There is a moment in every technology cycle when the narrative becomes so seductive that even the sharpest analysts forget to look at the bones. I had that moment last week while stress-testing a network topology for a decentralized compute protocol. The simulation kept failing at a single node: the optical interconnect layer. Not the laser. Not the fiber. The vendor lock-in that made the entire system brittle.
Then I saw the Goldman Sachs report on Zhongji Xuchuang. 65% profit growth in 2026. 108% in 2027. 119% in 2028. The kind of numbers that make you wonder whether the numbers are real or whether they are a carefully constructed narrative designed to sell a story. Based on my 2017 audit of Gnosis Safe—where I discovered 12 critical logic flaws in a multi-signature wallet that everyone trusted—I have learned that whenever a system is presented as inevitable and flawless, the truth is almost always buried in the assumptions we refuse to question.
Let me be clear: this is not a hit piece on Zhongji Xuchuang. The company is executing brilliantly. But as someone who has spent years bridging economics and code, I see patterns that the market euphoria is ignoring. This is the story of how the AI optical module boom might be repeating the same centralization mistakes that crypto tried to escape.
Context: The Architecture of Dependence
Goldman Sachs raised its profit forecast for Zhongji Xuchuang based on a single thesis: AI infrastructure expansion is relentless, and optical modules are the bottleneck. Faster modules mean faster training. Faster training means more GPUs utilized. More GPUs utilized means more revenue for hyperscalers. And Zhongji Xuchuang, as the leading supplier of 800G and soon 1.6T modules, is positioned as the "optical foundry" equivalent to TSMC.
This analogy is attractive but dangerous. TSMC manufactures chips for many customers, but the design is controlled by the customer. Zhongji Xuchuang designs its own modules, meaning it has control over the critical layer that connects every GPU in a cluster. If that layer fails—due to a design flaw, supply chain disruption, or geopolitical tension—the entire training job stops. That is not a foundry. That is a single point of failure dressed up as efficiency.
The report specifically highlights "silicon photonics" and "higher average selling prices for 1.6T/3.2T modules." The technology is real. The demand is real. But the narrative that this will translate into three years of triple-digit profit growth is built on assumptions that look a lot like the assumptions that fueled the 2017 ICO mania: a belief that the trend is linear, that there are no competing forces, and that the market will reward the first mover forever.
I have seen this before. In 2020, when Compound's governance token crashed, I watched my savings evaporate because the model assumed liquidity would always flow. In 2021, when NFT floor prices collapsed, I saw projects that had minted on hype alone vanish. The common thread is that markets always overestimate the durability of early advantages.
Core: The Technical Brittleness of Speed
Let me walk you through the hidden technical vulnerabilities that the Goldman report does not mention. I will do this the way I audit code: by looking at the edge cases.
First, signal integrity. As you push optical modules to 1.6T and beyond, the electrical interface between the module and the switch chip becomes a nightmare. We are already hitting the limits of PAM4 modulation at 800G. At 1.6T, we will need either coherent optics or more complex DSPs. Coherent optics are more expensive and consume more power. More complex DSPs increase latency. The industry is betting that these problems will be solved by engineering ingenuity. I believe they will be solved, but not on the timeline the market expects. Every six-month delay in 1.6T volume production directly undermines the growth assumptions in the Goldman model.
Second, thermal management. Optical modules produce heat. A single 1.6T module dissipates around 20 watts. A cluster with 10,000 modules—which is what a 100,000-GPU training cluster would need—represents 200 kilowatts of heat that must be removed just from the connectors. In my experience auditing DeFi protocols, the same pattern appears: everyone models the upside, but no one models the cooling bill.
Third, and most important, the switching bottleneck. The industry is focused on module speed, but the network is only as fast as its switches. Broadcom’s Tomahawk 5 supports 51.2 Tbps, but that bandwidth must be divided among ports. At 1.6T per port, you get only 32 ports per switch. A large training cluster requires hundreds of switches, and the latency of hop-by-hop communication grows faster than the speed of individual links. The real bottleneck is not the module; it is the protocol layer and the switching fabric. No amount of optical speed can fix a topology problem.
I remember a conversation in 2022 with a friend who had lost everything in the Terra-Luna collapse. He said, "We assumed the algorithmic stablecoin was a done deal. We never asked what happens if the oracle fails." The same question applies here: what happens if the switch fails? The optical module is the oracle of the physical network. We are betting our AI infrastructure on its uninterrupted performance.
Contrarian: The Three Assumptions That Will Break
Here is where I challenge the consensus. The Goldman report is not wrong in its direction, but it is fragile in its magnitude. Let me name three assumptions that, if tested, could turn those 119% growth numbers into a mirage.
First, the assumption that hyperscalers will not vertically integrate. Look at Microsoft's Lyra project. Google's own optical interconnects. Amazon's efforts with silicon photonics. Every major cloud provider is investing in self-designed optics. Why? Because they realize that relying on a single external supplier for a component that determines cluster performance is a risk they cannot afford. When self-designed modules reach volume production—likely within two years—the demand for third-party modules will flatten. This is exactly what happened to the GPU mining industry: when ASICs became accessible, GPU mining died. Optical modules are the ASICs of AI infrastructure.
Second, the assumption that capital expenditure remains linear. The market is discounting a 40% to 50% year-over-year growth in AI CapEx for the next five years. But history shows that technology investments come in waves. The 2024 wave was Phase 1: build the clusters. Phase 2 will be utilization and optimization. Once clusters reach capacity, the need for more modules does not disappear, but it decelerates. A deceleration from 50% growth to 20% growth destroys the compound interest that makes 119% profit growth possible.
Third, the assumption that the technology path is stable. The industry is betting on linear scaling of optics. But what if the next AI architecture uses neuromorphic chips that require local memory instead of distant GPUs? What if CXL-based memory pooling reduces the need for long-haul optical links? These are not science fiction; they are active research areas. When the connection pattern changes, the module count changes.
I have a personal framework I call the "blob saturation rule." In blockchain, I argued that after the Dencun upgrade, blob data would saturate within two years, causing rollup gas fees to double. The same logic applies to optical bandwidth: the number of bits you can send per fiber is finite. We are approaching that limit faster than anyone admits. The industry is treating 1.6T as the endpoint. In reality, it is the middle of a logistic curve that flattens.
The Ethical Dimension: What We Build and What We Ignore
Let me step back. Why does this matter beyond portfolio decisions? Because the way we build infrastructure reflects our values. Crypto taught me that trust is not a given; it must be earned and verifiable. The AI industry is currently building the most powerful computational infrastructure in human history on a foundation of proprietary, opaque hardware supplied by a handful of companies. There is no code to audit. No public roadmap. No recourse if a single vendor fails.
In 2020, I wrote about "The Psychology of Impermanent Loss." I interviewed 30 retail users who had lost money in DeFi crashes. The common thread was not poor strategy; it was a misplaced belief that the system was designed to protect them. The same misplaced belief is at work here. Analysts believe the system will continue to deliver because it has delivered so far. But past performance is not a guarantee of future returns—especially when the system itself is changing.
I founded Verifiable Truth in 2026 to use zero-knowledge proofs to audit AI training data. The motivation was simple: if we cannot trust the inputs, we cannot trust the outputs. The same principle applies to hardware. If we cannot trust the optical interconnect—because it is a single point of failure—then we cannot trust the AI model that depends on it. The ethical imperative is to decentralize the stack, not just the software.
Takeaway: Follow the Fear, Not the Chart
Goldman Sachs is not wrong about the opportunity. Zhongji Xuchuang is a great company operating in a high-demand market. But the specific numbers—65%, 108%, 119%—are the product of a financial model that assumes the world stays the same. The world never stays the same.
If you are an investor, ask yourself: what happens if hyperscalers self-source? What happens if the switching bottleneck becomes the new wall? What happens if AI CapEx decelerates by even 10 percentage points? The margin of safety is thin.
If you are a builder, ask a different question: how do we make this infrastructure resilient to single-vendor failure? How do we design networks that can route around a failed module, like a routed network routes around a failed node? The answer is in protocols, not hardware.
If you can see past the optical illusion—if you can follow the fear instead of the chart—you will realize that the real opportunity is not in riding the wave, but in building the decentralized alternative.
The future of AI does not depend on faster optics. It depends on trust that the optics will work, will remain independent, and will be accountable. And that is a problem no financial forecast can solve.