Goldman Sachs has turned its spotlight on China's AI hardware sector, identifying a set of local stocks that could ride the wave of export-driven growth. The report, flagged by Crypto Briefing, signals a pivot from domestic substitution to global supply chain integration. But beneath the bullish headline lies a complex web of technical dependencies, regulatory choke points, and structural fragility that demands a forensic audit.
Context
The investment bank's thesis is straightforward: China's AI hardware exports—spanning optical modules, AI servers, and networking equipment—are becoming indispensable to the global AI infrastructure buildout. With North American cloud giants spending over $200 billion in combined CapEx for 2024, Chinese manufacturers have embedded themselves deep into the supply chain. The shift from "assembled in China" to "engineered in China" is no longer aspirational; it's a matter of fact. Optical module players like Zhongji Innolight and Eoptolink now command over 50% of the 800G market. AI server ODM leaders—Foxconn, Wistron, Quanta—produce roughly 35-40% of global AI server units, much of it from mainland factories.
Yet the report's real edge is its timing. In a bull market where euphoria often masks technical flaws, Goldman is essentially re-pricing Chinese tech assets under a new narrative: from "catching-up victim of sanctions" to "unavoidable manufacturing partner." This is a classic re-rating trigger. But the code must be audited, not the pitch.
Core: Systemic Teardown of the China AI Hardware Export Thesis
Let's dissect the three layers of this export stack: optical modules, AI servers, and cooling/power infrastructure.
1. Optical Modules: The High-Margin Crown Jewel
Zhongji Innolight reported Q3 2024 gross margins of 33-35% and net margins above 20%. Order visibility extends into H2 2025. This is the cleanest play. The 800G to 1.6T upgrade cycle ensures pricing power and volume growth. However, the dependency on US-based DSP and laser chips (Coherent, Lumentum) introduces a single-point vulnerability. A BIS license denial for these components could freeze production lines. The sector's technical elegance masks a structural fragility: the high-value components are still imported.
2. AI Servers: Volume, Not Value
Industrial Foxconn's AI server revenue surged over 200% YoY in H1 2024, yet gross margins remained at a meager 8%. This is the "smile curve" in action—assembly yields low margins, while chip design and branding capture the profit. Goldman's report likely includes these ODM names, but investors must distinguish between "revenue growth" and "margin expansion." The latter remains elusive. Moreover, the server assembly line is rapidly migrating to Southeast Asia (Malaysia, Thailand, Vietnam) to circumvent tariffs. This geographic shift may dilute the "China exposure" that Goldman is pricing.
3. Cooling and Power: The Next Bottleneck
Single AI data center power draw has jumped from 50MW to 200MW+. Liquid cooling (cold plate and immersion) is now a necessity. Chinese vendors like Envicool and Gaolan are first-tier globally. This segment is less exposed to export controls—cooling is a physical, not silicon, product. But it is also lower margin (15-20%) and highly competitive. The bull case for cooling is a derivative of the cloud CapEx cycle, not a direct AI hardware play.
Contrarian: Where the Bulls Got It Right (and Wrong)
The bulls are correct that China's manufacturing ecosystem is difficult to bypass. The US would face a 15-30% cost increase and 6-12 month delays if it fully decouples. This is a genuine moat. However, the report's implicit assumption—that the current CapEx boom is sustainable—is the weakest link. AI investment is cyclical. If the Application Layer fails to monetize, cloud providers will trim spending. A 20% CapEx cut would cascade through the supply chain, turning today's high-beta winners into volatility traps.
Furthermore, the regulatory landscape is shifting. China's own export controls on AI chips (announced in 2024) and the US's latest rule (February 2025) expanding licensing requirements create a double squeeze. Goldman's "export-driven" narrative may be pricing in a smooth path, but the code of geopolitical risk is not visible in the P&L. Trust no one, verify everything.
Takeaway
Goldman Sachs has provided a useful signal, but it is a signal, not a guarantee. The optical module segment offers the most defensible thesis, but the server and cooling plays are commodity-style bets on a single variable: cloud CapEx persistence. Investors should audit the code—the granular supply chain dependencies, the margin profiles, the regulatory filings—not the pitch. The bull market rewards narrative, but the bear market settles in technical debt. Do your own math, not your own fear.