Hook
Wall Street just broke the silence with a whisper that sounds like a siren. Goldman Sachs—the same institution that once dismissed Bitcoin as 'retail's folly'—is now training its microscope on Chinese AI hardware. Not code. Not algorithms. The physical stuff: the metal, the silicon, the optical fibers that make the models breathe. The message is clear: the AI race isn't just about who builds the best model—it's about who builds the box that runs it. And right now, that box is made in China.
I caught the signal early this morning from a Crypto Briefing flash. It was short, barely a paragraph. But the subtext was deafening. Goldman Sachs has identified a set of Chinese stocks that could ride the wave of AI hardware exports. The analysts call it a shift toward 'export-driven growth' that might 'significantly boost' A-shares. The markets are already twitching. But the real story isn't the ticker—it's the tectonic shift underneath.
Context
Let me rewind the tape. The narrative around Chinese AI has been stuck in a loop for years: 'catching up,' 'restricted by sanctions,' 'domestic substitution.' But the hardware export story flips the script. China isn't just trying to build its own AI chips—it's already the world's factory for the infrastructure that makes AI possible. Think servers, optical modules, liquid cooling, power supplies. The components that NVIDIA and AWS buy in bulk are assembled, tested, and shipped from factories in Shenzhen and Suzhou.
I've been watching this from the crypto side for a decade. Back in the ASIC mining boom, I saw how Chinese manufacturers dominated the hardware supply chain for Bitcoin. The same pattern is emerging in AI, but the scale is orders of magnitude larger. Goldman's report is a validation that the world's largest investment bank sees this as a multi-year trend, not a quarterly spike.
But here's the kicker: Goldman used the term 'AI hardware' instead of 'AI chips.' That's a deliberate choice. Chips—especially the high-end GPUs from NVIDIA and AMD—are still largely off-limits to Chinese design houses due to US export controls. But hardware is broader. It includes the servers, the networking gear, the cooling systems. And in those segments, Chinese companies hold formidable market share. According to public data, Chinese firms supply over 50% of the world's high-speed optical modules (800G and beyond) and roughly 35-40% of AI server assembly. The 'hardware' label is a signal that Goldman is looking at the entire ecosystem, not just the bottleneck.
Core
Now let's drill into the numbers. The core of Goldman's thesis rests on three pillars: optical modules, server assembly, and the 'pick-and-shovel' suppliers like liquid cooling and PCB.
First, optical modules. Companies like Zhongji Innolight (mid-forecast) and Eoptolink have become the backbone of AI data center connectivity. Their 800G modules are shipping in volume to the hyperscalers—Microsoft, Google, Amazon, Meta. Margins are healthy: gross margins around 33-35%, net margins above 20%. Order visibility extends into late 2025. This isn't a speculative bet; it's a revenue stream that's already flowing. The switch to 1.6T modules in 2026 could be another leg up.
Second, server assembly. Foxconn Industrial Internet (FII) is the poster child here. Its AI server revenue grew over 200% year-on-year in the first half of 2024. But the margins tell a different story: single-digit gross margins, around 8-10%. This is the classic 'smile curve' of manufacturing—high volume, low margin. The profit is in the design and the components, not the assembly. But the volume is so massive that even thin margins add up to billions in profit.
Third, the 'hidden' layer: cooling, power, and packaging. Liquid cooling providers like Envicool and Gaolan are seeing demand explode as data center power densities climb from 50MW to 200MW+ per facility. These companies are now exporting to Southeast Asia and the Middle East. The supply chain is diversifying, but the core expertise remains in China.
Based on my experience auditing hardware supply chains for crypto mining operations, I can tell you that the bottlenecks in AI hardware are eerily similar to what we saw in the ASIC boom. The key is lead time: if you need a 200MW data center built, you're waiting 6-12 months for electrical transformers alone. Chinese manufacturers have the capacity to compress those timelines. That's a moat that's hard to replicate.
But here's the original insight most analysts miss: the AI hardware export story is not a 'China vs. US' narrative. It's a 'China embeds itself deeper into the global AI supply chain' narrative. The exports are going to the US, to Southeast Asia, to the Middle East. The US is buying Chinese-made servers even as it restricts Chinese chip access. That's a paradox built on specialization: the US designs the chips; China assembles the boxes. Goldman's report is essentially betting that this co-dependency persists.
Contrarian
Now let me flip the lens. The consensus is that this export boom is a golden ticket. But the contrarian angle is that it's a fragile house of cards—and the cards are all made of the same thing: speculative capital expenditure from US hyperscalers.
Look at the data: the 'Big Four' cloud providers (Microsoft, Google, Amazon, Meta) are expected to spend over $200 billion in capex in 2024, up 40% year-on-year. A significant portion goes to AI infrastructure. But what happens when the ROI on AI doesn't materialize? We've seen this movie before. In 2022, when crypto winter hit, the ASIC mining hardware market collapsed overnight. Orders were canceled, inventory piled up, and Chinese manufacturers were left holding the bag. The AI hardware market is larger, but the same dynamics apply. If the hyperscalers blink, the export orders will dry up faster than the supply chain can adjust.
Goldman's report, as a sell-side product, has a natural bias: it wants to create a narrative that attracts capital. The 'export-driven growth' story is compelling, but it conveniently ignores the risk of a capex cliff. The analysts likely assume that AI capex stays elevated for 3-5 years. But history shows that tech investment cycles are shorter than that. The dot-com bubble lasted about 4 years. The AI capex cycle might be even more compressed.
There's another blind spot: the 'overhype of the data availability layer' in crypto has a parallel here. Just as 99% of rollups don't generate enough data to need dedicated DA, 99% of AI hardware export narratives ignore the fragility of the capex cycle. The real story is that export growth is a derivative of US cloud spending. If that spending slows, the entire thesis collapses.
And let's not forget the regulatory risk. The US Department of Commerce is already eyeing the 'loophole' of Chinese-made servers. If the BIS (Bureau of Industry and Security) expands restrictions to cover server assembly or optical modules, the supply chain could be severed. The 'Hackers don't hack, they listen' principle applies here: the best way to disrupt the supply chain is to listen to the whispers of regulators. The whispers are getting louder.
Finally, the ethical dimension. AI hardware is a dual-use technology. The same servers that power chatbots can power surveillance systems. China's exports to the Middle East and Southeast Asia are already raising eyebrows. If the technology is used for mass surveillance or military applications, the reputational risk could spook Western investors. Goldman's report may have glossed over this, but it's a real overhang.
Takeaway
So where does this leave us? Goldman's signal is real, but it's not a buy signal—it's a risk signal. The hardware export story is a high-beta play on the AI capex cycle. If you believe the cycle has legs, then Chinese optical module and server suppliers are a screaming buy. But if you think the AI bubble is about to deflate, this is the trade to short.
The merge wasn't just about Ethereum—it was about the hardware that powered it. The same is true for AI. The next two quarters will tell us whether the hyperscalers are building for the long haul or just for the quarterly earnings call. Watch the capex guidance like a hawk. The answer is in the numbers, not the headlines.
Code is law, but hardware is the judge. The verdict is still out.
