Newsletter · · Ashutosh Agarwal

Memory Prices Soar as Memory Stocks Sink and Cash Rotates to Power - HBM & The Memory Supercycle - Week of July 25, 2026

HBM & The Memory Supercycle for the week of July 18-25, 2026. DRAM prices ripped roughly 20% higher in July even as Micron, SK Hynix, Samsung and SanDisk fell 20-40% off their June highs, Cerebras' CEO confirmed HBM, CoWoS packaging and 3nm foundry space are all sold out, and the marginal AI dollar rotated toward power.

HBM & The Memory Supercycle

Week of July 25, 2026: Memory Prices Soar as Memory Stocks Sink and Cash Rotates to Power

TL;DR

The strangest chart in the market right now: the price of memory keeps ripping, with DRAM up roughly 20% just in the month of July and up 300–500% over the past nine months, while the memory stocks have fallen 20–40% off their June highs. Micron, SK Hynix, Samsung and SanDisk all sold off even though nothing in their order books got worse.

We finally heard from a chip CEO who buys nothing from the memory makers, and even he says they're sold out. Cerebras' Andrew Feldman named the three companies that make high-bandwidth memory and said flatly, "they're sold out. That's the number one problem." He also said TSMC's advanced packaging (CoWoS) and its most advanced factory space are sold out too.

The money didn't leave AI; it just moved one shelf over, to power. The loudest new trade this week was electricity: Elon Musk quietly bought a mobile-gas-turbine company for about $1 billion, and several podcasts framed memory's wobble as investors rotating from "the memory trade" into "the power trade."

The bear tell to watch: even memory bulls now openly admit the 85% gross margins won't last forever. The debate is no longer whether prices break; it's when, and whether it's one year out or three.

What's new

1. A rival chip CEO confirms the shortage from the outside, and adds two more sold-out bottlenecks

The most valuable voice this week wasn't a memory executive at all. It was Andrew Feldman, CEO of Cerebras, on The MAD Podcast with Matt Turck (The MAD Podcast with Matt Turck, July 23). Cerebras builds a giant wafer-sized AI chip that deliberately avoids high-bandwidth memory (HBM), so Feldman has no reason to talk the market up. That's what makes his read useful. Asked about the "memory shortage people may have heard of," he laid out three bottlenecks in the AI supply chain:

"The first is all GPUs and all ASICs except us … use a type of memory called DRAM. And a particular flavor of that memory called HB[M]. And that memory is made by three companies in the world" (SK Hynix, Samsung and Micron [the audio garbles the names]) "and they're sold out. That's the number one problem. … The only way to increase supply would be to build new factories."

He then named the second bottleneck as CoWoS, the TSMC packaging process that stitches the logic chip and memory chips together on a silicon "motherboard": "That process is sold out. And so it is highly constrained." The third is 3-nanometer factory space at TSMC, where "the bulk of the GPU world" is jammed. Why it matters: this is an operator-grade confirmation that the squeeze isn't just DRAM; it runs straight through CoWoS packaging and leading-edge foundry space, exactly the read-through names (TSMC, hybrid bonding, the equipment sellers) our list cares about. Feldman also gave rare plain-English color on why HBM is the choke point: generating each word of an answer means shuttling the model's "weights" from memory to compute, over and over. "This exact step is where HBM is slow." That is the physical reason hyperscalers are paying anything to get more of it. (Operator caveat: Feldman is a competitor to the GPU-and-HBM stack, not a memory maker, so he's incentivized to stress that memory is a problem "we don't face.")

2. Prices up, stocks down, the decoupling nobody can un-see

On Limitless: An AI Podcast (Limitless: An AI Podcast, July 22), hosts Josh Kale and Ejaz (of Bankless) put numbers on the divergence that defined the week: "When you look at the month of July, the cost of DRAM is up 20% and the stocks are down 20%, even though the demand has not changed one bit." They pegged memory prices as up "on average three to 500% over the last nine months," covering both HBM and the NAND flash that SanDisk makes. On Micron specifically: down about 24% on the month, yet "7x forward earnings on 350% revenue and 85%, 85% gross margins." And on the sold-out theme, a concrete SK Hynix data point: "Between 13 to 15 customers secured around 40% of their projected profit for next year. So they've already sold out their supply for the entirety of 2027." Their conclusion, the fundamentals didn't change, the narrative did, is the through-line of this whole issue.

3. The pricing dynamic in plain, slightly brutal terms

On 20VC (The Twenty Minute VC (20VC), July 23), the hosts described the memory makers' current mindset with unusual candor: "All the memory guys are like, this is a commodity business. You all screwed us three years ago. We're going to screw you now for every dime we can. We're going to raise prices on you every quarter … We're going to make 80% operating margins in what [some] would call a commodity, because we know that two years from now, you're going to screw us." They also flagged the cost this pushes downstream: for a data-center builder like CoreWeave, "the cost of building the product you're trying to build has gone up by 2x because the suppliers are charging you more." The warning attached: "The big aha for me is when that pricing breaks, it'll be brutal to the downside. But maybe that's a year, two years from now." (These are venture investors, pundits, not operators, but the framing is sharp.)

4. The structural bull case got a cleaner number: memory went from 2% of a data center's cost to 18%

On Excess Returns (Excess Returns, July 21), the guest, an analyst who has been mapping the AI economy layer by layer, said that when they model AI demand five to ten years out, "the biggest, tightest bottleneck ends up being around memory, not compute." His cost slide: memory was 2% of a data center's cost in 2021 and is now 18%. And the HBM market is "dominated by three companies who have sold out a year in advance." But he was honest about the other side: "if you end up at the top of a cycle, we know what happens to memory stocks, they give up all the gains of the previous 5 or 10 years." That single sentence is the whole bull/bear fight in miniature.

5. Memory is now a swing factor in the hyperscaler capex story, and in inflation

Ahead of Alphabet's print, The Exchange on CNBC (The Exchange, July 22) featured analyst Rohit Kulkarni, who put Google's capex at a ~$200B base case this year and "at least $300B … could be as much as $400B" next year, and named memory as the reason the market fears diminishing returns: "the worry here in the market is that the next $100B are not going to be as efficient as the previous $100B, given how much more they need to spend on memory." Morgan Stanley's Andrew Slimmon countered with the bull's mantra: "Invest in scarcity. There's a scarcity of memory chips … some of the memory names got overhyped early in the year. But the big hyperscalers, they're not hyped at all." Separately, memory prices are now big enough to nudge the macro data. On Macro Hive (Macro Hive Conversations, July 24), former Fed vice-chair Richard Clarida discussed chip prices as a driver of core PCE inflation. When your product starts showing up in the CPI conversation, you've stopped being a niche.

The debate

Last week the fight was "is the valuation real." This week it sharpened into something more specific and more useful: the price of memory and the price of the memory stocks have come apart. Which one is telling the truth?

The bull steel-man (buy the shakeout). Demand didn't move; the tape did. DRAM is up ~20% in July while the stocks are down ~20% (Limitless). SK Hynix has effectively pre-sold 2027, 13–15 customers locking in ~40% of next year's profit (Limitless). Three companies control HBM and are "sold out a year in advance," and in a world where memory has grown from 2% to 18% of data-center cost, that's a structural bottleneck, not a fad (Excess Returns). Even a CEO who competes against HBM says it's sold out with no fix but new fabs (Feldman, The MAD Podcast). On this read, the selloff is mechanical: over-leveraged Korean retail unwinding (more below) plus a fashion-driven rotation into the next shiny trade (power). The fundamentals sit still; the buyers got bored.

The bear steel-man (this is the top tell). Everyone, even the bulls, now concedes the 85% margins are a temporary squeeze. On Forward Guidance (Forward Guidance, July 22), quant Steve Hou, who is broadly constructive on memory demand, said: "I would be very, very shocked if in 2 years the gross margin of the memory makers are still running at like 85%." His frame: price (P) comes down, volume (Q) grows enough that everyone still does fine, but the stocks discount that mean-reversion long before it lands. The 20VC crowd put it bluntly: "when that pricing breaks, it'll be brutal to the downside." And Hou added a demand-side crack the bulls hand-wave: Chinese models like Kimi are delivering DeepSeek-style efficiency gains (better KV-cache handling) so that "memory demand doesn't have to quite grow linearly with context length." If software learns to need less memory per token, the "infinite demand" story softens. Classic late-cycle setup: peak margins, a whiff of a supply/efficiency answer, and smart money selling strength.

Where they actually agree: nobody thinks the physical shortage unwinds this year. Even the bears' timeline is "one to three years" (20VC) or "not until ~2030" for new fab capacity (Limitless). The disagreement is entirely about what a cyclical, commodity business is worth when its margins are visibly at a peak. That's a valuation fight, not a demand fight, and valuation fights are won by whoever has the longer holding period.

Stocks in play

Micron (MU)

Bull: the volume-and-margin machine of the moment: ~7x forward earnings on ~350% revenue growth and ~85% gross margins (Limitless); Jim Cramer, who has liked it "for ages," notes it's up ~240% year-to-date (Mad Money, Mad Money w/ Jim Cramer, July 21). And it's winning allocation battles: on Tesla's call Elon Musk publicly thanked Micron for securing memory Tesla needed, echoed on Squawk on the Street (Squawk on the Street, July 23) as "a big shout out to San[jay] Me[hrotra] at Micron for being able to get something that nobody else has, which is DRAMs." Bear: down 24% on the month (Limitless); Cramer's July 20 warning (Mad Money w/ Jim Cramer, July 20): memory names are "ridiculously cheap on an earnings basis, but they're high" in absolute terms, and he'd rather diversify than "own nothing but semis." Watch: any hard DRAM/NAND contract-pricing print to confirm the "+20% in July" is holding, and Micron's own capex cadence ($200B pledged across Boise and New York, per Squawk on the Street, Squawk on the Street, July 21).

SK Hynix (000660 KS / US listing)

Bull: the clearest "sold out 2027" evidence in the group, 13–15 customers locking ~40% of next year's profit via long-term deals (Limitless). Bear: the poster child for the unwind. Cramer (nicknaming it "SK Hijinks") says the Nasdaq-listed shares "hijacked our trading," down ~22% from a high set the prior week, with the Korea-listed line down ~39% from its late-June peak (Mad Money, July 20). Market Mondays (Market Mondays, July 21) had it "down 40% from highs." Watch: whether Korea's suspension of new single-stock leverage ETFs (see below) actually drains the retail leverage that amplified both the melt-up and the melt-down.

Samsung Electronics (005930 KS)

Bull/Bear: mostly along for the ride, Cramer notes it's down ~34% from its mid-June peak (Mad Money, July 20). It remains one of the three HBM makers Feldman says are sold out, but no Samsung-specific memory-division catalyst surfaced on any podcast this week. Watch: any sign Samsung is closing the HBM qualification gap with SK Hynix, still absent from the conversation.

SanDisk (SNDK)

Bull: the parabola of the cohort, Cramer flags it up ~570% year-to-date (Mad Money, July 21); on Black Wealth Renaissance (Black Wealth Renaissance, July 18) the hosts tracked it from ~$36 in early 2024 to roughly $1,900 (they called it ~5,000%). NAND prices are riding the same up-only wave as DRAM (Limitless). Bear: a stock up thousands of percent in ~18 months carries no cushion, and it's a named defendant in a new lawsuit (below). Watch: NAND spot pricing specifically, which has been the softer of the two memory categories in prior weeks.

A legal overhang for all three IDMs. Black Wealth Renaissance highlighted a class-action suit filed in California by consumers and small businesses accusing Micron, SanDisk and SK Hynix, who together hold ~90% of the DRAM market, of "colluding to restrict the supply and drive up memory prices," alleged up 500–700% over four years. Antitrust suits like this rarely move fast in the US, but it's a tail risk worth logging, and it dovetails uncomfortably with the industry's own "sold out / we ramped down" messaging.

Read-throughs

Memory equipment (Advantest, BESI, Camtek, KLA, Lam, AMAT, ASML): another quiet week for named commentary; the standing gap persists. The closest thing to a read-through came from Feldman (The MAD Podcast): if HBM, CoWoS packaging and 3nm foundry space are all sold out, the tools that add capacity in each of those steps stay in demand. But no operator or analyst discussed the equipment names directly.

Packaging / substrates (CoWoS, hybrid bonding): this is where the week actually delivered. Feldman named CoWoS as sold out and "highly constrained," the first clean packaging data point in a few issues. Still no hybrid-bonding (BESI) deep dive.

GPU makers (NVIDIA, AMD): 20VC floated Gavin Baker's useful "cross-sectional" rule: whatever growth assumptions you use to value NVIDIA, you should apply the same to the DRAM makers, "to a rounding error," because they're bottlenecks in the same buildout. Feldman's frame flips it: HBM being sold out is "a real weakness in the GPUs" that Cerebras claims to sidestep with on-chip SRAM. Take that with the appropriate grain of competitor salt.

PC / handset / console OEMs facing rising memory costs: the pass-through is now visible to consumers. Black Wealth Renaissance noted Apple raising prices "because of the rise in memory costs" and even five-year-old game consoles getting more expensive. On Mac OS Ken (Mac OS Ken, July 23), the timing of new Mac Mini and Mac Studio models was reported (via Mark Gurman) to "hinge on the state of memory chip supplies." Memory is now setting the price of things people actually buy.

The rotation destination, power and energy: worth flagging because it is the story of the money leaving memory. On Limitless, the hosts tied memory's wobble directly to Elon Musk buying mobile-turbine company APR Energy for ~$1B (~1 gigawatt of power), calling electricity "the memory trade before it became the memory trade." If you own memory, know that at least some of your marginal buyer just left for the power names (GE Vernova, Bloom Energy, the neoclouds).

What changed vs last week

A real shift in the storyline. Last week (Issue #3) was about SK Hynix's record Nasdaq IPO round-tripping and the debate migrating to "is the valuation real." This week:

  • The selloff kept going and broadened, but the defining new fact is the price/stock decoupling: DRAM up ~20% in July while the equities fell 20–40% (Limitless). Prices are not confirming the fear.

  • A concrete regulatory action: Korea suspended new single-stock leverage ETF listings, and both Cramer and the Limitless hosts fingered those leveraged Korean retail products as the accelerant behind the June melt-up and the July collapse.

  • A new, credible operator voice (Feldman) filled part of the standing packaging gap by calling CoWoS sold out, the first clean read-through to TSMC's advanced packaging in a while.

  • The narrative rotation into power became explicit and loud: the marginal AI dollar is chasing electrons, not bits, at least for now.

  • The class-action lawsuit against the big three DRAM makers resurfaced with specific numbers (500–700% price rise, ~90% market share).

Still missing everything we've been missing: no direct memory-maker or hyperscaler executive on the record (Micron's Mehrotra only secondhand via Cramer/CNBC), no HBM3E/HBM4 qualification or yield/known-good-stack update, no equipment-name commentary, and no hard contract-pricing print, only directional "+20% in July."

Gaps

Honest absences this week.

  • No qualification/yield/KGS detail. Feldman explained why HBM is the bottleneck mechanically, but nobody discussed HBM3E or HBM4 qualification status, yields, layer counts (12-hi vs 16-hi), or bin splits at NVIDIA/AMD.

  • No on-the-record memory or hyperscaler executive. The one operator (Feldman) runs a company that avoids HBM. Everything about Micron, SK Hynix, Samsung and SanDisk this week was third-party (Cramer, CNBC desks, VC and macro pundits).

  • No hard pricing print. We got direction ("DRAM +20% in July," "+300–500% over nine months") but no TrendForce-style contract number for Q3.

  • Equipment remains a blind spot. No Advantest, BESI, Camtek, KLA, Lam, AMAT or ASML memory-specific commentary; only Feldman's indirect CoWoS/foundry read.

  • China stayed peripheral. CXMT/YMTC appeared only in passing (Feldman "we don't sell" to China; Steve Hou's point that Chinese memory "gets eaten up by Chinese demand by itself"). No dedicated China-supply analysis.

  • No SemiAnalysis, Asianometry, Sharp Tech, Acquired, All-In or BG2 memory episode landed in the window, and no Korean/Taiwanese business-press coverage surfaced.