메모리 병목 현상 뒤 또 다른 메모리 병목 현상

THE DAILY H.E.A.T. Torching Wall Street’s Obsolete Playbook. Delivered straight to your inbox 5 days a week. Subscribe Home Posts The Memory Bottleneck Behind the Memory Bottleneck The Memory Bottleneck Behind the Memory Bottleneck The HBM trade is correctly priced. The next layer down isn't. Inside: 11 names, 4 layers, 1 contrarian thesis. May 5, 2026 I’ve been a trader and investor for 44 years. I left Wall Street long ago—-once I understood that their obsolete advice is designed to profit them, not you. Today, my firm manages around $4 billion in ETFs, and I don’t answer to anybody. I tell the truth because trying to fool investors doesn’t help them, or me. In Daily H.E.A.T. , I show you how to Hedge against disaster, find your Edge, exploit Asymmetric opportunities, and ride major Themes before Wall Street catches on. Table of Contents H.E.A.T. News vs. Noise: What’s Moving Markets Today ETF News In Case You Missed It H.E.A.T. A Turing laureate just named the next AI hardware constraint. Wall Street is still trading the last one. The names that win this trade aren't the memory makers — they're the companies that build the machines that build the memory. THE SETUP On April 30, David Patterson — Turing Award laureate, Berkeley professor, and the co-architect of RISC, the instruction-set philosophy that has quietly become the default of every smartphone, datacenter accelerator, and embedded chip on earth — sat down with reporters in San Francisco after delivering a keynote at the Dreamy Next event. Asked what comes after High Bandwidth Memory (HBM), the supply-constrained DRAM stack that gates every NVIDIA accelerator shipping today, he gave a single answer: High Bandwidth Flash. HBF. The plain-English version: HBF marries HBM-style bandwidth to flash-style capacity. It is built for inference, where context is the real constraint. If HBM is the sports car — fast, expensive, almost no trunk space — and traditional SSD is the cargo ship — cheap capacity, painfully slow — HBF is the freight train: massive capacity, fast enough to keep an inference workload fed without bankrupting it. Patterson's track record is the relevant fact here. He is not a sell-side analyst. He is the person whose architectural calls — RISC in 1980, datacenter-as-a-computer in 2009, domain-specific accelerators in 2017 — preceded the commercial waves by five to ten years. When Patterson names a bottleneck, the engineering community treats it as a planning input. The investment community usually treats it as noise for another two years and then pays the premium to catch up. We think this is one of those windows. "HBF is highly likely to stand at the center of the next bottleneck — a surge in demand." — David Patterson, Turing Award laureate, UC Berkeley The technical picture explains why Patterson is so confident. AI workloads are bifurcating into two completely different problems. Training is bounded by compute and bandwidth — HBM is the right answer. Inference is bounded by context: the prior conversations, the long documents, the agentic task histories, the persistent state that a model needs to retrieve and reason over. That data does not fit in HBM, and even if it did, the cost would be prohibitive. The industry needs a tier between HBM's speed and SSD's capacity. HBF — stacked NAND flash, non-volatile, dense, low-power — is what fills it. In March, SK hynix and SanDisk jointly announced a global standardization push for HBF. KAIST professor Kim Jung-ho, speaking at a research and technology development briefing in February, summarized the architectural shift in a single line: "What determines speed is HBM. What determines capacity is HBF." Same memory hierarchy, two complementary tiers, two simultaneous bottlenecks. That standardization push is the signal — supply chains start ordering machinery before there is a ticker for the thing being built. BY THE NUMBERS ~3x Wafer capacity HBM consumes per bit vs commodity DRAM Industry estimates March 2026 SK hynix–SanDisk HBF standardization announcement Joint press release 2027–28 Industry consensus window for first commercial HBF parts Patterson, KAIST briefings THE CONSENSUS IS WRONG The consensus AI memory trade is HBM and only HBM. Three names — SK hynix, Micron, Samsung — are doing essentially all the work in institutional portfolios. The HBM trade has been correct, profitable, and crowded for eighteen months. We are not telling you to sell it. We are telling you that the same procurement signal Patterson is reading — that inference-era memory needs a second tier — is one Wall Street has not yet priced. Important to be precise: HBM is not dying. HBF is not a replacement. They are complementary tiers of the same memory hierarchy, and the leading vendors — SK hynix in particular — will sell both. The mistake is thinking the HBM trade is the entire AI memory trade. The structural argument is that the memory wall is widening, not shifting. Two bottlenecks form where one used to be. The capital implication is that the market is currently pricing the second bottleneck at zero. The cleaner way to position for this — and the way we want to lay out today — is not to chase the memory makers harder. It is to look one layer down, at the companies whose machines physically build the stacks. Stacking is stacking. Whether you are bonding DRAM die for HBM or NAND die for HBF, the equipment is structurally identical: thermo-compression bonders, wafer thinning, hybrid bonding metrology, known-good-die probe, advanced inspection. The equipment layer is technology-agnostic to the memory type. It wins on HBM today. It wins on HBF tomorrow. And it wins on the transition between them, because every fab building HBF capacity is buying tools, not licenses. This is the trade. Don't buy the bottleneck. Buy the tollbooth. SPOTLIGHT THE FOUR-LAYER MEMORY MANUFACTURING STACK The HBM supply chain — and by extension the HBF supply chain — is not one industry. It is four. Understanding the four layers is the entire investment framework for the rest of the AI memory cycle. Layer 1 — Direct memory producers. The names everyone owns: SK hynix, Micron, Samsung, plus Kioxia and SanDisk for NAND. This is where capacity sits and where revenue prints. It is also the most consensus, most volatile, and most cyclical layer. Layer 2 — Front-end equipment. ASML for EUV lithography (which gates HBM4 DRAM and will gate HBF NAND scaling). Applied Materials and Lam Research for the etch and deposition steps that make TSV — through-silicon vias, the vertical wiring that lets stacked die communicate — actually work. KLA for process control. Tokyo Electron and ASMI for atomic-layer deposition, which is the step that produces the thin films inside every stacked memory die. None of this is optional. None of this is substitutable. Layer 3 — Stacking, test, and inspection. This is the layer almost nobody outside the supply chain has heard of, and the layer where the bottleneck actually sits. Hanmi Semiconductor builds the thermo-compression bonders that physically stack HBM die — primary supplier to SK hynix, single-name exposure to the bonder bottleneck. Disco Corporation makes the wafer thinning and dicing equipment every HBM die passes through. BESI makes hybrid bonding tools — the technology that enables HBM4 and is structurally identical to what HBF stacking will require. Advantest tests the finished stacks. Onto Innovation and Camtek inspect them. FormFactor probes the known-good-die. This entire layer wins on HBM and wins again on HBF. Layer 4 — IP and advanced packaging. Rambus owns the controller IP that sits on more than a hundred HBM design wins and is structurally positioned for HBF as well. Amkor is the pure-play OSAT — outsourced semiconductor assembly and test — for advanced packaging, the company that physically integrates stacked memory into accelerator modules without GPU-customer beta dilution. The quiet point: Layers 2, 3, and 4 receive almost zero attention in the AI memory narrative — and in our view, the majority of the marginal capex dollar in the HBM-to-HBF transition flows through them, not through the memory makers themselves. That is the gap. THE TRADE, IN ONE LINE Long the equipment, stacking, and IP layers — not the memory makers — into the HBM-to-HBF transition. The memory names are correctly priced. The machines that make the memory are not. Below we tier the names by conviction. Layer 2 and Layer 3 are the contrarian core of this issue. WINNERS — TIERED BY CONVICTION Tier 1 names are structurally insulated from HBF timing risk — they win on HBM and on the transition itself. Tier 2 names require HBF to commercialize on the 2027–2028 window. Tier 3 names are the asymmetric, less-followed exposures. TIER TICKER / NAME THESIS TIER 1 Hanmi Semiconductor (042700 KS) Builds the thermo-compression bonders that physically stack HBM die. Primary equipment supplier to SK hynix. Underowned in US-mandate funds because of Korean listing. The single cleanest expression of the bonder bottleneck — wins on HBM, wins on HBF, wins on the transition. If you own one Layer 3 name, this is it. TIER 1 Disco Corp (6146 JP) Makes the wafer thinning and dicing equipment every stacked memory die passes through, regardless of whether the stack is DRAM or NAND. Technology-agnostic monopoly position in a step that has no substitute. Trades at a premium that is fully earned. TIER 1 BESI / BE Semiconductor (BESIY) Hybrid bonding tools — the technology that enables HBM4 and that HBF stacking will require by definition. The cleanest pure-play on the bonder transition from thermo-compression to hybrid. Single most leveraged Layer 3 name to the architectural shift. TIER 1 Advantest (ATEYY) Tests every HBM stack and will test every HBF stack. Memory test is one of the steps that scales linearly with stack height — and HBF stacks will be taller than HBM. Single most overlooked beneficiary of capacity ramps. TIER 2 AMAT / LRCX Applied Materials and Lam Research split the etch and deposition steps that make through-silicon vias work. Both win on HBM capacity expansion now and on HBF capacity buildout from 2027. Less differentiated than the Layer 3 names but more liquid and more diversified. TIER 2 ASMI / Tokyo Electron Atomic-layer deposition leadership. ASMI is the cleaner pure-play; Tokyo Electron has broader DRAM-capex correlation. Both are required equipment in any HBM or HBF capacity expansion. TIER 2 KLAC / ONTO / CAMT Process control and inspection — the unglamorous tier that wins on every wafer, every die, every stack. Onto Innovation has specific HBM4 inspection wins; Camtek is positioned on hybrid bonding metrology. KLA is the diversified anchor. TIER 2 FormFactor (FORM) Known-good-die probe — the specific test step that determines whether a stacked memory module is sellable. Higher stacks mean more probe touches per finished part. HBF mathematically expands FormFactor's TAM. TIER 3 Rambus (RMBS) Controller IP on more than a hundred HBM design wins. Royalty model means HBF design wins drop straight to gross margin. Less correlated to capex cycles than the equipment names. Underowned IP play. TIER 3 Amkor (AMKR) Pure-play OSAT for advanced packaging — the company that integrates stacked memory into accelerator modules. Wins on every HBM and HBF deployment without GPU-customer concentration. Lowest multiple in the entire stack. TIER 3 SanDisk (SNDK) Post-spinoff from Western Digital, now a pure-play NAND name that is co-developing HBF with SK hynix. The most direct US-listed expression of the HBF thesis if you must own a memory maker. Position size accordingly — this is the speculative line. PRESSURE POINTS — WHERE THE RISK IS TIMING, NOT FAILURE These are not businesses that fail. They are businesses where the risk-reward looks worse than the consensus assumes — either because they are over-loved, over-correlated, or out of scope for the trade we are actually making. TIER TICKER / NAME THESIS PP1 Pure-play HBM memory makers SK hynix, Micron, and Samsung are correctly priced for the HBM trade. None of them are mispriced for HBF — they will sell HBF too, but the marginal new dollar of capex flowing through them goes to equipment vendors, not to retained capacity. Hold them; do not chase them. PP2 Western Digital (WDC) and Seagate (STX) Hard disk drive companies frequently get conflated with memory in retail screens and even in some institutional baskets. They are not in this trade. WDC completed the SanDisk separation in February 2025 and is now a pure HDD company. Out of scope. PP3 GPU and accelerator names NVIDIA, AMD, Broadcom, Marvell are HBM customers, not HBM beneficiaries. Accelerator makers do not own the bottleneck. They rent it. A widening memory wall is a margin headwind for these names, not a tailwind. Frequent investor confusion worth flagging directly. PP4 Equipment names at peak multiples ASML, AMAT, LRCX, KLAC trade at multiples that already price the HBM cycle near peak. If HBF takes longer than 2027–2028 — and it might — these names give back the run-up before the HBF wave arrives. The risk is timing, not direction. Position sizing matters more than thesis quality at current multiples. THE HONEST BEAR CASE BEAR CASE This can absolutely be a 2029 story — and markets will punish you for being early even if you're right. The SK hynix–SanDisk announcement is a standardization announcement, not a product announcement. There is no shipping HBF, no qualified design wins, no revenue. The NAND industry has a long history of standards that took five years longer than expected to commercialize. If HBF slips to 2029 or 2030, the equipment names give back the multiple expansion they are currently enjoying before the HBF wave actually shows up to support it. Patterson is a credibility anchor, not an oracle. His architectural calls have been directionally right and timing-late more often than they have been timing-early. Kim Jung-ho's KAIST briefing pegged the HBF-overtakes-HBM crossover at 2038 — twelve years out. That is too long for a Tech & Capital position to wait. The investable window is the equipment buildout in 2026–2028, not the eventual HBF dominance scenario. Equipment names are also more cyclical than they look. Memory equipment capex collapses 30–40% in a downturn and the multiples compress accordingly. If 2026 H2 brings any sign of HBM oversupply — and a few sell-side desks are already calling for one — the equipment trade gets re-rated downward before HBF buildout begins to support it. Finally, hybrid bonding may not be the right answer for HBF. NAND die are denser, thinner, and more thermally sensitive than DRAM die. The tooling that works for HBM might require modification for HBF — or might be substituted by an entirely different bonding approach. We think this is unlikely; we cannot rule it out. WHAT TO WATCH NEXT THE SIX SIGNALS THAT SAY HBF IS REAL A thesis without a process is a guess. These are the concrete, datable signals — drawn from earnings decks, press releases, and capex commentary — that will tell you whether the HBF buildout is accelerating, on track, or slipping. We track these going forward; readers should too. 1. HBF mentioned by name on SK hynix or SanDisk earnings decks — not as "standardization activity" but as pilot production, qualified design wins, or revenue. The vocabulary shift is the signal. 2. Hybrid bonding capacity language from BESI on quarterly calls. If BESI starts talking about NAND-customer order activity alongside its HBM-customer book, the HBF tooling cycle has begun. 3. Memory test demand inflections from Advantest. HBF stacks will be taller than HBM stacks; if Advantest commentary shifts toward longer test times per finished part or expanded NAND-test capacity, the volume cycle is approaching. 4. Hanmi Semiconductor or Disco Corp announcements of dedicated HBF tool variants or NAND-stacking-specific equipment. These are the companies whose customers tell them the HBF roadmap a year before the press release. 5. Hyperscaler capex commentary that explicitly references "next-generation stacked memory infrastructure" rather than "HBM3E ramp." Vocabulary precedes orders. 6. Any explicit "HBF pilot line" language from a major fab — SK hynix, Micron, or Samsung. This is the single highest-conviction signal that commercialization is within twelve months. FIVE TAKEAWAYS 1. Patterson named the next AI hardware bottleneck on April 30. HBF — stacked NAND for inference-era context storage — is what comes after HBM. The market is pricing the second bottleneck at zero. 2. The cleanest one-line expression: don't buy the bottleneck — buy the tollbooth. Long the equipment, stacking, and IP layers, not the memory makers. The machines that make the memory are mispriced. The memory itself is not. 3. Stacking is stacking. The same thermo-compression bonders, wafer thinning tools, hybrid bonding metrology, and probe equipment that build HBM today will build HBF tomorrow. The Layer 3 names — Hanmi, Disco, BESI, Advantest — are technology-agnostic to the memory type. 4. If you own one name from this issue, make it Hanmi Semiconductor. Single cleanest expression of the bonder bottleneck, primary supplier to SK hynix, underowned in US-mandate funds because of Korean listing. Wins on HBM, wins on HBF, wins on the transition. 5. Bear case is timing, not direction. This can be a 2029 story, and markets will punish you for being early. Equipment multiples already reflect peak-cycle optimism. Use the six signals in the section above to track whether the buildout is accelerating or slipping — and size positions accordingly. Space Exposure Designed For Potential Weekly Payouts The space economy is no longer just a government program — it’s becoming part of a global industry that is gradually advancing. But why view long-term growth as the only potential path forward? SPCI seeks to offer approximately 100% concentrated exposure to the space industry while striving to generate weekly income through a disciplined put credit spread options strategy. Gain exposure to the “final frontier” and a potential weekly distribution while you wait for the theme to play out. Discover SPCI. Distributor: Foreside Fund Services | Investing involves risk including possible loss of principle. News vs. Noise: What’s Moving Markets Today The market has been shrugging off higher oil prices and higher rates, yesterday we saw the 10 year yield near the March highs….. Remember, the main reason rates are moving higher, when everyone knows the Fed would prefer to cut, is that bond traders are expecting higher inflation (partly due to higher oil prices). Why does the market continue to shrug this off? Morgan Stanley raised it’s 2026 hyperscaler capex forecast to $805 billion and $1.1 trillion for 2027. The 2026 spend equals total non-tech S&P 500 capex from 2025. What Iran Tells Us About UFO Disclosure When governments confront unknown threats in their airspace, defense budgets surge and the same aerospace and surveillance companies move hardest. On March 2nd, Northrop jumped 6% and Lockheed 3.3% on the Iran news — and President Trump has since ordered the formal release of government UAP files, with the Pentagon confirming compliance. So if a conventional conflict can move these stocks this fast, what happens when the bigger story breaks? See the UFOD holdings: [thetruthisoutthereufod.com ETF News $MEMY Holdings Update: We replaced $CF ( ▲ 1.71% ) and added $INFQ ( ▼ 1.47% ) All 5% positions. For a full list of MEMY holdings, visit: https://incomeblastetfs.com/etf/memy Distributor: Foreside Fund Services, LLC A Stock I’m Watching $INFQ ( ▼ 1.47% ) One of my favorite quantum stocks, just had an undercut and rally at the 50 day moving average. In Case You Missed It Great talk on with the Acquirers Podcast on markets, value investing, inverse Cramer, and Michael Gayed joins me to talk about taking income from your portfolio and how to get more than 4%…… The H.E.A.T. (Hedge, Edge, Asymmetry and Theme) Formula is designed to empower investors to spot opportunities, think independently, make smarter (often contrarian) moves, and build real wealth. The views and opinions expressed herein are those of the Chief Executive Officer and Portfolio Manager for Tuttle Capital Management (TCM) and are subject to change without notice. The data and information provided is derived from sources deemed to be reliable but we cannot guarantee its accuracy. Investing in securities is subject to risk including the possible loss of principal. Trade notifications are for informational purposes only. TCM offers fully transparent ETFs and provides trade information for all actively managed ETFs. TCM's statements are not an endorsement of any company or a recommendation to buy, sell or hold any security. Trade notification files are not provided until full trade execution at the end of a trading day. The time stamp of the email is the time of file upload and not necessarily the exact time of the trades. TCM is not a commodity trading advisor and content provided regarding commodity interests is for informational purposes only and should not be construed as a recommendation. Investment recommendations for any securities or product may be made only after a comprehensive suitability review of the investor’s financial situation.© 2026 Tuttle Capital Management, LLC (TCM). TCM is a SEC-Registered Investment Adviser. All rights reserved. Keep Reading Apr 20, 2026 • 10 min read The Last Undiscovered Bottleneck: Why the Next $NVDA Moment Belongs to Light, Not Silicon Apr 24, 2026 • 9 min read The Robot Has a Body Problem May 1, 2026 • 15 min read Wall Street Is About to Pay the Wrong People for the AI Defense Boom On December 22, the FCC quietly started a $25B replacement cycle. Three weeks later, the Pentagon picked its drone suppliers. None of them were Lockheed, Northrop, or RTX. View more 3