Nvidia Sees $1T in AI Orders by 2027, but Most U.S. Investors Can’t Own the Chinese Suppliers Behind 60% of Its Optical Modules

GTC 2026 wrapped up this week and there's an angle here that I think deserves more attention from an investment perspective than it's getting. Jensen Huang projected $1 trillion in combined Blackwell and Vera Rubin orders through 2027, doubling the $500 billion figure from last year. Goldman reiterated Buy immediately. The Vera Rubin platform is a genuine generational jump: the Rubin GPU has 336 billion transistors on TSMC 3nm, 50 petaflops of FP4 per chip, and the full POD reaches 60 exaflops. Inference token costs drop to a tenth of Blackwell. This is the hardware backbone for the next wave of AI capex from Microsoft, Meta, Google, and Amazon, who have collectively committed around $185 billion in orders already. But building these systems requires enormous quantities of 800G and 1.6T optical modules to connect GPUs together inside each rack. TrendForce projects 800G+ optical transceiver shipments jumping from 24 million units in 2025 to roughly 63 million in 2026. And this is where things get interesting for investors: about 60% of Nvidia's 800G optical module orders go to two Chinese companies, InnoLight (Zhongji Innolight, 300308.SZ) and Eoptolink (300502.SZ). Coherent, Lumentum, and Broadcom split the remaining 40%. Eoptolink posted 283% revenue growth in H1 2025, driven almost entirely by 800G volume. Both companies are already shipping 1.6T samples timed for the Vera Rubin ramp in H2 2026. The geopolitical layer makes this even more paradoxical. Nvidia's Q4 earnings showed zero H200 revenue from China. The CFO stated they are not assuming any data center compute revenue from China in Q1 guidance. Production was halted in early March and TSMC capacity redirected to Vera Rubin. Over 400,000 Chinese orders were effectively cancelled. Jensen mentioned at GTC that new export licenses have been secured and production is restarting, but the financial guidance still prices in nothing. At the same time, in the Stratechery interview right after his keynote, Jensen said 50% of the world's AI researchers come from China, called DeepSeek and Qwen "really, really good," and argued that trying to bundle everything into one locked down American stack is "a terrible mistake." BYD and Geely just joined Nvidia's DRIVE Hyperion platform for L4 autonomy. Chinese robotics companies like Unitree and UBTech are among the first adopters of Nvidia's Jetson AGX Thor. On the domestic chip side, Cambricon is targeting 500,000 AI accelerators in 2026, triple its 2025 output, with ByteDance as its biggest customer. So Nvidia can't sell its best chips to China, but Chinese companies are building the physical interconnects that make the trillion dollar buildout possible, adopting Nvidia's autonomy and robotics platforms, and scaling their own AI chip production to fill the gap. From an investment standpoint, the problem is access. If you want exposure to US optical players like Coherent or Lumentum, that's straightforward. But the Chinese side of this supply chain, which is the majority share, is basically invisible to most US listed funds. KWEB is pure internet names. CQQQ is mostly Hong Kong listed with no exposure to these A share companies. I've been digging around and the only US listed ETF I've found that actually captures this angle is CNQQ, which holds InnoLight at roughly 3.5%, Eoptolink at 1.3%, Cambricon at 2.3%, and BYD at 1.9%. All four showed up repeatedly in the GTC announcements and none of them appear in KWEB or CQQQ. Obviously there are real risks here. CPO (co packaged optics) is a medium term threat to pluggable module makers. Nvidia's own Spectrum X Photonics switches with integrated silicon photonics ship in H2 2026 with 3.5x power savings over pluggables. Analysts at Cignal AI say pluggable shipments won't see material CPO impact for at least three years, but the transition is coming. Cambricon's 500K chip target depends on SMIC yields that are currently around 20%. And the entire China exposure carries obvious geopolitical tail risk. But the core thesis is that Nvidia's trillion dollar AI infrastructure buildout has a massive Chinese supply chain dependency that most investors aren't positioned for, and the window where pluggable optical modules dominate (2026 through 2028 at minimum) lines up with the biggest hardware ramp cycle in computing history. Curious what others think about this angle and whether anyone has found other ways to get exposure.