The two-day bear market for AI is over. Why are funds buying back storage first?

TL;DR

• After a significant de-risking of US chip stocks on June 5, Micron rebounded nearly 10% on June 8, followed by a recovery in the Korean market on June 9, with SK Hynix and Samsung Electronics seeing notable gains.
• Based on earnings reports, DRAM/NAND price increases, and Korean chip export data, memory is currently more easily priced by the market based on EPS upgrades.
• Related assets: Micron, SK Hynix, Samsung Electronics, Western Digital, SanDisk, NVIDIA, Broadcom, Marvell Technology, Coherent, Credo Technology, SOXX Semiconductor ETF, SMH Semiconductor ETF.

Following the semiconductor sell-off on June 5, the market's focus quickly shifted from "why did it fall?" to another question: after the decline, which stocks will recover first?

The answer is not uniform. According to Reuters, US-listed chip stocks saw their market value evaporate by over $1 trillion at one point, with the Philadelphia Semiconductor Index falling nearly 8.5% intraday. At the individual stock level, Micron dropped about 13.25%, NVIDIA fell about 6.2%, AMD declined about 10.86%, and Broadcom fell about 7.92%. However, by June 8, Micron quickly rebounded nearly 10%; on June 9, SK Hynix and Samsung Electronics in the Korean market also strengthened in tandem.

Capital hasn't left AI semiconductors; it's re-screening within the sector. As valuations come under scrutiny, the market's focus has shifted from "who has an AI story" to "who can most quickly translate AI demand into profits." Compared to certain AI hardware sectors still trading on expectations of future product cycles, customer adoption, and capital expenditure expansion, the demand growth for memory is already more directly reflected in orders, prices, and earnings reports.

This is why memory is leading the capital inflow. The market isn't just buying back memory itself, but rather its underlying EPS growth logic, which is more easily verifiable.

The Sell-Off Means High-Expectation Trades Are Being Repriced

One trigger for this de-risking was the expectation gap following Broadcom's earnings report.

Based on absolute numbers, Broadcom's fundamentals are not weak. According to the company's announcement, FY2026 Q2 revenue was $22.2 billion, a 48% increase year-over-year. The company expects FY2026 Q3 total revenue of approximately $29.4 billion and projects AI semiconductor revenue to reach $16 billion, up over 200% year-over-year.

But the market chose to sell. The reason isn't that AI demand has suddenly vanished, but that AI semiconductor assets have accumulated high expectations over the past year or more. When a fundamentally strong company can trigger selling pressure because its AI revenue guidance falls short of some expectations, it indicates that the market's pricing threshold has changed. Just being part of the AI chain is no longer enough; the growth slope, profit delivery, and next quarter's guidance all need to keep pace with the valuation.

This is the meaning of the June 5 sell-off. It wasn't a test of demand collapse, but a stress test for high-expectation trades.

In the past, the main theme for AI semiconductors was more about "who is closest to AI CAPEX (capital expenditure)." GPUs, ASICs, high-speed optical modules, copper interconnects, equipment, and materials – anything integrated into the AI cluster expansion chain could command a valuation premium. But when the market begins to worry about crowded trades, high valuations, and the pace of guidance delivery, the question shifts from "who has an AI story" to "who can most quickly turn AI demand into earnings."

For the stock market, what ultimately determines valuation is not the orders themselves, but whether those orders can translate into earnings per share (EPS). Because in the long run, stock prices are essentially a pricing of a company's profitability. When the market starts focusing on next quarter's profits rather than stories three years out, changes in EPS often matter more than the narrative itself.

Broadcom's role is therefore symbolic. It's one of the core assets in the AI ASIC and networking chip chain. Precisely because it is strong, the market's reaction post-earnings shows that the AI semiconductor chain is now under a higher verification standard.

Why Memory: Prices and Profits Are Already in the Model

Memory's advantage lies in its shorter EPS transmission chain.

AI server demand first changes the supply-demand dynamics for high-value products like HBM (High Bandwidth Memory), server DRAM, and eSSD (Enterprise Solid State Drives). Cloud providers and AI system builders need more computing power, which in turn requires more GPU-companion memory, higher-capacity server memory, and larger-scale data center storage.

As memory manufacturers shift capacity towards HBM and high-end server products, the supply of traditional DRAM and NAND is further constrained, driving up contract prices. This chain doesn't rely entirely on distant imagination; it flows relatively quickly into revenue, gross margin, and EPS.

Micron's earnings already reflect this change. According to the company's announcement, FY2026 Q2 set multiple records in revenue, gross margin, EPS, and free cash flow. Data center-related revenue saw significant year-over-year growth, with guidance for FY2026 Q3 to continue reaching significant new highs. For Micron, AI storage is no longer a long-term vision but a revenue source entering the current quarter's financials.

SK Hynix's financials are even more direct. According to the company's announcement, 1Q26 revenue was 52.5763 trillion Korean Won, with operating profit of 37.6103 trillion Korean Won, resulting in an operating margin of 72%. The company attributes this growth to high-value products like HBM, high-capacity server DRAM modules, and eSSD. For investors, this margin profile reflects product mix, supply-demand gaps, and pricing power all flowing into the financial statements.

Industry pricing data also supports this logic. TrendForce expects 2Q26 conventional DRAM contract prices to rise 58% to 63% quarter-over-quarter, and NAND Flash contract prices to rise 70% to 75% quarter-over-quarter. Their report also shows that 1Q26 DRAM industry revenue grew 81% quarter-over-quarter.

Price isn't profit, but in a phase of tight supply, improving product mix, and strong demand, price increases will improve the market's modeling of EPS for the coming quarters. Korean export data provides an early industrial-level verification. According to Reuters and Korean media, South Korea's exports hit a record in May 2026, with semiconductor exports surging 169.4% year-over-year to approximately $37.16 billion, and chips accounting for over 40% of total exports for the first time.

This doesn't directly equate to EPS for SK Hynix or Samsung Electronics, but it shows that the memory boom is already reflected in accelerating revenue at the national export level.

Memory is Not a Stronger Narrative, but Faster Verification

In this revaluation, the difference between memory and other AI semiconductor sectors is not about whether there is growth, but how that growth is verified.

NVIDIA remains the main valve for AI demand. GPU platform iterations determine AI server architecture, HBM capacity requirements, and supply chain eligibility. However, the market is already highly familiar with NVIDIA's growth and profitability, and its valuation has long been concentrated as the strongest AI asset. In the short term, it is more susceptible to factors like export controls, supply chain constraints, platform transition pace, and expectation gaps.

The ASIC direction also has a real logic. Cloud providers' in-house chips, custom accelerators, and rising AI inference demand are all driving the long-term potential for assets like Broadcom and Marvell. But ASIC is more of a project-based business. Customer concentration, the pace of single-project adoption, production ramp-up windows, and next-generation platform transitions can all affect market perception of revenue visibility.

Optical modules and copper interconnects also have EPS realization paths. Companies like Coherent and Credo benefit from bandwidth upgrades within AI clusters. The shift to 1.6T/3.2T optical modules and changes in cluster interconnect architectures will generate demand. However, the pricing of these directions relies more heavily on future architectural roadmaps, customer qualifications, shipment pace, and capital expenditure cycles. When the market is willing to give a premium, they have strong leverage. When the market demands verification, they are also more likely to be questioned about when orders will translate into revenue.

In contrast, the current pricing basis for memory is more direct. HBM demand pulls high-end products, capacity shifts constrain traditional DRAM/NAND supply, contract price increases improve revenue, product mix upgrades boost gross margins, and all of this ultimately feeds into EPS.

This chain doesn't mean there are no risks, but it is more easily verifiable in the next quarter's earnings report than a "future generation architecture will bring massive orders." This is what it means for memory to be more easily modeled. It's not to say memory is more important than GPUs, ASICs, or optical modules, but rather that after this de-risking of AI semiconductors, the market prefers assets that can be jointly verified by prices, orders, profit margins, and export data.

The EPS Logic is Strengthening, But Not Yet a Consensus

A one or two-day rebound does not prove that AI semiconductor trading has completely transitioned from PE expansion to EPS verification.

The nearly 13% drop in Micron on June 5 and its nearly 10% rebound on June 8 likely include elements of technical correction, short covering, and improved risk appetite. SK Hynix's rise was also catalyzed by news related to its data center partnership with NVIDIA. News, positioning, and fundamentals often overlap in short-term market moves; not all gains can be attributed solely to EPS certainty.

Memory itself remains a cyclical industry. Rapid increases in DRAM and NAND prices will improve supplier profits but may also stimulate supply expansion or dampen procurement appetite from some end customers. HBM annual contracts, yield ramps, customer qualifications, and share allocation are still evolving. It cannot be simply assumed that all price increases will flow seamlessly into the income statement.

SK Hynix and Micron are already high-profile AI memory targets for the market, and stock price elasticity does not always align perfectly with fundamental elasticity. If the slope of future DRAM/NAND price increases flattens, HBM market share falls short of expectations, or customer double-ordering is debunked, the EPS upgrade logic could be challenged.

Similarly, this cannot inversely negate ASICs, optical modules, copper interconnects, and equipment materials. If these directions deliver stronger orders, clearer customer adoption, or better-than-expected guidance, the market could still reassign a valuation premium. AI semiconductors are not just about memory. It's that at this stage, memory can more easily explain, via financial reports, why it should be bought back.

A more prudent interpretation of this market move is that the June 5 sell-off raised the verification threshold for AI assets. The recovery on June 8 and 9 shows that within the AI chain, capital favors sectors with a shorter path to EPS delivery. Memory happens to be in a position where orders, prices, capacity, and profit margins are all simultaneously visible.