Newsletter · · Ashutosh Agarwal

Micron's HBM4 More Than Doubles Bandwidth While Arm Says Phones Now Pay - HBM & The Memory Supercycle - Week of September 12, 2026

HBM & The Memory Supercycle for the week of September 5 to September 12, 2026: Micron's core data center chief described HBM4 as more than doubling bandwidth while naming construction labor as the real bottleneck across five fabs being built at once, Arm's CEO confirmed from the buyer's side that memory prices have nearly doubled and are denting smartphone demand, and a second memory-only ETF launched citing a projected 70% net margin for Micron in 2027.

HBM & The Memory Supercycle

Week of September 12, 2026: Micron's HBM4 More Than Doubles Bandwidth While Arm Says Phones Now Pay

Two of the people who actually build and buy this stuff sat down for long interviews this week. There was no earnings print and no fresh contract-pricing number, Micron's next report doesn't land until September 30. So instead of chasing prices, this week we get something rarer: operators talking, on the record, in plain English, about why the shortage exists and how long it takes to fix.

The short version: the executive running Micron's data-center business says HBM4 more than doubles the bandwidth of the last generation, and the real bottleneck now is pouring concrete, Micron is building five new factories at once and still can't find enough tradesmen. The head of Arm, who has to buy all this memory, put it bluntly: memory prices have "nearly doubled," it's making phones more expensive, and "if it's a bump, it's a really, really big bump."

Meanwhile Wall Street keeps building products to sell you the theme, a second memory-only ETF launched this week, even as the old worry ("this is a classic overbuild that ends in tears") got a fresh, well-argued airing.

Let's dig in.

TL;DR

  • A memory maker is back on the record. Jeremy Werner, who runs Micron's Core Data Center business, gave the clearest operator description of HBM4 in weeks: bandwidth "more than doubled" versus the prior generation, energy "meaningfully" lower. The catch, the choke point is now building fabs, not designing chips. Micron has five under construction and can't hire fast enough. (Embedded Insiders)
  • The buyer's side confirms the pain. Arm CEO Rene Haas: memory prices have "obviously nearly doubled because of the AI data centers hoovering up all the supply," it's raising phone prices and hurting phone demand, and the industry is in "an absolutely supply-constrained environment... for a bit." (Big Boss Interview)
  • Retail keeps buying the story. A second memory-focused ETF launched (YRAM, after Kurv's KMEM last month). Its strategist flags Micron's projected 2027 net profit margin of 70%, "I don't think anybody could have forecast this," while warning those margins will eventually "mean revert." (Dividend Stockpile)
  • The bear case got a good week too. A widely shared "prisoner's dilemma" episode laid out the classic overbuild autopsy: fiber went 85% dark, railroads bankrupted their builders, and today's data centers are worse because half the cost is GPUs that go obsolete in roughly 5 years. (Capital Decanted)

What's new

1. A Micron operator finally explains HBM4, and the real bottleneck is a construction site

Podcast: Embedded Insiders, "The Basis of Battery Systems & Hitting The Memory Wall" (Sept 10). Speaker: Jeremy Werner, SVP and GM, Core Data Center Business Unit, Micron.

This is the most valuable thing on any podcast this week, because it's an actual memory-maker executive walking through the technology and the capacity plan, not a pundit guessing at it. Werner was recapping his keynote at the Future of Memory & Storage conference (the show formerly known as Flash Memory Summit).

On why the shortage exists at all, the "memory wall":

"The big change with integrating AI is that the bottleneck really shifts from computation to memory, whether it's bandwidth or capacity... the rate of flops [in a GPU]... has grown at a much faster rate than the ability of memory to service data to the components. So that creates a widening gap that props up the wall."

On HBM4, the new generation everyone's waiting on, this is the first clean operator read on it in weeks:

"An example of that is our HBM4. HBM4 doubled the bandwidth from the prior generation or more than doubled and brings down the energy meaningfully, allowing the system to run faster."

Worth noting: he gives the performance spec, bandwidth and power, but not the number the market really wants, which is yield and qualification status.

He also described what "memory is the new architecture" means in practice: inside an AI inference system there are now "five different layers of memory and storage" that data (the "KV cache") moves up and down, and the HBM sits right on the same slab of silicon as the GPU logic, "that requires a deep collaboration and co-design." When the host teased him about lock-in, did integrating the memory that tightly force customers onto Micron, Werner gave the line of the week:

"No one has to use our memory, but usually the smartest people do."

The part that actually moves the thesis, though, is capacity. Asked whether sky-high memory prices will ease anytime soon, Werner laid out the build:

"We're actually at the moment building five fabs around the world. Two 600,000 square foot fabs in Boise, Idaho. And the first of what eventually will be four 600,000 square foot fabs in Clay, New York, as well as a fab in Taiwan, an extension in Singapore and an extension in Hiroshima, Japan."

And why that supply still takes years to show up, the bottleneck has moved from the cleanroom to the job site:

"Everyone was caught off guard by the pace of the demand growth as we go through what is like a modern-day industrial revolution... What we found is there's just not enough of the supply chain, of the labor force, of the know-how to be able to build these things as fast as we'd like."

He put the scale in human terms: each building is 600,000 square feet, "10 football lengths," and has to be clean enough that there are "only 10 particles per million atoms of air." Why it matters: this is the same message Sanjay Mehrotra and Jensen Huang have been giving for two months, the shortage is real, and the fix is gated by physical construction, not willingness to spend. It's a structural-bull data point, straight from a memory maker. But note the tell for bears: a company racing to open five fabs at once is, by definition, adding a lot of future supply.

2. The Arm CEO, a buyer, confirms memory prices nearly doubled and are hitting phones

Podcast: Big Boss Interview, "Arm CEO: AI Will Cure Cancer" (Sept 7). Speaker: Rene Haas, CEO, Arm, an operator on the demand and buyer side.

Haas matters here because Arm sits on the other side of the table from Micron and Hynix, it designs the chips that consume memory, and this year it started having its own server CPUs built (the Meta-led "ARM GI" CPU, now also going to Cloudflare, SK Telecom and Oracle). So he's now a memory customer, and he's feeling the squeeze firsthand:

"We are in an absolutely supply-constrained environment, which I think we will be in for a bit because the demand for artificial intelligence... it touches everywhere and everything."

The consumer pass-through, the "memory tax" reaching your pocket, stated by a buyer, not a pundit:

"Memory chip prices have obviously nearly doubled because of the AI data centers hoovering up all the supply. It's making smartphones more expensive. It's affected demand for smartphones as well."

Is this a passing spike or something structural? Haas:

"If it's a bump, it's a really, really big bump. Because the demand for chips and artificial intelligence and memory is at a scale we've not worked on before."

His reasoning is the one bulls keep coming back to: modern AI models have "trillions of parameters... that's a lot of memory." And he flagged the circularity risk that bears love, hyperscalers spending "hundreds of billions of dollars of CapEx" on data centers, numbers "that are somewhat circular in terms of driving overall demand." Why it matters: independent, on-the-record confirmation from the buyer's side that memory prices roughly doubled, that it's now denting phone demand, and that the supply crunch has staying power. For anyone tracking the handset OEMs, that "affected demand for smartphones" line is the one to underline.

3. Wall Street launches a second memory ETF, and quietly flags a 70% net margin

Podcast: Dividend Stockpile, "Memory Stocks Are Booming, Is YRAM the Right Way to Play It?" (Sept 11). Speaker: Mike Ko, strategist, YieldMax.

A month after Kurv rolled out a memory ETF (KMEM), YieldMax launched YRAM (the "Memory and Storage Portfolio Option Income ETF") on August 24. When two firms package the same narrow theme into retail products inside a few weeks, that itself is a signal worth noting. Top holdings, per Ko: Micron (largest), then Western Digital, Seagate, and SanDisk, with smaller positions in Everpure, Penguin Solutions and Silicon Motion.

Two data points stood out. First, the volatility, this basket carries a one-month implied volatility "in the neighborhood of nearly 60%," versus about 15 on the VIX: "the memory space has got volatility as a group about four times larger than that of the S&P overall." Second, the eye-popping profitability number:

"Let's take a look, for example, at Micron's margins... their anticipated 2027 net income margin, and that is net of taxes. It's 70%... I don't think anybody could have forecast this."

Ko's own view lands squarely in the middle of our standing debate. His bull case is the "railroads versus chips" durability argument, but with a twist that favors memory:

"Track beds, ties and rails tend to last longer than chips do... I don't think we're going to be seeing a lot of the chips that are in use right now still running in the data centers 20 years from now. They're going to get replaced much sooner... So will the memory."

In other words, unlike railroad track that you lay once, memory wears out and gets replaced, which stretches the demand cycle out for years (helped along, he argues, by robotics as a second wave). But his bear case is honest:

"As the pace of growth in the space slows, I expect... the supply and demand constraint that we currently have, that that's going to come in. And when it does... some of the stellar margins that some of these companies have been seeing recently will likely compress. I do expect margins to at least mean revert."

Why it matters: the 70% projected 2027 net margin is a useful anchor for anyone building a Micron model, and a reminder of how much "peak" is baked into current numbers. The proliferation of retail memory products is a late-cycle tell worth respecting.

The debate

With no fresh prints this week, the argument returned to first principles: is this a durable, multi-year shortage, or a classic capacity boom that ends the way they always do?

The structural (bull) case, from the operators: Werner and Haas are singing the same song. Demand caught everyone off guard; the fix is gated by years of fab and packaging construction, not by capex appetite; and AI's appetite for memory ("five layers," "trillions of parameters," KV cache) is different in kind from past cycles. When the people building and buying the product both say "supply-constrained for a bit," that's not nothing.

The cyclical (bear) case, sharpened this week by Capital Decanted ("The Prisoner's Dilemma: Investing in the AI Build-Out," Sept 8, note this was a best-of replay of an earlier episode, and it's about the whole AI stack, not memory specifically). Hosts John Bowman and Aaron Philbeck, with guest work from Sparkline Capital's Kai Wu, laid out the overbuild autopsy every memory investor should keep in a drawer:

  • The "prisoner's dilemma": rational players know cooperation (disciplined capex) is best, but nobody dares slow down. As Google's Larry Page reportedly put it, "I'd rather go bankrupt than lose the AI race." That fear is exactly what produces gluts.
  • The historical pattern: telecoms laid roughly 80 million miles of fiber; roughly 85% went dark; WorldCom, Global Crossing, Level 3 and Qwest went bankrupt or through massive restructurings. Railroads, at their 1872 peak, were roughly 6% of GDP and bankrupted their builders too.
  • The AI twist that makes it worse than fiber or rail: "over half the cost [of a data center] comes from GPUs, and GPUs have a useful life of... five years, plus or minus." Rail track lasts decades; fiber lasts a decade-plus; the newest GPU is obsolete in a few years. Adjust 2025's roughly $400B hyperscaler capex for that fast depreciation and Wu's math takes it from roughly 2% of GDP to "nearly 8% of GDP," "the highest by far in history."

The punchline of that framework is uncomfortable for memory bulls: in every past cycle, "the long-term winners are very rarely those initial builders of the network." The overbuild subsidizes customers (the next Netflix or Google), not the companies that laid the pipe.

Where the two sides actually meet: even the bears in the podcasts concede the physical shortage doesn't unwind this year. The fight is about 2027 to 2028 and about how much "peak" you're willing to pay for. Ko (a bull) openly expects margins to mean-revert; the Capital Decanted framework says the crash always comes eventually. Nobody credible is calling the top for the next few quarters, they're arguing about the exit.

Stocks in play

Micron (MU)

  • Bull: An operator (Werner) confirming HBM4 more than doubles bandwidth and that supply is gated by multi-year fab construction, five fabs going up at once. Projected 2027 net margin cited at roughly 70% (Dividend Stockpile). Top holding in the newly launched YRAM ETF.
  • Bear: Those same five fabs are future supply. A roughly 70% net margin is, almost by definition, a peak that mean-reverts. Options-market color this week had the stock selling off toward roughly $977 (The Options Insider Radio Network, Sept 10), volatile, richly valued on peak earnings.
  • Next catalyst: Micron FQ4 earnings on September 30, the single most important print of the coming weeks, and the first hard read on HBM4 ramp, pricing and FY2027 guidance.

SK Hynix (000660 KS and US ADR)

  • Bull: ADRs jumped roughly 7% on Sept 8 and pushed back toward the post-IPO high near $194.80; a Schwab technician noted "it doesn't seem like this whole memory crunch has gone away" (Schwab Network, Sept 8).
  • Bear: This week's coverage was pure price action, no fresh fundamental datapoint, allocation update, or pricing print from the company.
  • Next catalyst: Any confirmation of the HBM4 qualification timeline at NVIDIA; the next Korea-listed monthly export and pricing data.

SanDisk (SNDK), Western Digital (WDC), Seagate (STX)

  • Bull: All three sit in the top of the new YRAM basket, the NAND and storage side of the trade is now a packaged retail theme (Dividend Stockpile).
  • Bear: No fresh NAND pricing print this week; the "margins mean-revert" warning applies most sharply to the historically most cyclical names.
  • Next catalyst: NAND contract pricing direction; any high-bandwidth-flash (HBF) product news.

NVIDIA (NVDA), as the memory buyer

  • Bull: One analyst framed the memory squeeze as a solvable, temporary cost: HBM, Hynix and Micron "all matter, yes, but those are more temporary. Those things look like they're getting solved," and NVIDIA "will mitigate their way through that" (Market News with Rodney Lake, Sept 10). Free cash flow trajectory cited at $3.8B (FY23) to roughly $127B (trailing) to roughly $192B (FY27E), roughly 24x forward earnings.
  • Bear: That "temporary" framing is exactly what memory bulls dispute, last week's news was NVIDIA guiding gross margin down (75% to 72%), "mostly memory." If the memory guys have the leverage, NVIDIA's margin is the loser.
  • Next catalyst: HBM4 supply terms into 2027; whether the memory line item keeps pressuring NVIDIA's guided gross margin.

Read-throughs

  • Memory equipment (Advantest, BESI, Camtek, KLA, Lam, AMAT): Almost nothing memory-specific this week. KLA got a mention only as a technical trading setup, up roughly 55% year-to-date, with "the ASML news" flagged as a possible booster for KLAC and AMAT (Schwab Network). Werner did note Micron is expanding "back-end packaging production capabilities" alongside the five fabs, a small, indirect positive for advanced-packaging tool demand.
  • Packaging and substrates (CoWoS, hybrid bonding): The closest signal is Werner's description of HBM sitting "on the same substrate integrated into a product with the logic" and requiring "co-design," a reminder that advanced packaging is now inseparable from the memory story.
  • GPU makers (NVIDIA, AMD): Covered above under NVIDIA. The tension is unchanged: memory pricing power flows from the GPU makers to the memory makers, and last week's NVIDIA margin guide-down is the proof. Arm's own server-CPU push ("north of $2 billion" of demand, "double what you could supply") is a reminder the compute buildout, and its memory pull, is broadening beyond the usual names.
  • PC and handset OEMs facing rising memory costs: This is the clearest read-through of the week, and it came from a buyer. Haas: memory prices have "nearly doubled," it's "making smartphones more expensive," and it has "affected demand for smartphones." Consumer electronics is where the memory tax is now visibly biting end demand.

What changed vs last week

Last week (Sept 5 issue, "Memory is now eating Nvidia's margins") was driven by NVIDIA's earnings and the hard proof that memory pricing power was showing up inside NVIDIA's gross margin (guided 75% to 72%, "mostly memory"), with hyperscalers increasingly buying memory direct from the makers.

This week shifted in three ways:

  1. Operators came back, but the mix changed. We got a genuine memory-maker executive on the record (Micron's Werner), an upgrade from last week, when the Micron color was secondhand, plus a marquee buyer (Arm's Haas) confirming the price pain. But there was no NVIDIA-style hard number: no fresh gross-margin guide, no pricing print.
  2. The debate rotated back to the glut question. Last week's fight was "who captures the margin." This week the conversation returned to the oldest bear argument, the prisoner's-dilemma overbuild and the fast-depreciating-GPU math (Capital Decanted).
  3. Retail productization deepened. A second memory-only ETF (YRAM) launched within a month of the first (KMEM). That's the theme reaching the retail-packaging stage, historically a late-cycle, not early-cycle, phenomenon.

Net: no thesis-changing news, but the operator commentary reinforces the structural-shortage view, while the ETF launches and the well-argued overbuild replay keep the cyclical warning firmly in the room. The next hard number is Micron's September 30 print.