Newsletter · 2026-10-03 · Ashutosh Agarwal

Micron's Record Quarter and Nvidia's Buyback Reopen the Memory Cycle Debate - Semiconductor Podcast Briefing - Week of October 2, 2026

Semiconductor Podcast Briefing for the week of Sep 26 – Oct 2, 2026. Podcast synthesis on Micron's $54.2B fiscal Q4 and 87% gross margin, Nvidia's record $150B buyback and the memory supercycle-versus-cycle debate, hyperscaler capex heading toward $1 trillion a year, CXMT's China supply ramp, AMD's $8.2B World Labs deal, and the equipment bottlenecks at ASML and TSMC.

Semiconductor Podcast Briefing

Week of October 2, 2026: Micron's Record Quarter and Nvidia's Buyback Reopen the Memory Cycle Debate

Micron's results took up most of this week's chip podcasts. The company reported a quarter that would have looked impossible two years ago: $54.2 billion in revenue and an 87% gross margin. The stock barely moved. Why it didn't move is this week's central debate: is memory still a boom-and-bust business, or has AI changed it? Nvidia's record $150 billion buyback came second, and it set off a related argument. Is the world's most valuable chip company cheap because its earnings are at a peak, or because the market doesn't trust the customers paying it?

TL;DR: Things That Mattered This Week

• Micron's fiscal Q4 (quarter ending late August) blew past estimates. EPS was $33.42 against $31.61 expected, revenue $54.23B against $51.0B, and gross margin 87% against 86.3% (Closing Bell). Revenue was nearly 5x the $11.3B of a year ago, and next quarter is guided to $61.5B (The Rundown).
• The one soft number was next quarter's margin: 86.25–86.3%, against roughly 86.6–86.7% expected. CEO Sanjay Mehrotra put the dip down to record employee bonuses (Squawk on the Street).
• Micron says memory supply will be tighter in 2027 and 2028 than it was in 2026, with "no line of sight" to when supply and demand balance (CNBC's Fast Money).
• Long-term customer deals went from 16 to 26, some running into 2031. About 75% of fiscal 2027 output is already committed (Futurum Equities Podcast).
• Price did the work this quarter, not volume. DRAM revenue rose about 27% from the prior quarter on shipments up only mid-single digits. NAND revenue rose about 42% on bit growth of about 10% (Futurum Equities Podcast).
• Nvidia added $150B to its buyback and expects to complete $235B of repurchases by fiscal 2028. Bernstein's Stacy Rasgon worked that out to about $40B a quarter, double what he had modeled (Squawk on the Street).
• Hyperscaler capex (spending on data centers and equipment by Alphabet, Amazon, Meta, Microsoft and Oracle) is about $780B in 2026, up from $416B in 2025, and expected to pass $1 trillion a year from 2027 (The a16z Show).
• AMD agreed to buy Fei-Fei Li's World Labs for $8.2B in stock, adding spatial-intelligence and "physical AI" software (Closing Bell).
• China is the swing factor in memory supply. Chinese DRAM maker CXMT is expected to end the year within about 20,000 wafer starts of Micron's own monthly output (Investing Experts). Apple was reported to have tested CXMT memory (The Circuit).

1. AI Chip Demand and Hyperscaler Spending

The short version: almost everyone on the podcasts agrees demand is still running ahead of supply. The disagreement is about what Nvidia's stock price is telling you.

Nvidia's buyback

• A buyback is when a company uses its cash to buy its own shares, which shrinks the share count and lifts earnings per share. Nvidia's new authorization is the largest on record.
• Stacy Rasgon (senior analyst, Bernstein; Outperform rating, $400 target), Sep 28:

"$235 billion by the end of fiscal 28 is over the next six quarters. It's like $40 billion a quarter, which is double what we were modeling... In the context of a $5 trillion market cap, it's still not that big." (Squawk on the Street)

• Jonathan Bloxham (senior media and tech analyst, Bloomberg Intelligence), Sep 28, read the buyback two ways: either the shares are undervalued, or "maybe they're running out of things to invest the money in. And it's probably a mixture of the two." He put Nvidia's cash flow at about $200B this fiscal year and roughly $330B next year by consensus, and said the message to investors was that the "forward pipeline, at least in the next one to two years, is looking incredibly healthy" (Bloomberg Intelligence).
• The Motley Fool's Hidden Gems hosts, Sep 28, said it was too small. Tyler Crowe noted that $150B is about one year of Nvidia's $134B free cash flow and only about 2% of its market value. Rachel Warren called Nvidia "essentially like a central bank for a lot of the AI industry": it holds 13 public stakes and over 200 private stakes, and Anthropic's contracted value across Nvidia-backed cloud providers is above $180B (Motley Fool Hidden Gems Investing).
• On Marketplace, Futurum analyst Daniel Newman said Nvidia's balance sheet gives it "the optionality to return to shareholders, and you have the optionality to invest in the ecosystem." Another speaker described hyperscalers as "filling the dump trucks with money and sending it to NVIDIA" (Marketplace).

How big the spending is

• The a16z Show, Sep 30: hyperscaler capex is $780B in 2026, up from $416B, and "all expectations point to them spending over a trillion dollars annually from 2027." Their sources say "every stage of the supply chain" reports demand outrunning supply, and some data center parts can't be had "until 2028" (The a16z Show).
• Odd Lots, Oct 1: a guest cited a Goldman Sachs note estimating that about half of S&P 500 earnings growth this year comes from hyperscaler capex alone. Wall Street's estimates for that spending "keeps getting revised higher," and a "solid chunk" of the current trade is a bet that this continues (Odd Lots).
• Adam Parker (Trivariate Research), Sep 28, recommends semiconductors, energy and power: "There's a shortage of compute. There's a shortage of power." (Squawk on the Street)

Nvidia versus everyone else

• Ben Pouladian (on Monetary Matters), Sep 30, a long-time Nvidia bull, argued that buyers now judge chips by cost per token, meaning what it costs to produce each unit of AI output. On that measure Nvidia's full system wins: "all roads and rockets lead to NVIDIA." His evidence: Elon Musk dropped his custom-chip plans and went "exclusive NVIDIA" on the Vera Rubin NVL72 platform. He sees AMD "squeezed in the middle" between Nvidia and the labs' own inference chips (OpenAI's, Meta's MTIA, Microsoft's Maia) (Monetary Matters with Jack Farley).
• Walter Goodwin (founder and CEO, Fractile, an AI chip startup), Oct 2, on why hyperscalers design their own chips:

"There's a bit of a joke today that the sort of first party efforts, their primary purpose is to reduce the price that people pay NVIDIA. And there may be some truth to that." Fractile is betting on 25x more memory bandwidth per chip than HBM-based designs. (HBM, high-bandwidth memory, is the stacked memory that sits beside AI chips.) Goodwin noted that over 20 years, chip compute has grown "like a million fold" while "memory bandwidth has gone up about 40x" (No Priors).

Contrarian view

• Jay Goldberg (Seaport Securities) keeps his Sell on Nvidia and pointed out that "NVIDIA is the worst performing stock in the SMH and the Sox index" since he started covering it in May 2025. He prefers Semtech ($200 target), ON Semiconductor ($100) and Marvell ($270). His top pick is Semtech, a company that "made a bad acquisition three years ago" and has since refocused on AI data centers (Power Lunch).

2. Memory: Micron's Quarter and the HBM/DRAM/NAND Picture

The short version: Micron says the shortage gets worse, not better. The market is pricing it as though the shortage will end.

What management said

• Sanjay Mehrotra (CEO, Micron), Oct 1, on long-term customer contracts:

"In fact, we cannot fulfill the demand of our customers. Customers are extending these agreements. Even now, into 2031 timeframe." He said supply in "27 and 28" will be "even tighter than 26." He expects "well over $100 billion in cash" by the end of the current quarter, and said investors should "stay tuned" on buybacks until after December 9, the second anniversary of Micron's CHIPS Act award. Capex in the first half of fiscal 2027 will be "around $25 billion. That's almost as much as what we invested in full fiscal year 26" (Squawk on the Street).

• He also cited $250B of planned US investment across Idaho, New York and Virginia. On the call, Micron said it had poured concrete on its first New York fab, with first output expected in about four years (CNBC's Fast Money).

The numbers underneath

• From the Schwab Network's live reaction: data center revenue was $18B against an $11.34B estimate, operating income $44.64B against $42.75B, and next-quarter revenue guidance $60–63B against a $56.77B estimate. Hosts flagged expected capex of $40–50B as the main risk to free cash flow. Paul Meeks, who has covered Micron since the 1980s, noted this is "the 14th quarter of the upcycle. Usually it's 8 to 9" (Schwab Network).
• Futurum Equities Podcast, Oct 2: the quarter was "extremely price driven." Long-term contracts now cover "more than 35% of the total revenue through 2030" and remaining contracted revenue is about $150B. One host said industry contacts see "no supply relief probably till 2030." He added a supply-chain anecdote: "there are accelerator companies giving wafers back to TSMC right now because they can't... secure the memory." In his view that favors Nvidia, which "has secured memory more effectively than any other company" (Futurum Equities Podcast).
• Matt Bryson (Wedbush), Sep 30: "All of my price checks have suggested that pricing is still moving up. It's at a little bit slower pace." He said a lower margin isn't a negative if it reflects more HBM in the mix (Closing Bell).

Where new supply comes from

• Jake Silverman (semiconductor analyst, Bloomberg Intelligence), Oct 1: the industry burned "tens of billions of dollars" in 2022–23 and stopped adding capacity. When AI arrived, "the supply and demand mismatch just balloons." New fabs from SK Hynix and others start up at the end of 2028, but it is "more like a 29 story." Until then, buyers are "just going to have to pay a little bit more." He noted Samsung had mentioned a $100 increase on some products. He thinks the stock's upside is limited because "you can only have so many price increases. You can only have so much margin expansion from here" (Bloomberg Intelligence).

Korea

• SK Hynix and Samsung fell about 5% on Sep 28, part of a broad chip pullback after five straight weeks of gains. Qualcomm, Intel and AMD fell 3–7% the same day (Closing Bell).
• Barron's Streetwise noted the Roundhill Memory ETF (DRAM) holds Micron, Samsung and SK Hynix at about 26%, 25% and 22%. Its host described Micron's earnings going from about $8 a share in fiscal 2025 to more than $70 expected for fiscal 2026 (Barron's Streetwise).

Hard drives: a related scare

• Western Digital and Seagate fell about 10% on Oct 2 after a Nikkei report that Toshiba would double drive capacity. Matt Bryson (Wedbush) said "the Nikkei got the story" wrong: the doubling at Toshiba's Philippines site runs from FY25 to FY27, and most of it is ordinary technology gains. "If you don't have the clean room space, you can't bring on new capacity. You can't blow up the industry." He also said drive makers have no new recording heads "until 29" (CNBC's Fast Money).

3. Chip Equipment and Manufacturing Bottlenecks

• Ritesh Jain (on Insights Now), Oct 1, explained why chip stocks held up during a week of AI-safety worries: the supply chain itself slows the buildout.

"There are only so many lithography machines that come out of ASML. There are only so many 3-nanometer, 2-nanometer wafers that come out of TSMC." Even small parts suppliers to the chip-equipment makers, he said, "just cannot physically" raise output 5–10x in a year. In his reading, the market sees those limits as a brake that keeps the boom orderly (Insights Now).

• Baillie Gifford's emerging markets team, Sep 28, described Chroma ATE, a Taiwanese maker of chip-testing equipment that has worked with Nvidia since 2006, as a quiet winner: chips "continue to get more and more complex, require more and more power." They named Advantest and Teradyne as bigger competitors. They also own Silergy, a mainland Chinese maker of analog chips, and said many Chinese chip companies "won't meet foreign investors at all anymore" (Short Briefings on Long Term Thinking - Baillie Gifford).
• Investing Experts (Tech Contrarians analyst), Sep 27, is bullish on ASML and TSMC. The concern is how the stocks react to results: last quarter both "reported great results" and "both sold off," because results acted "more as a test than a catalyst" (Investing Experts).

4. Foundry and Processors: Intel's September

• Patrick Moorhead and Daniel Newman (The Six Five), Sep 28, tied a September rally in processor stocks to Meta's new Muse AI assistant. Muse gives each user a virtual machine that runs on AMD EPYC processors. AI agents, software that carries out tasks on its own, need far more conventional processing than chatbots do. They said Arm was up about 30% for the month and Intel about 40%, and that Intel "can only meet about 50% of the demand." Much of the enthusiasm, they argued, is about Intel's foundry (contract manufacturing) business: "can you imagine had they... spun that off for parts?" (The Six Five)
• Pouladian also expects the processor market to reach $200–300B by 2030, split among AMD, Intel, Nvidia and Arm (Monetary Matters with Jack Farley).

5. China: Memory Supply and the Cost Gap

• CXMT's ramp. The Tech Contrarians analyst said CXMT, China's main DRAM maker, is "coming in aggressively with a lot of supply." Combined with weak PC and smartphone demand, "we could go very quickly from a shortage to a glut." He cited TrendForce data showing DRAM price increases slowing to 13–18% in Q3 from roughly 50–60% in Q2 (Investing Experts).
• Apple and CXMT. On The Circuit, Sep 28, the hosts discussed reports that Apple "had tested CXMT memory" and was "pushing very hard to get a waiver" from the US government to buy it. If Korea and Idaho can't supply memory, "China can." They warned that once a second supplier is in, "barn doors open," and urged Samsung, SK Hynix and Micron to look after long-standing consumer customers. They said scaling at CXMT and YMTC should bring relief to Chinese phone and PC makers first (The Circuit).
• The cost gap. Raja Koduri (founder of Oxmiq; former graphics chief at AMD and Intel, ex-Apple), Sep 29, put Western AI data centers at about $50–60B per gigawatt of capacity, roughly $45B of it for chips and hardware.

"The target for China Inc. for gigawatt is to be less than 10 billion dollars... I thought they would be able to do that in the next three years. I see indications that they will be able to do it much sooner than that." He called it a "five times to six times cost disparity." He also argued that an older TSMC 7-nanometer chip with memory 3D-stacked on top could deliver "10x the bandwidth than the current HBM" and "beat Vera Rubin." His point: in AI inference, how memory is arranged matters more than having the newest manufacturing process (TechSurge: Deep Tech Podcast).

6. Mergers and Deals

• AMD buys World Labs for $8.2B in stock. Kate Rooney (CNBC) reported that the deal brings in Fei-Fei Li as executive vice president and adds 3D "spatial intelligence" for robotics and simulation. AMD is "trying to close the gap with NVIDIA" in software. She also reported AMD had earlier looked at buying Hugging Face, which Nvidia acquired. Closing is expected by year-end, pending regulatory approval (Closing Bell).
• Bloxham put Nvidia's Hugging Face spend at roughly $18–20B (Bloomberg Intelligence).
• Patrick Boyle, Sep 27, walked through Nvidia's role in the SB Energy / OpenAI Ohio campus. Nvidia bought $1.5B of SB Energy shares at a 10% discount and agreed to guarantee up to $105B for the campus. In return, the site runs exclusively on Nvidia hardware for 20 years. "So, SoftBank borrows at junk rates to fund OpenAI, which rents the building that NVIDIA guarantees and fills with NVIDIA chips. It's a very tidy arrangement." (Patrick Boyle On Finance)

7. The Cycle Debate: Peak or a New Kind of Memory Market?

Memory has historically gone boom then bust: prices rise, everyone builds factories, supply floods in and prices crash. That history is why Micron trades at about 6–7 times next year's earnings, against about 18.5x for the S&P 500 (The Rundown).

The case that this time is different

• Mehdi Hosseini (senior equity research analyst, Susquehanna; Buy, $2,000 target), Sep 30:

"I make an argument that compute is actually becoming more commoditized. And it's memory that makes the difference." He sees Micron earning $30–40 a quarter, eventually $50 a quarter ($200 a year), with operating margins around 70%. His reason: demand now comes from many applications and customers, unlike the smartphone era when "there was only one demand driver" (Squawk on the Street).

• Ben Reitz (head of tech research, Melius; Buy, $2,200 target), Oct 1: the new information was 2028. Investors had assumed the capacity coming in 2028 would end the story, and management "basically are saying to you, yes, we know about all the capacity, and we're not rolling over." He expects buybacks of "epic proportions," possibly "15% a year, easy" of market value. "If memory was invented in 2023, it's basically an AI chip on every tier." (Squawk on the Street)
• Steve (CNBC's Fast Money), Sep 30, said he had changed his mind: "agents create token demand. Tokens are inference and inference runs on memory... Micron is probably going higher from here." (CNBC's Fast Money)
• Hamid Shojaee (Buy Hold Rant), Oct 1, argued the market's cyclical view "is actually looking into the past," and has been "wrong for at least five quarters." Micron is 34% of his portfolio (Buy Hold Rant).
• Ed Carson (IBD), Oct 1, noted that 75% of fiscal 2027 sales are "already baked in" and Micron is "already selling fiscal 2028." He saw the stock's turnaround that day as a possible buying point (Stock Market Today With IBD).

The skeptics

• Tim Seymour (CNBC's Fast Money), Sep 30: "the 11-fold growth in data center is not something you're going to have year over year out 2 years." He pointed to CXMT and said: "The story is not, I think, without turbulence ahead. And I'm not chasing Micron here." (CNBC's Fast Money)
• On the same show, Dan Nathan argued the size of Micron's beats has shrunk, to roughly 5–6% this quarter from 50–100% earlier in the cycle (per the episode summary).
• Joe Moore (Morgan Stanley), as relayed on Closing Bell: "it's not how good can it be, it's how long can it stay this good" (Closing Bell).
• Tech Contrarians analyst, Sep 27: "Memory is still cyclical. It's crazy to think that, you know, margins are just going to go up endlessly from here." (Investing Experts)
• Matt Bryson (Wedbush) took a middle position. At past cycle peaks, memory stocks traded at "mid-single-digit type earnings multiples. That's where we are right now." With a year or more before supply becomes a worry, plus big buybacks, "they should trade above mid-single digits" (CNBC's Fast Money).
• On why investors are cautious, Mike Santoli (CNBC) cited Morgan Stanley work showing that since the summer, the chip companies with the best upward earnings revisions "have actually underperformed." The market has "persistent... unwillingness to give these companies credit for anything beyond the next few quarters." (Squawk on the Street)

Is the whole AI buildout a bubble?

• Jim Chanos's argument, as relayed on RiskReversal, Oct 2: Nvidia trades at about 19x next year's earnings against about 22x for the market. If you believe the AI story, why should companies that depend on Nvidia's chips trade at higher multiples than Nvidia? Another guest said that if the AI cycle turned, he would short "the very heavily indebted names" and "data centers where the business models are questionable," not Nvidia (RiskReversal Pod).
• Stephen Yu (CIO, Blue Whale Growth Fund, about $4B in assets), Sep 28, separates a "valuation bubble" in private AI labs (OpenAI, Anthropic, SpaceX) from a possible "earnings bubble" at Nvidia, which trades at "16, 17 times earnings." The risk for Nvidia is that its earnings estimates are too high, not that the stock is overvalued on them (Merryn Talks Money).
• Patrick Boyle gave four reasons Nvidia looks cheap at under 17x forward earnings. One is that the market treats it like a cyclical company at its peak. Gross margin was 75% last quarter but is expected to slip below 72%, "partially because NVIDIA has to buy memory chips from other companies, and those companies have been raising their prices." The memory boom is now eating into Nvidia's margin (Patrick Boyle On Finance).
• Empire, Oct 2, worked through the data center math. One gigawatt costs $50–60B, "roughly half of that is like Nvidia chips," and at current rental rates earns $10–20B a year. Whether that pays back depends on rental prices and how long the chips stay useful, which so far "has really stood up" (Empire).

Earnings Reactions

What to Watch Next Week and Beyond

• ASML Q3 results, Oct 14 (before the open). The first read on equipment orders for 2027, and a test of whether strong results can lift the stock this quarter after last quarter's sell-off.
• TSMC Q3 results, Oct 15 (before the open). Leading-edge wafer supply, advanced packaging capacity, and any sign of the memory shortage limiting how many accelerators customers can ship.
• Lam Research (Oct 21) and Texas Instruments (Oct 21), then Intel (Oct 22), all after the close. Lam for memory-equipment spending after Micron's capex increase; TI for whether analog and industrial chips are finally recovering; Intel for processor supply against Muse-driven demand and foundry customers.
• Samsung Q3 results, Oct 28. The Korean side of the memory-pricing story, and HBM share.
• CXMT and Apple. Any decision on a US waiver for Apple to buy Chinese memory would be the first real dent in the three-company memory market.
• Hyperscaler results (late October). Capex guidance for 2027 is the input everything above depends on.
• Micron after December 9. Mehrotra told investors to "stay tuned" on capital returns once the CHIPS Act restrictions anniversary passes.