The data point arrived quietly, buried in a social media thread from a semiconductor analyst who calls himself the Photon Stock Guru. On August 9, he revealed he had sold his memory positions months ago and was now deploying capital into Co-Packaged Optics (CPO). The market immediately interpreted this as a bearish signal for memory giants like Micron and Samsung, while light-based infrastructure companies surged. But beneath the surface noise, this rotation tells a deeper story about the evolving bottlenecks in AI compute—and by extension, the blockchain networks that increasingly depend on that compute for validation, ZK-proof generation, and decentralized AI inference.
I have spent the last 19 years watching liquidity flows through the lens of macro technology cycles, and this particular shift feels like a repeat of the early 2020s when GPU shortages dictated the pace of Ethereum mining. The difference today is that the bottleneck is no longer just about memory bandwidth for GPUs—it is about the optical interconnects that stitch together clusters of tens of thousands of accelerators. The market is collectively betting that CPO, still in its infancy, will become the next scarce resource. But as with any rotation, the sold-off asset often carries a hidden asymmetry.
Context: The Memory Bubble and the Quiet Accumulation of Optical Bets
To understand the rotation, one must first map the liquidity terrain. Memory stocks—DRAM, NAND, HBM producers—have enjoyed a spectacular run from 2023 to mid-2025. The AI boom drove HBM3E prices to historic highs, and traditional memory followed on the coattails of data center expansion. But the market is now pricing in a classic cyclical peak: capacity expansions announced by Samsung, SK Hynix, and Micron during the 2024-2025 profit surge will hit the market in 2026-2027, potentially flooding a demand environment that is already showing signs of fatigue in consumer electronics. The "collective bearishness on memory" that the Photon Stock Guru references is not irrational—it is a reflection of the industry's own structural overconfidence.
CPO, by contrast, occupies a different position in the capital expenditure chain. It is a technology that replaces pluggable optical modules with an integrated photonic engine co-packaged directly on the switch ASIC. This reduces power consumption by 30-50% and increases bandwidth density, making it the natural successor to 800G and 1.6T optics in the next generation of AI clusters. The major cloud providers—Microsoft, Google, Meta—are already investing in CPO standards and pilot lines. The Photon Stock Guru's rotation reflects a bet that the penetration rate of CPO will accelerate from near-zero today to double digits by 2027, creating a new wave of high-margin component suppliers.
Core: The Structural Divergence Between Memory and Optics
From a technical architecture perspective, the two sectors are fundamentally different. Memory is a mature, capital-intensive industry where the key competitive advantage lies in process node shrinks (1β nm for DRAM, 200+ layers for NAND) and TSV stacking for HBM. The barriers to entry are astronomical: a single EUV lithography tool costs over $400 million and has a lead time of 12-18 months. The three existing players—Samsung, SK Hynix, Micron—control roughly 95% of the DRAM market, and their pricing power is directly tied to the supply-demand balance of a commodity-like product.
CPO, on the other hand, is a system-level innovation that spans multiple layers of the stack. The switch ASIC (Broadcom's Tomahawk 5 or Marvell's Teralynx) is the brain, fabricated on 5nm or 3nm advanced logic. The optical engine, or photonic integrated circuit (PIC), is typically built on a silicon photonics platform using mature process nodes. The packaging—2.5D CoWoS or EMIB—is the most critical bottleneck, as it requires precise alignment of lasers, fibers, and electronics. This is not a commodity business; it is a bespoke integration challenge where the value accrues to the companies that can solve the yield and reliability puzzles first.
Based on my experience auditing the Ethereum whitepaper and later stress-testing DeFi protocols, I recognize a pattern: the market often underestimates the time required for a complex systems-level technology to mature. CPO is currently at the stage where the first few proofs of concept have worked, but the yield curve for high-volume manufacturing is still steep. The Photon Stock Guru's rotation assumes that the adoption curve will follow a classic S-curve, but the reality of 2.5D packaging—where the CoWoS capacity is already fully booked by NVIDIA and AMD GPUs—suggests a more constrained timeline. The asymmetry is that CPO stocks may trade on hype before the real revenue materializes, while memory stocks may be oversold on cyclical fears that ignore the secular growth of AI-related memory demand.
Contrarian: The Decoupling Thesis That No One Is Talking About
The mainstream narrative today is that memory is a sinking ship and CPO is the lifeboat. But this binary framing ignores a crucial nuance: the two industries are not substitutes; they are sequential bottlenecks. AI compute first hits the memory wall, which HBM solved. Then it hits the network bandwidth wall, which pluggable optics partially solved. Now it hits the power and density wall, which CPO aims to solve. The order of the bottlenecks creates a natural phase shift: memory stocks peaked when the HBM shortage was resolved, and CPO stocks will peak when the first major CPO deployment is announced. But the market always rotates ahead of the fundamentals.
The hidden information here is that the "collective bearishness on memory" may be a self-fulfilling prophecy driven by the very rotation itself. If institutional capital flows out of memory stocks and into CPO stocks, the price action reinforces the narrative, creating a feedback loop that overshoots the fair value of both sectors. The contrarian trade would be to ask: what if the memory cycle is not as over as it appears? The data suggests that HBM demand is still growing at 50%+ year-over-year, and the traditional memory market's weakness is largely driven by consumer electronics, which is a separate demand vector. The market is pricing a uniform decline, but the AI-driven memory segment is structurally different.
Moreover, the CPO sector faces a hidden vulnerability: its reliance on the same advanced packaging capacity (CoWoS) that is already strained. If the cloud capital expenditure cycle slows down—as the market briefly panicked about two weeks ago over a rumored cut—the CPO narrative will collapse faster than memory, because memory has a more diversified end-market. The rotation from memory to CPO is a bet on continued AI infrastructure growth, but it is also a bet that the optical supply chain can scale faster than the memory supply chain. That is a bold assumption given the current yield challenges.
Takeaway: Positioning for the Chaotic Surface
The market is not wrong to rotate; it is wrong to be absolute. The structural integrity of the AI infrastructure thesis remains intact, but the vectors of value creation are shifting. Memory stocks may offer a more resilient entry point after the rotation exhausts itself, while CPO stocks will likely experience violent pullbacks as the technology matures. The question I ask myself, sitting in Milan with a macroeconomic map of global liquidity flows, is this: when the next capital expenditure scare hits, will the market remember that memory is the bedrock of all compute, or will it double down on the shiny new optical toy? The answer will determine the shape of the next cycle.
As I write this, I am reminded of the chaotic surface of 2021, when the NFT mania obscured the underlying value of the Ethereum network. The same pattern is repeating: the surface noise of rotation disguises the deeper structural currents. Those who can read the currents—who understand that CPO and memory are not adversaries but phases of a single, evolving bottleneck—will be the ones who position themselves before the next liquidity wave breaks.