Newsletter · · Ashutosh Agarwal

Memory Is Sold Out Through 2028 and Still a Commodity - HBM & The Memory Supercycle - Week of August 29, 2026

HBM & The Memory Supercycle for the week of August 22 to 29, 2026. Podcast synthesis on veteran analyst Jim Handy reporting from the Future of Memory & Storage conference that HBM is sold out through 2027 with DRAM contract prices up 80 to 95 percent, hybrid bonding spreading from premium HBM into commodity NAND, high-bandwidth flash shipping real product, and a new memory-only ETF signalling the theme has reached retail.

HBM & The Memory Supercycle

Week of August 29, 2026: Memory Is Sold Out Through 2028 and Still a Commodity

Issue #8, week of Aug 22 to 29, 2026.

The most useful thing this week didn't come from a boardroom or an earnings call. It came from the show floor of a memory-industry conference, narrated by the analyst who has covered this business longer than most of the people trading it have been alive.

His message, in one breath: memory is sold out into 2028, the physics of making more is genuinely hard, and none of that changes the fact that it is still a commodity that will, eventually, crash. Hold both of those thoughts. This week's whole debate lives in the space between them.

TL;DR

  • The best read of the week is a conference recap, not a news event. Veteran memory analyst Jim Handy (Objective Analysis), fresh off the Future of Memory & Storage conference, says HBM is sold out through the end of 2027 and will "very likely" be sold out through 2028 within a couple of months, and that DRAM contract prices are up 80 to 95%. He also warns, bluntly, that it's "still a commodity market" and "there is probably going to be a crash at some point."
  • A new memory-only ETF just launched, which tells you the theme has officially reached retail. Kurv's Howard Chan calls the big three memory makers "the mines, not the picks and shovels" of AI, and confirms SK Hynix's last quarter (revenue up 250%, profit up 500%) and a 25 to 30% gap between SK Hynix's US-listed shares and its Korean ones.
  • The bears got louder as the stocks pulled back. AI-hardware names fell on the week, and this issue's pundits leaned hard on the classic memory-cycle warning: everyone is adding capacity at once, which is "a tell, not a plan."

What's new

1. The conference floor view: sold out to 2028, DRAM contracts up 80 to 95%, and "still a commodity"

Podcast: Grey Beards on Systems, "176: GreyBeards talk recent news at Future of Memory & Storage Conference with Jim Handy, Objective Analysis" (Aug 22). Speaker: Jim Handy, Objective Analysis, a memory and SSD industry analyst rather than a company executive, but the closest thing to an insider voice on tape this week. He was on the show floor in Santa Clara last week.

This is the episode to read if you read only one. Handy walks through what he saw at the Future of Memory & Storage conference, and it is the richest single hour of memory content the podcasts have produced in weeks. The headline numbers:

HBM has already sold out through 2027. And there is a high likelihood that it's, you know, in the next couple of months, it's going to be sold out through 2028.

And a rare, concrete pricing figure, the kind these podcasts almost never give us:

The DRAM contract prices are up as much as like 80 to 95%.

Why it moves the thesis. For weeks the "sold out" language has been round-number hand-waving ("10 to 15x," "sold out 2027"). Handy attaches it to a mechanism and a real contract-price range. His logic on why the shortage is broader than HBM is worth quoting, because it undercuts the simplest bear argument, that this is a narrow AI-HBM bubble:

Even if HBM didn't exist, we would have a DRAM and NAND flash shortage right now.

His reasoning: the hyperscalers are buying so much standard server memory (DDR5) and so much NAND flash alongside their AI gear that the shortage would exist anyway. Then HBM stole wafer capacity on top of that, and it "snowballed," since even the hard-drive makers (Western Digital, Seagate, Toshiba) couldn't keep up with AI storage demand, which pushed buyers toward SSDs, which worsened the flash shortage in turn. And the wafers going into NVIDIA's chips are eating into the same foundry capacity everyone else needs.

But Handy is nobody's permabull. His closing warning is the sharpest bear line of the week precisely because he's the one who just told you it's sold out:

It's still a commodity market. And I think a lot of people aren't recognizing that. And there is probably going to be a crash at some point where hyperscalers decide that, okay, we've built enough.

He reached for the same analogy the skeptics always do, the late-1990s telecom build-out, where fiber was laid at a thousand-percent-a-year pace and then, when the dot-com bubble burst, "98% of it was dark."

2. What "hybrid bonding" actually is, and why NAND stopped bragging about layer counts

From the same episode. Two things worth understanding:

On HBM packaging. DRAM chips have barely shrunk in a decade, so the way you get more performance out of a stack is to shorten the wires between the memory and the processor. Hybrid bonding is the technique for physically fusing chips face-to-face to do that. It's also the technique behind the more exotic ideas on display: Samsung's "Z-HBM" (stacking the HBM directly on top of the AI accelerator) and its "SAINT" process (putting an SRAM chip on top of a processor), which rhyme with AMD's 3D V-Cache. The catch Handy flags is thermal: when you stack chips, the hot one in the middle expands while its neighbors stay cool, so the whole stack bends "like a bimetallic strip," a brand-new category of mechanical failure to engineer around.

On NAND flash. This is the more surprising bit. The industry has gone quiet about layer counts (200+ layers, once a bragging point) for a physical reason:

People aren't talking too much about layer count in NAND flash anymore. And it's for a very strange reason. It's because the staircases are getting too big.

Each layer needs a "step" on a staircase so a wire can reach it; double the layers and you double the staircase's footprint. Worse, each layer adds electrical capacitance, which eventually demands transistors so large they, not the memory bits, dictate the chip's size. The fix, pioneered by China's YMTC as "Xtacking" and now adopted by SanDisk/Toshiba (BiCS-8, BiCS-10) and Western Digital/Kioxia, is to build the memory bits on one wafer and the drive transistors on another, then hybrid-bond the two together. In other words, hybrid bonding is showing up in commodity NAND, not just premium HBM. That's a read-through for anyone modeling packaging demand.

3. High-bandwidth flash gets real, and a new front in the KV-cache war

From the same episode.

A year ago "high-bandwidth flash," flash storage fast enough to sit near the GPU the way HBM does, was a slide. This year, per Handy, SanDisk and SK Hynix both showed real product and the industry agreed on a spec. The performance bar is startling: on the order of 100 million I/O operations per second at 512 bytes, on a single module. The prize is the "KV cache," the fast-access memory that lets a model remember the earlier parts of your conversation. Right now that job is fought over by three technologies, high-bandwidth flash, CXL memory pooling, and plain DRAM, and the winner isn't settled.

That fight matters for a subtle, bullish-then-bearish reason. Handy relayed a Meta paper describing something all the hyperscalers are quietly doing: instead of scrapping old servers, they strip the older DDR4 memory out of retired machines, drop it onto CXL cards, and bolt it onto their DDR5 servers to fill the supply gap they can't buy their way out of. It's clever engineering, and a reminder that when memory is this expensive, buyers get very creative about not buying more of it.

4. The theme reaches retail: a memory-only ETF, and "the mines, not the picks and shovels"

Podcast: Behind the Ticker, "Why the Real AI Bottleneck Isn't Compute | Howard Chan, Kurv Investment Management" (Aug 23). Speaker: Howard Chan, founder and CIO of Kurv Investment Management (MIT engineering, ex-Goldman, ex-PIMCO). An investor talking his own book, a pundit rather than an operator, but a specialist one.

Chan launched a dedicated memory ETF (Kurv Memory Select) on July 1, roughly 85% concentrated in the three big memory makers. When single-sector ETFs launch, the theme has gone mainstream, which is worth noting for positioning. His framing is the sharpest packaging of the bull case:

The phrase I keep hearing over and over again is the picks and shovels. But you really referred to [memory] as the mines instead of the picks and shovels of the AI trade.

His point: 90% of the world's memory comes from three companies (Micron, Samsung, SK Hynix), fabs cost $30 to $50 billion and take years, and the one company that makes the lithography machines (ASML) already has "years of wait list." He calls it a two-to-four-year thesis and confirmed a couple of hard datapoints from SK Hynix's latest quarter:

SK Hynix just reported earnings. Their revenue went up 250% and their profit went up 500%. So they are price setters instead of price takers.

Two other useful nuggets. First, on China as a threat: Chan says the new Chinese entrant is pricing its chips "at the same level as the Korean company," so it is not yet the cheap disruptor the bears fear. Second, a practical arbitrage note for anyone playing SK Hynix: its US-listed shares trade at a 25 to 30% premium to the Korean-listed ones, which is why Kurv buys the Korean line where it can.

The debate

This week the two sides were argued by different people than usual, a conference analyst and a fund manager on the bull side, a stack of macro and market pundits on the bear side, but the shape is the same one that's defined this whole cycle.

The structural, multi-year-shortage case. Handy and Chan land in roughly the same place: the shortage is real, it's broad (not just HBM), and it's hard to fix fast because fabs and lithography tools take years. Handy's "even without HBM we'd have a shortage" argument is the strongest version, because it means the demand isn't resting on a single fragile AI product line. Chan adds that the memory makers have, for the first time, pricing power and a genuine moat, turning a commodity into something closer to a specialty product. Both peg relief no earlier than 2028.

The classic-cycle, it's-still-a-commodity case. The tell of the week is that the most credible bull, Handy, is also this issue's most quotable bear: "it's still a commodity market… there is probably going to be a crash at some point." Justin Klein on InvestTalk put the trader's version more bluntly, and it's the line to sit with:

All of them are trying to add capacity right now. So that's a tell. That's not a plan. That's a tell that they know that this is going to be very short-lived… These are the type of names that look the cheapest at the top and look the most expensive at the bottom.

And on Geopolitical Cousins, BCR Research's macro strategist "Cousin Marko," an AI bull on the fundamentals, conceded the historical base rate is brutal, citing the book Engines That Move Markets:

Every CapEx cycle that was based on the new technology ended in tears… We have never, as humans, said we have built enough canals, we have built enough railroads, let's stop. We always overbuild.

His only caveat, and it's the crux: the data-center revenue is so strong right now that the reckoning has been "pushed off into late '27, if not '28." Which is, notably, the same horizon the bulls give for supply relief. Both camps now agree the physical shortage does not unwind this year. The entire argument is about what happens in 2028, and what the stocks are worth in the meantime.

Stocks in play

Micron (MU), discussed as one of "the three."

  • Bull: One of three firms controlling about 90% of memory, now with pricing power on a product (HBM) sold out into 2027 and 2028 and DRAM contracts up 80 to 95% (Handy; Chan).
  • Bear: "Still a commodity market" with an inevitable crash when hyperscalers stop building (Handy); the whole group is racing to add capacity, which the bears read as a top signal (Klein).
  • Next catalyst: Micron's fiscal Q4 and full-year report. The company's fiscal year ends in late summer, so the print typically lands in late September. Watch HBM revenue mix and any 2027 allocation commentary.

SK Hynix (000660 KS), last quarter's numbers re-confirmed.

  • Bull: Revenue up 250%, profit up 500% last quarter; "price setter, not price taker" (Chan). Showing real high-bandwidth-flash product alongside SanDisk (Handy).
  • Bear: Same commodity-cycle risk as the group; a Chinese competitor is now pricing at Korean levels (Chan), which caps the pricing umbrella over time.
  • Next catalyst: the 25 to 30% premium of the US-listed shares over the Korean line (Chan), a positioning and arbitrage signal to monitor. Also watch the geopolitics: on Geopolitical Cousins, the hosts flagged a widening US and South Korea rift ("we like North Korea more than South Korea") as a slow-burn risk for both Korean makers.

Samsung Electronics (005930 KS), technology rather than news.

  • Bull: Pushing the packaging frontier with "Z-HBM" (HBM stacked on the accelerator) and the "SAINT" SRAM-on-logic process (Handy).
  • Bear: Group-wide commodity risk; the same US and Korea political overhang.
  • Next catalyst: any HBM4 qualification milestone at NVIDIA.

SanDisk (SNDK), this week's bear lightning rod.

  • Bull: Showing real high-bandwidth-flash product with an agreed spec (Handy); a name Kurv has been adding on pullbacks (Chan).
  • Bear: The single most-attacked stock of the week. Klein: "This is by far the riskiest part of the market right now. And I would run far, far away… SanDisk went bankrupt a number of years ago because of how bad, how cyclical the industry is."
  • Next catalyst: NAND spot and contract pricing direction, and whether its long-term supply agreements actually hold if demand softens.

Western Digital (WDC) and Seagate, read-through names.

  • Bull: Storage is being "eaten up" across the data center; hard-drive makers can't meet AI demand, which spills incremental volume into flash (Handy); both are on Kurv's shopping list on weakness (Chan).
  • Bear: Purest commodity exposure in the complex.
  • Next catalyst: WD and Kioxia's move to hybrid-bonded (wafer-to-wafer) NAND, an execution and cost milestone to watch.

Read-throughs

Memory, test and packaging equipment (Advantest, BESI, Camtek, KLA, Lam, AMAT). The only equipment name that surfaced was ASML, and only as the archetypal bottleneck, with Chan noting it is the sole maker of the high-end lithography tools and already has "years of wait list." If you're long the equipment complex, the podcasts gave you narrative tailwind (everything upstream is constrained) but no company-specific datapoints.

Packaging and substrates (CoWoS, hybrid bonding). The best read-through of the week. Handy's explanation makes clear hybrid bonding is now a two-market story: it's essential for premium HBM stacks and it's spreading into commodity NAND (SanDisk BiCS-8 and BiCS-10, WD and Kioxia) as the layer-count staircase problem forces wafer-to-wafer bonding. That broadens the addressable demand for hybrid-bonding tools well beyond the AI-accelerator narrative. Separately, Handy flagged that the buildout is straining humbler parts of the supply chain too, with printed circuit boards (more layers, higher frequencies), power-management chips, and even ceramic capacitors, resistors and inductors "starting to become short on supply."

GPU makers (NVIDIA, AMD). NVIDIA is still the gravitational center, since its chips are pulling wafers away from everyone else (Handy), but the concrete color this week was on AMD. The Grey Beards discussion (with an AMD-side co-host) detailed the "Helios" rack: 72 GPUs, 18 sleds on AMD's "Venice" CPUs, drawing about a quarter-megawatt each and cooled by what he described as "four fire hoses" of liquid. And in a genuinely interesting sign of where the smart money is going, they noted AMD recently bought a memory-management startup that uses AI to find the "hot blocks" in memory and tier the cold data off to cheaper CXL or SSD, that is, software to use scarce, expensive memory more efficiently. When engineers start optimizing around a resource, it tells you how tight it really is.

PC and handset OEMs facing rising memory costs. Chan reiterated the now-familiar pass-through: Apple, MacBooks, iPhones, and even Xbox have raised prices because memory costs are up, and, the more important second-order effect, because the big three are shifting capacity to high-margin HBM, they are crowding out production of the commodity memory that goes into consumer electronics. His COVID analogy: car prices spiked then "not because there wasn't enough metal… but because there were not enough memory chips." Expect the squeeze to keep filtering into anything with a chip in it.

What changed vs last week

Last week (Issue #7, Aug 22) was a corporate week: Micron's CEO Sanjay Mehrotra broke cover with a CNBC exclusive and a $10B research-lab announcement, SK Hynix unveiled a $29B buyback, JPMorgan slapped a $1,550 target on Micron, and the stocks had roared back above $1,000, so the beat-and-sink pattern had finally broken.

This week the mood cooled on three fronts:

  1. The substance shifted from news to engineering. With no earnings and no corporate announcements, the value this week was technical, with the hybrid-bonding, NAND-staircase, and high-bandwidth-flash detail from FMS, plus a datapoint that the theme has reached retail (a new memory ETF).
  2. The tape softened and the bears got the microphone. AI-hardware names fell on the week, and the debate slid back from "how long do 80% margins last" toward the older, more existential question: is this still a commodity that will crash?
  3. Even the credible bulls now concede the endgame. Both Handy and Marko openly accept a crash is coming eventually. They just argue it's a 2028 problem, not a 2026 one.