The Memory Supercycle: HBM, the AI Memory Wall, and the Companies Winning

Research date: June 13, 2026 | OSINT supply-chain research on the public companies powering the memory boom

Important disclaimer. This is OSINT-based research and educational analysis, not investment advice. I am not a financial advisor. Nothing here is a recommendation to buy or sell any security. Memory is a deeply cyclical industry; market caps, prices, and market-share figures are point-in-time (mid-June 2026), press-reported where noted, and move fast. Do your own due diligence and consult a licensed advisor.

Companion tool - jump to the interactive dashboard to sort and filter every company, browse the memory-technology explorer, and explore the supply-chain map, leaderboard, and demand analysis.

TL;DR

For thirty years, memory was the commodity nobody wanted to own: a brutal boom-and-bust business where three giants periodically flooded the market, crashed prices, and bled red ink. AI broke that pattern. A modern AI accelerator is starved not for compute but for memory bandwidth, and the fix - High Bandwidth Memory, or HBM - now costs roughly as much as the GPU silicon it feeds. Because HBM eats three to four times the factory capacity per gigabyte of ordinary memory, building it has starved the rest of the market, and in early 2026 memory prices spiked at the fastest pace on record. Three companies - SK Hynix, Samsung, and Micron - control almost the entire market, and all of them are sold out through 2026. The cleanest way to play it is the HBM leader (SK Hynix), the one US-listed maker (Micron), and the picks-and-shovels every maker has to buy from: the testers, the wafer tools, and the interface chips. The single biggest risk is the calendar: a China-driven glut in commodity memory is widely expected to arrive in 2027.

Explore it yourself: the interactive dashboard

Here is the whole analysis in one place. Use the dashboard to sort and filter all 40 companies by any score, flip through the memory-technology explorer (what each type of memory is and who leads it), and read the supply-chain map, the who-wins-where leaderboard, the demand-durability debate, and the risks, without leaving this page.

Open the dashboard in a full screen

Prefer a spreadsheet? Download the Excel scoring model with the full 40-company table, all eleven scores, and the composite, so you can re-sort and flex the weightings yourself.

The memory wall: why AI ran out of memory, not compute

Picture a brilliant chef who can chop, sear, and plate faster than any human alive. Now imagine the kitchen has only one tiny door to the pantry, and a single assistant carrying ingredients through it one handful at a time. It does not matter how fast the chef is - the meal comes out at the speed of that door. In a computer, the chef is the processor (it does the math) and the pantry door is the memory bandwidth (how fast data can be carried to the processor). For decades, chipmakers made the chef faster and faster; the pantry door barely widened. Engineers call the result the “memory wall,” and it is exactly what it sounds like: the processor finishes its work and then sits idle, drumming its fingers, waiting for the next batch of data to arrive.

AI made this gap the defining problem of the industry. Training a large model, and especially running one (inference, the token-by-token generation you see when a chatbot replies), is overwhelmingly about moving enormous quantities of data - model weights, the running context - in and out of the processor. The chip is not compute-bound; it is memory-bound. You can bolt on more arithmetic units, but if you cannot feed them, they are dead weight.

High Bandwidth Memory is the answer - it is a much, much wider pantry door. Instead of spreading memory chips flat across a circuit board (the way a normal PC does with its DDR5 sticks, like a single hallway to the pantry), HBM stacks the memory chips vertically - 8, 12, soon 16 high - and drills thousands of microscopic vertical wires straight down through the stack (these are the “through-silicon vias” you will see mentioned everywhere). The whole stack sits right next to the processor on a sliver of silicon, connected by an enormously wide highway: 1,024 lanes in today’s HBM3E, 2,048 in the coming HBM4. Think of it as replacing the one tiny pantry door with a hundred loading docks side by side. The result is terabytes of data per second, at low power, in a footprint the size of a fingernail.

This is why memory has become the story. On an Nvidia B200 - the workhorse AI chip of 2025-2026 - independent teardown models estimate that the HBM alone accounts for roughly half the cost of building the entire accelerator, and HBM plus its advanced packaging together approach two-thirds. A top GPU carries eight HBM stacks. The processor used to be the expensive part; now the memory is. And there is a cruel twist that keeps the bottleneck alive: every gain in HBM bandwidth gets immediately swallowed by a bigger model, so memory stays the next thing standing in the way.

How the money flows

flowchart TD
    CAPEX["AI + cloud capex (hyperscalers, >$1T by 2027)"]
    CAPEX --> GPU["AI accelerators (Nvidia, AMD, custom ASICs)"]
    GPU --> HBM["HBM demand: ~half the cost of a top GPU"]
    HBM --> MAKERS["Memory makers: SK Hynix, Samsung, Micron"]
    CAPEX --> SERVERS["AI + general servers"]
    SERVERS --> DRAM["Server DRAM (DDR5) + LPDDR (SOCAMM)"]
    SERVERS --> STORAGE["QLC enterprise SSD + nearline HDD"]
    DRAM --> MAKERS
    STORAGE --> NAND["NAND makers: Samsung, SK/Solidigm, Kioxia, SanDisk, Micron + WDC (HDD)"]
    MAKERS --> EQUIP["Equipment: ASML (EUV) - Lam/AMAT/TEL (etch/dep) - KLA (control) - DISCO (grind) - Besi/ASMPT (bonding)"]
    MAKERS --> TEST["Test: Advantest, Teradyne"]
    MAKERS --> FOUNDRY["HBM base die + packaging: TSMC"]
    MAKERS --> MATL["Materials: Entegris, photoresist, precursors"]
    SERVERS -.->|every DIMM| RCD["DDR5 interface: Montage, Rambus, Renesas"]
    STORAGE -.->|every SSD| CTRL["NAND controllers: Phison, Silicon Motion"]
    CHINA["China watch: CXMT (DRAM), YMTC (NAND) - the 2027 oversupply risk"] -.->|floods commodity| MAKERS

Follow the money from the top. Hyperscaler AI and cloud capex - projected to top $1 trillion a year by 2027 - is the source of everything. It buys AI accelerators, and those accelerators create the HBM demand that flows to the three memory makers. The same capex builds AI and general-purpose servers, which need ordinary server DRAM (DDR5), increasingly low-power LPDDR in Nvidia’s new SOCAMM modules, and a fast-growing tier of high-capacity QLC solid-state drives plus nearline hard drives for storing the data.

All of that demand funnels into a remarkably narrow base: essentially three companies make the DRAM and HBM, and a handful make the NAND. They, in turn, spend their windfall on the layer beneath them - the equipment makers whose tools etch and stack the chips, the testers that screen every HBM stack, the foundry (TSMC) that now builds the logic “base die” at the bottom of an HBM4 stack, and the materials suppliers. Two thin but lucrative side-streams run off to the edge: the interface chips that sit on every server memory module, and the controllers on every SSD - per-unit toll-booths that grow with the market. And the dashed line at the bottom is the thing that keeps everyone awake: China’s state-backed memory makers, ramping fast, poised to flood the commodity tier.

The shape of the chain is the whole investing point. The demand is enormous and the suppliers are few, which is why this cycle has been so violent - and why the safest exposure is often not the volatile makers themselves but the companies that sell to all of them.

A field guide to memory

“Memory” is not one thing. The single most useful distinction to hold in your head: some memory is working memory (fast, but it forgets everything the instant the power goes off - this is DRAM and HBM, the “desk” you spread your work out on) and some is storage (slower, but it remembers without power - this is NAND flash and hard drives, the “filing cabinet”). A computer needs both: a desk to work on and a cabinet to keep things in. Everything below is a variation on those two jobs, each tuned for a different device. Here is the plain-language tour. (The interactive dashboard has a card for each.)

DRAM is the working memory - fast, volatile (it forgets everything when the power goes off), and the main system memory of every computer. Each bit is one tiny capacitor holding a charge, refreshed thousands of times a second. The server and PC version is DDR5. The mobile, low-power version is LPDDR5X, which is now creeping into AI servers because it sips power - Nvidia’s Vera Rubin systems pair their HBM with 54 terabytes of LPDDR in compact “SOCAMM” modules. The graphics version is GDDR7, used in gaming cards and some cheaper inference accelerators. DRAM’s problem is that shrinking the capacitor has gotten extraordinarily hard, so density gains are slowing - which is a big reason prices are so sensitive to demand.

HBM is the AI star described above: stacked DRAM, enormous bandwidth, about half the cost of a top GPU. The current generation, HBM3E, delivers around 1.2 terabytes per second per stack. The next, HBM4, doubles the interface to 2,048 bits and adds a clever wrinkle: the bottom die of the stack becomes a logic chip built on a foundry process. SK Hynix is having TSMC build its HBM4 base die; Samsung is using its own foundry. That single change pulls TSMC directly into the memory supply chain for the first time.

NAND flash is storage, not working memory: it is non-volatile (it keeps data without power), cheaper per bit, and slower. It is the basis of every SSD and every phone’s storage. Since flat scaling ran out around 2013, NAND has gone vertical, stacking layers like a parking garage - the state of the art in 2026 is roughly 330 to 400-plus layers (Kioxia’s BiCS10 hits 332). The AI angle here is QLC enterprise SSDs: giant, dense drives (122 terabytes shipping, around 245 on the roadmap) that are starting to replace nearline hard drives for AI data lakes, because they offer far better density and power per terabyte.

CXL is the newest idea: a way to add and share memory beyond the slots on a server’s motherboard, attacking the memory wall by letting data centers pool expensive DRAM instead of stranding it in idle machines. It is early but strategically important, and it is the home turf of companies like Astera Labs.

Then there is the long tail. NOR flash is an old technology that survives because it is perfect for one job - storing the boot code and firmware in cars, industrial gear, and medical devices - and that automotive demand is quietly booming. Emerging memories like MRAM and ReRAM combine persistence with speed but have never displaced DRAM or NAND at scale; their real home is embedded memory inside automotive microcontrollers. (Intel and Micron’s famous 3D XPoint, sold as Optane, died - its last shipments were in late 2025 - because it was stuck in the middle with no cost advantage.) And on the horizon is 3D DRAM and processing-in-memory, the industry’s answers to the end of flat scaling: stack the DRAM cells vertically like NAND, and put simple compute directly inside the memory chip so you stop wasting energy shuttling data back and forth.

Who wins where

The competitive map is unusually clean, because so few companies matter.

The makers. SK Hynix is the clear winner of this cycle. It owns the HBM market with roughly 57 to 62 percent share, it was the first and largest HBM supplier to Nvidia, and the financial results are staggering: a record quarter in early 2026 with an operating margin around 72 percent, and full-year 2025 operating profit that overtook Samsung’s for the first time. Samsung remains the overall heavyweight - number one in commodity DRAM (around 39 percent) and number one in NAND (around 29 percent) - but it stumbled badly in HBM, repeatedly failing Nvidia’s qualification tests through 2024 and 2025 before finally passing its 12-layer HBM3E around September 2025; it is now racing to leapfrog on HBM4. Micron is the rising challenger and, crucially, the only memory maker listed in the United States. It is number three in everything, but its 2026 HBM is sold out, its margins are at records, and its US fabs in Idaho and New York are backed by the CHIPS Act.

The NAND specialists. Beyond the big three, Kioxia (the former Toshiba memory unit, freshly public in Tokyo and filing for a US listing) and SanDisk (spun out of Western Digital in 2025) are NAND pure-plays riding sold-out 2026 capacity; both stocks have run enormously. Western Digital, now an HDD pure-play after the split, is a quieter beneficiary - hard drives are enjoying a renaissance as cheap bulk storage for AI data lakes.

The picks-and-shovels - often the better risk-adjusted bet. Because the makers are so cyclical, some of the most attractive exposure is the layer that sells to all of them. Advantest is the standout: it dominates the testing of HBM and memory chips, making it close to a pure play on the AI-memory ramp. Lam Research is the most memory-levered of the big equipment makers (roughly half its system sales go to NAND and DRAM), with Applied Materials, Tokyo Electron, and the process-control near-monopoly KLA close behind. ASML sells the EUV lithography machines that advanced DRAM increasingly needs (memory is about a third of its sales and a rising share of its backlog). DISCO has a near-monopoly on the wafer grinding and dicing essential to thin HBM stacks. And TSMC, as noted, now builds the HBM4 base die.

The toll-booths. Two niches deserve special attention because they tax the market on a per-unit basis. Every DDR5 server memory module needs an interface chip called a registering clock driver, and that market is a near-oligopoly of Montage, Rambus, and Renesas (for the newest variants, just Montage and Renesas). Rambus sweetens its position with a high-margin patent-royalty stream. And every SSD needs a controller, where Phison and Silicon Motion are the merchant leaders. These businesses grow with the volume of memory shipped and the content per module, without taking on the makers’ brutal capital intensity.

The China watch. The names not on most Western screens are the ones that could end the party. CXMT has rocketed from nothing to roughly 8 percent of the DRAM market - now number four globally - launched competitive DDR5, and cleared a roughly $4 billion domestic IPO. YMTC is around 13 percent of NAND and targeting 15. Both are state-backed, both are expanding aggressively, and both are constrained (but not stopped) by US equipment export controls. The one cleanly listed proxy is GigaDevice, a specialist in NOR flash, niche DRAM, and microcontrollers.

Company by company: who’s who in the memory boom

The summary above is the map. This is the territory - the individual companies, what they actually make, how they are doing, and the one-line bull and bear for each. (All figures are mid-2026 and move fast; the dashboard has the full scored table.)

The makers - the cyclical core

SK Hynix (000660.KS / OTC: HXSCL) is the company this cycle made. It is the HBM leader at roughly 57 to 62 percent share, the first and largest HBM supplier to Nvidia, and the financials are almost hard to believe: a record quarter in early 2026 with revenue around 53 trillion won and an operating margin near 72 percent, and full-year 2025 operating profit that overtook Samsung’s for the first time ever. Through its Solidigm subsidiary it is also number two in NAND and the leader in enterprise SSDs. Bull: the dominant HBM franchise, sold out for years, with the best margins in the industry. Bear: peak-cycle margins and an HBM lead that Samsung and Micron are both spending billions to erode.

Samsung Electronics (005930.KS) is still the heavyweight - number one in commodity DRAM (around 39 percent) and number one in NAND (around 29 percent), with a fortress balance sheet. But it badly fumbled HBM, failing Nvidia’s qualification tests repeatedly through 2024 and 2025 before finally passing its 12-layer HBM3E around September 2025, and it claims to be first to mass-produce HBM4 for Nvidia’s Rubin. Its memory division is enormously profitable again; the drag is its foundry and mobile businesses. Bull: number one in DRAM and NAND with HBM4 finally landing, so the memory upside is huge. Bear: the HBM stumble showed execution risk, and the non-memory units dilute the story.

Micron (MU) is the rising challenger and the single most important name for US investors, because it is the only memory maker listed in the United States. It is number three in HBM, DRAM, and NAND, but its 2026 HBM is entirely sold out, it guided to record gross margins around 81 percent, and its Idaho and New York fabs are backed by the CHIPS Act. Bull: the cleanest US-listed way to own the cycle, with HBM sold out and a domestic-manufacturing premium. Bear: number three everywhere with the smallest HBM4 allocation, fully exposed to a 2027 commodity correction.

Kioxia (285A.T) is the former Toshiba memory business, freshly public in Tokyo and reportedly filing for a US listing. It is a NAND pure-play (number three in flash) with its entire 2026 capacity sold out, a leading-edge 332-layer product, and an Nvidia AI-memory partnership; the stock has run several hundred percent. Bull: a pure NAND play on sold-out capacity with a US-listing catalyst. Bear: single-product cyclicality with no DRAM or HBM to cushion the next downturn.

SanDisk (SNDK) is the NAND pure-play that Western Digital spun off in 2025, and it has been one of the cycle’s biggest winners - the stock is up roughly 500 percent in 2026, with fiscal-year revenue heading toward $20 billion, data-center revenue up 233 percent, and more than $42 billion in multi-year supply contracts. Its UltraQLC platform targets the AI-storage tier. Bull: a pure NAND play riding the AI-SSD wave with a huge contract backlog. Bear: the lowest-tier NAND share, no HBM, and fully exposed to the flash cycle after a 500 percent run.

Then there is the specialty fringe. Nanya Technology (2408.TW) is a Taiwan niche-DRAM maker posting record profits (April revenue up more than 700 percent year on year) with enormous leverage to DRAM prices but no HBM. Winbond (2344.TW) makes niche DRAM plus NOR flash. Macronix (2337.TW) is a NOR-flash and ROM specialist that one analyst called “an unlikely winner of the supercycle - without AI,” as legacy non-volatile memory tightens. GigaDevice (603986.SS) is the cleanest listed China memory proxy, top-ten in NOR flash, SLC NAND, niche DRAM, and microcontrollers. All four are small, commodity-exposed, and the most vulnerable to Chinese flooding of the low end.

The storage tier - the cold-data half of AI

Not all AI data lives on expensive memory. The vast archives of training data sit on cheap, dense storage, and that has revived a business everyone had written off: the hard drive. Seagate (STX) and Western Digital (WDC) are a duopoly (more than 95 percent of hard-drive shipments with Toshiba), and tight supply has handed them real pricing power. Seagate is the leader on the new HAMR recording technology, has shipped more than a million HAMR drives, posted record gross margins around 47 percent and revenue up 44 percent year on year, and has nearline visibility into 2027. Western Digital, now an HDD pure-play after shedding SanDisk, is the other half of the duopoly, with cloud at roughly 89 percent of revenue, though its HAMR ramp trails Seagate to about 2H 2027. Bull (both): duopoly economics and record margins as AI data lakes balloon. Bear: still cyclical, capital-intensive commodity storage with long-term substitution risk from giant QLC SSDs.

Test, probe, and inspection - the HBM yield gatekeepers

This is the most underappreciated corner of the chain, and often the best risk-adjusted exposure, because these companies sell to every memory maker and get paid more with each harder-to-build HBM generation.

Advantest (6857.T) is the standout: it dominates the testing of memory and HBM chips (around 58 percent of the memory-test market), which makes it close to a pure play on the AI-memory ramp - its results have been at records on HBM tester demand. Teradyne (TER) is the duopoly partner, with memory-test revenue that more than doubled in a quarter, though it is more diversified into chip testing broadly. FormFactor (FORM) is the leader in probe cards - the dense electrical interfaces that touch every die during testing - and HBM has grown to roughly half of its DRAM revenue, with every new HBM generation forcing a fresh probe-card design. Camtek (CAMT) is, in its own description, the tool-of-record for HBM4 3D inspection and measurement at all the major HBM makers, with around half its revenue AI-related - close to a toll-taker on every HBM stack built. Onto Innovation (ONTO) plays the same advanced-packaging inspection game with its Dragonfly platform (packaging revenue growing more than 50 percent) but is more diversified into front-end metrology. And Aehr Test Systems (AEHR) is the small-cap option: it is pivoting from electric-vehicle chips to AI, developing HBM burn-in hardware and winning AI-processor test orders, but its memory revenue is still mostly prospective and lumpy.

Wafer equipment and materials - the cost of building it all

Every maker, to expand or upgrade, must buy from this layer. ASML (ASML) holds the EUV-lithography monopoly that advanced DRAM increasingly needs (memory is about a third of sales and a rising share of backlog). Lam Research (LRCX) is the most memory-levered of the big equipment names - roughly half its system sales go to NAND and DRAM - and is critical for the high-aspect-ratio etch that 3D NAND requires. Applied Materials (AMAT) is broader and logic-led but with DRAM revenue accelerating. Tokyo Electron (8035.T) is a broad tool-maker guiding to double-digit DRAM growth. KLA (KLAC) owns process control (around 57 percent share), and HBM stacking demands more inspection per wafer. DISCO (6146.T) has a near-monopoly on the wafer grinding and dicing needed to thin HBM stacks. Lasertec (6920.T) is the near-monopoly in EUV mask inspection, a secondary beneficiary as DRAM adopts more EUV. Kulicke & Soffa (KLIC) makes the bonding tools (including thermo-compression bonding) that stack HBM dies. ACM Research (ACMR) supplies the specialized cleaning for HBM through-silicon vias, though it carries heavy China exposure. MKS Instruments (MKSI) sells the subsystems (RF power, vacuum) inside those tools, with the NAND-layer upgrade cycle as its memory angle. Entegris (ENTG) provides the consumable materials (slurries, precursors, filters) that scale with every wafer. And TSMC (TSM) has entered the memory chain for the first time, building the logic “base die” at the bottom of an HBM4 stack - strategically important, though tiny against its total revenue.

Interface, controllers, and custom silicon - the per-unit toll-booths

These businesses tax the market on a per-module or per-drive basis and grow with volume, without the makers’ brutal capital intensity. Every DDR5 server memory module needs a registering clock driver, and that market is a near-oligopoly of Montage Technology (688008.SS), Rambus (RMBS), and Renesas - for the newest variants, just Montage and Renesas. Montage is the near-pure-play (revenue up around 50 percent) but is China-listed; Rambus pairs its interface chips with a high-margin patent-royalty stream and easy US access. Every SSD needs a controller, where Phison (8299.TWO) (record ~61 percent gross margin) and Silicon Motion (SIMO) are the merchant leaders. And Marvell (MRVL) is the adjacent custom-silicon name - it designs custom AI accelerators and has a custom-HBM architecture co-developed with all three makers - though it is really an AI-silicon play (data center is about three-quarters of revenue) rather than a memory pure-play.

The modules - maximum torque, maximum risk

Module makers buy finished memory chips and assemble them into the sticks and drives that ship to customers, which gives them violent leverage to the price cycle. Kingston is the world’s number-one third-party module maker but is privately held, so it cannot be bought directly. The listed plays are ADATA (3260.TWO) and TeamGroup (4967.TWO), both posting revenue up well over 100 percent on the price spike - but both have borrowed heavily to stockpile chips (ADATA’s inventory topped NT$40 billion, roughly $1.3 billion), which is wonderful if prices keep rising and a trap if they reverse. Penguin Solutions (PENG), the former SMART Global, is a hybrid: its SMART Modular memory unit grew around 63 percent while it also runs an AI-infrastructure services business. And Netlist (NLST) is the wild card - a module maker (revenue up 262 percent) whose real story is patent litigation: it has won roughly $866 million in jury verdicts against Samsung and Micron over DRAM and HBM patents, but none is final cash (the largest Samsung award is being unwound on appeal, while the $445 million Micron verdict is the most likely to stick), and it trades over-the-counter. Treat it as a binary IP option, not a memory investment.

The China watch - the risk, not the trade

The names that could end the boom are mostly not buyable in the West. CXMT has gone from nothing to roughly 8 percent of the DRAM market - now number four globally - launched competitive DDR5, and cleared a roughly $4 billion domestic IPO. YMTC is around 13 percent of NAND and targeting 15. Both are state-backed, expanding hard, and constrained but not stopped by US equipment controls. Their capacity ramp through late 2026, landing in 2027, is the single most likely trigger for the next downcycle. The one cleanly listed proxy, as noted, is GigaDevice.

Memory beyond AI

It is easy to forget, amid the HBM frenzy, that most memory still goes into things that are not AI accelerators - and these markets are part of why the shortage is so acute. When the makers reallocated capacity to HBM and AI-server DRAM, they pulled it away from everything else, so the squeeze landed on consumers too.

Smartphones remain a huge consumer of low-power DRAM and flash storage, and on-device AI is pushing content up - premium phones now ship with 12 to 16 gigabytes of LPDDR5X. PCs run on DDR5, and although AI PCs carry more memory, in 2026 the story was actually demand destruction: notebook shipments were revised down sharply (around 15 percent) because memory got so expensive. Automotive is the quiet growth engine - an electric vehicle with advanced driver assistance can use up to five times the DRAM of an old combustion car, plus automotive-grade NAND and NOR, and that market is compounding at mid-teens rates toward roughly $70 billion by the mid-2030s. Gaming consoles and graphics cards lean on GDDR. General-purpose servers (the non-AI kind) still buy enormous quantities of DDR5. And data-center storage - the enterprise SSD market, around $47 billion in 2026 - is booming on its own as AI data lakes and the shift from hard drives to flash play out.

The takeaway: AI is the marginal driver and the headline, but it sits on top of a broad, growing base of everyday demand. That breadth is part of the bull case - and the part of the market most exposed if China floods it.

Is the boom durable?

This is the question that matters, because memory has burned investors many times before. The honest answer is: durable through 2026 and probably into early 2027, then genuinely at risk.

The bull case is that this time is structurally different. The key new fact is HBM cannibalization. Because HBM consumes roughly three to four times the wafer area per gigabyte of standard DDR5, every line a maker converts to HBM removes more ordinary bits than it adds AI bits - so the AI build-out tightens the entire market at once, including the PC and phone memory that has nothing to do with AI. TrendForce estimates AI will consume about 20 percent of global DRAM wafer capacity in 2026. On top of that, the makers have shown real supply discipline after the carnage of the last downturn, aiming their 2026 capital spending at process upgrades and HBM rather than brute capacity, and no major new cleanroom comes online in volume before late 2027 or 2028. Memory content per device is rising everywhere. And everyone is sold out under multi-year contracts. The result has been the steepest upcycle in the industry’s history: DRAM contract prices rose an unprecedented 90-plus percent in a single quarter in early 2026, Gartner coined the word “memflation,” and a 32-gigabyte DDR5 kit that cost around $80 in mid-2025 was fetching $400-plus by early 2026.

The bear case is that cyclicality is never repealed. Memory is the most cyclical business in technology, and peak margins above 70 percent are exactly the conditions that historically invite overbuilding. The specific threat is China: CXMT and YMTC are installing new fab equipment through the second half of 2026, with the resulting supply wave widely expected to hit in 2027, concentrated in the commodity and legacy tiers. Layered on top is the AI-capex question - much of the demand is funded by a handful of hyperscalers spending more than a trillion dollars a year on infrastructure whose returns are not yet self-sustaining - and the ever-present risk that today’s panic-buying and long-term contracts are partly double-ordering that will unwind.

The most likely outcome is not a clean answer but a split one: a bifurcated 2027 where premium AI memory (HBM, high-capacity server DRAM) stays tight and profitable while commodity DRAM and NAND correct as Chinese capacity scales. For investors, that argues for the HBM-levered names and the picks-and-shovels over the pure commodity makers - and for treating today’s record earnings as the top of a cycle, not a new normal.

Companies to watch (bull / base / bear)

SK Hynix (HXSCL / 000660.KS) - the HBM leader

Bull: HBM demand keeps compounding, SK Hynix holds its ~60 percent share through HBM4, and its margins stay extraordinary.
Base: It remains the HBM leader but Samsung and Micron claw back share, and commodity DRAM softens in 2027.
Bear: A China-driven glut plus an HBM share loss compress its peak margins hard. Memory is cyclical; this is cycle-peak.
Watch: HBM4 qualification and share at Nvidia, DRAM contract prices, China DRAM output.

Micron (MU) - the only US-listed maker

Bull: HBM ramp and sold-out capacity drive record results; the US listing and CHIPS-funded fabs are a scarcity premium.
Base: Captures the cycle as the strong number three without leading it.
Bear: Smallest HBM4 allocation, number three in everything, fully exposed to the 2027 commodity correction.
Watch: HBM4 ramp and allocation, gross margin, Idaho/New York fab timing.

Advantest (6857.T) - the HBM test pure play

Bull: Every HBM stack must be tested, Advantest dominates that, and AI/HBM test demand is exploding - a cleaner, less capital-intensive way to own the ramp.
Base: Grows with HBM but at a more measured pace as test intensity normalizes.
Bear: A memory-capex pause cuts tester orders quickly; the stock already prices in a lot.
Watch: HBM bit growth, tester order trends, memory-maker capex.

Montage Technology (688008.SS) / Rambus (RMBS) - the DDR5 interface toll-booth

Bull: A near-oligopoly taxing every server memory module, with content rising each DDR5 generation; Rambus adds a royalty annuity.
Base: Steady growth tied to DDR5 server penetration.
Bear: A server-DRAM downturn cuts module volumes; Montage carries China-listing and geopolitical risk.
Watch: DDR5 server penetration, MRCD/MDB ramp, CXL adoption.

The China watch (CXMT, YMTC; listed proxy GigaDevice 603986.SS) - the risk, not the trade

Why it matters: Their capacity ramp is the single most likely trigger for the next downcycle. Watch CXMT’s DRAM share and DDR5 progress and YMTC’s NAND share as the clock on the supercycle.

Risk controls

The memory cycle always turns; current margins above 70 percent should be treated as cycle-peak, not run-rate. The clearest structural threat is Chinese oversupply, concentrated in commodity DRAM and NAND and expected to bite in 2027, which argues for favoring HBM-levered and picks-and-shovels names over pure commodity makers. Demand leans heavily on a handful of hyperscalers spending more than a trillion dollars a year, so a capex pause or a double-order unwind would hit memory first and hardest. Many names have already run 200 to 600 percent into the cycle, so valuation is a real risk and the high-torque small caps should be sized as options. Geography is concentrated - roughly 60 percent of all memory is made in South Korea, with a meaningful share of NAND capacity inside China - so export controls, the proposed MATCH Act, and Taiwan risk all bear on the chain. HBM leadership can shift with each generation, as Samsung’s HBM3E stumble showed, so base-die, packaging, and yield execution decide the HBM4/HBM4E winners. And many of the best pure plays are listed in Korea, Tokyo, Taipei, or mainland China with thin ADR or OTC access - confirm the tradable line and currency exposure before sizing anything.

Methodology, sourcing, and data-quality flags

This report synthesizes live web research (June 2026) across five streams - the HBM and AI-memory story, the memory manufacturers, the underlying technologies, the equipment/test/controller supply chain, and the demand/macro outlook - prioritizing company filings and earnings, TrendForce, Gartner, Yole, Counterpoint/Omdia market-share trackers, JEDEC standards, and credible trade press. Load-bearing figures were re-verified against primary sources.

Data-quality flags:

Market-share percentages bounce quarter to quarter and differ by analyst. The safe framing used here: SK Hynix is the clear HBM leader (~57-62 percent); Samsung and Micron contest number two; Samsung leads commodity DRAM and NAND.
HBM market-size forecasts vary. Micron’s roughly $35 billion in 2025 growing toward $100 billion by 2028 is the best-sourced; the 2026 figure is around $45 billion (some earlier estimates of $55-60 billion now look high), and a widely cited “$170 billion” figure refers to the total memory market, not HBM.
The wafer-intensity multiplier for HBM versus DDR5 is cited as roughly 3x (Tom’s Hardware) to 4x (TrendForce); both point the same direction.
Profitability nuance: SK Hynix overtook Samsung on full-year 2025 operating profit, but Samsung’s memory division margin was comparable or slightly higher in early 2026 - SK Hynix is the most profitable as a pure-play, not on every single metric.
Total memory market size differs by definition: Yole’s roughly $190 billion (2025) and WSTS’s roughly $295 billion (2026) use a “memory IC” basis; some houses publish far larger numbers using broader scopes.
The China oversupply timing (2027) and the AI-capex forecasts (>$1 trillion by 2027) are forecasts, not facts.
Scores are directional judgment, not a mechanical formula; the composite rewards quality and diversification, so it favors the equipment and test names over the highest-torque commodity makers - read the Upside column and the leaderboard for the pure plays.

Key sources: TrendForce (pricing, wafer intensity, HBM share, China), Gartner (“memflation” forecasts), Yole and WSTS (market size), JEDEC (HBM4 standard), Counterpoint and Omdia (market share), Micron, SK Hynix, Samsung, Kioxia, SanDisk, Western Digital, Montage, Rambus, Advantest, ASML, Lam Research, Applied Materials, KLA, TSMC filings and investor materials, and reporting from CNBC, Reuters, Tom’s Hardware, Blocks & Files, and Caixin.

Prepared June 13, 2026. Figures are point-in-time and will change. This is research and analysis for educational purposes - not investment advice, not a recommendation, and not a solicitation. Memory is a highly cyclical industry and involves real risk of loss. Verify all figures independently and consult a licensed financial advisor before making any decision.