Newsletter · 2026-10-10 · Ashutosh Agarwal

Micron Posts Record Margins as the Memory Supercycle Debate Heats Up - Semiconductor Podcast Briefing - Week of October 9, 2026

Semiconductor Podcast Briefing for the week of October 3 to 9, 2026. Podcast synthesis on Micron's record 87% gross margins and the memory supercycle debate, the trillion-dollar AI capex question, surging chip-equipment spend, and the debt financing the buildout, ahead of ASML and TSMC earnings.

Semiconductor Podcast Briefing

Week of October 9, 2026: Micron Posts Record Margins as the Memory Supercycle Debate Heats Up

Micron just reported a quarter with $54 billion of revenue and gross margins of roughly 87%. Samsung's operating profit rose about ninefold to around $80 billion. TSMC's sales were up 51%.

Then the shares fell. Samsung slid after its results, Micron opened lower and TSMC dropped more than 3% in the week before it reports.

That gap is what the podcasts kept circling this week. The numbers coming out of the chip industry are the biggest it has ever posted. A growing group of voices is asking how long they can last, who is paying for it all, and what happens when the borrowed money runs out. The bulls have the numbers on their side right now. The skeptics have better questions than they did a month ago.

TL;DR: Things That Mattered This Week

• Memory is the story. Micron posted about $54B in revenue, $32.87 in earnings per share and margins around 87%, and guided to $61B and $38 a share for next quarter. All of the growth came from higher prices, not from selling more chips. The Six Five with Patrick Moorhead and Daniel Newman
• The sell side keeps raising its targets. DA Davidson's Gil Luria set a $3,000 price target on Micron, about 3x the current price. He argues it trades at about 6x earnings while chip peers trade at 20x to 40x+. Squawk on the Street
• Bernstein's Stacy Rasgon says AI chip demand is "off the charts." Nvidia and Broadcom are talking about 70% to 100%+ growth next year. He sees the top five cloud giants spending "probably… a trillion and a half dollars next year." Monetary Matters with Jack Farley
• Chip-equipment spending could nearly double in two years. Rasgon forecasts about $154B of wafer-fab equipment spend in 2026, about $204B in 2027 and $250B+ in 2028. Monetary Matters with Jack Farley
• SpaceX is reportedly borrowing about $40B to buy Nvidia chips. Bloomberg counts about $360B raised so far in 2026 for data centers, chips and power, roughly 3x last year's pace. Bloomberg Tech
• Marvell doubled its ambitions. CEO Matt Murphy said sell-side estimates now imply over $30B of data center revenue in 2028, against the $15B target the company set two years ago. Squawk on the Street
• The cycle warning is getting louder. Fidelity's Jurrien Timmer says semiconductor earnings are up 178% year over year, and the index's roughly 40-month earnings cycle is "right at the part where it should start to roll over." The Compound and Friends
• Next week matters: ASML reports Wednesday, Oct 14, before the open. TSMC reports Thursday, Oct 15.

1. AI Chip Demand and Cloud Spending: "Nobody Has Enough Compute"

The most bullish and best-informed voice this week was Stacy Rasgon, Bernstein's senior analyst for US semiconductors and chip equipment, on Monetary Matters (Oct 8). He was blunt:

"Demand is off the charts… Nobody has enough compute."

Here are the numbers he gave:

• Nvidia and Broadcom have said they can grow 70% to 100%+ next year. Broadcom "suggested they could even grow 100% again in 28."
• He puts Broadcom's AI revenue at about $115B next year.
• He expects the five largest cloud providers (the "hyperscalers") to "probably spend a trillion and a half dollars next year."
• Nvidia guided to about 70% rather than 100%. Rasgon says that reflects a shortage of land, power and buildings, not a shortage of demand: "They're not guiding higher than that because they're not sure if they can get it."
• Jensen Huang's earlier forecast of $3–4 trillion a year in AI infrastructure spending "doesn't sound so crazy anymore."

He also pushed back on the idea that customers are hoarding chips:

"Nobody's really stockpiling AI hardware. Everything that's actually getting built and deployed is being used immediately, right? So that's a good sign."

Monetary Matters with Jack Farley

Stephanie Link (Hightower Advisors) gave the same picture on CNBC's Halftime Report (Oct 9). She sees hyperscalers spending about $1 trillion this year and $1.2–1.6 trillion next year. She noted that Nvidia has made $99B of equity investments (versus $7B a year earlier) and signed supply agreements worth almost $400B:

"There's no indication whatsoever that these companies are going to reduce spend… They are not decelerating."

Halftime Report

Marvell's CEO made the case for the "connective tissue." Matt Murphy argued on Squawk on the Street (Oct 7), the day after Marvell's investor day, that AI spending has moved in waves:

"You had the three big compute companies… That was followed by a huge memory wave… what we're seeing now is there's a gap because there's not enough in the connectivity layer to connect all of those things."

• Two years ago Marvell talked about a $15B data center business for 2028. After the investor day, he says sell-side estimates imply more than $30B, "fundamentally 2X'd."
• Marvell cites a $235B custom-chip market by 2030, up from $40B when it framed the opportunity in April 2024. "Custom chips" means processors designed for one cloud customer rather than sold off the shelf.
• Custom revenue is targeted at "$12 billion plus" in 2028 and $30B in 2030.
• Next year's revenue view was raised from $18B to $20B.
• On outside risks: "we see no impact from any of these externalities right now on our market potential."

Host Scott Wapner noted the stock has "more than tripled this year." He also relayed that Morgan Stanley questioned whether Marvell needed to set the bar this high. Squawk on the Street

Nvidia's own numbers on AI usage. On a sponsored Heavy Strategy episode (Oct 6), Nvidia's VP of AI Factory said the company itself is processing "trillions, trillions of tokens a month now and growing 40% month over month." A "token" is a unit of text an AI model reads or writes. He said agent-style workloads now involve inputs of about 142,000 tokens, compared with about 1,000 in 2023. Treat this as company promotion, but it explains why agents need so much more hardware. Heavy Strategy

Bloomberg equities reporter Ryan Vlaselica described how investors are thinking. Agent products are a "hot potato" with no clear winner, so investors are buying the companies that supply the hardware instead:

"People are expecting a lot more demand for chips, for memory, for networking, for optical components, for all of the different parts of the ecosystem that make up AI infrastructure."

Bloomberg Tech

The dissenting view: growth is no longer accelerating. Goldman Sachs chief US equity strategist Ben Snyder said on Squawk on the Street (Oct 8):

"AI CapEx growth will remain strong, but it's not accelerating anymore."

He estimates about half of S&P 500 earnings growth comes from the AI buildout. He still thinks "the data infrastructure stocks are quite attractive." Squawk on the Street

Power is the practical limit. Rasgon called power "a huge constraint." He said the backlash against data centers is "a really good political talking point" heading into the midterms, but "I don't think it's slowing anything down." Monetary Matters with Jack Farley

Former Intel CEO Pat Gelsinger, on The a16z Show (Oct 9), was more pointed:

"Why build a new data center and buy, you know, the million GPUs if I can't power them… I think you're going to see more and more defaults happening on many of those data center projects because the energy won't be there."

Gelsinger said gas turbine lead times are now about eight years. In his words, "energy capacity somewhat dampens how euphoric we can get on AI." He also pointed out that power use per unit of compute has "flatline[d]" across the last five generations of Nvidia chips, which he called "not okay." The a16z Show

Morgan Stanley's head of US public policy research, Ariana Salvatore, added a policy angle: "For data center pushback, absolutely midterms are a catalyst." She still thinks a federal ban or moratorium is "not super likely." Squawk on the Street

2. Memory: The Best Cycle in History, or the Top of One?

This was the week's biggest debate, and the most interesting conversation was on The Circuit (Oct 5). Ben Bajarin and Jay Goldberg argued over whether memory has stopped being a boom-and-bust business.

Bajarin's bull case rests on a simple point: the industry's floor has moved up. At past peaks, memory was about a $100 billion market. His model now puts it at "$600 or $800 billion depending on ASP floor" (ASP means average selling price). His argument is that investors are pricing in a collapse that may not come:

"It seems like the way the market is valuing these is that they're assuming that crash comes in 29."

His report this week was titled "Bet on Bits, Not History":

"Bit growth demand is inevitably going up… Everything else in between, price, vendor share, et cetera, is negotiable."

He expects margins to settle in the 60s to low 70s, not fall back to the 30s. Historically they averaged single digits to low double digits.

Goldberg's response was more cautious. Micron itself says the amount of memory produced will grow only in the low 20s% for DRAM and mid-20s% for NAND into 2027–28, even though "prices are basically doubling this year." Meanwhile every maker is adding capacity:

"I think the bear case would be, we get to supply demand equilibrium in 28. The market sees that and starts forecasting… much more pessimistic in sort of some point next year."

He also made a technical point that explains why supply is so tight. Making HBM, the high-performance memory stacked next to AI chips, uses far more factory output than ordinary memory: "for every four sort of trailing edge memory, you get one HBM leading edge memory." He called it "a reverse node shrink." The most advanced product is less dense per silicon wafer, not more. He also noted reports that Tesla is "de-spec-ing memory" in its robots because it "just can't get enough." The Circuit

Rasgon put it in historical context:

"We're in the best memory cycle in history right now, but we are coming off the worst memory cycle since the tech bubble."

• HBM takes "three or four times as many wafers" per gigabyte as standard DRAM.
• In Nvidia's NVL72 rack, "80% of the silicon wafer area inside that rack is memory."
• Memory makers say supply will be tighter in 2027–28 than in 2026. Meaningful new capacity arrives around "27, 28."

Monetary Matters with Jack Farley

Gil Luria's $3,000 Micron call (DA Davidson, Squawk on the Street, Oct 8) is the strongest bull view out there:

"It's the same cycle. The AI cycle is the same cycle for GPUs, CPUs, and memories. When the market wraps its head around that, the multiple should expand very dramatically."

• Micron trades at about 6x earnings. AMD and Intel trade at 40x+ and GPU makers at about 20x.
• Micron is about 3% of the Nasdaq 100 by size but about 10% of its expected earnings over the next 12 months.
• More than half of its output will be sold under five-year contracts.
• Its buyback restriction ends on December 9, and it plans to spend 100% of excess cash flow on buybacks: "Those are the worst things you can hear as a short seller."

Host Sarah Eisen pushed back that the shortage "is not going to go on like this forever." Luria answered that if supply catches up, GPUs, CPUs and memory would all be hit together. Squawk on the Street

Industry chatter from The Six Five (Oct 5). Patrick Moorhead and Daniel Newman passed along several notable details:

• Customers paid Micron $12 billion in deposits to fund cleanrooms that won't arrive until late 2028.
• Micron's cash is about $70 billion. Newman said it "is going to be able to go NVIDIA on the market" and start investing in the ecosystem.
• One insider said HBM is now Micron's lowest-margin product, which Moorhead admitted he can't explain.
• "Some of the accelerator makers are starting to have to give wafer allocations back because they can't secure the memory." Both hosts think Nvidia is the likely buyer of those freed-up wafers, because "he bought up the most memory."
• A "very credible and smart person" told Newman the shortage lasts until 2029.

The Six Five with Patrick Moorhead and Daniel Newman

Pricing data points:

• Research firms expect memory prices to rise 10–20% in the third quarter.
• Citi forecasts HBM4 prices could rise "as much as 100 percent in the next year."

Bloomberg Tech

An Asia fund manager who is selling. On Equity Mates (Oct 8), the manager said Nvidia's DRAM demand is up about 30-fold in four years and prices are up about sevenfold. SK Hynix is up roughly 20x since 2023 and trades near 4x earnings. He has sold about 90% of his position but still holds nearly 7% of the fund, and thinks the market still underestimates memory profits. Equity Mates Investing Podcast

The skeptics on memory:

• RenMac's Jeff DeGraaf on stock trends: "AMD and NVIDIA still look good… But there are other ones that are still questionable, like, you know, SanDisk or SK Hynix." RenMac
• Options activity on Micron on the Halftime Report was bearish. About $270M was spent on puts versus $230M on calls, including a roughly $14.5M position in deep in-the-money puts expiring June 2028. CBOE's Oliver Renick described it as a synthetic short (a bet that the stock falls). Halftime Report
• Jim Cramer flagged a labor risk: "There's a union that organized Micron. 99% vote that they have the right to strike." Squawk on the Street

Consumers are starting to feel memory prices. Nikkei reported that Apple cut October iPhone 18 Pro production orders by at least 15%, and the Pro models now start at $1,200. Odyssey Capital's Jason Snipe: "I think memory costs are a big deal." Halftime Report

Gelsinger summed up the change: memory "was a terrible industry… you would make money one year and then you'd lose money for the next four." He estimates the memory industry has added $2.5 trillion of market value in four years. He has just funded a new memory startup that is still in stealth, and says "memory innovation for the first time in 30 years is nigh upon us." The a16z Show

3. Chip Equipment: A $300 Billion Market by Decade-End?

Rasgon covers Applied Materials, Lam Research and KLA, and he is bullish. "WFE" means wafer fab equipment, the machines used to make chips. His Bernstein forecasts:

"I have no problem… signing up for a 300 billion plus WFE into the end of the decade."

• The limit isn't demand. It's buildings: "There's a clean room short."
• Micron says facilities spending will grow faster than equipment spending next fiscal year.
• He slightly prefers Applied Materials for its DRAM exposure.
• Lam Research is weighted toward NAND, capturing about 25–30 cents of every NAND equipment dollar. He pointed out that Lam earned through the COVID-era NAND slump and held up through the downturn.

Monetary Matters with Jack Farley

ASML reports Wednesday. Investor's Business Daily's Rachel Fox noted that ASML is "the sole global manufacturer of extreme ultraviolet lithography equipment," the machines required for the most advanced chips. Consensus expects earnings per share up 89% to $12.07 on revenue up 51% to $13.16B, compared with 27% growth last quarter. Ed Carson called ASML's and TSMC's spending outlooks "really important for the whole chip gear sector." Investing With IBD

University of Illinois professor Rakesh Kumar, on The Dynamo Show (Oct 7), put the cost of a single leading-edge fab at "roughly around $30 billion." He argued that the roughly $52B CHIPS Act is too small to change many decisions, and that China's domestic equipment capability is now "pretty good," which he sees as a structural loss for US vendors. The Dynamo Show: a Dynamo Ventures Podcast

4. Foundries: Intel Gets a Second Look, TSMC Gets a Shakeout

Intel is getting more respect from analysts who used to dismiss it. Rasgon, a long-time critic, said:

"I don't hate it right now, which is new for me."

He still thinks meaningful foundry revenue (making chips for other companies) is "not going to be for quite a few years." Packaging revenue may come sooner. He also noted that Intel had written down server products to zero and customers bought them anyway, adding about 200 basis points (2 percentage points) of gross margin upside. He views CEO Lip-Bu Tan favorably. Monetary Matters with Jack Farley

Wedbush's Matt Bryson walked through the TerraFab confusion on The Information's TITV (Oct 7). TerraFab is the SpaceX/Tesla plan to build chip fabs in Texas. Musk floated a TSMC role, then walked it back. Lip-Bu Tan reaffirmed that Intel is still involved. Bryson's best guess at the structure:

"It would almost look like a fab that is funded by SpaceX, but running on Intel and in part operated by Intel."

That means Intel's next-generation 14A process. TSMC, he said, would not normally let a third party run one of its fabs.

On Intel's turnaround, he said demand from inference (running AI models, as opposed to training them) took Intel from "mothballing fabs back in the middle of last year to not having enough fab capacity." Lead times on Xeon server chips are now about six months. The hard part is telling how much of Intel's improvement is the market and how much is Intel:

"If we see gross margins go up a bunch in Q4, Q1, and we don't see pricing go up, that's a pretty good sign that Intel's getting its fabs [in order]."

The other signal he is watching is signed agreements from large chip designers to use 14A. The Information's TITV

Bloomberg's global tech editor Peter Elstrom said one analyst estimates the full TerraFab vision would cost $5 trillion to $13 trillion. He was cautious:

"The chip industry is terribly cyclical. These companies invest billions and billions of dollars and then they find themselves with too much capacity."

Bloomberg Tech

TSMC reports Thursday. Consensus expects earnings per share up 55% to $4.41 and revenue up 40% to $45.44B. Monthly sales already came in "slightly stronger than expected." Carson's view: "we know the earnings are going to be great… it's the expectations going forward. The capital spending outlook, will they raise that further?"

Rachel Fox flagged two questions:

• How much of TSMC's growth is AI versus a recovery in phones and PCs.
• Whether next-generation chip processes pressure gross margins in the short term.

The stock fell more than 3% on Friday. Carson attributed part of the pullback to the TerraFab rumors fading. Investing With IBD

5. China, Export Controls and Huawei

Brent Wilsey, president of Wilsey Asset Management, argued that export controls may be creating a competitor (Smart Investing, Oct 3):

"The chipmaking technology Huawei is using today, now it's still behind NVIDIA's, but the company is moving aggressively to close that gap."

Huawei plans two new chips next year and more for 2028–29, with Beijing's backing. The risk he sees isn't the US market but third markets:

"Nvidia might have to compete against Huawei in Europe… the question for Nvidia is how quickly can Huawei catch up and whether restrictions on China ultimately create a long-term competitor rather than simply limiting China's access to NVIDIA's technology."

Smart Investing with Brent & Chase Wilsey

On Better Offline (Oct 7), Ed Zitron and GamersNexus questioned how well export controls are enforced. Micro Center now requires government ID and an end-user certification to buy Nvidia's flagship RTX 5090 consumer GPU. Zitron called that "slapping a Band-Aid on the side of the Titanic." He argued chips are reaching China through Singapore, Malaysia and Taiwan. He also said Nvidia's switch to reporting sales by the buyer's headquarters rather than by destination makes this harder to see. GamersNexus's take: "if we're going to have them, they should make sense and probably be actually enforced." This is a skeptic's view, not established fact. Better Offline

On the other side, Cramer said on Squawk on the Street (Oct 5) that the "China doesn't matter" worry has faded on Wall Street because orders are strong. Squawk on the Street

6. Deals and Financing: Who Is Paying for All This?

The SpaceX borrowing was the week's headline deal. Bloomberg reported that SpaceX is in talks to borrow about $40B to buy Nvidia chips: about $10B syndicated through banks and roughly $30B from private lenders led by Apollo, with PIMCO possibly involved. Bloomberg chief Wall Street correspondent Sridhar Natarajan:

"It's $60 billion one week, $40 billion next week… the tech story today is the debt story… We are talking about trillions of dollars in requirements, and these companies will keep coming back to the market."

He noted that lenders are "basically betting on the underlying creditworthiness of the borrower," not on what the chips will be worth later. SpaceX's five-year CDS (credit default swaps, the price of insurance against default) jumped to its highest since trading began in June. Bloomberg Tech

CNBC's David Faber asked the obvious question: "why would NVIDIA's stock price not moving even higher on what's going to clearly be a significant order for its chips?" Cramer called Nvidia's 3% opening jump "totally unsustainable… as if it's some sort of meme stock." Squawk on the Street

Qualcomm is moving into the data center. CEO Cristiano Amon told Alex Heath (Oct 8) why Qualcomm bought Chris Lattner's Modular, a software platform that lets AI run on any company's chips. It is a direct challenge to Nvidia's CUDA software lock-in:

"We will celebrate every single modular deployment on NVIDIA, on AMD, on Broadcom, on any other chip, on Intel, because we actually believe the industry will benefit from a standard, common open platform."

Modular will be open-source under an Apache 2.0 license. Qualcomm's bet is that once software is open, it can compete on hardware, specifically energy-efficient CPUs and accelerators. Sources with Alex Heath

7. Is This the Peak? The Cycle Debate

The bull case: valuations are falling because earnings are growing so fast. Rasgon:

"I'm not really worried about a cycle peak yet... It's not this year. It's not that next year."

He compared today with 2000. Cisco traded at about 200x earnings then; Nvidia trades at about 12x today:

"We're actually building out with demand versus building out ahead of demand."

He admitted his cycle tracker shows high channel inventories, but "they've been high for like four or five years now," so the old warning signs are less reliable. He also flagged the classic late-cycle risk of double ordering: "If I'm not getting my allocation, I want to order maybe twice as much." Monetary Matters with Jack Farley

The warning from Fidelity's Jurrien Timmer (The Compound and Friends, Oct 9):

"The semiconductor index has a very clear 40-month earnings cycle. And we're right at the part where it should start to roll over. And of course, 178% is not sustainable."

Even so, he isn't worried about near-term earnings, because suppliers "cannot keep up with demand." His concern is the money. Hyperscalers used to return about 90 cents of every dollar they earned to shareholders. Now they spend it all on capex with an unknown return. They borrowed about half a trillion dollars this year, and his team expects another half trillion next year. He added a caution of his own: "It's no longer cyclical, it's secular, but it's different this time are not good words to live by." The Compound and Friends

Market breadth is narrowing. RenMac's Jeff DeGraaf said only about "a third of the names are… making relative strength highs," meaning outperforming the market. "A lot of them are still languishing." RenMac

The skeptics:

• Steve Hsu (Manifold, Oct 8) said plainly: "I am calling a kind of AI bubble right now." AI revenue needs to reach "multi-trillions… within the next few years" to justify the spending, roughly 10x today's $150–200B. Ramp data shows company spending on AI tokens has "leveled off," which "could be a negative signal for semiconductors memory chip makers people who are building data centers." Manifold
• Peter Boockvar (RiskReversal, Oct 9) said the "money tree" of free cash flow has run out, with up to "$500 billion worth of debt this year… Double of last year." His view: "I think the CapEx portion is [a bubble], but the technology that's going to be created will all benefit from." RiskReversal Pod
• Rakesh Kumar (The Dynamo Show): "These trillion dollar valuation AI companies do not have money to pay for all the chips that they're buying." And: "You get a chip and three years later, it's completely useless." The Dynamo Show: a Dynamo Ventures Podcast
• Jason Snipe (Odyssey Capital, Halftime Report): "I'm not calling the end of the cycle right now. But if you're not worried about it happening at some point, then you truly are, you know, whistling by the graveyard." He named circular funding between AI companies as the worry. Halftime Report
• Nick Rossolillo (Chip Stock Investor, Oct 6) sees a "mini bubble" in neoclouds (GPU rental companies) rather than in chips themselves, and expects a washout "within the next couple of years. But the trend remains intact." Chip Stock Investor Podcast

Earnings Reactions

Ed Carson's warning applies to the whole group: "On the things like NVIDIA, you can have amazing earnings. And you can go sideways for six months, nine months." Stock Market Today With IBD

What I'm Watching Next Week

• ASML results (Wed, Oct 14, before the open). The key number is the equipment-order outlook. It is the first real test of Rasgon's roughly $204B WFE forecast for 2027.
• TSMC results (Thu, Oct 15). Watch whether 2026 capex goes up again, how much growth comes from AI versus phones and PCs, and gross margins on the newest processes. Any comment on TerraFab would also matter.
• Memory pricing. Watch fourth-quarter DRAM and HBM contract prices against the 10–20% third-quarter increases researchers expected, and for any sign that consumer memory prices have flattened, which Goldberg raised on The Circuit.
• The financing pipeline. Watch final terms on the SpaceX borrowing, and SpaceX and Oracle CDS levels as a gauge of lender appetite.
• Apple's foldable iPhone Duo pre-orders open Oct 16. It is a real-world test of whether higher memory costs are hurting device demand.
• Further out: Micron's buyback restriction ends Dec 9, and the midterms are increasingly seen as a test for data center politics.