By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
While there's consensus on cyclical pricing pressure and potential ASP compression, the panel is divided on the timing and extent of the impact. Bulls argue that demand remains strong and supply constraints may persist longer than currently priced in, while bears warn of a potential cliff-edge effect due to inventory stockpiles or demand pauses.
Risk: Inventory stockpiles leading to a cliff-edge effect on ASPs in late 2026
Opportunity: Premium HBM segment staying tight with robust backlog and yield improvements
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Micron, Sandisk, and SK hynix have crashed 30 to 35% despite record revenues and HBM supply sold out well into future production quarters.
Expanding HBM capacity could let a manufacturer ship 30% more chips yet earn less money if average selling prices drop 20%.
Hyperscalers are redirecting AI budgets toward power infrastructure, liquid cooling, and custom chips, shrinking memory's share of incremental AI investment.
Artificial intelligence remains the biggest force driving the stock market in 2026. The world's largest technology companies are on pace to spend more than $700 billion this year building AI infrastructure, according to company guidance and earnings releases. New data centers continue breaking ground, Nvidia (NASDAQ:NVDA) can't build enough cutting-edge GPUs to satisfy demand, and cloud providers are racing to expand capacity.
Yet one corner of the AI supply chain is telling a very different story. Memory stocks have stumbled despite AI demand showing few signs of slowing. That disconnect looks puzzling on the surface, but the numbers suggest the market is already looking beyond today's boom and pricing in tomorrow's risks.
Down 35% after a roughly 600% rally earlier this year
Those declines aren't being driven by collapsing AI demand. Quite the opposite. Micron's latest earnings release showed record revenue, while management said high-bandwidth memory (HBM) remains sold out well into future production. SK hynix has likewise reported strong HBM demand fueled by Nvidia's latest AI accelerators.
July 16 is the Final Day to Tap Into the Lithium Boom (sponsor) General Motors, POSCO, and 50,000+ everyday investors have already backed lithium producer EnergyX.
Here's why you should do the same before their July 16 investment deadline: lithium prices are up 75% this year, with demand projected to grow a staggering 5X by 2040.
Here is what Wall Street is really worried about. Memory has always been a cyclical business. Unlike software, where each additional sale carries high margins, DRAM and NAND chips behave much more like commodities. Prices rise when supply is tight, then fall once manufacturers expand production.
That's exactly where investors think this cycle is heading.
Over the past two years, AI created an unprecedented shortage of HBM, the specialized memory used alongside Nvidia's GPUs. Tight supply allowed Micron, Samsung, and SK hynix to command premium pricing while expanding margins. That shortage, however, is expected to begin easing.
Each major manufacturer is ramping HBM production through new fabrication capacity and better manufacturing yields. More supply is good news for customers, but it isn't always good news for shareholders.
A memory company can sell 30% more chips and still earn less money if average selling prices decline 20%. Historically, pricing has mattered more than shipment volume. Ironically, AI demand can remain healthy while memory profits begin shrinking.
AI Spending Is Also Changing
Another reason investors have become cautious is that AI spending itself is evolving.
During the first wave of generative AI, spending centered on GPUs and HBM memory because those were the biggest bottlenecks. Today, hyperscalers are directing more capital toward:
Power infrastructure
Data center construction
Liquid cooling
Optical networking
Custom AI chips
Memory remains indispensable, but it represents a smaller share of incremental AI investment than it did two years ago.
Granted, the bullish case hasn't disappeared. Every new AI server still requires far more HBM than traditional enterprise servers, and larger AI models continue increasing memory requirements. Company earnings releases from Micron and SK hynix indicate much of their premium HBM production is already committed to customers.
The debate is no longer about whether AI demand exists. It's about whether supply growth eventually catches up.
Key Takeaway
In short, the recent sell-off says more about expectations than it does about AI itself.
Wall Street isn't betting that the AI boom is ending. It's betting that memory pricing may have already peaked. If HBM and DRAM prices remain firm while hyperscalers continue investing hundreds of billions of dollars, today's weakness could prove to be an attractive buying opportunity. Conversely, if new capacity pushes prices lower, memory stocks may struggle even while Nvidia and the broader AI ecosystem continue growing.
Ultimately, smart investors shouldn't judge memory companies by AI headlines alone. The numbers that matter most are memory pricing, production capacity, and inventory levels. In this industry, those figures usually determine where the stocks go next.
Four leading AI models discuss this article
"Memory stocks are pricing in an imminent HBM oversupply that is plausible but not yet certain given exploding per-cluster memory intensity."
The article correctly flags memory's commodity-like cyclicality: even with HBM sold out into 2026, Micron (MU), SK hynix, and Samsung are each adding substantial capacity. A 30% volume increase paired with a 20% ASP decline can still shrink revenue, and hyperscalers' shift toward power, cooling, and custom ASICs is diluting memory's share of the >$700B AI capex. Yet the piece underplays how rapidly model sizes and inference demands are scaling; if training clusters grow from 100k to 1M+ GPUs as projected, HBM attach rates could rise faster than supply, sustaining pricing power longer than the market now assumes. Memory stocks' 30-35% pullback from 2025 highs already prices in a material ASP erosion.
The strongest case against is that hyperscalers are already signaling preference for custom silicon (Google TPU, Amazon Trainium, Microsoft Maia) that uses far less third-party HBM per FLOP; if this architectural shift accelerates, even explosive model-size growth may not rescue traditional HBM pricing.
"The market is correctly discounting memory stocks because the transition from supply-constrained scarcity to capacity-driven commodity pricing will compress margins despite continued volume growth."
The market is correctly pricing in the 'commodity trap' inherent to memory. While HBM is currently a premium product, it is rapidly transitioning to a capacity-constrained commodity as manufacturers like Micron and SK hynix aggressively expand capex. The article correctly identifies that hyperscalers are shifting budget toward power and cooling, which acts as a tax on memory margins. Investors are looking at the 2026-2027 supply-demand curve; once HBM yields stabilize and production scales, the pricing power that drove recent record revenues will evaporate, leading to margin compression regardless of volume growth. This is a classic cyclical peak disguised as a secular growth story.
If the transition to specialized, custom-built AI chips requires increasingly proprietary memory architectures, the commoditization of HBM may be slower than historical DRAM cycles suggest, preserving premium pricing longer.
"Memory stocks are repricing on cyclical ASP risk, not AI demand risk, but the article omits whether hyperscalers have pricing power or flexibility to delay purchases—the actual determinant of whether this is capitulation or justified caution."
The article conflates two separate dynamics: cyclical pricing pressure (real) and demand destruction (speculative). Micron and SK hynix are down 30-35% despite sold-out HBM through multiple quarters—that's a *valuation reset*, not a demand signal. The real risk is ASP (average selling price) compression from 20-30% capacity additions, which the article correctly identifies. But the article undersells a critical countervailing force: hyperscalers' willingness to pay premium for HBM hasn't cracked yet, and custom AI chips *increase* memory intensity per unit compute. The shift toward power/cooling infrastructure doesn't cannibalize memory spend; it's additive. What's missing: actual HBM contract pricing trends, inventory levels at hyperscalers, and whether capacity ramps are actually hitting timelines or facing yield issues.
Memory stocks could be crashing because the market has already priced in a 25-30% ASP decline that won't materialize if hyperscalers' capex discipline forces them to absorb higher memory costs rather than delay AI deployment—meaning current valuations are a gift.
"HBM backlog and long-term supply deals can cushion margin pressure, making a price-driven collapse less likely than the article implies."
The article frames memory stocks as a buy-the-dounce risk: demand remains intact, but pricing pressure will crush margins as supply expands. My take: AI demand is complex and memory pricing dynamics depend heavily on product mix. If hyperscalers continue locking in long-term memory deals and the premium HBM segment stays tight (backlog, higher ASPs), the profit shock may be milder than feared. Also, a robust backlog and yield improvements could offset ASP declines; capacity ramp timing and mix shifts (HBM vs. DRAM) matter more than headline volume. The missing context includes contract pricing, inventory levels, and capex timing—all of which could prove the downside risk less severe than the article implies.
If capacity ramps faster than AI demand grows, or if ASPs drop sharper than 20% due to aggressive pricing, memory margins could deteriorate and the stock re-oversell.
"Power/cooling bottlenecks extend HBM tightness 12-18 months beyond consensus supply ramps."
Claude correctly flags that custom silicon may raise memory intensity, but everyone underweights second-order power density: 1M-GPU clusters aren't feasible without liquid cooling breakthroughs that add 18-24 months delay. This timeline mismatch means HBM attach rates stay supply-constrained through 2027 even with aggressive capex. The 30-35% stock pullbacks price in 2026 ASP collapse that likely arrives later.
"Hyperscaler inventory stockpiling creates a 'shadow supply' that will trigger a sharper ASP collapse if infrastructure bottlenecks delay GPU cluster deployments."
Grok, your timeline mismatch argument is compelling, but you ignore the 'inventory overhang' risk. Hyperscalers aren't passive; they are stockpiling HBM to mitigate supply-chain volatility. If the 1M-GPU cluster rollout stalls due to cooling bottlenecks, those stockpiles become a massive 'shadow supply' that will hit the market simultaneously. This creates a cliff-edge effect for ASPs in late 2026, regardless of whether the underlying demand is structurally sound. The market isn't just pricing in capacity; it's pricing in a potential inventory liquidation event.
"Inventory overhang matters less than demand pause risk—hyperscalers' ability to throttle orders creates a sharper ASP cliff than gradual supply-driven compression."
Gemini's inventory stockpile thesis is plausible but needs stress-testing: hyperscalers have sophisticated demand forecasting and wouldn't irrationally hoard if cooling delays were visible. More likely, if stockpiles exist, they're contractually locked at negotiated ASPs—not 'shadow supply' that floods spot markets. The real cliff risk isn't inventory liquidation; it's if hyperscalers simply pause orders mid-2026 to burn through existing stock, creating a demand vacuum that compresses ASPs faster than capacity growth alone would. That's the scenario memory stocks should price in.
"Stockpile unwind is unlikely to be abrupt due to long-term contracts and pricing floors, so the memory-margin risk is more about gradual compression than a sudden cliff-edge."
Gemini's 'shadow supply' cliff-edge hinges on a liquidation that isn't likely to hit overnight. Hyperscalers' HBM stockpiles are typically governed by long-term deals with minimums and take-or-pay clauses; pricing floors and supply agreements dampen sudden dumps. Even if volumes rise, ASPs are more likely to compress gradually than crash, as vendors renegotiate contracts and new capacity comes with contracted volumes. The real risk is a drawn-out margin squeeze, not a sudden market flood.
While there's consensus on cyclical pricing pressure and potential ASP compression, the panel is divided on the timing and extent of the impact. Bulls argue that demand remains strong and supply constraints may persist longer than currently priced in, while bears warn of a potential cliff-edge effect due to inventory stockpiles or demand pauses.
Premium HBM segment staying tight with robust backlog and yield improvements
Inventory stockpiles leading to a cliff-edge effect on ASPs in late 2026