The Memory Selloff Is a Mispriced Option: Why the Real AI Bet Is Co-Packaged Optics

The tape is loud. Every timeline in crypto and equity-land is declaring the memory trade dead. DRAM spot prices rolled over. NAND is headed for glut. The smart money has already rotated into co-packaged optics, the new shiny object in AI infrastructure. On August 9, a notorious photon-stock guru wrote that he sold his memory positions months ago and is now bottom-fishing for CPO. The crowd sees this as a clean two-way trade: sell memory, buy optics. It's not. It's a mislabeled volatility surface. The market is pricing memory as if all of it is commodity, and optics as if all of it is deployable. Both assumptions are wrong, and that mismatch creates the kind of opportunity that pays for early retirement. Let's ground this in actual silicon. Memory is mature. DRAM is at 1-beta nanometer, shipping DDR5 and LPDDR5X to every server and phone maker on the planet. NAND is pushing past 200 layers, with TLC and QLC as the workhorses. HBM3E is in mass production using TSV stacking, and HBM4 is already in development with a qualification cycle that will eat up fab capacity for the next 18 months. Yields are high because these processes have been refined for decades. That means the product works, but it also means we are deep into the commodity cycle. When yields are high and capacity is full, supply catches up. We saw the same pattern in 2021, and we're seeing it again. Now look at CPO. Co-packaged optics takes an optical engine, a silicon photonics die with an indium-phosphide laser, and places it on the same substrate as a switch ASIC. This shortens electrical paths, cuts power, and opens the door to 1.6T and 3.2T ports. The networking world loves it because pluggable modules are hitting a power wall. But CPO is not a product yet. It is a lab innovation with low yields, thermal mismatches, and coupling losses that still need to be solved. The industry expects meaningful scale only in 2026-2027. The supply chain is a tight oligopoly: Broadcom owns the switch ASIC, TSMC owns the CoWoS packaging, and both are already stretched by NVIDIA's GPU orders. Chinese module makers like Innolight and Eoptolink have packaging engineering talent, but they still depend on US and Japanese lasers and switch chips. So the CPO narrative is not a democratic infrastructure wave. It's a concentrated bet on a few companies that can actually deliver. Here is the core of my argument. The bear case for memory is built on the commodity stack: PC DRAM, mobile LPDDR, consumer NAND. That part of the market is rolling over, and prices will keep sliding as supply catches demand. But the AI stack is different. HBM was the tightest constraint in 2025, with fab utilization above 95%. SK hynix controls more than 50% of the HBM market, Samsung has about 35%, and Micron is the fast follower. HBM4 will require new TSV processes and new packaging capacity, which means the biggest memory players will allocate more capital to HBM and less to commodity DRAM. In other words, the oversupply in 2026 narrative is true for commodity memory, but it is not true for AI-specific memory. The market is treating the memory complex as one trade, just as it treated DeFi as one trade in 2020. That is a mistake. Let's talk numbers. The cloud giants are spending roughly $300 billion on capex in 2025, and the majority is AI infrastructure. That number is the fulcrum of the entire trade. If AI returns disappoint, capex will be cut, and both memory and optics will face a demand shock. But here is the key: two weeks before that August 9 post, the optical names panicked over imagined capex cuts. That panic created an overshoot. The market was pricing an outcome that had not happened. As an options trader, I live for these moments. You want to sell puts on the HBM and enterprise SSD names that are directly tied to AI buildout, and buy calls on CPO companies with actual design wins, not the ones that just issued a press release about a proof of concept. The market is indiscriminately dumping memory and indiscriminately buying optics. Dispersion is your friend. Now let me be contrarian. The memory is dead narrative hides a rotation within memory. Samsung and SK hynix are not building new commodity DRAM fabs. They are building HBM fabs and converting existing lines to serve AI accelerators. Micron's new plants in New York and Idaho are designed for HBM and specialized memory, not for commodity DIMMs. So even if DRAM spot prices fall, the AI-memory producers will maintain pricing power. Meanwhile, the CPO is the future narrative hides a brutal reality: CPO has not even solved the yield problem. The industry needs two to three more years of learning before it can replace pluggable optics across a data center. And during that time, it depends on TSMC's CoWoS line, which is already allocated to AI GPUs. If TSMC has to choose between packaging a GPU or packaging an optical engine for CPO, the GPU wins. That reality is not in the price. The capex cycle also cuts differently. Memory makers are spending heavily. Micron is building fabs in New York and Idaho. SK hynix is expanding in Yongin. Samsung is pushing new capacity in Pyeongtaek. Capital intensity runs at 30-40% of revenue, and depreciation schedules of five to seven years mean new capacity will hammer gross margins in the early years. That is another reason the memory bear case has legs for commodity parts. But the same capex is being aimed at HBM and DDR5 for AI servers, not at the low-margin NAND that powers budget SSDs. The supply that comes online in 2026 will be disproportionately high-end, which is not the same as oversupply. CPO companies, in contrast, are asset-light. Most are design houses or packaging specialists. They don't need billion-dollar fabs. They need precision die-attach machines and coupling alignment stations. The lead time for those tools is six to nine months, not 18 months like EUV. That means CPO capacity can scale faster than memory capacity, but only if the underlying yield curve cooperates. And it hasn't yet. Geopolitics adds a layer of optionality that the market keeps ignoring. The US has tightened export controls on HBM to China, and the 2025 rules directly restrict sales of advanced HBM to Chinese AI chipmakers. That is bearish for China's AI ambitions, but it is bullish for HBM prices globally because it removes a chunk of demand and keeps supply tight. On the CPO side, the export control list is still empty. But Washington is watching. If CPO becomes classified as critical AI infrastructure, the entire ecosystem, from InP lasers to silicon photonics to switch ASICs to advanced packaging, becomes a national security football. The market is not pricing this tail risk. Greeks don't price export-control surprises. I learned that the hard way in the 2017 ICO cycle: I audited a token called CryptoGem and found an integer overflow that could mint infinite tokens. I published the audit, shorted the token through Bitfinex, and made a small fortune when the rug was pulled. The lesson was simple: read the mechanism, not the marketing. The mechanism in this market is the interaction between memory product mix and packaging capacity. Everyone is staring at the DRAM spot index. The real signals are the HBM4 contract price, the qualification milestones for TSMC's CoWoS, and the quarterly capex guidance from Microsoft, Google, and Meta. If HBM4 pricing stays firm into 2026, the memory bear case collapses. If CoWoS capacity is diverted to optics at the expense of AI GPUs, the entire AI stack slows down, and no derivative strategy will save you from that systemic bottleneck. So let's get practical. Stop trading memory versus CPO as if they were enemies. Start trading the components. Buy the AI-memory names, the HBM producers and enterprise SSD makers, but sell the commodity memory names that are exposed to PC and mobile demand. Buy the CPO names that have real switch ASIC partnerships and clear packaging roadmaps, but avoid the pure-play hype. The market's narrative floor is a feeling, not a number. NFT floor is a feeling, not a number; so is the DRAM contract price. Watch the actual order flow: cloud capex guidance, HBM4 qualification cycles, and CoWoS pricing. That's where the truth lives. And remember, the trade is never the thing you think you're trading. It's the mispriced optionality underneath.