High yield is a warning, not a welcome. Goldman Sachs just doubled its price target on Zhongji Innolight, a Chinese optical module maker, from 1,187 to 2,581 RMB. The market cheered. I read the report and saw something else: a confirmation that the AI industry is about to hit the same scaling wall that Ethereum hit in 2021. The problem is not compute. It is interconnect. And the same structural flaws that plagued blockchain’s layer-2 solutions are now repeating in AI networking.
Let me be clear. I am not here to praise Zhongji Innolight. Code does not lie; people do. The report, published on April 2026, centers on three technical trends: silicon photonics shipments ramping, the market shifting from scale-out to scale-up networking, and the upgrade to higher-speed optical transceivers (800G → 1.6T). On the surface, this is a bullish story about a hardware supplier riding the AI capex wave. But forensics don’t stop at the headline.
Context: The Industry Hype Cycle
First, understand the context. For the past two years, the narrative has been GPU-First. Everyone tracked NVIDIA’s H100/B200 shipments, packed flops into clusters, and assumed the network just needed to connect them. That assumption is now dead. The reality is that a 100,000-GPU cluster is only as fast as its slowest link. When you run a trillion-parameter model, intra-rack bandwidth becomes the binding constraint. This is the same dynamic we saw in blockchain: as Ethereum moved from monolithic to modular, the data availability layer became the bottleneck, not the execution layer. AI’s scale-up network is its data availability layer. Investors who missed that analogy are about to get a painful lesson.
Core: A Systematic Teardown
Let me break down the report’s three pillars and expose the hidden asymmetry.
First, silicon photonics. The report highlights that the company is ramping silicon photonics shipments. This is code for “we are reducing reliance on expensive III-V compound semiconductors (like InP) and moving to CMOS-compatible photonics.” Good. But here is the catch: silicon photonics at scale requires mature silicon photonic foundry capacity, which is still heavily concentrated in TSMC and a few US-based fabs. If the US imposes export controls on high-speed optical engines—which they already discussed in 2024—Zhongji Innolight’s silicon photonics ramp could be throttled overnight. The bull case assumes supply chain immunity. The data says otherwise.
Second, the scale-up vs. scale-out shift. The report says the market is expanding from scale-out (connecting many servers) to scale-up (connecting GPUs inside a rack). This is technically correct. The NVIDIA DGX GB200 NVL72 rack requires 36 to 72 800G transceivers per unit for NVLink spine connectivity. But let me ask the question the report conveniently avoids: who owns the NVLink ecosystem? NVIDIA. And NVIDIA is actively backing Coherent as a second source. The risk here is not about technology—it is about counterparty concentration. If NVIDIA decides to pull Zhongji’s share below 30%, the revenue projection collapses. Audit the promise, not the poster.
Third, higher-speed, higher-value products. The report extrapolates that each generation (800G → 1.6T) multiplies unit value 4-5x. This is true, but it ignores a critical feedback loop: as speeds increase, power consumption per port also increases. A 1.6T transceiver can consume 20-30W, meaning a cluster with 100,000 ports adds 2-3 MW of thermal load just for optical modules. That heat must be dissipated, adding to data center costs. The bull case assumes frictionless deployment. The reality is that many cloud operators are already hitting power constraints for AI clusters. The demand for 1.6T may peak earlier than predicted because the total cost of ownership (TCO) becomes prohibitive. This is the same mistake DeFi protocols made in 2020: assuming yield is sustainable without accounting for systemic cost.
Contrarian: What the Bulls Got Right
To be fair, the bulls are not entirely wrong. Zhongji Innolight has demonstrated superior manufacturing yield and customer relationships with Google, Amazon, and Meta. Their 800G market share is real. And the shift to scale-up networking is structurally sound—it is not a fad. The report correctly identifies that AI clusters will require denser interconnects. Where the bulls miss is the asymmetry. The upside case requires all three trends to play out perfectly: no export controls, no customer concentration shifts, no power ceiling. The downside scenario only needs one of those to fail. This is a binary bet disguised as a linear growth story.
I also note a historical pattern. In 2018, I audited a DeFi protocol that promised infinite liquidity through rebasing mechanisms. The yield looked identical to today’s optical module revenue projections: high, predictable, and backed by a leading brand. The protocol collapsed when the assumptions behind the liquidity curve broke. The same logic applies here. If AI capex slows—even by 10% due to regulation or model diminishing returns—the optical module demand curve inverts. The stock price would correct faster than the underlying business. Forensics don’t lie.
Takeaway
The question is not whether Zhongji Innolight is a good company. It is whether the market is pricing in a perfect execution scenario that ignores tail risks. I have seen this movie before. The exit strategy is not a price target; it is a stress test. Until someone models a 20% tariff on Chinese-made 800G modules, or an NVIDIA-driven supply shift, or a power cost spike, the current valuation is a thesis in search of a flaw. High yield is a warning, not a welcome. I am watching, not buying.