Newsletter · · Ashutosh Agarwal

The Memory Tax Arrives, Strong Prints and Sinking Stocks - HBM & The Memory Supercycle - Week of August 8, 2026

Memory names beat and sold off across the board for the week of August 1 to 8, 2026, as podcast operators reframed the trade around a new idea: the memory tax that the whole AI economy now has to pay.

HBM & The Memory Supercycle

Week of August 8, 2026: The Memory Tax Arrives, Strong Prints and Sinking Stocks

HBM & The Memory Supercycle, Issue #6 · Week of August 1–8, 2026

The best-performing corner of the entire market just reported some of the best numbers you will ever see on a spreadsheet, and got sold anyway. SanDisk earned $39.25 a share against 29 cents a year ago, at an 84.6% gross margin, and the stock fell. Western Digital beat and guided up and fell harder. Seagate beat and raised last week and it barely mattered. If you want the one idea that ties this week's podcasts together, it's this: memory has become so expensive, so fast, that the market has stopped treating it as a growth story and started treating it as a tax: a levy the whole AI economy now has to pay, and one that investors are betting can't stay this high forever.

That reframe, the memory tax, is genuinely new this week, and it cuts both ways. To the bulls, it's proof memory is now the single most valuable ingredient in an AI data center. To the bears, a tax is exactly the kind of thing that gets legislated away by new supply. Let's get into it.

TL;DR

  • Beat-and-sink, again. SanDisk and Western Digital both crushed estimates and both fell: SanDisk ~16%, WDC ~12% on the print. The market's message: 84–85% gross margins are a bottleneck, not a business model, and bottlenecks ease.
  • "The memory tax." The week's dominant framing: on one macro podcast, the estimate that ~75% of 2027 hyperscaler capex will go to memory chips, making memory "way more valuable than GPUs" right now. Two on-record chip CEOs (AMD, AWS) confirmed "sold out" through 2027–28, but on their own silicon, not HBM.
  • China moves from shock to slow burn. After last week's CXMT IPO earthquake, the China story narrowed to a specific question, will Apple qualify Chinese DRAM maker CXMT, with two senators warning them off. Everyone agrees CXMT still can't make HBM.

What's new

1. SanDisk's blowout, and why "great" wasn't good enough. On the Schwab Network earnings desk (Aug 5), the analysts walked the numbers live as they crossed: revenue $8.97 billion (versus ~$8.48B expected), up 51% sequentially, and, crucially, "roughly one-third from higher volumes, two-thirds from higher pricing." Non-GAAP EPS of $39.25, gross margin 84.6%. Next-quarter guidance of $44–46 EPS on $10.3–10.8B revenue, a beat, and still the stock slid. Analyst Alex Coffey put his finger on why the good news didn't help: "When you've already sold out everything you can possibly make, that companies like Apple can't secure the supply they need... two-thirds of the pricing and the growth and the margins is driven by the fact that they are just raising prices. It's not because they're selling more stuff, because they can't make it fast enough to sell it." When the entire bull case is "sold out," there's no upside surprise left, only room to disappoint.

Bill Gunderson gave the eye-popping year-over-year math on Best Stocks Now (Aug 6): SanDisk's revenue went from $1.9 billion to $9 billion year on year (+372%), EPS from $0.29 to $39.25, with guidance for next quarter of $45.64 versus $1.22 a year ago. The stock is up ~469% year-to-date and roughly 3,077% over twelve months, and fell 6.5% on the day anyway. (Plain-English translation of a memory "spin-off nobody wanted" a year ago now printing a ~$200 billion market cap.)

2. The "memory tax": memory is now worth more than the GPU. The single most important new idea this week came from Andreas Steno Larsen on Real Vision's Macro Mondays (Aug 3):

"Next year, 2027... of the more than a trillion spent in CapEx, the projected spend in CapEx from the hyperscalers, around 75% of that will go to memory chips. So memory chips are currently, and I stress currently, way more valuable than GPUs."

His reasoning is worth spelling out because it's the crux of the whole bull case. AI agents, he argues, "need memory a lot more than they need logic", to make good decisions with lots of background information, a model has to hold that context somewhere, and that somewhere is memory. So as the "agentic economy" scales, the mix inside a data center keeps tilting toward memory and away from pure compute. Yet NVIDIA is priced as the stable blue-chip and Micron, SK Hynix and Samsung are priced as here-today-gone-tomorrow cyclicals. Larsen thinks that's backwards: he called it "the stupidest market I can recall having participated in," noting Samsung is on track to generate more than a trillion dollars of free cash flow over the next three years against a market cap of roughly a trillion (i.e., it could theoretically buy itself back in three years), and the market still prices it as a cyclical.

The counter came the same week from Pat Moorhead and Daniel Newman on The Six Five (Aug 3), whose episode was literally titled around "the memory tax." Moorhead's take: "I don't think there's a memory tax at all. This is really the memory folks finally able to make some money." His reminder, Micron ran a negative 33% gross margin back in 2023 and nobody cared, is the whole bull retort in a sentence. He also relayed the scale of the demand pull: the memory content of a gigawatt of AI infrastructure was about $10 billion 12–18 months ago, jumped to $30–50 billion, and "Jensen's throwing the idea that it could go to 80 to 100." Newman added the structural point that matters for durability: memory is "pivot[ing] from being truly a commodity to becoming strategic... which gives them increasingly more pricing power."

3. Two chip CEOs on record, both "sold out," neither a memory maker. This was a good week for operator voices, if not quite the memory-maker exec we keep wanting. On Bloomberg Talks (Aug 3), AWS CEO Matt Garman said Amazon's own AI chip business is "largely sold out through the end of next year for Tranium capacity," with capacity "already spoken through the end of 27 and even through much of 28" on five-year customer commitments, and "today demand still significantly outstrips supply." He also confirmed Amazon's $220B capex (up $20B) reflects "higher memory pricing," and that Trainium saves customers "20%, 30% off of their inference costs" versus NVIDIA GPUs, the clearest signal yet of why hyperscalers are racing to build their own silicon around the memory they're forced to buy.

On Squawk on the Street (Aug 5), AMD CEO Lisa Su, whose stock is up 142% this year against NVIDIA's 13%, said the data center business is growing "well over 100% as we go into 2027," that customers now plan "12, 24, 36 months in advance," and that AMD is investing across the supply chain so it can "supply at the much higher rate that our customers are asking for." She raised the total compute market forecast to "over $2 trillion... through 2030" (up from $300B three years ago, $500B last year), and flagged the shift toward "better than 1 to 1" CPU-to-GPU ratios as agents proliferate. Both CEOs describe the same world memory investors live in: demand outstripping supply, multi-year visibility, prices going up. Neither runs an HBM fab, so treat their "sold out" as corroboration of the demand tide, not a direct HBM read.

The debate

Is this a multi-year structural shortage, or a classic cycle about to roll over? This week the two sides were unusually well-argued.

Structural / buy-the-shakeout. Morgan Stanley CIO Mike Wilson, on RiskReversal (Aug 7), gave the most credible bull frame. His view: this buildout is "much bigger than the cloud buildout... also bigger than the mobile buildout," and we're only at "97, 98", not 99, 2000, in the dot-com analogy. The tell that gives him comfort is that the selloff is already doing the work: "multiples have derated 30% for the hyperscalers... 15% for the S&P 500. Semiconductor stocks have derated by 50%. So that gives me comfort that the market is well on to this." And critically on supply: "I don't hear that there's an excess of compute at the moment... We're hearing that there's a shortage of compute." He is emphatic there will eventually be a "massive hangover... malinvestment... bankruptcies", just a couple of years out, not now.

"There will be a massive hangover. There will be malinvestment here... But I just don't think that we're at that stage yet." (Mike Wilson, RiskReversal)

Cyclical / the bottleneck breaks. The clearest bear voice was the plainest. On The Morning Market Briefing (Aug 6), the host framed the WDC/SanDisk selloff not as a reaction to results but to fundamentals: memory makers are "making 85% gross margins... they're just kind of taking advantage of a bottleneck. And I think that bottleneck is going to ease because that's what the market does. When the price goes up, demand goes down." He'd be happy to see memory profits normalize.

The sharpest structural bear case came on Smart Investing with Brent & Chase Wilsey (Aug 1), whose episode was titled "Chip Deals May Not Be Secure." Their argument attacks the load-bearing beam of the bull case: the long-term contracts. Lock a customer in at $100 a unit, they said, and if spot falls to $20, "that is not a good working relationship", customers will honor the letter of the deal and then walk, exactly as advertising and supply contracts got quietly renegotiated during COVID. Their timing warning is the one to internalize: "It's not when the crisis starts. It's when people start to feel that the crisis is starting that the stock falters." They noted Michael Burry used Thursday's rally to increase bearish positions in Micron, NVIDIA and a semis ETF.

And the credit-market bear, Jeff Snider on Eurodollar University (Aug 1), gave the cleanest capital-cycle logic of the week: "Record profits encourage record investment. Record investment increases supply. Increased supply eventually pressures prices and returns." He pointed to SK Hynix reporting a six-fold profit jump at 80%+ margins and falling 19%, because it also announced "at least $31 billion in capital spending... approximately 50% above its previous level," which a bear reads as "another supplier expanding capacity near the top."

Where they agree: the physical shortage does not unwind this year, new fabs are years away. The fight is entirely about the multiple and about 2028+.

Stocks in play

Micron (MU). Bull: near-monopoly on US-made HBM, on allocation to NVIDIA, riding a demand wave no one disputes. Mike Wilson calls it "appropriately valued" at ~5x: "the market fully understands that when it gets that crazy, you can't pay a full market multiple." Bear: on the same RiskReversal show, Guy Adami and Dan Nathan asked the uncomfortable question, "Why is it that Micron did not have the visibility to really start ratcheting up capacity like 14 months ago?", and flagged that Micron is "seeing gross margins they will never see again." Watch: fiscal Q4 earnings in late September; any sign of Micron finally adding HBM capacity (which would be its own bear signal on future pricing).

SanDisk (SNDK) / Western Digital (WDC). Bull: NAND is genuinely sold out; SanDisk's 84.6% margins and $9B quarter are real, and management has locked multi-year customer agreements. Bear: the reaction is the thesis: a beat-and-raise that still fell ~16% (SNDK) and ~12% (WDC) says expectations have finally caught reality, per Schwab (Aug 5). Watch: NAND contract pricing direction into Q4, and whether Apple's hedging (see below) shows up as softening demand.

Seagate (STX). Bull: the quiet winner, beat and raised last week with supply deals reportedly out to 2029, up ~200% YTD, per the Schwab desk. HDD is having an AI-storage moment. Watch: whether HDD strength holds as a tell for total AI storage demand.

SK Hynix / Samsung. Bull: Samsung's operating profit rose 1,814% year over year (TechLinked, Aug 1), and its memory VP said the shortage "will keep widening through 2027 and extending into 2028." Samsung is building a second Texas site with the windfall (per The Six Five). Bear: SK Hynix's 6x profit jump was met with a 19% share drop on capacity-expansion fears (Eurodollar University). Watch: Samsung's HBM share gains at NVIDIA; any Korean capex acceleration read as glut-seeding.

AMD / NVIDIA (demand side). AMD is guiding "well over 100%" data-center growth into 2027 and is raising GPU kit prices "by at least 10% starting in August," following NVIDIA (TechLinked). Watch: NVIDIA earnings at month-end, the tone-setter for whether the HBM demand pull is still accelerating.

Read-throughs

  • Memory equipment (Advantest, BESI, Camtek, KLA, Lam, AMAT): quiet this week. The only equipment thread was on Market Mondays (Aug 4), where the hosts flagged caution on ASML: both the China domestic-lithography scare and TSMC signaling it wants to slow machine orders to squeeze more efficiency out of existing tools.
  • GPU makers (NVIDIA, AMD): both raising prices into the memory-cost environment; AMD +142% YTD has now blown past NVIDIA +13% YTD (Squawk). The pricing power is flowing downstream from memory to accelerators.
  • PC / handset OEMs: the consumer squeeze is now explicit and named. On Big Technology (Aug 8), the hosts dissected "Apple's memory problem": Tim Cook complaining about memory costs while Apple raises device prices. The Six Five was blunter: Apple "could print 90% margins" and is using rising memory costs as cover to raise prices and increase net margin dollars: "opportunistic price gouging based on the fact that they feel they're having their prices gouged." Qualcomm raised prices with handsets down ~20%. TechLinked's line of the week: "in 2026, MSRP stands for memory shortage ransom pricing."

The China thread

Last week's CXMT IPO was an earthquake; this week it settled into a specific, slower question: will Apple qualify CXMT's DRAM for its China devices? On Best Stocks Now (Aug 3), Gunderson framed CXMT as "the fifth one in the world out there that makes memory chips," now building a second Beijing fab, but "severely constrained in the more difficult applications like high bandwidth memory. Now, this is where the big shortage is. This is where Micron and SanDisk eat their lunch, and SK Hynix is in HBM." His caution for the bulls: "Memory, even though in a major shortage now, is still a commodity... it will return to commodity status at some point in the future." On The Six Five, Moorhead noted CXMT is on a US Department of Defense watch list tied to the PLA, that two senators have warned Apple off qualifying them, and drew the historical parallel to Japanese steel and memory dumping. Net: China can pressure commodity DRAM, not HBM, for now.

The capex-and-credit backdrop (why the stocks moved)

The memory selloff didn't happen in a vacuum. Three threads matter for the book:

  • Debt is the new tell. Jeff Snider (Eurodollar University) laid out that AI-related debt issuance hit ~$270 billion by early July, roughly double all of 2025, with Amazon/Alphabet/Microsoft/Oracle/Meta accounting for ~$194B of it. Amazon's own $25B bond saw orders fall from $62B to $41B; a BlackRock/Meta Texas data-center bond priced at 7.53%; Oracle's CDS hit ~215bps (from ~145 at year-end). Alphabet posted its first negative free cash flow since 2004 (-$5.9B).
  • The circular-financing question. On Equity Mates, Loftus Peak's Alex Pollak noted NVIDIA is "on the hook for half a trillion dollars worth of guarantees" for OpenAI chip purchases, and that five hyperscalers carry ~$1.6T of commitments due over the next 3–5 years. His two-year visibility keeps him constructive, and he made the bulls' core point cleanly: "The compute becomes more important than the models. That is 100% correct."
  • Accounting sleight of hand. Both Real Vision and Big Technology dug into Microsoft extending its data-center useful-life assumption from 15 to 25 years, which shifts spend off the capex line and made Microsoft look disciplined (its stock had a record day), even as skeptics note "GPUs are not going to last 25 years" and chips are ~50% of build-out cost.

What changed vs last week

Last week's issue was the blockbuster: SK Hynix's record quarter, a violent round-trip, the CXMT IPO shock, and Jensen Huang on record. This week was the digestion: the last two memory names in the reporting cluster (Western Digital, SanDisk) beat and still sold off, confirming that the price-versus-stock decoupling is now the defining feature of the trade. The framing sharpened from "is the valuation real" into the crisper "memory tax": memory as the dominant slice of 2027 capex. We got two on-record chip CEOs (Su, Garman) but no memory-maker executive this week (a step back from Jensen last week and SK Hynix's Chey two weeks ago). And China shifted from acute IPO panic to the slower-burning Apple-qualification question. Michael Burry is back to shorting the group.