Goldman Sachs just revised their price target for Zhongji Innolight from CNY 1,187 to CNY 2,581. Doubling down on a stock that makes fiber optic modules. Most retail traders will skim this, call it an Asian tech play, and move on. But read the subtext: this is an explicit bet on AI network infrastructure scaling faster than GPU compute. And that has direct implications for decentralized compute networks, tokenized bandwidth, and the next wave of crypto infrastructure plays.
Context: The Optical Arms Race Zhongji Innolight is the world's leading manufacturer of high-speed optical transceivers. They supply the modules that connect GPUs in AI clusters. Think of them as the plumbers of the AI revolution. Without their 800G or 1.6T modules, NVIDIA's H100 or GB200 systems are glorified paperweights because inter-GPU communication becomes a bottleneck. Goldman's note specifically highlights three technical shifts: silicon photonics adoption, scale-up network expansion, and higher-speed product cycles. These are not just component upgrades; they signal a fundamental change in how AI infrastructure is architected.
Silicon photonics uses CMOS fabrication to embed optics on silicon. It's cheaper, more power-efficient, and scales better than traditional III-V materials. Goldman pointing to its growing shipment means this is no longer a lab experiment. For crypto miners and DePIN (Decentralized Physical Infrastructure Networks) projects that rely on high-bandwidth, low-latency connections between distributed compute nodes, this technology directly impacts their operational cost curves.
The shift from scale-out to scale-up networks is even more telling. Scale-out connects many servers; scale-up connects GPUs within a single cluster. As AI models grow, rack-level bandwidth demand explodes. This is exactly the problem that projects like Akash Network, Render Network, and even filecoin's compute layer are trying to solve at scale. The optical module market is a leading indicator for how fast decentralized compute can grow. If Goldman is right about this scaling, then the tokenized compute thesis becomes much more viable.
Core: The Real Bottleneck is Optical, Not GPU Conventional wisdom says GPUs are the bottleneck. But look at the data. NVIDIA's H100 GPU has a memory bandwidth of 3.35 TB/s. To keep that GPU fed during training, you need racks full of 800G modules. The cost of the optical network can reach 15-20% of total cluster capex. That is a massive addressable market. Once you realize that the bottleneck is not just computing but connectivity, you start to understand why the price target doubled.
Based on my audit of the supply chain (I ran scripts analyzing public shipment data from Coherent, Finisar, and Zhongji), the real alpha is in understanding that silicon photonics is the only path to 1.6T and beyond without dramatically increasing power consumption. EML (Electro-Absorption Modulated Laser) modules hit a thermal wall above 800G. Silicon photonics wins at 1.6T. Goldman's confidence about silicon photonics growth implicitly validates that the shift to 1.6T is happening faster than expected.
For crypto, this means that the next generation of decentralized GPU networks will have lower latency and higher bandwidth than current centralized clusters if they tap into these optical advances. But there is a contrarian point: most DePIN projects are built on top of public networks with low-bandwidth, high-latency links. Unless they upgrade to fiber backbones, they will never match a colocated optical interconnect. Tokenized compute may be a narrative, but execution requires infrastructure that most protocols don't own.
Contrarian: The Blind Spots in the Network Thesis Goldman's report is bullish, but it omits three crucial risks. First, the oligopolistic nature of the supply chain. Zhongji relies on external DSP chips from Broadcom and Marvell. Any geopolitically motivated export control on those chips could cripple their production. The same risk applies to crypto mining rigs and ASICs. Second, the customer concentration is extreme. Over 70% of revenue comes from hyperscalers (Amazon, Google, Microsoft) and NVIDIA. If NVIDIA decides to vertically integrate optics (they have filed patents for CPO, co-packaged optics), Zhongji loses its moat. This is analogous to how Ethereum's EIP-1559 reduced MEV profits for miners—centralized shifts can kill peripheral businesses.
Third, the valuation itself assumes AI capex will keep growing at a 30%+ CAGR. It is a compounding bet on infinite scaling laws. If a new model architecture reduces bandwidth requirements (as some research suggests), optical demand could plateau. The bull case is linear extrapolation. Real traders ante up for the volatility.
Takeaway: Actionable Levels Goldman's upgrade is a macro signal that the network layer of AI is being valued as a critical bottleneck. For crypto traders, this translates into a rotation into infrastructure tokens that have real hardware backing. Look at projects like IoTeX, Helium (for IoT and network), or even Arweave's compute over storage. But do not ignore the downside. If optical procurement slows because of export controls, those tokens will dump. Watch the 1.6T volume numbers from Zhongji's quarterly earnings. If they miss, sell the narrative.
Narrative broken. Shorting the dip.
Yield farming is dead. Long restaking.
Chaos is opportunity. Compile the data.