Newsletter · · Ashutosh Agarwal

Micron Sells Out 2027 as Memory Rations the AI Buildout - Foundry & Chip Equipment Weekly - Week of October 4, 2026

Foundry & Chip Equipment Weekly for the week of October 4, 2026. Micron's record fiscal Q4 ($54.2B revenue, 87% gross margin, $61.5B guide) and its CEO's warning that memory stays tighter in 2027 and 2028, plus the operator and analyst debate over whether rationed memory now caps the AI buildout, China's CXMT ramp, Nvidia and Broadcom vendor financing, and the read-through to the equipment chain.

Foundry & Chip Equipment Weekly

Week of October 4, 2026: Micron Sells Out 2027 as Memory Rations the AI Buildout

Last Sunday we said Micron's results would be the test of the quarter, and that the podcasts were split down the middle on whether memory was heading for a glut. The results came in on Wednesday night, September 30. On the numbers, the bulls won easily. In the stock market, the result was closer to a draw.

Micron did $54.2 billion of revenue in a single quarter, up from $11.3 billion a year earlier. It kept a gross margin of about 87%, a level that would have looked like a typo a few years ago. It guided to $61.5 billion for the current quarter. Then its chief executive went on CNBC the next morning and said the shortage gets worse from here, not better: supply will be tighter in 2027 and 2028 than in 2026, customers are signing contracts out to 2031, and about three-quarters of next year's output is already spoken for.

The stock barely moved. It was flat after hours, down about 3% the next morning, and then closed the day up 3%. That gap between spectacular results and a shrug from investors is the most interesting thing that happened in chips this week. It tells you the debate has moved on from "is memory tight?", because everyone now agrees it is. The live question is how long it stays tight, and who gets squeezed while it does.

The squeeze is already spreading. One podcast host said, citing industry contacts, that some AI chip designers are handing wafers back to TSMC because they can't get enough memory to go with them. If that's true, memory has stopped being one input among many. It's now the thing that decides how many AI chips get built.

A note on sourcing, as always. This week brought the most operator testimony in a while. Micron's CEO spoke at length, and we also heard from the founders of an AI-chip startup and a data-center builder, plus an IBM Quantum product lead. TSMC, Intel Foundry, Samsung, ASML and the big US and Japanese toolmakers were silent again. Everything else here comes from analysts, investors and commentators, and we weight it that way.

TL;DR

  • Micron's quarter was huge, and its CEO says the shortage gets worse. Revenue was $54.2 billion, the gross margin about 87%, and the guide $61.5 billion. Sanjay Mehrotra expects supply to be tight in 2027 and 2028, "even tighter than 26," and said "we cannot fulfill the demand of our customers."
  • Long-term contracts are the new story. Micron now has 26 strategic customer agreements, up from 16 in August, with some running into 2031. About 75% of fiscal 2027 supply is already committed.
  • Capex steps up, but new supply is a 2028–29 story. Micron will spend about $25 billion in the first half of fiscal 2027, mostly on building construction. That's almost as much as all of fiscal 2026. Bloomberg Intelligence's Jake Silverman says meaningful new industry capacity arrives at the end of 2028: "It's more like a '29 story."
  • The margin dip is about pay. Next quarter's gross margin guide is about 86.25%, down from 87%. Mehrotra put that down to record bonuses for factory workers flowing into the cost of goods sold, not to falling prices.
  • Memory may now be limiting how many AI chips get built. The Futurum Equities hosts say accelerator companies are "giving wafers back to TSMC" because they can't secure memory. They think that favors Nvidia, which has locked up memory "more effectively than any other company."
  • The bear case moved to China. The Circuit's Jay Goldberg says CXMT is on track to add 100,000 wafers a month of DRAM capacity by the end of next year, about a 33% increase, with yields improving. Reports that Apple tested CXMT memory have "caught a lot of attention" at the incumbents.
  • A memory scare that didn't hold up. Western Digital and Seagate fell about 10% on a Nikkei report that Toshiba would double hard-drive capacity. Wedbush's Matt Bryson says the Nikkei "got this one wrong," and that drive-head capacity is fixed until 2029.
  • Nvidia is acting more and more like a bank. It announced a new $150 billion buyback, for $235 billion in total through fiscal 2028. It also disclosed $180 billion of Anthropic hardware commitments. Separately, Broadcom agreed to lend Anthropic up to $42 billion to lease its chips.
  • Quantum: IBM restated its 2029 target. An IBM Quantum product lead said the company expects its first fault-tolerant machine, Starling, "in about three years or so in 2029."

What's New

Micron's CEO: "We cannot fulfill the demand of our customers"

The most important conversation of the week was on Squawk on the Street, "9AM HOUR: Playbook for Q4 and October, Micron CEO Exclusive, Bond Sell-off Accelerates 10/1/26" (October 1). Jim Cramer interviewed Sanjay Mehrotra, Micron's chairman, president and CEO. This is operator testimony from the biggest US memory maker, the morning after its results.

The year in one sentence. "We delivered an exceptional quarter, exceptional year with $133 billion in revenue, more than 250% up year over year, with 87% gross margins in FQ4." He guided to "86.25% gross margin, approximately that for our FQ1, the current quarter."

The contracts. Cramer pointed out that Micron had 16 strategic customer agreements when he visited Boise on August 20 and now has 26. How did it add ten in about six weeks? Mehrotra's answer is the line every memory investor should read twice:

"Not only do we have more customers and customers wanting to do long-term supply agreements with us, but customers are also... bringing upsides to us for demand. In fact, we cannot fulfill the demand of our customers. Customers are extending these agreements. Even now, into 2031 timeframe."

These are not the loose one-year volume deals the industry signed in past cycles. Bloomberg Intelligence semiconductor analyst Jake Silverman explained the difference on Bloomberg Intelligence, "Micron Gives Bullish Forecast, Even as Pay Raises Crimp Margins" (October 1). He called them "take-or-pays" (the customer pays whether or not it takes the chips), unlike "the contracts of a few years ago where it was just one-year volume commitments. It didn't really mean a lot." Zaid Admani of Public.com described the same deals on The Rundown, "Micron Posts Monster Quarter, Apple Readies Smart Home Devices" (October 1): five-year terms that "include price floors, customer deposits, and commitments to take the chips or pay anyways."

The outlook. Mehrotra: "we see going forward in an environment where supply is tight in 27 and 28, even tighter than 26, we see an environment of continued, strong, durable, predictable financial performance." On the Buy Hold Rant, "Ep 58: $MU Q4 Earnings, $META Connect, OpenAI Dots vs Muse, $HOOD Summit" podcast (October 1), co-host Dustin Alper read the matching line from Micron's own results deck: "Industry demand has strengthened since our last earnings call. And we expect memory and storage supply demand conditions to be much tighter in fiscal 2027 and 2028 than they were in 2026."

Why the margin slipped. The guide calls for a gross margin about three-quarters of a point lower next quarter, and the stock dipped on it. Mehrotra explained: "we rewarded our team members in a record fashion as well with their incentive compensation that impacted the OPEX in our fiscal fourth quarter. But as the manufacturing workers' incentive compensation in fiscal fourth quarter carries into the first quarter's cost of goods sold, it does impact our gross margin." Put simply, Micron paid its factory staff record bonuses, and some of that cost lands in next quarter's production costs. Prices aren't falling. Cramer's response: "What is it supposed to have, gross margins of 100%? It's unheard of."

Spending and cash. "In the first half of our fiscal 27, we will be... in terms of CapEx, mostly going into construction... around $25 billion. That's almost as much as what we invested in full fiscal year 26." He expects "well over $100 billion in cash, by the end of our current fiscal quarter." On buybacks, which CHIPS Act terms have limited, he pointed to "December 9th... the second anniversary of the CHIPS Act" and said "we will, you know, share more details... after December 9." He also repeated Micron's US plan: "$250 billion of investment in Idaho, New York, and Virginia... create 90,000 jobs."

Why it matters for the equipment chain. About $25 billion in six months, mostly for buildings, is the strongest sign yet that memory capex for 2027 is going up, not just staying flat. Buildings come before tools, so this is the early stage of the cycle. The orders for deposition, etch, lithography and test tools that go into those buildings come later. CNBC's Seema Modi, reporting from the call on Fast Money, "Micron Reports Results… And Sights & Sounds From Robinhood's Hood Summit 9/30/26" (September 30), added that Mehrotra said "CapEx will increase in 2027 as it adds more capacity here in the US and overseas in markets like Singapore," and that Micron "just poured concrete on its first New York fab with initial wafer output expected in about 4 years."

The wafers going back to TSMC

The most provocative claim of the week came on the Futurum Equities Podcast, "Payrolls Rose 29,000. Micron Guided $61.5B. AMD Bought World Labs. | Ep 59" (October 2). Treat it as a host relaying what industry contacts told him. It's not a disclosure, and it's not confirmed:

"I was talking to some insiders this week and literally there are accelerator companies giving wafers back to TSMC right now because they can't supply. They can't secure the memory. Which by the way is extremely bullish for Nvidia because Nvidia has secured memory more effectively than any other company."

If that's accurate, it matters a lot. An AI accelerator is a processor die from TSMC packaged next to stacks of high-bandwidth memory (HBM). Without the memory, the processor is useless. If smaller chip designers are giving up TSMC wafer slots because they can't get HBM, then memory supply now sets how many AI chips get built, and the companies that locked up memory early (Nvidia above all) gain share from the ones that didn't.

The same hosts set a high bar for the bull case. They see "no supply relief probably till 2030," and one host put Micron's calendar-2027 earnings at "$200 of EPS" with "a 10X multiple, which is not drastic." They also think Micron's cash pile will turn it into an industry investor in the way Nvidia has become one: "MU will be the NVIDIA... of the next couple of years in terms of their VC... corporate... investment portfolio." (One host said cash "swelled to $70 billion this quarter." Mehrotra's own figure was "well over $100 billion" by the end of the current quarter. Those are different dates, but use the CEO's number.)

Bloomberg Intelligence: expectations have caught up

Silverman gave the clearest explanation of why such a strong quarter got a flat reaction: "expectations have kind of caught up to the company to some extent." The long-term contracts were already announced last quarter, and HBM price increases were already expected, so "there's not a lot of incremental that the company can offer." His conclusion: "the upside is somewhat limited now because you can only have so many price increases. You can only have so much margin expansion from here."

He was more useful on supply timing, which is what matters most for the toolmakers. The industry, he said, "was burning tens of billions of dollars" in 2022–23 and "decided not really to invest in additional capacity." Now Micron, SK Hynix "and Chinese competitors" are all adding fabs, "but the way we're kind of viewing it is now like a lot of these fabs are really coming online at the end of 2028 for meaningful capacity... It's more like a '29 story." Until then, buyers of anything with memory in it are "just going to have to pay a little bit more." He noted that "Samsung mentioned $100 increase in at least some of their SKUs this morning."

The sell-side bull: memory, not compute, makes the margin

Before the results, Mehdi Hosseini, senior equity research analyst at Susquehanna (Buy rating, $2,000 price target), made the boldest structural argument of the week on Squawk on the Street, "10AM Hour: Contrarian Market Call, Micron Results Preview, New Eli Lilly Weight-Loss Data 9/30/26" (September 30):

"I make an argument that compute is actually becoming more commoditized. And it's memory that makes the difference."

He frames the AI hardware chain as a fight over who keeps a 70% margin: "Is that memory? Is that foundry? Is that customers like NVIDIA and AMD?" His answer is memory. His target is "a $50 quarterly earning or $200 annualized," supported by an operating margin "in the 70%." What's different from past cycles, he says, is the spread of demand. The old memory cycle "had only one demand driver and that was a smartphone." Now there is "diversification in application and customers." He was candid about why the stock trades so cheaply: "there's always been this big divide between sell side and buy side," and the sell side is "struggling in how to justify a margin sustaining by 70 percent."

That claim (memory over compute, and arguably over foundry) is the first direct challenge we've heard in a while to the idea that TSMC is the natural home for AI profits. It's one analyst's view, but a serious one.

The skeptics on the desk

The best of the doubt came from the Fast Money panel on results night (September 30), all commentators and traders rather than insiders:

  • Dan Nathan said the size of the beats is shrinking: "they basically beat consensus by about 5, 6% of earnings and sales... the quarters of having this like 50%, 100% kind of beats, they're gone."
  • Guy Adami quoted the old Amazon line: "your margins are our opportunity. And although competition is not coming tomorrow, competition is coming." He also noted that a stock "supposed to move 7 or 8% given the options market is not moving at all."
  • The host, Brian (sitting in for Melissa Lee), pointed out that 49 of 51 covering analysts rate it a buy, with no sells: "I'm worried about the next 12 months, not the previous 12 months."
  • Panelist Steve pushed back: "we are talking about a three-horse race... there's plenty of room for Micron to gobble up someone else's share."

On valuation, the Saxo host made a striking comparison on Saxo Market Call, "Burnham declares end of Thatcherism. Micron earnings on tap." (September 30). Micron's quarterly operating income is running "on the order of $33 billion... more than half of what NVIDIA does," yet it is "valued at less than a quarter of NVIDIA's $5.x trillion." The market, he said, is pricing a future in which "this incredible profit growth could be just a temporary phenomenon."

China's memory makers are getting "bigger pretty quickly"

The sharpest look at the supply risk came from long-time semiconductor analysts Ben Bajarin and Jay Goldberg on The Circuit, "Memory Pricing Pressures, Meta's Muse AI, and the Agentic CPU Crunch" (September 28). Bajarin opened by calling the Chinese players "the fourth and fifth horsemen of the memory apocalypse."

Goldberg's specifics: YMTC (China's NAND flash maker) "has pulled forward some of its capacity expansion," while "CXMT is on track to add another 100,000 wafers per month by the end of next year. So that's a 33% increase for them year on year in capacity." CXMT's "yields are improving," and it is "starting to branch out in different flavors of DRAM." He was careful: "not like the sky is falling, China is coming kind of news. But just a periodic reminder that these are serious providers of memory and they are getting better and bigger pretty quickly."

The detail that should worry the incumbents most concerns Apple. Goldberg recalled reports "two, three weeks ago" that Apple "had tested CXMT memory and was pushing very hard to get a waiver... from the US government to be able to buy it." He added: "If Korea and Idaho can't provide the memories, China can." Bajarin's follow-up: once a customer like Apple qualifies a second supplier, "that would hurt... barn doors open."

Two more points from the same conversation:

  • Consumer electronics are taking the hit. Bajarin cited two forecasts that week "downgrading consumer devices in volume, really across the board... because of the memory price bug." Goldberg said he is hearing, "mostly coming from interested parties like memory company CEOs," that consumer memory pricing may be plateauing even as "data center pricing... is just going to keep going up."
  • Pushback on ever-taller HBM. At Hot Chips, Goldberg said, memory presenters "faced a pretty hostile crowd," with engineers asking "why are you moving to hybrid bonded memory, why are you going higher... that is just going to be more and more wafers." That points straight at the equipment chain. Hybrid bonding and taller HBM stacks are among the most tool-intensive steps in advanced packaging. If chip designers rebel against them, that's a real risk to that part of the toolmakers' demand.

A memory scare that wasn't

On Fast Money, "Nasdaq Hits Record After Soft Jobs Report… And The Fed's Next Move After Jobs Report 10/2/26" (October 2), Wedbush hardware analyst Matt Bryson took apart what the show called "today's memory meltdown." Western Digital and Seagate fell about 10% on a Nikkei report that Toshiba would double hard-drive capacity. His verdict: "The down 10% is too much because the Nikkei got the story." The actual Toshiba plan is that "the capacity in the Philippines is going to double from FY25 to some point FY27," and normal technology improvement already adds "25% plus bits" a year.

His larger point applies to memory too. What really stops a glut is time: "the real thing that prevents overcapacity is just the amount of time to bring on fabs." Drive makers "don't have any more heads until '29." For memory: "If you don't have the clean room space, you can't bring on new capacity. You can't blow up the industry." On valuation, memory names at the peak of a cycle have "historically" traded at "mid-single-digit type earnings multiples," but "because we have a year plus before you have to worry about there being a cycle, because you're going to have these massive buybacks, there should be some premium to that," perhaps "seven times instead of five times or eight times instead of six times."

Nvidia as the industry's bank

Nvidia's news on Monday, September 28 matters to the foundry chain mostly through financing. On Squawk on the Street, "10AM Hour: Nvidia's $150B Buyback, Mortgage Rate Outlook, Meta Pulls Back 9/28/26", Bernstein senior analyst Stacy Rasgon (Outperform, $400 target) did the math. The new $150 billion authorization means about $235 billion of buybacks by the end of fiscal 2028, "over the next six quarters. It's like $40 billion a quarter, which is double what we were modeling," and "about 50% of my free cash flow across fiscal 27 and 28."

On Bloomberg Tech, "Nvidia Goes Big on Buybacks as SpaceX Reaches Orbit" (September 28), the host noted Nvidia had disclosed that "Anthropic is a case study to date $180 billion of NVIDIA hardware commitments." Bloomberg Intelligence credit analyst Robert Schiffman described what Nvidia is doing with its cash: "they're making credit-style investments where they're both leasing facilities. They're providing residual value guarantees. They're providing support behind the value of their chips beyond their depreciable life... they're just acting as a bank for everybody." He also cited Nvidia's deal "with the biggest alternative asset managers out there to help raise $500 billion."

Broadcom is doing the same thing. On the Bloomberg Intelligence episode (October 1), one of the hosts reported that "Broadcom has agreed to lend Anthropic... as much as $42 billion to lease its chips," and Silverman said Anthropic's prospectus shows "just north of $500 billion in terms of commitments over several years." The host, a former analyst, was openly uneasy: "my radar goes up big time." Silverman's historical comparison was "vendor financing agreements in 2001."

On Odd Lots, "Everything in Markets Is Now Moving Incredibly Fast" (October 1), the guest strategist described this as "NVIDIA trying to effectively credit wrap, make everyone the safest AI borrower possible." He also cited a Goldman note finding that "about half of earnings growth this year is expected to be just driven by hyperscaler CapEx," and a striking market statistic: only two S&P 500 semiconductor companies made 52-week highs in the third quarter, "AMD and Skyworks."

Two operators on what is actually scarce

Walter Goodwin, founder and CEO of AI-chip startup Fractile, gave a chip designer's view on No Priors, "The Future of Frontier Model Architectures with Walter Goodwin, Fractile Founder and CEO" (October 2). His main point: almost every AI chip on the market makes the same bets. "All of these chips, you have HBM... You have the same kinds of bets on tensor cores... the same advanced packaging with TSMC." He sees memory bandwidth (how fast data moves between memory and processor) as the bottleneck the industry has neglected: "we've scaled flops like a million fold in the last 20 years. Memory bandwidth has gone up about 40x in the same time frame." Fractile is working "with memory vendors, as well as our sort of logic foundry partners" to get "much, much, much higher bandwidth to DRAM," targeting "25 times more bandwidth per chip than an HBM based chip," with a platform "ramping in the second half of next year." He also put the hyperscalers' in-house chips bluntly: "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."

Chase Lochmiller, co-founder and CEO of data-center builder Crusoe, spoke on 20VC, "The Future of Datacentres: What You Need to Know..." (October 3). He added a useful data point to last week's argument about how fast AI chips lose value. When Crusoe bought Hopper GPUs in 2023, lenders worried "we don't even know if this is going to be valuable after year three." Now: "Here we are three years later and the prices being charged for utilizing hoppers is higher than the rates that were being charged three years ago when they were brand new." Crusoe depreciates over six years, which he called "sort of the standard across the industry," and expects its managed services to "extend the depreciation cycle beyond six years." His list of the real constraints is power and labor ("a finite amount of skilled trade workers"). He also said customers are now being asked "to commit to compute in 2028," which he called "an eternity from now."

Power is the constraint that keeps coming up. Michelle Weaver, head of US thematic equity research at Morgan Stanley, said on The Wall Street Skinny, "How NVIDIA Is Secretly Coming for Anthropic & Open AI" (September 26) that her team is "forecasting a 57 gigawatt shortfall for power needed through 2028," and that even with fuel cells, gas turbines and converted bitcoin-mining sites, "we still end up with around a 30 to 40 percent shortfall."

The rest of the supply chain is tight too

Investor Ben Pouladian made a point on Monetary Matters with Jack Farley, "Token Bill Dwarfs CPU Bill..." (September 30) that's worth noting for anyone tracking bottlenecks. Farley showed a TrendForce table rating each part of the hardware stack: CPUs, DRAM, NAND, hard drives, ABF substrates (the base boards advanced chips sit on) and MLCCs (tiny capacitors). All were rated "tight to very tight except for GPUs, which was balanced." That may explain why Nvidia's stock has lagged other chip names this year even as it reports record results. On foundry capacity specifically, Pouladian said CPUs are "supply constrained based on the two fabs that are making these... either coming out of TSMC or Intel. So Intel really needs to up its fab game and get it online." He was also skeptical of AMD's first rack-scale system, Helios: it's "basically twice the size of a Blackwell system," it "weighs like eight tons," and "already I'm hearing that there's potential delays in scaling it." He is a self-described long-time Nvidia bull, so weight that view accordingly.

An Intel footnote, via chip-design software

Intel Foundry had no direct coverage this week. An indirect data point came from the Chip Stock Investor Podcast, "Why Synopsys Lagged Cadence (And Why It May Be Cheap)" (October 1). Intel was Synopsys's largest customer, "low teens percent of revenue in fiscal 2023 and 2024," falling to "under 10%" by 2025. Hosts Nick and Kasey Rossolillo suggested Cadence "picked up some extra business from Intel" after Lip-Bu Tan, Cadence's former CEO, moved to Intel. The same episode noted that GlobalFoundries paid $450 million for Synopsys's ARC processor IP, pairing it with MIPS to build a RISC-V design platform for automotive and industrial customers. That's a small but real example of a foundry moving up into chip design.

The Debate

The bull case had its best week since June. Micron's CEO says the shortage tightens through 2028. Customers are signing take-or-pay contracts into 2031. A Susquehanna analyst argues memory, not compute, keeps the 70% margins. Bloomberg Intelligence puts meaningful new supply out to "a '29 story." Wedbush's Bryson argues the real protection against a glut is that clean rooms and components take years to build. A podcast host relays that chip designers are returning TSMC wafers for lack of memory. Data-center operators say power, not demand, is the limit, and that three-year-old GPUs rent for more than they did new. Taken together, these voices describe an AI hardware cycle limited by supply, with orders visible for years.

The bear case was voiced too, just more cautiously. Mainly:

  • China supply. CXMT is adding 100,000 wafers a month by the end of 2027, and Apple has reportedly tested its memory. That's the clearest threat to the "three-horse race" (Micron, SK Hynix, Samsung).
  • Peak beats. Dan Nathan's point that beats have shrunk from 50–100% to 5–6%, and Silverman's that "you can only have so many price increases," together say the rate of improvement has peaked, even if the level hasn't.
  • Vendor financing. Nvidia's guarantees, Broadcom's $42 billion loan and Anthropic's $500 billion of commitments drew explicit comparisons to 2001.
  • Consumer damage. Phone and PC forecasts are being cut because of memory prices, and Samsung is raising some prices by $100.
  • Engineer pushback. At Hot Chips, designers objected to ever-taller, wafer-hungry HBM stacks.

What wasn't voiced on the podcasts this week: no one made the case for Intel 18A or Samsung Foundry catching up to TSMC. No one discussed China's toolmakers (Naura, AMEC, SMEE, Piotech) taking share from US, Japanese or European equipment makers; China came up only through its memory makers. No one argued that mature-node chip demand is digesting. Taiwan geopolitical risk didn't come up. We won't fill those gaps from memory.

Read-Throughs

Advanced packaging and HBM. Rationing memory makes advanced packaging more valuable, not less. Every accelerator that does get built needs HBM stacked and bonded next to it. Fractile's Goodwin confirmed that "the same advanced packaging with TSMC" sits under nearly every AI chip on the market. The risk to watch is Goldberg's Hot Chips report: if engineers push back on hybrid bonding and taller stacks, the most tool-intensive part of packaging could grow more slowly than the current plans assume.

Wafer-fab equipment. Micron's roughly $25 billion first-half spending, "mostly going into construction," is building now, with tool orders to follow. Silverman's end-2028 capacity timeline means those tools need to be ordered and installed through 2027–28. SK Hynix and "Chinese competitors" are adding fabs too. That supports memory-driven demand for deposition, etch and inspection tools, the core of Applied Materials, Lam Research and KLA. No toolmaker spoke this week, so this is an inference from the customer side, not company guidance.

TSMC and fabless customers. Two opposing signals. On one side, if accelerator companies are returning wafers for lack of memory, some of TSMC's leading-edge AI demand is being limited by its partners' supply chains, which reinforces Nvidia's advantage. On the other, Hosseini's claim that "compute is actually becoming more commoditized" goes after the profit pool foundry bulls count on. Neither is confirmed by TSMC itself. Its own monthly sales and quarterly results will be the test.

China's domestic chipmakers. This week the China risk ran through memory (CXMT, YMTC), not tools. The Apple-CXMT waiver story is the one to follow. A major Western customer qualifying a Chinese DRAM supplier would change Micron's pricing power more than any capacity number.

Quantum hardware (frontier view). On The Post-Quantum World, "Qiskit and Quantum Utility – with Sanket Panda of IBM Quantum and Izhar Medalsy of Quantum Elements" (September 30), Sanket Panda of IBM Quantum restated the company's target: "we expect that to be in about three years or so in 2029 when we expect to deliver IBM Starling, which we believe will be the first fault tolerant quantum computer." (The transcript renders it "Starlink." IBM's roadmap name is Starling.) Izhar Medalsy, co-founder of Quantum Elements, said its error-suppression software raised the share of accurate results on a benchmark of 4,000 circuits from "effectively zero" to "a quarter," and that it is "the first to show" error suppression at the logical-qubit level. That matters because logical qubits, built from many physical ones, are what fault tolerance depends on.

On the money side, Marie Lepske, general partner at Swiss deep-tech fund Constructor Capital, said on The New Quantum Era, "Quantum Venture Investing from the Lab to the Market with Marie Lepske" (September 28) that the field has seen "already, I think, 10 IPOs and SPACs in total" in 2025–26 and that the "market is becoming overhyped." She warned that non-specialist investors find it hard to tell "if this is real roadmap or just it's written for the SPAC." Host Sebastian Hassinger noted that portfolio company QuEra closed a round "of over $230 million, led by Google Quantum AI and SoftBank, with NVIDIA participating." Her fund is $110 million, with "more than 50, even 60 percent" going to next-generation computing including quantum. The frontier view for chip hardware hasn't changed: 2029 is still the date that matters, and until then specialized packaging and control hardware stay a research-budget business.

What Changed

  • Micron settled last week's debate on the numbers, not in the stock. Last week, Sara Awad on Investing Experts, "What will Micron's gross margin guide be?" called the gross-margin guide the deciding number, and the hosts of the Futurum Equities Podcast, Ep 58 said there was "not one" sign of pricing power fading. The guide came in at about 86.25% versus 87% actual, and management blamed bonuses, not prices. Futurum's side was right on the fundamentals. The stock's flat reaction suggests Awad's wider worry, that the market won't pay up for peak margins, still holds.
  • The operator drought broke, but only at the memory layer. For months the newsletter has noted that no foundry or toolmaker executive was speaking. This week a top memory CEO gave a detailed account, and two infrastructure founders added theirs. TSMC, ASML and the toolmakers remain silent.
  • The China story moved from tools to memory. Last week CXMT came up only as a capacity number: Awad said on Investing Experts that it was expected to end the year within about 20,000 monthly wafer starts of Micron. This week the specifics were a 100,000-wafer-a-month expansion, better yields, and the Apple waiver story.
  • The GPU-value argument gained an operator data point. The debate over how fast AI chips lose value has so far been argued mostly by investors and analysts. This week an operator weighed in: Crusoe's CEO said Hopper rental rates are higher now than when the chips were new.
  • Nvidia doubled down on financing the buildout. A $235 billion buyback plan, $180 billion of Anthropic commitments and residual-value guarantees, all discussed on Bloomberg Tech, made the circular-financing question more concrete. That question has hung over this newsletter since August.

What to Watch

  • Micron buybacks after December 9. Mehrotra said he would "share more details" after the CHIPS Act anniversary.
  • Confirmation of the TSMC wafer give-backs. If true, it should show up in what TSMC says about AI demand and packaging use when it reports.
  • The Apple-CXMT waiver. Any decision would be the biggest single change to memory pricing power.
  • Toolmaker results. The equipment makers' fall results should show whether Micron's construction spending is turning into tool orders yet.