Goldman Sachs just projected a 65% earnings surge for a fiber optic component maker in 2026, followed by 108% and 119% in the two subsequent years. The numbers are so aggressive they border on the absurd — a triple-digit growth trajectory that would make any DeFi yield curve blush. Zhongji Xuchuang, a Chinese manufacturer of optical modules used in AI data centers, has become the unlikely star of a market that ordinarily chases token narratives. But the market is buying it. And for once, the math might hold.
Let’s strip away the hype and look at the code. These modules — 800G today, 1.6T and 3.2T tomorrow — are the connective tissue of AI training clusters. Every GPU node in a massive parallel compute array talks to every other node through a fabric of lasers and glass. The faster the module, the less time GPUs spend waiting for data. In my years auditing DeFi protocols like Aave v2, I learned that bottlenecks always migrate. In DeFi, it was oracle lag and MEV. In AI, it’s the network pipe. The optical module upgrade cycle is not a luxury; it’s a necessity if anyone wants to run trillion-parameter models.
Logic holds until the ledger bleeds. But here, the ledger is the income statement of every hyperscaler. Amazon, Microsoft, Google, and Meta collectively spent over $200 billion on capex last year, much of it on AI infrastructure. Zhongji Xuchuang sits directly in that flow. It’s the ‘water seller’ of the AI gold rush — no exposure to model risk, no alignment debates, just a physical product that must be replaced every two to three years as bandwidth demands double. The Goldman report assumes that capex continues compounding at a similar rate through 2028. That’s a big if, but not an unreasonable one given the current competitive dynamics in AI.
From a technical perspective, the shift from 800G to 1.6T involves a non-trivial change in modulation schemes, laser density, and thermal management. I’ve seen similar transitions in blockchain consensus mechanisms — from PoW to PoS, from L1 to L2. The engineering always looks straightforward on the whitepaper, but the failure modes are subtle. In optical communications, the challenge is signal integrity at higher baud rates. The industry is moving from EML (electro-absorption modulated laser) to silicon photonics, which trades raw performance for better integration and lower cost. Zhongji Xuchuang has bet heavily on silicon photonics. If the yield curve for their 1.6T silicon photonic modules shows any kinks, the entire profit forecast unravels.
Trust is a variable, not a constant. During the Terra-Luna collapse, I traced the depegging back to a circular dependency in the minting algorithm — stablecoin print runs that assumed infinite demand. This Goldman forecast suffers from a similar circularity: it assumes AI capex grows because big models demand more compute, but big models are funded by the same capex cycles. If any hyperscaler pulls back — say, due to antitrust pressure or a recession — the entire demand curve shifts lower. Optical modules are not like DeFi protocols that can be forked. They are physical goods with 12-week lead times. Inventories can bloat overnight.
Now let’s talk competition. The report barely mentions it. Coherent, New Ease, Huagong Tech — all are chasing the same 1.6T market. The optical module industry has historically seen razor-thin margins during glut periods. The only reason margins are high today is because 800G demand surged faster than supply. That will normalize. I’ve run stress tests on liquidity pools that looked great in bull markets but failed when yields dropped 20%. Same principle here. The ASP uplift that Goldman touts is a fragile premium that evaporates as more fabs come online.
We coded the escape, but forgot the exit. The biggest threat isn’t competition from other module makers; it’s vertical integration by the customers themselves. Microsoft’s Lyra project, Google’s internal optical interconnects, and even Meta’s custom networking teams are all working on proprietary solutions. If any of these reach production scale, Zhongji Xuchuang loses its largest buyers. The risk is asymmetric: a single major customer defection could shave 30% off revenue. In my DeFi audits, I always flagged single-source dependencies. Here, the dependency is on Nvidia’s reference designs and hyperscaler procurement cycles.
That said, the near-term setup is compelling. The market is sideways, with BTC and ETH stuck in chop. Capital is rotating into real assets — equities tied to AI infrastructure. Zhongji Xuchuang’s stock has already doubled this year, but Goldman’s target implies another 160% upside. For crypto investors accustomed to 10x token pumps, that sounds modest. But the risk-adjusted return is far superior: no smart contract risk, no regulation tail risk, no exit scams. Just a manufacturing cycle with a known cadence.
Silence is the only audit that matters. The real test will come in Q3 2024 earnings. If Zhongji Xuchuang reports actual numbers that align with the Goldman model — strong 800G shipments, early 1.6T sampling, stable gross margins — then the narrative is confirmed. If they miss, the whole house of cards wobbles. I’ve seen this pattern before in DeFi: a token with strong fundamentals can trade sideways for months, then explode on a single protocol upgrade announcement. The optical module equivalent is a design win with Nvidia’s Blackwell architecture.
The takeaway here is not to buy the stock blindly. The takeaway is structural: in a market desperate for yield and direction, the safest bets are those with the least narrative fragility. AI infrastructure hardware plays — Zhongji Xuchuang, but also Coherent, Lumentum, and the upstream silicon photonics foundries — offer a kind of yield that doesn’t depend on the next meme coin or the next L2 TPS war. They depend on a single variable: the number of GPUs deployed multiplied by the bandwidth per GPU. That number has only one direction for the next three years.
The algorithm saw the crash, not the pain. We can model the capex curves and simulate the competitive responses. But we cannot model the emotional pendulum of the market. When AI hype fades — and it will, as all narratives do — the optical module makers will still be selling modules, but earnings will reset. The key is to exit before that reset. For now, the math holds. But as I learned from Terra, math only holds until the next black swan.
I’m holding a small position in the sector through a basket of suppliers. Not because I believe Goldman’s forecast to the letter, but because in a sideways market, boring infrastructure is the only thing that pays the bills. The rest is just noise.