The Optical Module Trap: Goldman Sachs Re-Rates the Physical Layer, but the Code Remains Centralized

The code reveals what the pitch deck conceals. This morning, Goldman Sachs doubled its price target for Zhongji Innolight, a Chinese optical module manufacturer, from 1,187 to 2,581 yuan. The rationale: silicon photonics ramp, scale-up network expansion, and higher-speed transceivers. The report reads like a love letter to the AI infrastructure buildout. But for those of us who audit the soul of decentralized systems, the real story is not the upgrade—it is the failure mode embedded in the physical layer. Smart contracts do not care about your narrative. The narrative, in this case, is that the AI boom is uplifting every link in the compute chain. But the Goldman report, despite its quantitative rigor, suffers from a selective blindness that is dangerously familiar: it treats the supply chain as a black box, ignoring the geopolitical and technical leverage points that can, and will, break the system. Context: The blockchain industry is now in a sideways market, but capital is flowing heavily into AI-crypto hybrids. Decentralized physical infrastructure networks (DePIN), tokenized compute markets, and AI inference platforms all claim to democratize access to compute. Yet their underlying hardware—the GPUs, switches, and optical modules—is produced by a handful of companies vulnerable to export controls and single points of failure. Zhongji Innolight supplies optical transceivers to hyperscalers like Google, Amazon, and Nvidia. Its silicon photonics technology is seen as a cost-efficient path to 800G and 1.6T interconnects. The Goldman thesis: the shift from scale-out to scale-up networking, driven by Nvidia’s DGX GB200 racks, will multiply demand for high-speed optics. This is true. But it is also the trap. Core: Let me stress-test the assumptions. First, the scale-up network argument assumes that the bottleneck is communication within a compute node. That is correct. In a 72-GPU rack, the bandwidth between GPUs is the limiting factor for training. Optical modules are the enablers. But what happens when the optical module vendor itself becomes the bottleneck? Zhongji Innolight relies on imported optical chips and DSPs from U.S. and Taiwanese suppliers. The silicon photonics they are shipping likely uses external lasers and modulators. The chip-level dependency is the same as for every other module maker. The report glosses over this with a single phrase: “silicon photonics shipments growing.” It does not specify the percentage of self-sourced dies versus purchased ones. Based on my audit experience with a decentralized AI dataset marketplace that attempted to use verifiable computation to prevent data poisoning—a project that later folded due to supply chain constraints—I can tell you that the physical layer is the most audited and yet least understood risk in crypto-AI convergence. Second, the report highlights the transition from scale-out to scale-up networks. This implies a shift in the addressable market from traditional data center interconnects to high-density rack interconnects. But scale-up networks are proprietary. Nvidia’s NVLink, InfiniBand, and the upcoming Spectrum-X are closed ecosystems. The optical modules must be validated and qualified by Nvidia. Zhongji Innolight’s position is strong today, but Nvidia has every incentive to diversify suppliers and even bring interconnect design in-house. The “customer concentration risk” is not just a footnote; it is the entire thesis. If Nvidia decides to favor Coherent or a U.S. domestic supplier for geopolitical reasons, the revenue stream evaporates. The Goldman report does not model this scenario. It assumes linear growth. Third, the valuation jump from 1,187 to 2,581 yuan embeds an implicit assumption about sustained AI capital expenditure. In a sideways market for crypto but a bull market for AI, that assumption might hold. But the crypto world has seen this playbook before: during the 2017 ICO boom, everyone assumed infrastructure spending would compound forever. Neo’s whitepaper promised an “Economy of Value” with a Byzantine Fault Tolerance variant that I analyzed for weeks. The code had vulnerabilities in the consensus mechanism that the marketing team never disclosed. The lesson: theoretical elegance fails under practical stress. The optical module market is no different. The actual failure mode is not the product; it is the dependency chain. If the U.S. expands export controls to cover 800G optical modules or the DSP chips used inside them, Zhongji Innolight loses access to its core inputs. The Chinese government may respond with local alternatives, but the performance gap is still years wide. The report’s optimism is a projection of current momentum, ignoring the fragility of the underlying network. Now let me address the contrarian angle: what the bulls got right. The demand for high-bandwidth, low-latency interconnects is real and growing. The AI compute clusters being built by hyperscalers require orders of magnitude more bandwidth than traditional data centers. Optical module vendors are indeed the “pick and shovel” suppliers of this gold rush. The upgrade cycle from 400G to 800G to 1.6T is accelerating, and each generation commands a higher price. Zhongji Innolight’s execution has been strong—they shipped millions of 800G modules in 2024 and are sampling 1.6T. Their manufacturing scale and yield improvements are legitimate competitive advantages. The silicon photonics ramp could, in theory, reduce dependency on III-V compound semiconductor chips, which are expensive and subject to supply constraints. If the company succeeds in integrating all photonic components onto a CMOS-compatible silicon platform, they could achieve both lower cost and higher reliability. The report’s emphasis on silicon photonics is therefore not misplaced. It is the single most important technical variable for the next three years. But the contrarian view must go deeper. Even if silicon photonics adoption accelerates, the network effect works against decentralization. The optical module industry is consolidating, not fragmenting. The top three players—Coherent, Zhongji Innolight, and Fabrinet—control over 60% of the high-speed market. This concentration contradicts the core promise of blockchain: distributed resilience. Every DePIN project that claims to build a global, decentralized compute network ultimately depends on these few suppliers for its backbone. The physical layer is not permissionless. It is permissioned by chip fabs, export licenses, and qualified vendor lists. The most critical audit you can perform on any crypto-AI project is not of its smart contract, but of its supply chain. And when you do, you will find that the code is often the least vulnerable part. Reproducibility is the highest form of respect. The Goldman thesis is reproducible only if all assumptions hold: no trade war escalation, no Nvidia supplier shift, no technology disruption like co-packaged optics (CPO) that could render pluggable modules obsolete. CPO integrates optical engines directly onto switch ASICs, eliminating the need for separate optical modules. If CPO reaches commercial viability within five years—as many industry roadmaps suggest—the entire market for high-speed pluggable transceivers could shrink. The report does not mention CPO. The silence is telling. Takeaway: Logic is the only currency that never inflates. The Goldman upgrade is a signal, not a truth. It signals that the market is finally recognizing the importance of the physical interconnect layer for AI. But for crypto builders and investors, it should also signal a warning: the infrastructure you rely on is more centralized than the ledger you study. We audited the soul, and it was hollow. The next time you evaluate a DePIN or AI-crypto project, do not stop at the smart contract audit. Trace the supply chain. Ask where the lasers come from. Ask whether the DSP is subject to export controls. Ask if the vendor can survive a geopolitical shock. If the answer is fuzzy, the risk is real. The code may be open, but the physical world is closed. And that is the vulnerability no pitch deck will ever show you. A bug in the contract is a feature in the exploit. The bug here is the assumption that hardware supply chains are elastic and geopolitically neutral. They are not. The exploit is the moment when a single export license revocation freezes the network. That moment is coming. We just do not know when.