Subscribe to The Value Anchor Topic / The Trend Catcher Topic —unlock the full historical archive and never miss a weekly pick again.

Investment Thesis: Rotating from Memory Bottlenecks to Optical Bottlenecks

The AI infrastructure trade is entering a new phase.

For much of the past year, investors focused on high-bandwidth memory (HBM), DRAM, and NAND as the primary constraints limiting AI system expansion. Memory suppliers benefited from strong pricing power, tight supply, and aggressive AI-related demand.

However, the market narrative is beginning to shift.

Recent industry discussions suggest that while memory fundamentals remain structurally strong, the marginal pricing momentum is slowing, creating risks of a crowded positioning unwind. At the same time, high-speed optical connectivity, lasers, and optical interconnect infrastructure are emerging as the next critical bottleneck for AI data centers.

This has led to an increasingly discussed trading framework:

Short Memory, Long Optical.

The strategy is not based on a collapse in memory demand. Rather, it reflects a potential rotation in AI infrastructure value capture — from companies benefiting from memory shortages toward companies supplying the optical infrastructure required to connect increasingly powerful AI systems.

I. Memory: From the Strongest Cycle to a Crowded Deleveraging Trade

1.1 Fundamentals Remain Strong, But Marginal Momentum Is Slowing

Wall Street remains constructive on the long-term memory cycle.

The structural argument is clear:

• AI server demand continues to expand;
• HBM adoption remains strong;
• New memory fabs will take years to fully ramp;
• Supply constraints are expected to persist into 2027.

Several analysts still believe the memory industry could remain tight as supply additions from major manufacturers are limited.

However, the debate is shifting from “how strong is the cycle?” to “how much upside remains in pricing?”

Recent signals indicate that pricing momentum may be approaching a peak.

Citigroup recently maintained a Buy rating on Micron Technology (MU) but reduced its price target from $1,400 to $1,150 after discussions with memory suppliers and industry participants.

The key conclusion:

DRAM and NAND pricing momentum remains positive, but the rate of improvement is slowing.

Citigroup expects memory prices to rise sequentially over the next four quarters but at a declining pace, with DRAM and NAND potentially reaching a pricing peak around Q2 2027.

The revised quarterly DRAM price growth outlook:

For NAND:

Citigroup expects DRAM prices to decline approximately 3% in 2H27 and NAND prices to decline approximately 5%.

1.2 The Main Risk: Peak Pricing and Crowded Positioning

The biggest concern is not demand destruction.

The issue is that memory stocks may have already priced in much of the upside.

Several warning signs have emerged:

DRAM profitability has reached extreme levels

DRAM prices have increased dramatically over the past year, and gross margins at major suppliers have expanded significantly.

However, when margins exceed historical peaks, additional price increases become increasingly difficult.

Customers may resist further pricing increases, limiting upside surprise.

China Capacity Expansion Could Change the Supply Outlook

Citigroup highlighted increasing capacity expansion from Chinese memory companies, particularly:

• CXMT (ChangXin Memory Technologies)
• YMTC (Yangtze Memory Technologies)

While these companies are unlikely to immediately disrupt advanced HBM supply, their expansion could pressure conventional DRAM and NAND markets over time.

The market is increasingly watching whether supply additions arrive during 2027–2028, potentially creating a new oversupply cycle.

Investors Are Overcrowded in Memory Trades

The memory rally has become one of the most popular AI infrastructure trades.

Large leveraged positions, especially through Korean semiconductor ETFs and derivatives markets, have created additional vulnerability.

A forced deleveraging event could amplify downside moves even if fundamentals remain healthy.

2. The Bull vs Bear Debate

The current debate is not about whether AI demand exists.

It is about where the next dollar of AI infrastructure spending will flow.

Bearish View: The Cycle Is Nearing a Peak

Several investors argue that:

• memory pricing expectations are already extremely optimistic;
• AI infrastructure spending is shifting away from memory capacity;
• crowded positioning creates downside risk.

Citrini Research analyst Jukan, previously considered a memory bull, recently argued that investors should consider a tactical:

“Short memory, long optical”

His reasoning:

Korean leveraged ETF pressure

After liquidity issues emerged in Korean leveraged ETF markets, forced redemptions could create additional selling pressure in semiconductor stocks.

NVIDIA Corporation(NVDA.US)’s next-generation architecture may reduce memory intensity

Nvidia’s future AI systems, including Rubin Ultra, may optimize system architecture by reducing HBM requirements per rack while relying more heavily on optical interconnects between racks.

Memory price expectations are already peaking

The market increasingly expects memory pricing to reach a high point over the next two quarters.

Bullish View: Fundamentals Have Not Changed

Investor Serenity disagrees with the bearish interpretation.

His argument:

“The fundamentals have not changed. Only the stock prices and narratives have changed.”

He believes the market is rotating between bottlenecks rather than abandoning memory.

His key points:

• Optical companies were already supply constrained before the recent rally;
• Memory demand remains structurally supported by AI;
• Nvidia optimizing memory usage is normal in every generation;
• Current memory valuations underestimate long-term profitability.

According to Serenity:

“AAOI at $140 and AAOI at $75 were the same company. Samsung at a $1.5 trillion valuation and Samsung at $980 billion were the same company. The only thing that changed was valuation and narrative.”

3. Trading Implications: How to Position Memory Exposure

The investment approach is not simply “short everything related to memory.”

The market needs to distinguish between different memory categories.

Traditional DRAM and NAND

Potential risks:

• pricing momentum slowing;
• valuation compression;
• crowded positioning.

Potential candidates for reduction:

• Micron Technology, Inc.(MU.US) 
• SK hynix Inc. Sponsored ADR(SKHY.US)
• Samsung Electronics

However, risks remain:

• HBM demand could exceed expectations;
• supply constraints may persist;
• short positions could face squeeze risk.

II: Optical Communication: The Next AI Infrastructure Bottleneck

If the first phase of the AI trade was driven by HBM and memory shortages, the next phase may focus on a different constraint:

High-speed connectivity inside AI data centers.

As AI clusters scale, the bottleneck is shifting from compute capacity to system efficiency. Optical interconnects, high-speed transceivers, lasers, and InP (Indium Phosphide) materials are emerging as critical components of the next AI infrastructure cycle.

AAOI Results Shift the Market Narrative

The recent catalyst for the optical sector was a strong earnings report from Applied Optoelectronics (AAOI).

Key highlights:

• Q2 revenue surged 86% YoY to $191.9 million;
• Data center revenue exceeded $100 million for the first time;
• 800G revenue grew more than 10x YoY;
• Management indicated demand is currently 20%-40% above supply capacity.

The market is shifting from: Is AI demand real?” to "Can optical supply chains support AI growth?”

InP: The New Bottleneck

High-speed optical modules rely on advanced lasers, and lasers depend on: Indium Phosphide (InP)

The InP market is facing severe supply constraints:

• Global capacity: roughly 600k–750k wafers;
• Demand: roughly 2.6–3.0 million wafers;
• Potential supply gap: ~70%.

As AI data centers move toward 800G, 1.6T, and CPO architectures, InP shortages could become a key limiting factor.

Nvidia’s Architecture Shift: From Memory to Connectivity

Long-time memory bull Jukan recently argued for a tactical: Nvidia’s next-generation AI systems may optimize architecture by reducing HBM intensity per rack and relying more on optical interconnects between GPU clusters.

The implication is not that HBM demand disappears, but that AI infrastructure value is expanding:

GPU → HBM → Optical Interconnect → Laser → InP

Why Long Optical Could Be the Next AI Trade

Compared with memory:

• DRAM/NAND pricing has already surged;
• margins are near cycle highs;
• positioning is crowded.

Optical remains attractive because:

• demand is accelerating;
• supply remains constrained;
• bottlenecks are moving upstream.

The market may be rotating from: "More memory capacity” to: "More efficient AI connectivity.”

Key U.S. Optical Names

Coherent Corp.(COHR.US)

Core thesis: InP lasers and optical components

• Strong InP manufacturing position;
• AI data center laser demand;
• Nvidia investment support.

Lumentum Holdings, Inc.(LITE.US)

Core thesis: Optical laser leader

• Nvidia strategic investment;
• AI networking exposure;
• CPO and optical switching opportunities.

Applied Optoelectronics, Inc.(AAOI.US) 

Core thesis: High-growth optical transceiver play

• Strong 800G/1.6T momentum;
• Data center growth;
• Demand exceeding supply.

Companies exposed to legacy memory supply tightness may outperform traditional memory names.

Disclaimer: The content is provided as general information only and should not be taken as investment advice. All the contents shall not be taken as a recommendation to buy or sell any security or financial instruments. Any action you take resulting from information, analysis, or commentary on this article is your responsibility. Please consult your investment advisor before making any investments.