The Memory Wall

Nvidia’s HBM Memory Wall

The Thesis

On August 6, The Information reported that Nvidia was weighing lower-memory Rubin Ultra GPUs because it might not secure enough HBM. It had tested at least three configurations. Normally, a buyer with Nvidia’s purchasing power would secure a scarce component, not test a flagship with less of it. That is the constraint Dylan Patel of SemiAnalysis puts at about a third of 2026 big-tech capex, on his estimate. The AI buildout has four walls: the money loop that funds the labs, power, chips gated by a single lithography monopoly, and memory — the input this page tracks.

Memory was supposed to be the opposite of scarce. DRAM and NAND are the textbook boom-bust commodity — makers flood capacity into every upcycle, prices collapse, and the record fills with glut stories. What AI did was pull the scarce part, high-bandwidth memory, out of that cycle. An accelerator cannot fall back to the cheap commodity part, and the fabs that make the expensive part take about two years to build. Announced capacity cannot answer a shortage today. That shortage gives Micron, SK Hynix and Samsung leverage with buyers. Of the three, only Micron has a primary US listing.

That shift is what this page tracks: how fast memory went from commodity to constraint, which numbers are the makers’ own claims rather than settled facts, and which listed names sit on the short side of the wall. What it costs the rest of the chain — and who is now raising prices on whom because of it — is the chokepoint rotation. It does not forecast where any of these stocks go — memory has broken every writer who called the top or the bottom, and the honest position is to watch the record, not predict it.

The Evidence

The reported workaround is less HBM. The Information reported on August 6 that Nvidia had tested at least three lower-memory Rubin Ultra configurations because it might not secure enough HBM. The report establishes tests, not a shipping specification or a production cut.

The cheap part cannot replace it. Patel puts an HBM stack at about 2.5 terabytes per second of bandwidth, against 64–128 gigabytes per second for equivalent commodity DDR. On his figures, an accelerator cannot substitute ordinary DRAM and keep the same bandwidth. HBM production also takes wafer starts that could make ordinary DRAM. A buyer cannot solve that physical trade-off with a different purchase order.

TEXXR coverage provides a measure of attention, not supply. Micron articles in its archive ran at 2 per calendar quarter through early 2025. They reached 17 in 2026 Q1, 19 in Q2 and 15 in completed Q3, against an eight-quarter average of 8.5 per quarter. A 2023 glut also drew a burst of coverage before prices fell. These counts do not measure memory output, shortages or pricing power.

Maker revenue rose; new capacity lagged. The Wall Street Journal reported that memory makers were locking buyers into longer-term agreements. CNBC reported Micron fiscal Q3 2026 revenue of $41.46 billion, up 346% from a derived $9.30 billion in fiscal Q3 2025. That is company-wide revenue, not HBM shipments. CNBC also reported that Micron crossed $1 trillion in market value in May 2026, after passing $700 billion earlier that month. Neither valuation measures memory output.

Micron has committed $250 billion to US spending through 2035 and begun a roughly $9.3 billion Hiroshima expansion, which Bloomberg said would ship HBM from summer 2028. SK Hynix is seeking about $29.4 billion through a US listing; capital raised is not installed wafer capacity. China’s CXMT could add competition, though memory vendors disputed claims that its DDR5 was cheaper, saying its prices matched theirs. The chip group SEMI, which includes Micron and Samsung, also warned Washington against measures it said would distort the market. Those records leave future supply and policy open, not settled.

Price action, June–July 2026. On June 23, Micron and Sandisk each fell more than 13% from the prior close; memory and storage shares fell again on July 16. On July 28, South Korea’s KOSPI fell more than 11% from its prior close, with Samsung and SK Hynix each down 11–12%. On July 29, SK Hynix reported Q2 2026 revenue of about $54.34 billion, up 257% from a derived $15.22 billion a year earlier but below the reported $57.56 billion estimate. Operating profit of about $41.48 billion was up 557% from a derived $6.31 billion, but below the reported $43.85 billion estimate. The July 30 Samsung report put Q2 2026 revenue at about $118.1 billion, up 130% from a derived $51.35 billion a year earlier, and operating profit at about $61.46 billion, up 1,814% from a derived $3.21 billion. The prior-year figures are calculated from the reported growth rates. The selloff continued on July 29, before Samsung and SK Hynix shares rose more than 20% from their prior closes on July 31. October’s chip selloff, detailed below, extends the price record; none of these moves measures HBM output.

After Nvidia’s reported tests, Bloomberg reported an SK Hynix plan to spend $38 billion on new DRAM and NAND facilities in South Korea. Two weeks later, Bloomberg reported a roughly $29 billion share buyback, which SK Hynix framed as a response to worries about the durability of AI spending. The spending plan is not present supply; the buyback is not a measure of future demand.

The Companies

The supply side of this wall has three memory makers, plus the equipment companies that tool their fabs. Micron is the only one with a primary US listing.

Micron is the US-listed memory maker in this group, and DeadRisk’s second-order archetype: a company the AI trade did not create, but now depends on when memory turns scarce. Its dossier tracks whether long-term contracts hold or the cycle wins, quarter by quarter.

The other two scale makers are Korean. The cited earnings and capacity reports do not establish a dated, combined HBM market share for them. SK Hynix is the one Patel singles out: Nvidia is already its largest customer, alongside being TSMC’s — there is no adjacent phone-and-PC capacity to borrow from, which is the whole reason the wall holds. Its own dossier is still being checked before publication. Samsung is the third, and the widest — memory plus a foundry business that puts it on more than one wall at once, and the record now traces its HBM qualification fight against SK Hynix. Both sit in the memory basket, an equal-weighted category page that reports its constituents’ internal correlation.

Around the makers sits the supply chain that tools them. The HBM-equipment basket holds the back-end specialists that grind, stack, bond and test the memory dies — the steps that turn ordinary DRAM into an HBM cube, and a different set of names from the ones who make logic chips. On Dylan Patel’s count, EUV is “in the teens” as a share of a DRAM wafer’s cost, against about 28% for a leading-edge logic wafer, so less of a memory wafer’s cost passes through ASML. He names Applied Materials and Lam as “expanding capacity a lot” too, but does not split the rest of the wafer’s cost among them. Each basket page reports internal correlation among its constituents; listing several names does not make their price moves independent.

The demand side needs no introduction. Nvidia is negotiating three-year memory deals and carrying roughly $90 billion of long-term supply contracts — the clearest sign that the buyer, not the seller, is the one afraid of being caught short.

The Lenses

The Helmer lens reads HBM as a candidate cornered resource — one of the seven powers, preferential access to a coveted asset on terms rivals cannot match. Three makers, a two-year lead time on new fabs, and a buyer who cannot substitute the cheap part: on paper that is the textbook case. But the honest version states what breaks it, and here the thing that breaks it is memory’s own history. A cornered resource is supposed to be durable; a commodity that has mean-reverted every cycle for thirty years is not. The pricing power is real today and contested by construction — the tell will be whether the makers’ long-term contracts survive the supply that Micron, SK Hynix and Samsung are all now building toward, or melt the way memory prices always have.

The Dalio lens asks the question the cycle forces. His debt-cycle template separates a buildout funded out of cash flow from one funded on debt, because they fail differently — and it asks, when the revenue inflection comes late, who is forced to sell. Memory is the most cyclical link in the whole chain, the one with the longest record of boom turning to glut. If any part of the AI trade is going to test Dalio’s question first, it is this one. For now the makers are printing cash and the contracts are holding. The record’s job is to notice the quarter that stops being true.

What Moved

On October 22, Counterpoint reported that global PC shipments fell 21% from a year earlier in Q3 2026, to 57 million units, and said memory and other component costs pushed up prices and held back demand. IDC counted 62.7 million units, down 20.1% after a 3.8% fall in Q2, and said channels had bought early in the year to get ahead of price rises. Counterpoint added that Q3 2025 was a strong base, lifted by the Windows 10 deadline and tariff-related inventory moves; the two firms’ unit counts differ and cannot be combined. Phones fell the same way: Counterpoint on October 12 put Q3 shipments down 20%, IDC on October 14 put them down 22%, against a 4.1% IDC fall in Q1. Shipments count units sent to sellers, not what buyers wanted, so these reports show fewer devices moving at higher prices; they do not say how much is lost demand and how much is early buying or short supply. On October 25, Bloomberg relayed a DongA Ilbo report that SK Hynix gained about $15 billion on its Kioxia stake this year and plans a special dividend of about $12 billion; Nikkei Asia reported on October 16 that Kioxia’s market value had shrunk to about a third of its peak, and neither report says whether the gain is realized. The record supports both halves of the question, falling device units and cash returned to holders after the August buyback above, but it does not show that either causes the other.

On October 18, Reuters relayed a Seoul Economic Daily report that Nvidia had asked Samsung, SK Hynix and Micron to make base dies for its custom NVHBM memory using TSMC’s 3-nanometer process. The paper said Samsung agreed; it would be Samsung’s first use of TSMC for HBM. This reported manufacturing request is different from the August 6 Information report of lower-memory Rubin Ultra tests. Neither report establishes a shipping specification, delivered NVHBM or more HBM supply.

On October 21, Reuters reported, citing sources, that Samsung and SK Hynix were upgrading China fabs to make NAND flash with more than 200 layers as they faced competition from YMTC. That is a NAND upgrade, not a reported addition to HBM capacity.

The US factory funding route also changed. On October 23, the Wall Street Journal reported that SK Hynix’s SSD subsidiary Solidigm had halted its review of a potential US IPO to fund a US chip fab. SK Hynix cited “market conditions and share price trends.” That review is separate from Bloomberg’s June report that SK Hynix was seeking about $29.4 billion through a US listing. On October 15, Reuters reported, citing two people, that SK Hynix was seeking approval to export US chipmaking equipment to Dalian as a condition for building a US factory. Neither report establishes export approval, a factory start or new HBM output. The halted IPO review does not establish that the factory was canceled.

Micron’s Taiwan labor record changed within a day. On October 21, Reuters reported that Micron and the union at its Taichung facility had signed an agreement, without disclosing its terms, resolving tensions there. On October 22, Nikkei Asia reported that workers at Micron’s Taoyuan plant staged a strike seeking regular profit-sharing. These are two unions: the agreement covered Taichung, not Taoyuan. Neither report gives a production-loss figure or identifies HBM shipments affected. The October 7 and 8 Reuters report had established a strike authorization, not a walkout.

The price clock still differs from the factory clock. On October 13, the Financial Times reported SK Hynix down 9.2%, Micron 4% and Samsung as much as 8.8% from their prior closes in a chip-stock selloff. On October 15, CNBC reported SK Hynix down more than 10% and Samsung about 10% from their prior closes. At the October 23 close, the DeadRisk market layer put SK Hynix at 1,277,000 won per share, up 4.4% from the prior close but down 31.4% from 30 days earlier. Samsung closed at 215,500 won, up 0.5% on the day and down 24.5% over 30 days. Micron closed at $877.55, up 2.5% on the day and down 18.1% over 30 days. Those moves do not measure HBM output.

Micron’s fiscal Q4 2026 print raised the revenue baseline. CNBC reported on September 30 that quarterly revenue reached $54.23 billion, against a derived $11.32 billion in fiscal Q4 2025 and CNBC’s $51.07 billion estimate. Net income reached $37.7 billion, against a derived $3.20 billion a year earlier. The prior-year figures are calculated from CNBC’s reported growth rates. These are Micron-wide results, not a count of HBM units shipped.

The DeadRisk market layer recorded $17.5 billion of Micron operating cash flow in fiscal 2025, against $8.5 billion in fiscal 2024. Capex was $15.9 billion against $8.4 billion; net debt fell to $1.9 billion from $4.3 billion across those fiscal years. These annual figures show how Micron funded its prior buildout. They are not a fiscal 2026 cash-flow update.

On October 1, The Register reported that Micron CEO Sanjay Mehrotra expects RAM shortages into 2028 or beyond and customer prices much higher than in 2026. That is Mehrotra’s forecast, not proof of future supply; The Register’s account gives no per-unit price baseline.

The DeadRisk market provider’s October 24 calendar capture schedules SK Hynix earnings for October 27 and Samsung earnings for October 28. Its August 26 capture lists Micron for December 23; Sandisk’s own release sets October 29. The provider’s dates are not issuer filings and can move.

• 2026-09-25Micron closes at $1,082.28 in the DeadRisk market layer, up 16.2% from a derived $931.39 close 30 days earlier and 591.2% from a derived $156.58 close one year earlier. Corrected 27 October: our archive's closes give 15.3% and 590.1%.
• 2026-09-18Reuters reports that CXMT is preparing to expand from DRAM into NAND, challenging Samsung, SK Hynix, YMTC and Micron. Corrected 27 October: date.
• 2026-09-14Micron shares fall 5% from the prior close; this timeline entry does not provide either dollar close. Corrected 27 October: date.

Sources

• Dylan Patel on the compute supply chain: EUV’s share of logic and DRAM wafer cost — Dwarkesh Podcast. Original
• Micron reports fiscal Q3 revenue up 346% YoY to $41.46B, above estimates — CNBC, 2026-06-24. Original · texxr.com/1171494
• Micron hits a $1T market value for the first time — CNBC, 2026-05-26. Original · texxr.com/1169645
• AI has made memory chips more valuable than oil; makers lock in long-term deals — Wall Street Journal, 2026-05-30. Original · texxr.com/1169894
• Micron zooms past $700B market cap in a memory-stock rally — CNBC, 2026-05-05. Original · texxr.com/1168379
• Micron boosts US spending to $250B through 2035 amid memory demand — Bloomberg, 2026-07-09. Original · texxr.com/1172317
• Micron breaks ground on a ~$9.3B Hiroshima plant expansion — Bloomberg, 2026-07-04. Original · texxr.com/1172040
• Chip group SEMI, incl. Micron and Samsung, warns the US against distorting the memory market — Bloomberg, 2026-07-02. Original · texxr.com/1171994
• SK Hynix looks to raise ~$29.4B in a US listing — Bloomberg, 2026-06-24. Original · texxr.com/1171465
• Micron and Sandisk lead a sharp tech selloff, down 13%+ — MarketWatch, 2026-06-24. Original · texxr.com/1171425
• Memory vendors say CXMT’s DDR5 prices match Samsung, SK Hynix and Micron — Wccftech, 2026-06-06. Original · texxr.com/1170429
• US chip, memory and storage stocks fell sharply — Financial Times, 2026-07-17. Original · texxr.com/1173141
• South Korea’s KOSPI drops 11%+, led by chip stocks; Samsung falls 11%+, SK Hynix 12%, on China chipmaking-progress concerns — Bloomberg, 2026-07-28. Original · texxr.com/1173805
• SK Hynix reports Q2 revenue up 257% YoY to ~$54.34B, below ~$57.56B est., and operating profit up 557% YoY to ~$41.48B, below ~$43.85B est. — CNBC, 2026-07-29. Original · texxr.com/1173885
• Investors continue to dump chipmaker stocks following SK Hynix’s results; Sandisk down 7.5%+, Arm 6.5%+, AMD 6.5%+ — Financial Times, 2026-07-30. Original · texxr.com/1173988
• Samsung reports Q2 revenue up 130% YoY to ~$118.1B, operating profit up 1,814% to ~$61.46B, on AI demand — CNBC, 2026-07-30. Original · texxr.com/1173970
• SK Hynix and Samsung shares surge 20%+ in Seoul, a sharp reversal from the week’s selloff — CNBC, 2026-07-31. Original · texxr.com/1174063
• Nvidia is considering lower-memory versions of its Rubin Ultra GPU due to potential issues securing enough HBM, and has tested at least three versions — The Information, 2026-08-06. Original · texxr.com/1174494
• SK Hynix plans a $38B chipmaking expansion in South Korea, including a ~$24.7B DRAM facility in Yongin and a ~$13.3B NAND plant in Cheongju — Bloomberg, 2026-08-07. Original · texxr.com/1174562
• SK Hynix plans a ~$29B share buyback, repurchasing and canceling up to 24M treasury shares, to calm worries about the durability of AI spending — Bloomberg, 2026-08-19. Original · texxr.com/1175245
• Global PC shipments fell 21% YoY in Q3 to 57M units, as memory and other costs pushed prices up; Lenovo retained its market leadership despite a 23% decline — Counterpoint Research, 2026-10-22. Original · texxr.com/1297747
• Global PC shipments dropped 20.1% YoY in Q3 to 62.7M units due to higher prices and supply issues; Lenovo dropped 22.6% YoY, HP 30.9%, Dell 25%, and Apple 11.3% — IDC, 2026-10-08. Original · texxr.com/1296699
• Global smartphone shipments fell 20% YoY in Q3 to a 14-year low amid memory shortages; Apple’s share crossed 20% for the first time in any Q3, reaching 22% — Counterpoint Research, 2026-10-12. Original · texxr.com/1296849
• Global smartphone shipments fell 22% YoY to 255.3M in Q3, a record quarterly drop, as Samsung and Apple gained while Xiaomi, Oppo, and Vivo saw deeper declines — IDC, 2026-10-14. Original · texxr.com/1297042
• Global smartphone shipments fell 4.1% YoY in Q1, the first drop since mid 2023, amid memory constraints; only Samsung and Apple grew shipments, at 3.6% and 3.3% — IDC, 2026-04-14. Original · texxr.com/1167020
• DongA Ilbo: SK Hynix gained ~$15B on its Kioxia stake earlier in 2026, and plans to return ~$12B to shareholders via a special dividend — Bloomberg, 2026-10-25. Original · texxr.com/1297768
• Kioxia has lost two-thirds of its market cap since its peak as Chinese rivals emerge — Nikkei Asia, 2026-10-16. Original · texxr.com/1297233