Newsletter · · Ashutosh Agarwal

SK Hynix Record IPO Cracks as the Memory Selloff Spreads - HBM & The Memory Supercycle: Issue #3 - Week of July 11–18, 2026

HBM and memory newsletter for the week of July 11 to 18, 2026. SK Hynix's record $26.5 billion Nasdaq IPO round-tripped and dragged Micron and the whole memory complex lower, even as the podcasts agreed memory stays tight through 2028 and the real fight shifted from demand to valuation.

HBM & The Memory Supercycle: Issue #3

Week of July 11–18, 2026: SK Hynix Record IPO Cracks as the Memory Selloff Spreads

The biggest memory company you couldn't easily own three weeks ago just became the biggest memory company Americans can't stop trading. SK Hynix rang the Nasdaq bell on Friday July 10 with the largest foreign IPO in US history, $26.5 billion, seven times oversubscribed, popped, and then fell down the stairs. By Monday it was having its worst trading day ever in Seoul, and by mid-week anyone who bought the US shares after the first print was underwater.

Here's the strange part: nobody on the podcasts this week could point to a single thing that changed about the actual memory business. No demand cut. No pricing crash. No lost customer. What broke was the stock, not the story, a lesson every memory investor eventually pays tuition to learn. This issue is about the gap between those two things, because that gap is now the whole trade.

One housekeeping note up front: this was a loud week on the tickers and a quiet week on the fundamentals. Almost everything below is commentator and analyst reaction to a stock move. We did not get a single memory-company executive on a microphone this week. Where a real operator voice slipped through, mostly secondhand, I've flagged it.

TL;DR

  • SK Hynix's $26.5B Nasdaq debut turned into a rout. After a strong first day, US shares fell more than 9% on day two and the stock was "down nearly 40% off the high" by mid-week, dragging Micron, SanDisk, and the whole memory complex down with it. Nobody could tie the drop to memory pricing. The consensus culprit: over-levered Korean retail, "sell-the-news," and Korea's own market plumbing.

  • The real fight is no longer "is memory tight?", it's "what are the stocks worth?" Multiple hosts landed on the same framing: memory stays short through 2028 and possibly 2030, but the market doesn't believe today's earnings are the durable earnings. That's a valuation debate, not a demand debate.

  • Micron kept doing Micron things: it lifted its US manufacturing pledge to $250 billion through 2035, poured concrete on its New York fab a quarter early, and is running ~879% higher free cash flow than a year ago, even as the stock fell another 8% on the week. Meanwhile the loudest bear of the cycle, Jim Chanos, called the whole AI build a "Finance 101" duration mismatch and used DRAM as his prime example.

What's new

1. The IPO that went up, then face-planted: The Circuit, "EP 183: SKHynix IPO, AI Model Wars, and NVIDIA Supply Chain Rumors" (July 14)

This was the closest thing to an inside view all week: hosts Ben Bajarin and Jay Goldberg are semiconductor industry analysts, not TV pundits. Their walk-through of why SK Hynix listed in the US is the useful part. In plain terms: the US is simply the deepest pool of capital in the world, SK Hynix had gotten so big it was "trading in line with the Korean economy rather than their own fundamentals," and a lot of big US funds literally aren't allowed to buy Korean shares, so the company was leaving money and shareholders on the table.

The nugget that matters for the thesis came from the roadshow. Relaying the chairman's comments, Bajarin said: "we don't think this is going to end until 2030, but we're still committed to being... the largest memory provider in the world via... the capacity we're going to bring on." And on spending: "they said they're going to spend a hundred billion dollars on CapEx in the next few years." That's the operator (secondhand) telling you they see no cyclical top before the end of the decade, and are pouring $100B into that view.

Goldberg drew the sharpest line of the week between the two arguments people keep confusing:

"There's a pretty broad consensus that memory is going to be tight at least through 2028 and very possibly much longer than that into 2030. The question I see much more debate is what are memory stocks worth... It's more around the valuation than the very raw fact that memory is scarce."

Why it matters: if you're arguing about whether memory is short, you're fighting last year's war. The live question is whether an 80% gross-margin, cyclically-juiced earnings number deserves a normal multiple, and the market is voting "no" for now. Bajarin even flagged the number that scares him: he'd seen a claim that memory could eat "like 65 to 70%" of hyperscaler capex and said, "that would be nutty."

2. Micron gets paid for building while its stock gets punished: Telltales, "Weekend Update - W2628" (July 12)

(A note on this source: Telltales openly states it's produced entirely with AI-generated voices reading their written "cash flow memo", so treat it as an automated read of their research, not human commentary.) The facts it laid out are concrete and checkable. Micron raised its US manufacturing commitment to $250 billion through 2035, $50 billion more than the figure it gave earlier this year, and "poured first concrete at its new fab in Clay, New York, more than a full quarter ahead of schedule," plus put another $3 billion into the domestic supply chain including financing to GlobalWafers for a Texas wafer facility.

The line that captures the whole Micron setup:

"Micron just repriced itself without anyone noticing, because the earnings caught up to the multiple instead of the multiple getting cut... That's a company whose cash generation exploded faster than the stock price did."

Their number: roughly 41 times trailing free cash flow on about $26 billion of trailing FCF, "up about 879% from a year ago." And the bet, stated plainly: "The pledge increase and the ahead-of-schedule fab are Micron telling you it believes this isn't a one-quarter DRAM spike. Watch whether pricing holds through the next print. That's the whole bet."

3. The best technical color of the week on why memory is short: Monetary Matters with Jack Farley, "The Semiconductor Earnings Boom Is Just Getting Started | Ben Pouladian" (July 14)

Analyst Ben Pouladian (BEP Research) gave the closest thing this week to a real explanation of the HBM yield problem, the manufacturing reason the shortage is sticky. His analogy is worth keeping:

"It's almost like making a lasagna that you have to stack multiple layers exactly on top of each other. And then the wires that go through that connect all these stacks of memory have to be like nanometers precise... not all of them go perfect every time and you throw away a little bit and you have your yield isn't there."

Then a genuine datapoint on how hard the next generation is: "they wanted to go up. I think highest is like 14. They wanted to go to 16, and they weren't able to do it. Adding that extra layer was just too hard aligning everything." That is the single most "qual and yield"-flavored comment we got all week: HBM stack height hitting a physical wall around 14 layers.

His timing call is the one to write down. Asked when enough memory comes online to push prices down "prior early 2028," Pouladian said:

"You'll probably see something mid-2027 where people freak out... Everyone's basically like, there's like 3 doors and then everyone's just going to try to like rush out of these 3 doors and you'll see like this huge stampede... by then it's like, okay, we've done all the CapEx, we built the fabs, they're online, we're running 24/7... The market is flooded."

Farley's follow-up is the investor takeaway: Micron, SK Hynix and Samsung are "almost guaranteed to just be printing absolute hundreds of billions of dollars in profit" next year, but "the most money is made in investing, buying before that actually happens," and a company whose "pricing power is gradually going down" is "a far less exciting opportunity." In other words, the clock on this trade may be a mid-2027 sentiment event, not a 2028 fundamental one.

4. Cramer's read on the crash, it's the market, not the memory: Squawk on the Street, "SK Hynix Shares' Record Decline" (July 13)

Jim Cramer was blunt that the selloff isn't a fundamentals story: "The decline is not really related to the price of DRAMs, which is typically what it would be linked to." He called Friday's US pricing "a little bit peculiar" and "a gift to Americans," and described the Korean tape as a market that "seems to lack true pricing discovery." On the secular-vs-cyclical debate he planted his flag: "I continue to think it is a secular because you're getting longer-term agreements. That's what Micron's got. That's exactly what Sanjay Mehrotra, the CEO of Micron, said on my show", the week's one (secondhand) operator citation of the long-term-agreement thesis. He also flagged the split inside the complex: "NAND pricing, which is more Seagate and Western Digital, is down a little bit... NAND is a little less dear... than DRAM." And the sequel everyone now expects: he thinks Samsung follows SK Hynix with its own US issuance.

5. CoreWeave wants to hedge memory prices: Daily Stock Picks, "Memory Bottleneck, Mag 7 and Steve Cress's New H2 List" (July 15)

Small but telling: the host noted "you've now got CoreWeave eyeing derivatives to hedge memory chip price risk." When a big AI compute buyer starts looking for financial instruments to lock in memory costs, that tells you two things: memory is now a material line item, and buyers expect it to stay expensive enough to be worth hedging. The demand logic came via a Jensen Huang quote relayed on the show: everyone on Earth will eventually have an AI agent, and each agent needs "a memory in a data center that is identified for each individual... it's got to remember everything that you do." More agents, more remembered context, more memory. That's the structural bull case in one sentence.

The debate

The structural bull case got its most disciplined articulation from Morgan Stanley's Brian Nowak on Power Lunch (July 16): hyperscaler capex grows another ~60% next year to $1.2 trillion, then $1.4 trillion by 2028, with "no sign of letting up." His explanation for why the memory names are getting hammered anyway is the bull's whole thesis, that this is a rotation, not a top: "It's very normal... you make a disproportionate amount of return as an investor in the early part of the phase, in the semis, in the silicon... And then as time goes on... you move up the stack" to software and internet. He argued the spenders "have been disproportionately penalized" and the return "is not in the price." Dan Loeb, quoted on the same show, likened the memory selloff to "forced selling in credit markets where securities tank but value remains intact."

The cyclical/glut case got its heaviest hitter on RiskReversal, "Jim Chanos: The Math Ain't Mathing for the AI Data Center Build" (July 17). Chanos's core charge is a duration mismatch, building 20-year assets on 1–2 year contracts:

"People are making decisions on long-term projects based on spot prices. And that's a terrifying thing. We saw it in the shale business... So that's one of the biggest sort of financial crimes I'm seeing right now in terms of Finance 101."

And he named memory as the poster child: "the best example of that is actually DRAM, right? And so this is obviously Micron." He reminded listeners Micron "had negative gross margins three years ago," went "from like $100, like 18 months ago" to a peak of "$1,250," and asked why anyone should "assign to a company... this future expectation for revenues based on an 80% gross margin." His deeper worry is the customer side: return on incremental invested capital for the hyperscalers "has gone from as a group 40% a year and a half ago to about 20% today," and if spend keeps up "it's going to be moving toward 10%", at which point the C-suites ask whether the next trillion dollars beats just buying Treasuries. That's the mechanism by which a demand cut actually arrives.

A gentler version of the same worry came from Brent and Chase Wilsey on Smart Investing (July 11), who walked through SK Hynix's staggering build-out, up to $720 billion in South Korea (including a ~$390 billion cluster in Yongin), a $4 billion advanced-packaging plant in West Lafayette, Indiana by 2028, and ~$7.8 billion for EUV machines from ASML by end-2027, and then delivered the memory investor's oldest warning: "supply caught up, prices collapsed, profits disappeared, and investors who arrived late learned just how brutal that memory cycle can be." Their tell to watch: "The stock will fall when questions start to come around the demand falling," not when demand actually falls.

Where the debate actually sits: notice that the two sides barely disagree on the near-term facts. Both camps accept memory is tight into 2028. The bull says today's price is a gift; the bear says today's earnings are a mirage you shouldn't capitalize. Pouladian's "mid-2027 stampede" is the bridge between them, the moment sentiment turns before the numbers do.

Stocks in play

Micron (MU)

  • Bull: Revenue went "from 13 billion to 23 billion, up 75% quarter over quarter, to 41 billion up another 75%," with over $50 billion guided for this quarter, up ~350% year over year, per Buy Hold Rant (July 16). The host bought 50% more at $913. On The MoneyFlows Show (July 16) it was pitched as "the cheapest stock in the S&P 500 with a PEG ratio of 0.05" and a forward P/E that "looks like six and a half [but] could be somewhere around four," with "46 analysts" recently raising full-year EPS. US-based, 26% HBM share, "closing in much faster than people have expected."
  • Bear: Chanos's whole segment. Negative gross margins three years ago; an 80% gross-margin assumption baked into the LTAs he doesn't trust; a stock that ran from ~$100 to $1,250 and now sits ~$950. Buy Hold Rant catalogued the running list of times Micron was "declared dead" in six months, the Google memory-efficiency algorithm, Chinese DRAM flooding, hyperscaler pullback, DRAM prices "falling to the floor." Any one of those becoming real is the bear trigger.
  • Next number to watch: whether contract pricing holds "through the next print" (Telltales). Also flagged: Micron fell another ~8% on July 16 alone.

SK Hynix (000660 KS / new US listing)

  • Bull: The world's #1 HBM maker at ~56% share and NVIDIA's lead partner (MoneyFlows), now finally buyable by US funds. On MoneyFlows a host pegged it at "a PE of 4.7" and called it "trading cheap as chips," with a stated (pundit-projected) revenue path from $68.3 billion in 2025 to $237 billion in 2026 and $355 billion in 2027, treat those forward figures as the speaker's estimates, not company guidance. The chairman's roadshow line, no end "until 2030", is the operator anchor.
  • Bear: The ADR/foreign-ownership complexity is real. Buy Hold Rant's host explicitly prefers Micron because SK Hynix's US shares are an American Depositary Receipt (he cited a 10-to-1 ratio), adding a management-fee layer plus FX and geopolitical risk "and there isn't a substantial discount." And the market-structure risk is now on full display: Samsung and SK Hynix are "50 to 60% of Korea's stock market" (DHUnplugged, July 15), traded heavily through 2x/3x leveraged ETFs, the plumbing that turned a hot IPO into a "four or five standard deviation" move (CNBC Fast Money, July 13).
  • Next catalyst: whether index funds (SMH, SOXX) add it and at what weight, on Earn Your Leisure (July 12) the hosts called an 8%+ index allocation "a home run." And whether Samsung follows with its own US listing, which multiple shows now treat as a when-not-if.

Samsung Electronics (005930 KS)

  • Barely discussed on its own merits this week, mostly as "the next one to do a US issuance." MoneyFlows flagged the biggest implied upside to analyst targets in the whole group ("Samsung, if you can believe it, has an implied upside of 98%" to price targets, again, a pundit's price-target math, not a forecast). Watch for: any confirmation of a US listing plan.

SanDisk (SNDK)

  • Bull: MoneyFlows highlighted analysts continuously raising estimates ("2027 EPS estimates are just about $207... $242 in 2028"), an ~8x forward P/E on a ~$260B market cap, and "210%" expected EPS growth this year, a pure-play NAND/flash proxy for the storage side of the AI data build.
  • Bear: NAND is the softer corner of memory right now. Cramer: NAND "is a little less dear... than DRAM" and down over the last seven days. Flash lacks the HBM scarcity story. And it fell ~13% in the broad memory selloff (Fast Money).
  • Next number: NAND spot/contract direction, the one pricing series that's not clearly rising.

Seagate (STX) / Western Digital (WDC)

  • Grouped by MoneyFlows as the hard-drive/mass-storage beneficiaries of AI data hoarding (Seagate's HAMR "hammer drives," WDC's archival cloud storage). Net income projections roughly doubling out to 2027–28 on their (pundit) numbers. Same caveat as SanDisk: this is the NAND-adjacent, softer-pricing end of the complex, and both sold off with the group.

Intel (INTC), the cautionary read-through

  • Not memory, but Telltales's side-by-side is the week's cleanest warning about what "losing a cycle" looks like: reports that Intel's 18A/18AP nodes "won't hit profitable yields until late 2026 at the earliest, more likely 2027" drove a "21% decline in seven trading days," and AMD out-earned Intel in data-center revenue for the first time ever ("$5.8 billion to Intel's 5.1"). The lesson memory bulls should sit with: process/yield problems, not demand, are what actually breaks a chip stock.

Read-throughs

Memory equipment (ASML, Lam, Advantest, BESI, Camtek, KLA, AMAT). Thin again this week, but two real prints:

  • ASML beat and guided up hard, revenue "$10.8 billion" vs "$10.3 billion" estimated, next-quarter guidance of "$12.8 billion to $13.4 billion" against ~$11.3 billion expected, yet the stock barely moved (Buy Hold Rant). On Full Signal (July 16), macro strategist Jack noted ASML "said we're going to grow our volumes by 30%."
  • Lam Research got the week's only real equipment pitch: up ~84% year-to-date, with "roughly 30% of their revenue... recurring revenue from servicing the machines and selling spare parts." The bull framing is the oligopoly: "how oligopolized, or... for ASML, for EUV, it's a literal monopoly, that these companies are." The bear framing is twofold: China exposure (with the Chinese memory maker CXMT "going to be IPO-ing pretty soon"), and the ceiling on growth: a semicap name "could only grow their volumes 30%, whereas memory... could grow more just by pricing."
  • Still absent: Advantest (test), BESI (hybrid bonding), Camtek (inspection), KLA, AMAT. No HBM-specific equipment commentary this week.

Packaging / substrates (CoWoS, hybrid bonding). Largely a gap. The only signal was SK Hynix and Earn Your Leisure both flagging that memory makers are now investing in advanced packaging themselves (SK Hynix's Indiana plant is specifically an advanced-packaging facility), a hint that packaging capacity is becoming part of the memory bottleneck, not just a TSMC/CoWoS story. No hybrid-bonding or CoWoS deep dive this week.

GPU makers (NVIDIA, AMD). The relevant read-through is Pouladian's: "There's not an actual GPU shortage anymore", NVIDIA "spent $110 billion on their supply chain, basically buying everything up." The bottleneck has moved to "powered land and tradesmen to build data centers." For memory, that's double-edged: it confirms the GPUs (and their attached HBM) are getting built, but it also means the choke point, and the next disappointment, may be power and construction, not chips. On the competitive side, AMD's data-center revenue passing Intel's (Telltales) is the standout.

PC / handset OEMs facing rising memory costs. The pass-through is now explicit and, per one strategist, resented. Full Signal: "We had Apple raise prices, Microsoft raise prices on their consumer-facing products, laptops, consoles... in direct response to rising memory prices", "100%." The same host put the customer's frustration bluntly: if he ran Meta or Microsoft "spending $100 billion... these people need to stop ripping us off." That resentment is the seed of the bear case, it's exactly the pressure that eventually funds a Chinese supplier or an in-house workaround. (Recall the running list from Buy Hold Rant: Google's claimed 6x-more-memory-efficient algorithm, and Cerebras designing AI chips that skip HBM entirely because it's "too expensive and supply constrained.")

What changed vs last week

Last week was operator week: we had SK Hynix chairman Chey Tae-won on tape directly, Samsung's record quarter, hard TrendForce pricing guidance (DRAM +13–18%, NAND +10–15% for Q3), and SanDisk's $42B backlog. This week was the reaction shot. The listing that dominated last week's preview actually happened, and then it broke: the story shifted from "record fundamentals" to "record volatility." Three concrete deltas:

  • The narrative flipped from "sell the news" on earnings to "sell the news" on the IPO itself. Last week it was Samsung's record profit getting shrugged off; this week it's SK Hynix's blockbuster debut round-tripping and dragging the group down ~30–40% from highs.

  • The debate migrated from "structural vs. cyclical demand" to "is the valuation real." Multiple hosts (The Circuit most clearly) now say the demand argument is basically settled through 2028, the fight is entirely about the multiple. That's a meaningful reframing.

  • We lost the direct operator voice. Last week we filled the standing "no IDM exec interview" gap with Chey. This week it's back: the only insider commentary was secondhand (SK Hynix's roadshow relayed by The Circuit; Cramer relaying Mehrotra).

New this week that we didn't have before: a real (if secondhand) HBM stack-height/yield datapoint, the industry hitting a wall trying to go from 14 to 16 layers (Pouladian), and the first concrete sign of buyers hedging memory prices financially (CoreWeave's derivatives).

Gaps (what the podcasts did not give us this week)

  • No direct memory-company or equipment executive on tape. Everything operator-level was relayed secondhand. Treat the SK Hynix "no end until 2030" and "$100B capex" lines as roadshow paraphrase, not a primary quote.

  • No hard contract-pricing print. Last week had TrendForce's Q3 numbers; this week we got only qualitative "NAND softer than DRAM" color from Cramer. No fresh DRAM or NAND contract percentage.

  • HBM3E/HBM4 qualification at NVIDIA/AMD, KGS, bin splits: still essentially absent. The closest was Pouladian's 14-vs-16-layer yield comment, useful, but not a qual update.

  • Equipment names beyond ASML and Lam: no Advantest, BESI, Camtek, KLA, or AMAT commentary.

  • CoWoS / hybrid bonding: no dedicated discussion, only the passing note that memory makers are building their own advanced-packaging capacity.

  • China (CXMT/YMTC): mentioned only in passing (a coming CXMT IPO; "China is going hard for this"), no real supply analysis.

  • Networks that usually carry this beat, SemiAnalysis, Odd Lots, Asianometry, Acquired, and the Korean/Taiwanese business press, did not surface an on-topic episode in the trailing 7-day window.