A few weeks ago, Goldman Sachs dropped a report that sent shockwaves through the traditional tech investing world: they raised profit forecasts for Zhongji Xuchuang, a Chinese optical module manufacturer, by 65%, 108%, and 119% for 2026 through 2028. The narrative was simple—AI infrastructure is exploding, and the companies that make the cables and connectors between GPUs are the ultimate pick-and-shovel plays. But as a decentralized protocol PM who’s spent years watching centralized supply chains fail under their own weight, I see a different story. This isn't a story about inevitable growth; it's a story about fragility, centralization, and the silent bottleneck that blockchain-based infrastructure is designed to solve.
Let me give you the context. Optical modules are the glass-and-silicon bridges that connect servers in AI training clusters. When you have ten thousand GPUs trying to talk to each other, the network becomes the bottleneck. That’s where 800G and 1.6T modules come in. They convert electrical signals into light pulses and back, enabling the insane data rates needed for large model training. Zhongji Xuchuang is the dominant supplier, especially to Nvidia. Goldman’s bet is that as AI models get bigger, demand for these modules will keep rising, and the company will enjoy higher average selling prices on next-gen products. It’s a textbook bull case.
But here’s where my engineer instinct kicks in. The core insight Goldman misses—and most mainstream analysis ignores—is that the entire optical module supply chain is a house of cards built on a handful of specialized foundries and a single dominant customer: Nvidia. When you look at the numbers, over 60% of Zhongji Xuchuang’s revenue comes from one or two hyperscaler clients. That’s not diversification; that’s a single point of failure wrapped in a glowing profit forecast. In the blockchain world, we call that a “centralized validator set” with a staking dominance that could lead to a fork or collapse. The same principle applies here. If Nvidia decides to vertically integrate its optical interconnects—which it has every incentive to do—Zhongji’s growth story evaporates overnight.
Let’s go deeper. The technical challenge isn’t just about speed; it’s about yield, wavelength stability, and thermal management. Every time you double the data rate, you roughly quadruple the signal integrity complexity. Goldman’s model assumes smooth, linear scaling from 800G to 1.6T to 3.2T. But based on my own experience auditing hardware supply chains for a DePIN project, I can tell you that 1.6T modules are still in early engineering samples, and yield rates are below 40% for some critical components like the EML lasers and the DSP chips from Broadcom. The profit growth Goldman predicts depends entirely on these components becoming cheap and reliable at scale within two years. That’s a bet on physics, not just market demand. And as any engineer knows, physics doesn’t negotiate with financial models.
Moreover, there’s a hidden assumption that the demand for AI compute will keep growing at triple-digit rates for years. This is taken as gospel in the current bull market. Every conference speaker talks about “exponential growth.” But I remember the crypto winter of 2022, when everyone thought DeFi was inevitable, and then total value locked dropped 70% in six months. The froth we see in AI capital expenditure today mirrors the ICO mania of 2017. Back then, I ran workshops in Prague for developers who were confused by the hype. I told them: build for humans, not for nodes. The same applies here. The AI infrastructure buildout is real, but the timeline and the sustainability of that growth are deeply uncertain. If the market realizes that training GPT-5 requires 100 exaflops and a billion dollars in networking gear, and the returns on that investment are diminishing, the optical module orders will slow faster than Goldman can revise their target price.
Now, the contrarian angle that most blockchain natives will appreciate: The very centralization that makes Zhongji Xuchuang a “safe” investment is exactly what decentralized physical infrastructure networks (DePIN) are built to replace. Projects like Helium, Filecoin, and newer compute marketplaces (Akash, io.net) are creating distributed networks of idle compute and bandwidth. Instead of building a massive data center with a hundred thousand optical links, you tap into millions of devices with lower-speed, but far more resilient, peer-to-peer connections. The optical module boom is a direct consequence of the hyperscaler model—big boxes, big networking, big single points of failure. DePIN offers an alternative: small boxes, meshed networking, and economic incentives that align with geographic distribution. Education is the ultimate yield—the more we teach new developers about open, permissionless infrastructure, the less dependent we become on these fragile centralized supply chains.
I saw this first-hand during my work on the “Prague Consensus” workshops. We had 150 developers and operators come together to discuss how to build resilient systems. Many of them later contributed to the Polkadot and Cosmos ecosystems, where inter-chain communication is handled through relay chains and light clients, not 3.2T optical modules. The point is not that optical modules are bad—they’re amazing engineering. The point is that putting all your trust in a single company and a single technology path is a fundamentally non-blockchain way of thinking. We need to apply the same skepticism to AI hardware that we apply to centralized exchanges.
Look at the mental health impact too. During the last bear market, I started a peer support network called “Reclaim” for burnt-out devs. The stress came from two things: volatility and dependency. Devs who built on a single protocol or a single supply chain were the ones who burned out first. The same will happen to the engineers at Zhongji Xuchuang if their main customer shifts strategy. They will be left with 1.6T modules that no one wants to buy, and a workforce specialized in a dying standard. The human cost of centralization is rarely accounted for in Goldman’s spreadsheets.
Finally, let’s talk about the takeaway. The Goldman report is a snapshot of a specific moment in a bull market. It’s useful as a data point, but it’s dangerous as a gospel. For those of us building in the decentralized space, the lesson is clear: don’t bet your future on a company that exists because of a single customer and a single technology. Instead, invest your energy in protocols that diversify risk across geography, ownership, and hardware. In a world where AI models are becoming commodities, the infrastructure that supports them must be open and resilient. Build for humans, not just nodes. Because when the market turns—and it always does—the optical illusion will shatter, and only the truly decentralized will survive.
