On August 9, a widely-followed photon-stock account posted the kind of thesis that moves markets: "I sold all memory stocks months ago. I am rotating into co-packaged optics. Memory is cyclical. CPO is exponential." The market cheered. The logic felt clean. But that logic is built on a misread of the stack. CPO does not replace memory. It depends on it. And the memory cycle is not where the market thinks it is.
I spent nearly two decades tracing failure modes in infrastructure. I audited 0x protocol in 2017 and found integer overflow bugs that would have drained exchange balances. I spent three months simulating Curve Finance's constant-product curves to map liquidity fragmentation. I reverse-engineered the Terra/Luna collapse to find the exact point where the peg becomes irreversible. Every one of those systems looked simple from outside. Every one of them hid complexity in the abstraction layer.
Memory and co-packaged optics are no different. The market sees two narratives: an aging cyclical business and a shiny new silicon-photonics future. That is not a thesis. That is a chartist's ghost. Let's reverse the stack to find the original intent.
Context: Two Layers, One Machine
Memory and CPO are not competing technologies. They are complementary layers of the same AI infrastructure machine. Memory โ DRAM, NAND, and especially HBM โ feeds the compute engine. CPO โ co-packaged optics โ connects compute engines to each other. One solves the storage wall. The other solves the bandwidth wall. Saying you are rotating from memory to CPO is like saying you are abandoning concrete foundations to buy glass curtain walls. The building needs both.
The source of the rotation is real. DRAM and NAND prices spiked through 2024 and early 2025. HBM became the most constrained commodity in tech. Sam-sung, SK Hynix, and Micron all ran at full utilization, with HBM lines above 95% utilization. Then, by mid-2025, spot prices for conventional DRAM and NAND started to soften. Contract prices wobbled. The market smelled a peak. Meanwhile, CPO โ co-packaged optics โ is still in pilot production. The yield is not great. The packaging ecosystem is immature. But every hyperscaler โ Microsoft, Google, Meta โ is funding CPO projects. Broadcom and TSMC are jointly developing CPO lines. The narrative writes itself: memory is over, optics are next.
That narrative is false. Not because CPO is a bad bet, but because the rotation is built on a low-confidence reading of both cycles. Let me show you why.
Core: The Code-Level Analysis of Two Cycles
1. Process and Yield: The Hidden Asymmetry
Memory is a mature process. DRAM is at 1-beta nanometer nodes. NAND is past 200 layers. HBM3E uses TSV stacking and is in mass production. Yield is high, and the learning curve is thirty years old. CPO is still fighting basic physics. Coupling a photonic engine to an exchange ASIC on the same substrate means aligning sub-micron waveguides, managing thermal mismatch, and controlling loss. The industry does not have a yield number for CPO that would support a widescale deployment. My estimate โ based on open literature and conversations with packaging engineers โ puts CPO yield at 70% or less in early production. That is not a replacement technology. That is a research project with a capex slide deck.
Here is the overlooked detail: the exchange ASIC in a CPO system uses 5nm or 3nm logic. The photonic integrated circuit uses mature silicon-photonics processes. The packaging โ CoWoS or EMIB โ is the real bottleneck. And that packaging capacity is currently consumed by NVIDIA's GPU orders. If you want CPO, you need CoWoS. If you need CoWoS, you are waiting in the same queue as every AI accelerator. The queue is not getting shorter. That is infrastructure-crunch, not this-that. The memory side has its own packaging challenge โ HBM also uses CoWoS. But HBM is already at scale. CPO is not.
2. Supply Chain: Who Actually Owns the Value?
Memory is a vertical oligopoly. DRAM: Samsung 40%, SK Hynix 30%, Micron 25%. NAND: Samsung 35%, SK Hynix 20%, Kioxia/Western Digital 30%, with YMTC around 5%. HBM: SK Hynix above 50%, Samsung 35%, Micron 15%. These are IDMs. They own the fab, the process, the test, and the pricing power. The upstream equipment and materials are equally concentrated โ ASML for EUV, Sumco/Shin-Etsu for wafers, AMAT for deposition and etch. This is a classic walled garden. The supply chain is secure for those inside the garden.
CPO is a fragmented mess. The exchange ASIC is dominated by Broadcom, with Marvell and NVIDIA circling. The silicon photonics engine is split between Intel, Broadcom, GlobalFoundries, and a few Taiwanese and Chinese fabs. The packaging is TSMC, Amkor, ASE, and JCET. The modules are built by a different set of players โ Zhongji Innolight, Eoptolink, Coherent, AAOI. This is a nightmare for coordination. In a memory shortage, the memory vendor can raise prices and allocate supply with a phone call. In a CPO shortage, no single vendor controls the bottleneck. There is no verifiable code. There is no unified order book. That is a market with a governance failure built into its architecture.
What does this mean for investors? The CPO narrative assumes that value will flow to the most differentiable component โ the optics engine or the packaging. But the actual value capture is ambiguous. Broadcom, as the ASIC leader, will take a large cut. TSMC, as the packaging leader, will take another cut. The optical engine vendors are fighting for scraps. The module makers โ who are the Chinese champions โ are pushed down the stack. If you are buying a CPO stock, ask yourself: are you buying a landlord or a tenant? The answer changes the price you should pay.
3. Capex and Capacity: The Boom-Bust Playbook
Memory capex is the most reliable cyclical signal in semiconductors. In 2024 and 2025, memory vendors generated record free cash flow. They responded like every oligopolist: they invested. Micron built a fab in New York. SK Hynix expanded HBM capacity in Yongin. Samsung scaled Pyeongtaek. Industry capex intensity reached 30-40% of revenue. That is the classic top-of-cycle signal. The market is right to be wary. New fabs take 18-24 months from move-in to full production. Those fabs come online with heavy depreciation, dragging gross margins by 5-10 percentage points. If demand does not grow at the same exponential rate, supply will outpace demand by 2027. That is the deterministic failure mode I map out when I audit protocols. The mechanism is not if a collapse; the mechanism is when.
But here is the twist: the market is also wrong to be uniformly bearish. The 2024-2025 storage upcycle was not driven by consumer PCs or phones. It was driven by AI servers consuming HBM and enterprise SSDs. That demand is still growing. HBM4 is in development. NVIDIA's next-generation accelerators will need more memory per GPU, not less. The total content of memory per AI server is rising. The bear case for memory assumes that AI capex will slow. It assumes that the cloud consumers โ Microsoft, Google, Meta, Amazon โ will cut their combined $300 billion in annual capex. That is possible, but it is not the base case. The current market sentiment is pricing memory as if the AI cycle is over. That means either memory is mispriced downward, or the entire AI trade is overpriced. You cannot have both.
CPO capex is lighter. The capital intensity is much lower because the heavy capital sits in TSMC and the foundries. The module and engine vendors are asset-light design houses. That is appealing in a bear market. But light assets do not mean zero risk. CPO companies are spending heavily on R&D โ Broadcom and Marvell both spend over 20% of revenue on R&D. The yield ramp is uncertain. And the production capacity for CoWoS is not magically expanding just because CPO wants it. The same old economics of scarcity apply.
4. Demand and Inventory: Reading the Signals
The source article correctly notes that memory entered a period of inventory concern. Spot prices for DRAM and NAND started to fall in mid-2025. The market interprets this as the beginning of a downcycle. But there is a more precise reading. The conventional memory market โ DDR4, older NAND โ is still suffering from weak consumer demand. The AI-driven segments โ HBM, high-end eSSD โ are still tight. The divergence is the real story. The market is treating memory as a single commodity when it is actually two separate markets. HBM prices remain sticky. Conventional memory is the one facing oversupply. When the market says "memory is over," it is really saying "consumer memory is over." That is a different thesis, and it does not justify a wholesale exit from the entire sector.
This reminds me of the Curve Finance analysis I did in 2020. Stablecoin pools looked uniform from the front-end, but liquidity was fragmented. The slippage vector was different in every pair. The same is true here. HBM is the USDT pool. Consumer DRAM is the USDC pool. They are both memory, but they do not behave the same. Failing to separate them means you are trading a blended average that does not exist in reality. Truth is not consensus; truth is verifiable code.
5. Geopolitics: The Unseen Root Cause
The rotation out of memory has a geopolitical shadow. In 2025, the United States expanded export controls to restrict HBM shipments to China. This is not a small footnote. China is a massive consumer of AI accelerators, and HBM is a critical component. The controls force Chinese AI companies to either stockpile before the deadline, or accept stranded compute. Western memory vendors โ Samsung, SK Hynix, Micron โ lose a substantial addressable market. That is a real demand shock. But the market is misreading the effect. The ban does not just reduce demand; it also constrains supply. The same export controls apply to equipment used to make advanced memory in China. Chinese fabs cannot build competitive HBM without American lithography tools. So the supply response from China is minimal. The net effect is that Western memory vendors lose some Chinese demand, but they also gain pricing power in the non-Chinese world because Chinese supply cannot replace them. The 2025 controls are a tax on the Chinese AI ecosystem, not a death sentence for memory.
CPO is being treated as a greater geopolitical safe haven. The logic is that photonic components are not yet on the export control list. That is true today. But the list is a dynamic document. If CPO becomes critical to AI scale-up networks, the same governments that restricted HBM will restrict silicon photonics. The risk is latent, not absent. The market has a strange habit of assuming the current regulatory regime is permanent. That is an abstraction leak.
Contrarian: The Contrarian Truth Is that Memory Is Undepreciated
The consensus view is that memory is at the peak, and CPO is the future. My contrarian view is that memory is underpriced on a risk-adjusted basis, and CPO is overpriced on a certainty-adjusted basis. Let me map the failure modes.
For memory, the bear case is a self-fulfilling prophecy. Memory vendors are so spooked by the macro and geopolitical headwinds that they have already signaled disciplined capex. They are not building massive new greenfield DRAM fabs. They are upgrading existing nodes and allocating more capacity to HBM. This is the opposite of the 2018 oversupply disaster. The industry learned from the last collapse. The new normal is cartel-like behavior: reduce conventional DRAM output, push price increases for HBM and high-value products. The market is pricing memory as if the old cycle will repeat. But the old cycle โ massive identical DRAM capacity builds โ is not repeating. The next two years will see memory supply growth far below demand growth in AI segments. That is a textbook setup for continued strong margins for HBM leaders.
For CPO, the bull case assumes that bandwidth will hit a wall and that optical interconnects must be co-packaged with the switch ASIC to solve it. That is partially true. But the wall is not as close as the narrative suggests. Pluggable optical modules at 800G and 1.6T are still advancing. The industry has a habit of extending existing technologies far longer than expected. The cost per bit of pluggable optics continues to decline by 20-30% per year. Co-packaging only becomes economically necessary when the power density of pluggable modules exceeds the server rack limit. That point may be two or three generations away. Meanwhile, the yield curve for CPO is steep but slow. When every investor is cramming into CPO before the technology has passed its first production milestone, that is a crowded trade.
The deeper contrarian blind spot is the power wall. CPO reduces interconnect power, but it does nothing for compute power. Data center power is the real bottleneck for AI. Companies are not delaying deployments because of interconnect bandwidth; they are delaying because they cannot get grid power. If CPO does not solve the power availability problem, its adoption timeline will shift right. That is not a technology failure; it is an infrastructure dependency failure. Abstraction layers hide complexity, but not error.
There is also a market structure issue. The CPO rotation has been pushed by a handful of small-cap optical names. These stocks are highly sensitive to any rumor of hyperscaler capex cuts. The source article mentions a two-week-old panic in optical communications because of "imagined capital expenditure cuts." That panic is a red flag that the positioning is not fundamental; it is momentum. When the next capex rumor hits, CPO names will bleed harder than memory names. Memory has real earnings support. CPO has only a narrative support.
Takeaway: The Rotation Will Correct When the Stack Reasserts Its Dependencies
Here is my forward forecast. Memory has a 70% chance of a sustainable rebound in HBM and high-end eSSD over the next 6-12 months, driven by AI inference demand that the market is ignoring. CPO has a 50% chance of failing to meet the first large-scale deployment deadline in 2026-2027, due to packaging yield and CoWoS capacity. When that happens, the market will rotate right back into memory. The game is not about picking the better narrative. It is about identifying the tightest binding constraint.
I have spent my career failing to predict human sentiment but succeeding in predicting code behavior. The code of the semiconductor supply chain is unambiguous: memory is a mature, high-yield, high-margin oligopoly. CPO is a pre-revenue prototype with a dependency on a packaging supply that is currently owned by someone else. Reversing the stack to find the original intent of the rotation, I see a simple mispricing. The market is confusing cyclicality with secularity. That is a bug, not a feature. I am not saying memory will outperform every quarter. I am saying the current wholesale rotation out of memory and into CPO is trading a real cash flow stream for an unverified future proof. Based on my audit experience, I would rather hold the verified asset with a 40% margin and a clear failure mode than the speculative asset with a 0.1% market share and an unsolved yield problem.
The question is not whether CPO will eventually dominate. It will. The question is whether the market is early. And early is the same as wrong, when the capital is at risk.