The market is selling memory stocks while buying CPO, but the code doesn't lie—the real bottleneck is shifting from storage bandwidth to optical interconnect bandwidth. Over the past seven days, several memory-heavy funds lost 40% of their exposure to DRAM plays, while CPO-linked names gained 15% on speculation alone. This is not a narrative shift. It is a structural reallocation driven by the same logic that governs DeFi exploits: the next attack vector is always the one you underestimated.
Context: The Two Fabrics of AI Infrastructure
Memory chips—DRAM, NAND, HBM—are the working memory of AI. They store model weights, cache activations, and feed data to compute units. The market is dominated by three oligopolists: Samsung, SK Hynix, and Micron. They operate on mature processes with high yields, but their cycle is tied to commodity pricing. HBM, the high-bandwidth stack, is the exception—it is a premium product where SK Hynix holds over 50% share.
Co-packaged optics (CPO) is the opposite. It is a nascent technology that integrates an optical engine directly onto the same substrate as the switch ASIC. This eliminates the electrical bottleneck between the switch and the transceiver. The key players are Broadcom (switch ASIC), TSMC (CoWoS packaging), and a handful of optical module makers like Zhongji Innolight and Coherent. CPO is not yet a product; it is a promise.
Core: The Technical Audit of the Switch
Let me disassemble the thesis. The author claims they sold memory months ago to bet on CPO. Based on my audit experience of semiconductor supply chains, I can trace the logic to three technical layers.
First, memory capacity is approaching a cyclical peak. The storage industry went from a 2023 downturn to a 2024-2025 boom driven by AI demand for HBM. Now, capital expenditure is ramping: Micron is building a fab in New York, SK Hynix is expanding in Yongin, and Samsung is investing in Pyeongtaek. These investments take 12-24 months to come online. The risk is oversupply in 2026-2027, exactly when the market is already pricing in a correction. The 'collective bearishness on memory' is not irrational—it is a reflection of the known cycle.
Second, CPO addresses a different bottleneck. AI clusters are scaling from tens of thousands to hundreds of thousands of GPUs. Electrical interconnects between switches consume power and generate heat. The industry standard is pluggable optical modules, but at 1.6T and beyond, their power density becomes unsustainable. CPO reduces the electrical distance by placing the optical engine millimeters from the switch die. This is a physics problem, not a marketing problem. The energy savings are real, but the yield is not.
Third, the supply chain vulnerabilities are asymmetric. Memory manufacturing depends on EUV lithography from ASML, advanced etching from Tokyo Electron, and high-purity silicon wafers from Shin-Etsu. These are geographically concentrated in Japan, the Netherlands, and the US. Export controls on HBM to China, imposed in 2025, have already reduced addressable demand for Korean memory makers. In contrast, CPO's bottleneck is the CoWoS packaging line, which is dominated by TSMC. CoWoS capacity is currently consumed by NVIDIA and AMD GPU accelerators, leaving little room for CPO pilot lines. The equipment for CPO—die bonders, alignment tools—has a 6-9 month lead time, but the real constraint is the coupling yield between the optical engine and the silicon photonics chip. Current yield estimates are below 60%, which is insufficient for mass deployment.
Contrarian: The Blind Spots in the Collective Narrative
The market is selling memory as if the AI demand is collapsing. That is a misread. The demand for HBM and enterprise SSDs is still growing at 30%+ YoY. The weakness is in consumer DRAM and NAND, where PC and smartphone recovery is tepid. The 'memory crash' is a sector-specific correction, not a systemic collapse. The front-runners inside the block are the ones who understand that the AI storage story is intact—they are just rotating out of the cyclical tail.
On the CPO side, the hype is ahead of the technology. The author's switch from memory to CPO implies a belief that CPO will penetrate rapidly. But the cost per bit of CPO is currently 2-3x higher than pluggable modules. The first adopters will be hyperscalers like Microsoft and Google, but they are also the ones funding the pluggable roadmap. The risk is that CPO becomes a niche solution for the largest clusters, while the rest of the market stays with pluggable optics for another generation. The contrarian angle is that the collective excitement about CPO could be a decoy from the real issue: the memory cycle might be overblown, but the CPO cycle hasn't even started.
Takeaway: The Vulnerability Forecast
Reentrancy is not a bug; it is a feature of greed. The same recursiveness applies here: every cycle in semiconductors is a reentrancy of the same pattern—boom, oversupply, bust, then innovation. The switch from memory to CPO is a bet on the next bottleneck. But the timing is everything. If CPO yields fail to ramp above 80% by 2027, the capital expenditure on AI infrastructure will face a reality check. The best audit is the one you never see—the one that catches the failure before it hits the market. Watch the CoWoS capacity allocation and the yield reports from Broadcom and TSMC. Those numbers will tell you whether the front-runners are already inside the block, or just running in place.
