Newsletter · · Ashutosh Agarwal

Memory Is Now Eating Nvidia's Margins - HBM & The Memory Supercycle - Week of September 5, 2026

HBM & The Memory Supercycle, Issue #9, for the week of August 29 to September 5, 2026: Nvidia guided gross margin down to the low 70s and blamed memory, a TechInsights analyst put DRAM up more than 200 percent year-on-year, and the bear case finally found a supply number with a Chinese flag on it.

HBM & The Memory Supercycle, Issue #9

Week of August 29 to September 5, 2026: Memory Is Now Eating Nvidia's Margins

Here is the sentence that should stop you cold this week. The most powerful company in the history of technology just told Wall Street its gross margin is going down, and blamed it, mostly, on memory.

For a year the story has been simple: AI needs memory, memory is scarce, so Micron and SK Hynix and Samsung print money. This week the story got more interesting. It turns out the memory makers now have enough leverage to squeeze Nvidia itself. Jensen Huang went on Bloomberg and said out loud that "the bottleneck is still in memory, pricing and supply." His own margins are the proof.

And in the background, the bears finally found a number that scares them, not a demand number, a supply number, and it has a Chinese flag on it.

TL;DR

  • Nvidia is now paying the memory tax. On its earnings call it guided gross margin down from ~75% to ~72% and said it will stay in the low 70s for another three to four quarters, "mostly memory." Translation: the memory guys have pricing power over the most important customer on earth (The Circuit).
  • A real memory analyst went on the record. Dan Kim of TechInsights (former chief economist at the US Chips Program Office) says DRAM prices are up over 200% year-on-year, calls it "historically unprecedented," and says the market is now so tight that "even when you're willing to pay the price," makers can only fill a fraction of your order (The Economics Show).
  • The bear case got a supply number. Paul Kedrosky says two big Chinese memory makers each want to add "a Samsung's worth of capacity" by early 2028, a "tsunami of supply" he compares to the solar-panel crash. It's why the stocks trade badly even as the numbers are spectacular (RiskReversal Pod).

What's New

1. Nvidia Admits It Can't Fully Pass On the Memory Bill

Podcast: The Circuit, "EP 189: NVIDIA and Marvell Earnings, Hot Chips Hot Takes" (Aug 31), with the show's semiconductor analysts, deeply plugged into the supply chain but analysts rather than operators.

This is the most important read-through of the week, and it comes straight out of Nvidia's own numbers. On its earnings call Nvidia guided gross margin down from 75% to 72% for its January quarter, and said margins would stay "in the low 70s for the next three, four quarters." The hosts were blunt about the cause: "They said it's partly mix and it's partly memory. It's mostly memory."

Why does that matter? Because Nvidia has always been the company that passes its costs straight through. As one host put it: "It sounds like they're not able to pass on as much memory markup as they used to. So even they're feeling the pressure there. So the memory guys have the leverage."

There's a second, subtler shift underneath it. The hyperscalers, Amazon, Google, Microsoft, are increasingly buying memory directly from Micron, SK Hynix and Samsung rather than letting Nvidia buy it, mark it up, and bundle it into a rack. That high-margin memory markup business is quietly leaking away from Nvidia: "they're just seeing more of the memory business sort of go around them. And that's really high-margin business for them."

Why it moves the thesis: For a year the bull case for memory has been abstract ("AI needs bits"). This is concrete. The single most powerful buyer in tech is eating a two-to-three-point margin hit because it can't push all of the memory cost onto its customers. That is what real pricing power looks like, and it is showing up in someone else's income statement.

2. A Memory Analyst Puts Hard Numbers on the Mania

Podcast: The Economics Show, "Memory chip mania isn't going away, with Dan Kim" (Sep 4), with Dan Kim, chief strategy officer at TechInsights and former chief economist at the US Commerce Department's Chips Program Office. He is an analyst, but the closest thing to an insider the tape has offered in weeks.

This was the single most useful memory episode of the week, a whole show, by someone who does this for a living, in plain English. Highlights:

  • How crazy is it, 1 to 10? "Right now it's an 8. Next quarter it'll be a 9, and the quarter after that it'll be a 10. It'll stay at a 10 until the end of 2027 at least."
  • DRAM prices: "from last year to this year, we're anticipating an over 200% price increase of DRAM. That is historically unprecedented."
  • What "sold out" actually means now: the price signal has stopped working. "The price signals are so high that even when you're willing to pay the price, you might actually get a reply back from the manufacturer saying, we could only give you X% of what you're looking for." That squeeze has spread from AI data centers down to "the legacy demanders, automotives, medical device manufacturers."
  • How tight? "Effectively there is no supply out there that hasn't already been spoken for. So practically most of the supply has been sold out", for this year, and increasingly for next.
  • Why no relief until 2028: memory makers cut production in the 2023 down-cycle (when they were losing money), right as ChatGPT demand was about to explode. Building new capacity from scratch "could take somewhere between 2 to 6 years." Everyone, Korea, Japan, the US, China, is now "racing as quickly as humanly possible," but the new fabs don't come online "until late next year or early 2028."
  • The consumer hit, quantified: in a smartphone, memory used to be "somewhere between 15% and 25%" of the bill of materials. Now "it has the potential to be somewhere between 30% and 40% and even 50%."

Why it matters: This is the cleanest framing yet of why this cycle is different from every prior memory head-fake. Demand is real, supply was deliberately cut, and new supply is physically years away. Kim's "it'll be a 10 until at least end of 2027" is about as close to a dated call as the industry gives.

3. Jensen and MediaTek: Memory Is The Bottleneck, and the Fix Is Custom

Podcast: Bloomberg Tech, "Special Edition: Nvidia CEO Jensen Huang & MediaTek CEO Rick Tsai Talk New Partnership" (Aug 31), with Jensen Huang (Nvidia CEO) and Rick Tsai (MediaTek CEO), both operators speaking for their own companies. It also aired as Bloomberg Talks.

The biggest operator on the tape confirmed the choke point unprompted: "the bottleneck is still in memory, pricing and supply." Rick Tsai backed him up from the manufacturing side: "supply chain does present challenges, you know, not just HBM, gold, gold down to substrate."

Two things worth a portfolio manager's attention:

  • Nvidia's own HBM is going custom. "Our next generation HBM is a custom HBM. And the HBM stacks are built on an Nvidia custom base layer", a base layer Nvidia is now extending to its NVLink Fusion partners (like MediaTek) so their custom chips can be stacked "into a CoWoS package." This is the same theme the memory makers themselves are pushing (more below): HBM is drifting from a commodity part toward a semi-custom, design-locked component.
  • The content-per-gigawatt number keeps climbing. Huang laid out Nvidia's economics per gigawatt of AI factory: Hopper was ~$18 billion, Grace Blackwell ~$25 billion, and the new Vera Rubin generation "about $40-plus billion per gigawatt." More silicon per gigawatt means more memory per gigawatt. The demand pull isn't flattening, it's steepening.

4. China Stops Being a Footnote

Podcasts: TechLinked, "CXMT Beats Samsung to LPDDR6" (Sep 1) with host James Shrive, and Best Stocks Now with Bill Gunderson (Sep 1), an RIA talking his own book.

Two separate China datapoints landed in the same week, and together they're a genuine change of tempo:

  • LPDDR6: China's CXMT "officially started mass production of LPDDR6, the next generation of smartphone RAM, beating both Samsung and SK Hynix to market", with the chips going into a Xiaomi foldable. The caveat is real and worth keeping: the phone "isn't even expected to ship 500,000 units," and US trade restrictions force CXMT to use "older manufacturing equipment that lowers yields." So it's flag-planting more than a flood, for now.
  • HBM3E: separately, CXMT "has started producing small quantities" of HBM3E, the high-bandwidth memory used in AI chips, "with plans to expand their production in 2027" (Gunderson, citing The Information). Gunderson's own read stayed measured: SK Hynix "remains Nvidia's main HBM supplier," and Micron is "expanding much faster than CXMT, because CXMT is starting from a standing start."

Why it matters: China can't make leading-edge HBM at scale yet. Export controls and yield are real walls. But "small quantities of HBM3E" and "first to LPDDR6" are the first cracks in the "China is years behind" story. Which sets up the debate.

The Debate: Structural Shortage, or the Setup for a Glut?

This week the two sides finally argued about the same thing, supply, instead of talking past each other.

Bull, the shortage is physical and years long. Dan Kim's "10 until end of 2027" is the anchor: demand is real, supply was cut, new fabs are 2–6 years out. Gavin Baker, the well-known tech investor, went further on The a16z Show (Aug 31): "Everybody's worried about oversupply. I'm like more worried about undersupply. Massively undersupply", through 2028, with "no capacity available" even on the builds currently forecast. His provocative corollary: prices to access AI could rise, not fall. Daniel Pilling of Sands Capital, on Pitch The PM (Sep 2), said memory is one of his firm's two biggest AI holdings, DRAM, NAND, spinning disk and the equipment makers, precisely because "memory is a commodity" that "could reprice," and it did, up "4X at least year to date."

Bear, the discounting mechanism already sees the flood. Paul Kedrosky, on RiskReversal Pod (Sep 4), made the sharpest bear case in weeks, and, crucially, it's a supply argument, not a demand one. He showed client data suggesting "some of the largest Chinese memory manufacturers, two of them," each "essentially want to add a Samsung's worth of capacity over the next 12 months… between now and early 2028." He thinks the market is already discounting that: "I see this tsunami of supply coming to market… kind of what happened to solar panels… this really export-driven collapse in prices as a sovereign sees it in their best interest to sort of take over the market. But in this case, it'll be in memory." That, he argues, is why the memory names and the SOX "trade horribly" even as the companies beat and raise. The tape is looking through today's shortage to tomorrow's glut.

There's a second bear leg that isn't about chips at all, it's about how the buildout is financed. Kedrosky and Dan Niles (on Excess Returns, Sep 3) both hammered the same point: more than half of hyperscaler data-center spend is now externally financed, off balance sheet, through SPVs, through debt. Niles watches hyperscaler credit-default swaps every morning ("the cost of insuring that debt") and notes even Nvidia's CDS now trade above the average North American investment-grade level. His historical caution cuts both ways, though: in 1999 the Nasdaq rose 86%, then another 24% in early 2000, so "you can know we're in a bubble, but still have a lot of opportunity to make money before that bubble breaks."

Where they agree: nobody, not even the bears, thinks the physical shortage unwinds this year. The whole fight is about 2028 and about what multiple to pay today for earnings everyone admits are near a peak.

Stocks in Play

Micron (MU)

  • Bull: the only US memory maker, gaining HBM share, and still absurdly cheap on forward numbers. Gunderson pegged it at roughly a "six forward multiple" on ~$155 of next-year EPS, and noted the stock has "more than tripled this year." Kedrosky's own aside: Micron's stock ran "from $100 to $1,200" in a year (before coming off ~25%).
  • Bear: it's a commodity maker at a cyclical peak; if Kedrosky's China "tsunami" is right, today's 85%-type margins are a memory, not a run-rate.
  • Watch: any dated contract-pricing print, and Micron's Idaho fab timeline (Kim named it as one of the few new-supply sites, and it doesn't help until late 2027 or 2028).

SK Hynix

  • Bull: still "Nvidia's main HBM supplier" (Gunderson); the HBM share leader as Nvidia moves to custom base-die stacks.
  • Bear: Korean-retail leverage remains the swing factor. Niles again flagged how "SK Hynix and Samsung imploded" earlier in the cycle when leveraged retail unwound.
  • Watch: whether Hynix keeps the top HBM4/HBM4E slot as Nvidia's custom-HBM design-lock deepens.

Samsung Electronics

  • Bull: the natural share-gainer as the industry goes custom; deep pockets to out-build.
  • Bear: the one who has to prove it can hold HBM share against Hynix.
  • Watch: Samsung's own custom-base-die HBM disclosures (it and Hynix gave platform detail at Hot Chips; Micron pointedly didn't).

SanDisk (SNDK) and NAND

  • Bull: Bernstein's Mark Rubin, on Bloomberg Surveillance (Sep 4), values SanDisk on "11 times forward earnings" versus a stock "currently trading around five times", and notes "all these memory companies are trading three, four, five times forward earnings right now."
  • Bear: same problem as always. "Earnings have been doubling every quarter," so nobody knows the normalized number. Rubin's honest fix: 11x through-cycle earnings, ~9x peak earnings.
  • Watch: NAND is a step behind DRAM in this cycle; the read-through from any DRAM price roll would hit here first.

Nvidia (NVDA)

  • Bull: still put up a ~$96–100 billion quarter and guided full-year revenue growth of ~70% (The MoneyFlows Show, Sep 3); trades below a market multiple (~17.5x earnings, per Kedrosky) despite that.
  • Bear: the margin guide-down is the tell. Memory and the rest of the supply chain now take a bigger slice, and the high-margin memory-markup business is leaking to hyperscalers buying direct.
  • Watch: whether that low-70s gross margin is really a floor, or the first step down.

Read-throughs

Memory equipment, by name. The MoneyFlows Show (Sep 3) built a whole episode around three wafer-fab-equipment names as the memory build-out plays, with numbers:

  • Applied Materials (AMAT): down ~37% from its high, forward P/E below 25; Q3 revenue "just above $9.1 billion" (a beat), Q4 EPS guide $3.82–4.22 vs ~$3.71 Street; management flagged "leading-edge foundry logic, DRAM, advanced packaging" as the fastest-growing segments.
  • Lam Research (LRCX): ~$290, P/E under 30; raised Q1 revenue guide to "$8.1 billion plus or minus $400 million" against a Street ~$7 billion; investing "more than $3 billion" to expand R&D. The one-liner: "more memories, more layers, more complexity… that means more LAM."
  • KLA (KLAC): ~$168 and in a downtrend, forward P/E ~29; guided gross margin to stay in the 60–65% range and pegged 2027 industry WFE spending at "around $190 billion" (~20% growth). KLA does the metrology and inspection: "more complexity means more inspection."

Packaging and substrates. The custom-HBM theme is now the packaging theme. Nvidia's HBM stacks onto a custom base layer "into a CoWoS package"; the memory makers (Samsung, SK Hynix at Hot Chips) are moving to custom base dies, which The Circuit flagged as "a thermal problem… hotspots" and, importantly, a design-lock: Micron's CEO reportedly argued it's "harder for one vendor to support three memory players" once you customize the base die that deeply. Watch this, it could turn HBM from a three-way commodity race into customer-by-customer sole-sourcing.

GPU makers. Covered above, the memory tax is now visible inside Nvidia's margin. AMD's relative advantage, per The Circuit, is that it "might not have to increase prices… as much given what they've already secured for a two-year ramp."

PC and handset OEMs. The clearest consumer read yet. Bernstein's Rubin: memory prices rose "60%, 70% quarter on quarter" for two straight quarters, and "DRAM and NAND are both something like 400%, 300% up in one year." Apple "cannot not increase the price of the iPhone… it has to increase them pretty significantly," having already raised Macs and iPads; expect a "widening of the price spans," with a foldable pushing the high end toward ~$2,500. Dan Kim's bill-of-materials math is the same story from the cost side: memory heading from ~15–25% of a phone's parts cost toward 30–50%.

What Changed This Week

  • Operators came back on the record. Jensen Huang and Rick Tsai, both CEOs, both naming memory as the bottleneck.
  • A dedicated memory-analyst interview landed (Dan Kim, TechInsights) with hard, plain-English numbers: DRAM up 200% year-on-year, "an 8 going to a 10," sold out through 2027.
  • Equipment names showed up by name in a memory context (AMAT, LAM, KLA) for the first time in this newsletter's run, even if the source is a retail-flow show rather than an equipment executive.
  • China went from footnote to storyline. CXMT first to LPDDR6 mass production and small-quantity HBM3E, plus Kedrosky's "two Samsungs of new capacity by 2028" supply-flood thesis. The debate has a supply number now, not just a demand narrative.
  • The margin proof arrived. Nvidia's own guide-down is the hardest evidence yet that the memory makers hold the pricing whip. Note the sourcing caveat on the custom-base-die detail: the Micron comment reached the tape secondhand, not from a memory-maker executive directly.