Hook
Look at the numbers. Goldman Sachs just raised its profit forecast for Zhongji Xuchuang – a Chinese optical module manufacturer – by an average of 97% annually for fiscal years 2026, 2027, and 2028. That’s not a projection. That’s a narrative. A story about AI infrastructure that can never slow down. The target price implies a 163.6% upside from current levels. But I’ve seen this pattern before. In 2017, I spent 120 hours auditing the Groth16 proof verification logic in Zcash’s private Discord. The silence in the circuit constraints was louder than any noise. What Goldman is selling is an extrapolation of growth that assumes no side-channel failure. No governance fracture. No hidden bug in the supply chain. “Following the ghost in the side-channel shadows.”
Context
Zhongji Xuchuang, traded on the Shenzhen Stock Exchange (300308), is the world’s largest supplier of high-speed optical modules – the fiber-optic transceivers that connect GPU clusters in massive AI training runs. Its core product today is the 800G module, but the narrative is all about the next generation: 1.6T and eventually 3.2T silicon-photonics-based modules. Goldman’s bull case rests on three pillars: (1) silicon photonics volume shipments accelerating, (2) higher average selling prices (ASP) for 1.6T modules as they replace 800G, and (3) a compounding effect of AI capital expenditure growth that does not plateau. This is the same structure as the “supercycle” narrative we saw in DeFi during 2021 – Curve Wars, token emissions, liquidity as a political construct. Only now the asset class is not a governance token but a piece of hardware with real production lines. The market has crowned Zhongji Xuchuang the “TSMC of optical communication.” But I remember hearing similar words in the DAO governance debates: “This token is the Google of decentralized identity.”
Core
Let’s dissect the narrative mechanics. Goldman’s model assumes that AI-related capex by hyperscalers (Microsoft, Google, Amazon, Meta, and Nvidia) grows at a compound annual rate of 40%+ through 2028. From my experience analyzing the Lido stETH decoupling in 2022 – where I built a Python simulation to stress-test the protocol against a 40% ETH price drop – I know that stress scenarios rarely follow linear paths. What Goldman is doing is taking the current demand signal from Nvidia’s GPU shipments and extrapolating it exponentially. The hidden variable is network topology. A Fat-Tree architecture for a 100,000-GPU cluster requires roughly 12,000 optical modules per rack for 800G. But the next-generation Dragonfly+ topology reduces that by 30% to 50% because it uses fewer long-haul links. If the hyperscalers shift to more efficient topologies, the demand for modules per GPU could drop even as total compute grows. The sentiment data from chatrooms and institutional flow trackers shows a massive consensus that “optical is the most deterministic winner.” That’s exactly when I get nervous. My Curve Wars thesis – that liquidity is a political construct, not a mathematical function – applies here. The optical module supply chain is controlled by a handful of companies, and the concentration risk is extreme. Zhongji Xuchuang is the primary supplier to Nvidia for 1.6T, but if Nvidia decides to dual-source (as it did with memory vendors), the ASP premium collapses. The pre-mortem deduction: assume failure first. How does this narrative break? The most likely vector is not demand destruction but ASP compression. In 2021, the “stablecoin hegemony” narrative broke when CRV whale concentration triggered a liquidity crisis. Here, the fragility lies in the commoditization of high-speed optics. If a rival – say, Coherent or a new silicon-photons startup – achieves cost parity at 1.6T before 2027, Goldman’s margin assumptions dissolve. I’ve mapped this type of regulatory arbitrage before. In 2024, I spent 200 hours cross-referencing SEC no-action letters with CFTC commodity definitions to map the legal gray zone of spot Bitcoin ETFs. That dossier revealed that institutionalization often neuters the ideological core. Similarly, the optical module narrative obscures that the real bottleneck is not the fiber but the electrical die-to-module interface – a silicon engineering problem that moves slowly.
Let me quantify. Goldman’s implicit model suggests that Zhongji Xuchuang’s revenue will grow from approximately $4 billion in 2024 to over $20 billion by 2028. That implies it captures 40% of the entire optical module market by volume at ASPs that barely decline. Historically, optical module ASPs have decreased 15% per year per generation. If that holds, the only way to offset is unit volume growth of 100%+ per year – requiring the global annual production of 1.6T modules to reach 30 million units by 2028. That’s roughly 20 times the current 800G production. To put that in perspective, the entire global semiconductor back-end capacity for optical packaging is currently sufficient for about 5 million units. Building the rest requires capital expenditure of $8-10 billion and 3-4 years of construction. That’s the side-channel signal. The silence in the order books from upstream photonics component suppliers (like AAOI, Lumentum) tells a different story. Their capacity expansion plans are cautious. The narrative of infinite growth is clashing with the physical reality of fab construction cycles. “Auditing the fragility of synthetic stability.”
Contrarian
The contrarian angle is not to argue that AI infrastructure will collapse – I’m not a doomer. It’s that the narrative is already priced in for the first two years of growth, leaving zero room for error. The market discounts the risk that hyperscalers will self-manufacture optical engines – Microsoft’s Lyra, Google’s optics-in-PIC – which would decouple their growth from Zhongji Xuchuang’s. My work on the Zcash side-channel debate taught me that trust in a single supplier is the first thing to break under stress. The real blind spot is that Goldman’s forecast assumes a “steady-state” of AI training demand, ignoring the possibility that inference-on-the-edge (e.g., on-device AI agents) shifts the networking bottleneck from fat pipes to edge aggregation nodes. If AI moves from centralized training to federated inference, the optical module demand profile flattens. I’ve been tracing this vector in my current research on sovereign AI identity. In 2026, I designed a framework where AI agents use zero-knowledge proofs to prove competence without revealing weights. That architecture requires machine-to-machine trust, not high-bandwidth interconnects. The narrative contagion from “training scale is everything” to “inference efficiency is everything” is already visible in the research community. Goldman’s model is 18 months behind the data. “Tracing the vector of narrative contagion.”
Takeaway
The optical module narrative is not wrong – it’s fragile. My advice: follow the side-channels. Watch the upstream photonics component orders, the capital goods import data, and the hiring plans at optical packaging foundries. When the silence breaks – when a single capacity delay or ASP miss signals a fracture – the crowd will reprice in days. “Decoding the silence between the blocks.” I’ll be reading the logs.