cambricon at ~25x PS (with negative earnings). Even SMIC’s EV/EBITDA of ~25x is a premium to TSMC’s ~15x. The only justification is a strategic scarcity premium—investors are pricing in a future where China’s AI chip market reaches $800–1000 billion by 2027 and local vendors capture 60%+. If the actual market is only $500 billion, stocks would need to halve.
Contrarian Angle: The Self-Destruction of Government Dependency
Macquarie’s “top pick” thesis is built on a foundation that most analysts overlook: the possibility that excessive government support may actually destroy long-term competitiveness. When 80% of revenue comes from state buyers who prioritize “domestic content” over performance, companies have little incentive to innovate on cost or capability. They become rent-seekers, not global competitors.
Consider the precedent: China’s solar panel industry. After massive state subsidies, it became the world’s largest manufacturer—but also suffered a decade of negative industry margins, serial bankruptcies, and trade wars. The pattern is repeating in AI chips. The current top-down demand creates artificial scarcity, but once the government super-cycle ends (around 2027 when most Eastern Data-West Computing projects are complete), the industry will face a demand cliff. By then, NVIDIA will have moved to 1nm-class chips, and Chinese vendors will be stuck at 7nm with no export market to absorb excess capacity.
Moreover, the “software stack” moat is fragile. Huawei’s CANN is impressive, but it’s a walled garden. If ByteDance or Alibaba scale their own chip ecosystems, they will not use CANN. The fragmentation of domestic AI software could actually accelerate desertion by enterprise developers, who will prefer CUDA emulation or direct NVIDIA access if regulations loosen.
Macquarie’s pick may be the strongest horse in a weak stable, but the stable itself is vulnerable. The real contrarian trade is not to buy Chinese AI chips, but to sell them into the government-procurement rally—because the government, like any single-entity counterparty, eventually tightens its budget.
Takeaway
The market believes China’s AI chip industry is being built from bedrock; I see a skyscraper on a liquefaction zone. Code has conscience, and so does capital. The “top pick” may double from here on policy tailwinds, but the underlying tensions—technological stagnation, supply chain fragility, and valuation irrelevance—will eventually surface. Trust is the new token, and in China’s AI chip story, the trust is being extended by government decree, not earned through market competition. When that decree fades, so will the thesis.