Hook
Macquarie Bank just signaled its top pick in Chinese AI chips. The name remains unconfirmed—but the data points are clear. State-directed demand is the only demand. Government procurement drives 50-60% of revenue for these companies. That is not a market. That is a policy channel.
Context
China's AI chip sector is a tale of two realities. On one side, explosive domestic demand: government "East-West Computing" projects, state-owned enterprise AI server deployments, and a forced migration from NVIDIA hardware. On the other, a brutal technical ceiling: 7nm production via SMIC's N+2 process, with yields estimated at 50-60%—far below TSMC's 90%. The gap is 2.5 process nodes, equivalent to 3-4 years. Chiplet stacking and 2.5D packaging attempt to bridge the gap, but the software ecosystem remains CUDA-bound.
Macquarie's pick likely falls into one of two camps: a foundry like SMIC (capital expenditure heavy, strategic premium) or a design house like Hygon (x86 license, government lock-in). The investment thesis hinges on one variable: continued US export restrictions. Relax them, and these stocks face a Davis double-kill.
Core Analysis (Order Flow & Structural Risk)
Let me route this trade through my own risk framework. I run statistical arbitrage models for a living. I don't chase narratives. I trace liquidity.
First, revenue concentration. Top customers for these chip firms account for 70-80% of sales. That is counterparty risk on steroids. One policy shift—say, Beijing pivoting to support CSP self-developed chips (Alibaba's Yitian, Baidu's Kunlun)—and the revenue stream dries up. The order book is not a market signal. It is a directive.
Second, capital expenditure intensity. SMIC's capex-to-revenue ratio sits at 60-70%, versus TSMC's 35-45%. That means every dollar of revenue requires almost two dollars of spending. The depreciation drag will compress gross margins from 15-20% down to single digits once new fabs ramp. The only reason SMIC survives is that the Chinese state views it as a national security asset, not a profit center. Investors banking on margin expansion are reading the wrong playbook.
Third, the technology roadblocks. US export controls on DUV lithography (ASML NXT:1980i) and Japanese restrictions on photoresist materials have created a bottleneck that no amount of policy money can immediately fix. SMIC's N+2 yield is stuck at 50-60%. Without EUV, 5nm is effectively unreachable before 2027 at the earliest. Chinese AI chips today match NVIDIA's A100 (2020 vintage). They are 2.5 generations behind H100. The gap is not closing—it is being papered over by Chiplet integration and government weighting of "domestic equivalence" in procurement scoring.
Contrarian Angle: The Valuation Disconnect
Conventional wisdom says Chinese AI chip stocks trade at a "strategic premium." I see a different pattern: they trade on narrative beta, not fundamental alpha.
Hygon's trailing P/E exceeds 80x. SMIC's EV/EBITDA is 25x versus TSMC's 15x. These are not growth premiums—they are insurance premiums against decoupling. But insurance premiums are paid for protection against a low-probability event. Decoupling is already happening. The question is whether the market has priced in the risk that decoupling stops or reverses.
Consider: If a new US administration in 2025 relaxes export controls (even modestly), Chinese AI chip designers face immediate competition from NVIDIA's compliant products like the H20. Domestic market share could drop from 30% to 15% within two quarters. That is a 50% revenue shock. Current valuations assume perpetual policy support. They do not account for change.
Retail investors see "China's answer to NVIDIA." Smart money sees a single-customer, single-policy, single-foundry dependent derivative. Counterparty risk is the largest unhedged exposure in this trade.
Takeaway
I am not short these names. I am not long either. The position I run is a volatility sell—collecting premium on the VIX-like swings these stocks produce. Macquarie's pick may outperform over a 12-month horizon if export controls tighten further. But the asymmetry is clear: upside capped by technical ceilings, downside exposed by policy reversals.
Liquidity vanishes. Lessons remain. Calculate. Execute. Repeat.
Data over drama. Numbers don't lie, narratives do. Exit strategy is the only strategy.