On August 9, a semiconductor-focused social media account made a confession. Sold memory stocks months ago. Now bottom-fishing CPO. The author did not provide a model. Did not cite a balance sheet. The entire post was a vibe check with a timestamp.
That is enough to start a forensic review.
I have spent my career inside failure modes. Smart contracts. Exchange reserve proofs. AI agent platforms that write their own bytecode. The one lesson that survives every engagement is this: trust is a variable, not a constant. A market narrative is not an audit report. When an influential trader flips from memory to co-packaged optics, he is not describing a technical fact. He is describing a capital flow. And capital flows are evidence, not conclusions.
The post, if the timestamp refers to 2025, sits inside a strange moment. Memory companies are printing money. HBM is sold out. DRAM prices are rolling over. Meanwhile, CPO is an engineering promise with a yield problem. The rotation says the promise is worth more than the profit. My job is to test that claim against the physical layer.
Context: Two Corners of the Same Machine
Let's define the two corners. Memory here means DRAM, NAND, and particularly HBM. The main makers are Samsung, SK hynix, and Micron. DRAM has moved to 1-beta nm-class nodes with DDR5 and LPDDR5X. NAND is in 200-plus-layer 3D stacks. HBM uses TSV to stack DRAM dies and sits next to AI accelerators through CoWoS packaging.
CPO means co-packaged optics. The optical engine moves onto the same substrate as the switch ASIC. No pluggable module. No long electrical trace. The point is to shorten the interconnect, cut power, and raise density. It relies on silicon photonics and advanced 2.5D/3D packaging.
These are not technology sectors. They are two competing answers to the same question: how do you move data fast enough to feed a GPU cluster? Memory solves the capacity and bandwidth wall at the chip level. CPO solves the network power wall at the rack level. The blogger's rotation implies that the first wall has been climbed and the second wall is now the critical path.
I don't believe that. I believe the wall has moved, not disappeared. And in crypto, we already know what happens when a critical path changes: the old infrastructure becomes exit liquidity.
Core: A Cold Dissection of the Rotation
Technical Teardown: Two Different Maturity Curves
Memory is old. CPO is young. That may be the entire trade.
Memory fabs have spent decades perfecting yield. DRAM and NAND yields are high. HBM is trickier because TSV stacking adds a geometry problem, but HBM3E is in volume production. HBM4 is already in development. The technology is not failing. It is maturing.
CPO is not mature. The hard part is coupling light between the silicon photonics die and the switch ASIC. You need sub-micron alignment. You need thermal stability. You need a package that doesn't crack when the laser heats up and the substrate expands. Coupling yield is still below where pluggable modules operate. Reliability data is thin. The roadmap says 2-3 years before CPO becomes a dependable volume product.
That is not a contrarian point. It is the mainstream view. But mainstream views usually miss the derivative: CPO's yield problem is not a reason to avoid it. It is the reason to demand evidence. Anyone can sell a technology transition. Very few can ship it at scale.
In my audits, I look for the same line. A codebase can have a brilliant design and an exploitable deployment script. The code does not lie, but it does hide. The bugs are present before the deployment. If you audit CPO as if it were a smart contract, you start with packaging yield. You don't trust the whitepaper. You trace the alignment process.
I remember an exchange in 2022. The headline balance sheet looked perfect. The cold wallet matched the ledger. But my team spent three weeks cross-referencing on-chain transactions against internal SQL databases. We found misappropriated funds buried inside DeFi yield positions. The severity of the fraud only spoke through the sheer volume of discrepancies. The same method applies here. You don't ask what the analyst thinks. You ask what the yield curve says, what the material properties allow, and what the assembly line can actually produce.
The Supply Chain Is a Stack of Choke Points
Memory's upstream is controlled by a few equipment and material makers. ASML for EUV. Applied Materials and Tokyo Electron for deposition and etch. Shin-Etsu and SUMCO for high-purity silicon. The suppliers are concentrated. The buyers are concentrated. The memory oligopoly has pricing power, but only within the cycle.
CPO's upstream is worse. The switch ASIC comes from Broadcom or Marvell. The laser dies come from a handful of InP specialists. The advanced packaging comes from TSMC. Every layer has a single dominant supplier. The module makers, who do the packaging integration, have weak bargaining power. They sit between monopolists and hyperscalers.
I have seen this pattern before. In crypto, the custodian with the most balance sheet is not the one with the most security. Here, the module maker with the most revenue is not the one with the most control. Power sits upstream.
The geopolitical layer is also different. Memory equipment is heavily export-controlled. The US has already restricted HBM shipments to China. That creates a floor under memory prices, but it also removes a demand center. CPO is not yet on the restrictions list. That makes it convenient. The absence of export control is not a safety feature. It just means the product hasn't become strategic yet.
Capex: The Hangover Is Already Booked
Memory companies used the 2024-2025 price surge to open their wallets. Micron is building a megafab in New York. SK hynix is expanding HBM capacity in Yongin. Samsung is investing in Pyeongtaek. Capital intensity is running 30-40% of revenue. Depreciation on new fabs will hit gross margins by 5-10 points in the first few years. That is the shape of the next oversupply.
The timing is uncomfortable. If the market is already turning toward CPO, memory capex has probably peaked. The new capacity will land in 2026-2027, right when AI HBM demand might shift from shortage to balance. Classic cycle top behavior.
CPO is asset-light. The costs are R&D and packaging equipment. No giant fab. No five-year depreciation curve. That makes the model easier to sell to equity investors. But asset-light also means the CPO companies don't control their own capacity. They rent TSMC's CoWoS line, and CoWoS is already oversubscribed by AI GPUs. The most important production constraint is not in the CPO maker's hands.
Optimization is just risk wearing a disguise. A light balance sheet looks efficient until you need to pivot and discover your entire output depends on one foundry.
Demand: One Capex Pool, Two Narratives
The demand story comes down to one number: hyperscaler capex. Microsoft, Google, Amazon, and Meta have guided toward hundreds of billions of dollars in AI infrastructure spending. 2025 budgets are massive. Memory and optical are both paid from that pool.
Memory has two demand layers. HBM and enterprise SSD are tied to AI. Regular DRAM and NAND are tied to PC and mobile, which have been weak. The market's mood is not a pure AI trade. When the blogger says the memory cycle is near the top, he is looking at the traditional memory side. HBM may still be tight. The divergence matters.
CPO has one demand layer: AI network build-out. The ports on switch ASICs are moving from 800G to 1.6T and 3.2T. Pluggable optics get physically and thermally impractical at those speeds. CPO is a candidate. Penetration is near zero today. If it goes to 10% by 2028-2030, the revenue swing is enormous. But CPO does not create demand. It only captures a share of existing capex. If the AI capex pool shrinks, CPO is just as exposed as memory.
Flash loans expose the geometry of greed. AI capex is the flash loan of the equity market. It looks like infinite liquidity until someone checks the collateral.
The inventory signals support a cautious read. Memory went through a restocking cycle in 2024 and early 2025. Spot prices for DRAM and NAND began to soften around mid-2025. HBM contract prices remained strong, but the forward curve implies concern. If the market is already shifting from active restocking to passive destocking, the commodity segment is heading into a downcycle. The counterfeiter's trick is to smooth the signal and hide the inventory. The analyst's trick is to replace actual inventory checks with an emotional peer-pressure read.
Geopolitics: The Divorce Changes the Ledger
US export controls are the second-order variable. HBM restrictions push Chinese AI chip designers toward domestic HBM development. ChangXin Memory Technologies and others will try. They are years behind. In the meantime, China cannot build high-end accelerators without memory.
CPO is more complicated. Chinese module makers make half of the world's pluggable optics. They have strong engineering in optical packaging. If CPO goes mainstream, they could be a first-tier packaging partner. But the switch ASIC stays with Broadcom or Marvell. The laser dies stay largely with American, Japanese, and European suppliers. China could lead the integration layer while still being blocked from the core chip.
The Dutch and Japanese export regimes reinforce the chokepoint. ASML cannot sell EUV to China. Tokyo Electron and Shin-Etsu control key deposition and material niches. If Japan follows the US further, both memory and CPO manufacturing are impacted. China's response, export controls on gallium and germanium, touches the InP and GaAs substrates used in optical components. But it is not symmetric. A material restriction is not as decisive as a lithography restriction.
Every exit liquidity event is a forensic scene. If US policymakers decide that co-packaged optics is critical to AI, the current safe narrative can flip in a single Federal Register notice. The market will not be ready. It never is.
Competitive Structure: Oligopoly Versus Lottery
Memory is a clean oligopoly. Samsung, SK hynix, and Micron control nearly all DRAM and HBM. The moat is process R&D and sheer fab scale. The margins are cyclical but the competitive position is stable.
CPO is a multi-player race. Broadcom controls switch ASICs. Intel and TSMC have packaging and silicon photonics assets. Marvell and NVIDIA want to offer alternatives. Chinese module makers bring cost discipline. It is not obvious who captures long-term value.
My forensic habit forces me to ask who benefits if CPO succeeds. The ASIC owner? The TSMC packaging line? The optical engine integrator? Or the hyperscaler who buys at cost? In crypto, the same question applies to layer-2 infrastructure. Sometimes the value accrues to the base layer, not the application.
Let's put numbers on the market structure. In DRAM, Samsung holds roughly 40%, SK hynix around 30%, Micron around 25%. In NAND, Samsung leads at around 35%, Kioxia and Western Digital combine near 30%, and SK hynix has about 20%. In HBM, SK hynix is above 50%, Samsung is near 35%, and Micron is around 15%. That is a fortress. New entrants need years of qualification cycles to break in.
CPO has no fortress. Broadcom's Tomahawk and Jericho ASIC families dominate switching. Intel and GlobalFoundries offer silicon photonics. TSMC controls CoWoS. Chinese module makers like Innolight and Eoptolink are strong in optical packaging. Each layer has a different leader. Each layer has a different bottleneck. A bullish CPO thesis requires all five layers to mature simultaneously. That is a low-probability event in a single year.
Contrarian: What the Bulls Got Right
Now let me take the other side, because the blogger is not entirely wrong.
Memory faces a real inventory problem. The post moves at a moment when the market has been positioning for a memory downcycle. The sell-off is self-fulfilling. If everyone believes memory is over, the order book tightens, and the correction becomes a self-cleaning event. That is the bullish case for the bearish trade.
But the collective bearishness on memory hides a split. HBM is not ordinary memory. It is a custom logic-and-memory package with a multi-year qualification cycle. The cloud GPU build-out is not slowing. If HBM supply remains tight, memory manufacturers still have a growth engine. The 2025-2026 disappointment may be confined to commodity DRAM and NAND. The blogger may have thrown out the AI memory with the bathwater.
CPO also has a hidden risk. The blogger calls it a safer AI infrastructure play because it lacks the memory cycle. But a technology that depends on TSMC's packaging capacity and Broadcom's ASIC roadmap is not diversified. It is a concentrated bet on two players, plus a vapor yield curve. In my audit reports, I call that a single point of failure. The bug was there before the deployment, and the deployment is still in pilot.
The bigger blind spot: both sectors live and die on hyperscaler capex. If the AI capital expenditure cycle turns down, the rotation from memory to CPO is just jumping from the second-floor balcony to the first-floor balcony. The distance is not as large as the narrative suggests. The correct response to a falling building is not to change apartments. It is to check the foundation.
Let me also defend the memory bulls. Traditional memory demand could surprise. PC and mobile have been suppressed for two years. The replacement cycle does not go away. It only compounds. If consumer demand returns at the same time as HBM demand stays sticky, the oversupply scenario loses its urgency. The market is pricing a pure cyclical collapse. That assumes the old demand structure is dead. I am not sure financial analysts have proven that.
The CPO bulls also have a legitimate point about latency. Electrical interconnect does not scale cleanly at 1.6T and beyond. Signal integrity degrades. Power density goes up. Lasers are moving closer to the switch for a reason. The physics is genuinely pushing in the CPO direction. The question is timing, not direction. If hyperscalers are willing to absorb early-field failures, the adoption curve could be faster than most yield models suggest.
So the rotation is not stupid. It is early. And early trades make the best news headlines and the worst risk-adjusted entries.
Takeaway: Read the Evidence Like an Auditor
The chain remembers what the ledger forgets. The ledger of financial narratives records sells and buys, but it does not record physical constraints. Memory is not dead. CPO is not born. The only trustworthy signal is a verified data point, whether it comes from a yield report, an ASML shipment, or an on-chain reserve proof.
For crypto builders, the lesson is direct. So much of this industry is now collateralized by AI infrastructure. Mining, zk-proof generation, and decentralized compute all depend on the same chips, the same packaging, and the same export controls. When the market rotates from memory to CPO, it is telling you which part of the stack is expected to fail first. That is a comment about risk, not a forecast of value.
The next bear market will not be announced by a whitepaper. It will be announced by a capex guide cut. Watch the monthly shipment data. Watch the order cancellations. Watch the HBM inventory. The evidence is public. The question is whether you are willing to read it like an auditor, not a fan.
