Newsletter · 2026-07-11 · Ashutosh Agarwal

Samsung Out-Earns Nvidia as SK Hynix Lands a Record US Listing - Semiconductor Podcast Briefing - Week of July 11, 2026

Samsung briefly out-earned Nvidia on a memory-driven quarter, SK Hynix priced the largest first-time US listing by a foreign issuer (roughly $27 billion), and Micron lifted its US spending to $250 billion, even as Nvidia shed about $1 trillion in market value and Apple locked Broadcom into custom ASIC chips through 2031. Our synthesis of the semiconductor podcast tape for the week ending July 11, 2026.

Semiconductor Podcast Briefing

Week of July 11, 2026: Samsung Out-Earns Nvidia as SK Hynix Lands a Record US Listing

This was a memory-dominated week on the podcasts. The single loudest story: a South Korean memory maker (Samsung) briefly out-earned Nvidia, its cross-town rival SK Hynix priced the biggest-ever first-time US share sale by a foreign company, and Micron lifted its US spending plan to $250 billion, even as Nvidia quietly shed roughly $1 trillion in market value and chip stocks whipsawed between "the trade is dead" and "chips trade back on" inside of 48 hours. Below is the full debrief, organized by topic, with who said what, the specific numbers, and the pushback.

A few plain-English definitions used throughout: hyperscaler = a giant cloud/AI company (Amazon, Microsoft, Google, Meta, plus OpenAI/Anthropic) that buys chips at enormous scale; capex = capital expenditure, i.e., the money these firms spend building data centers; HBM = high-bandwidth memory, the specialized memory stacked next to AI chips; DRAM = the fast "working" memory in devices; NAND = the slower "storage" memory (SSDs, phone storage); ASIC = a custom chip built for one specific job; WFE = wafer-fab equipment, the machines used to manufacture chips; foundry = a contract chip factory.

TL;DR, Things That Mattered This Week

• Samsung out-earned Nvidia on a quarterly basis. Samsung posted Q2 operating profit of about $58.5 billion (about 89.4 trillion won), up ~19x year-over-year, on record revenue that more than doubled to ~$112 billion, with roughly 94–96% of profit coming from memory. For comparison, one host noted Nvidia's most recent quarterly profit at ~$53 billion. Yet Samsung's stock still fell ~7%, because it beat estimates by only ~6% and the stock had already roughly doubled this year.
• Memory pricing is going vertical. Two shows put hard numbers on it: HBM prices were hiked ~90% in Q1 and a further 50–60% in Q2, with Samsung reportedly planning another ~20% in Q3; separately, Bloomberg data cited DRAM +40% and NAND +50% last quarter. Micron's DRAM average selling price was up ~60% YoY and NAND ~80%, per Hightower's Stephanie Link. Gross margins are now extraordinary, Samsung ~52%, SK Hynix ~72–80%, and Micron's fiscal Q3 gross margin jumped to ~85% from 39% a year earlier.
• SK Hynix priced a record US listing. It set its US ADR at $149 (1 ADR = 1/10 of a Korean share), raising ~$26–28 billion (Jim Cramer cited $29B), in a deal reported more than 7x oversubscribed (~$160B of demand), the largest first-time US listing by a foreign issuer, ahead of Alibaba and behind only Saudi Aramco's domestic deal.
• Nvidia's ~$1 trillion slide. Nvidia lost roughly $1 trillion of market value in under two months and now trades at its cheapest since before the AI boom, variously cited at ~14x, 18x, and ~20x forward earnings (Cramer: "cheaper than Hershey," cheaper than half the S&P). The bear worry: everyone is building their own inference chips, and ~80% gross margins won't last.
• Apple locked in Broadcom through 2031 for custom ASIC chips powering a new Apple AI server ("Baltra"), plus Meta signed what one host called "the biggest AI deal in history" to buy AMD chips.
• Capex direction of travel is still up, not down. Street estimates for 2026 hyperscaler capex growth jumped from 17% to 23% in a few weeks, and Amazon's fresh $25 billion bond raise plus Google's comments point some analysts to ~37% growth. Amazon's capex path is modeled going from ~$200B to ~$300B by 2027.
• The bear case is cyclicality. Samsung alone plans ~$70 billion of capex this year; Micron is going to $250B (from $200B). History says memory booms end in gluts, but this week's consensus was that new fabs don't arrive until late 2027, and unusual five-year customer contracts now give memory makers rare forward visibility.
• China keeps leapfrogging under export controls. A delivery-app parent (Meituan) trained a frontier open-source model ("Longcat," 1.6 trillion parameters) on an estimated 40,000–50,000 Huawei Ascend chips and zero US GPUs, and China now controls the specialty-chemicals supply chain so completely that Taiwan is reportedly 100% reliant on Chinese NF3 gas.

1. AI chip demand & hyperscaler capex (NVDA, AMD, AVGO, MRVL)

The core debate all week: is AI spending still accelerating, or are we near a plateau?

The "still early" camp had the louder voices. On Closing Bell (Jul 7), Gene Munster of Deepwater Asset Management laid out the capex math directly: "The street's looking for 23% capex growth from the hyperscalers for next year. And now a few weeks ago that was 17%." Factoring in Amazon's new $25 billion bond sale and Google's capital-raise comments, he said, "that probably suggests that the numbers are going to be up around 37%, so much higher." His framing: Amazon tapping the debt market "proves that the AI buildout is in its early innings," and the real question is "do you believe that we are so early in AI that this can power a continued higher numbers for the next several years? And I think the answer is yes." (Closing Bell)

On the same show, HSBC's chief multi-asset strategist Max Kettner dismissed the recurring monetization scare: "This is the third wave where we have that sort of AI overspending... after that, you're kind of done, right? Because suddenly you start realizing that the numbers are just there... and the numbers actually don't justify that sort of scary narrative." His trade: lean less into semis and build-out names, more into the hyperscalers after their ~20% drawdown from mid-May to end-June. (Closing Bell)

Nick Rossolillo of Chit Chat Stocks, on Investing Unscripted (Jul 8), sized the flow: hyperscalers (AWS, Azure, Google Cloud, Oracle) are deploying $750–800 billion of capex this year, heading "toward a trillion dollars or more next year," money that flows down the whole chip supply chain "from design software through fabrication to deployment." (Investing Unscripted)

On Morning Call (Jul 7), analyst Dan Newman went bigger still, forecasting "over 10 trillion dollars in AI infrastructure spending between now and 2030," calling the current moment "the pregame," noting most systems still run chips "two to three generations old," with Nvidia's newest Blackwell only a small share of the installed base. (Morning Call)

Hightower Chief Investment Strategist Stephanie Link, on Money Rehab (Jul 6), argued the $800 billion of annual hyperscaler spend is "real and accelerating," pointing to picks-and-shovels backlogs "growing 34% year-over-year (versus historical 5% average)" at companies like Quanta Services, GE Vernova, and Vertiv. On Nvidia specifically: "Because it hasn't done anything in the last 6 months, it's actually trading at about 14 times forward estimates for a company that's going to grow 50%... It's crazy cheap for what you're getting," though she flagged rising competition from Amazon and Alphabet's custom chips, and that "everyone owns NVIDIA," so she prefers Marvell, Broadcom, and AMD around it. Her honest caveat elsewhere: the whole thesis "breaks if capex drops to $500 billion next year." (Money Rehab)

The bear/contrarian voices. On Bloomberg Talks (Jul 6), Morgan Stanley's Mike Wilson made the case that semiconductor stocks could correct 30–40% even while the structural capex bull market stays intact, describing three "mini-cycles" since ChatGPT's launch and noting semis' earnings-revision breadth near a peak (~75%) now set to decelerate. (Bloomberg Talks) On Animal Spirits (Jul 1), the hosts cited a Nomura chart showing hyperscaler free cash flow (Amazon, Microsoft, Google, Meta, Oracle) projected to fall from roughly $700 billion to essentially zero over the next 12 months as capex swallows it. (Animal Spirits Podcast)

Marvell got a pointed skeptic. Nvidia's Jensen Huang has publicly called Marvell "the next trillion-dollar company," but on Chip Stock Investor (Jul 7), Nicholas and Kasey Rossolillo pushed back hard. They walked through Marvell's Q1 FY2027: record revenue "just over 2.4 billion... up 28% year over year," guidance for "2.7 billion at the midpoint, which would be 35% year over year," and CEO Matt Murphy's target of "$16.5 billion in annual revenue" in FY2028. But their reverse DCF (a way to back out what growth the stock price already assumes) at ~$245 implied "an average per share profit growth rate of 47%... over the course of the next five years." Their verdict: "this is a point where we would pass on a company like Marvell and look for some other more compelling values," adding that "it's probably AMD and or Intel that are the next one trillion dollar companies." They also flagged dilution: shares outstanding rose from 856 million to over 890 million as Marvell issued stock for acquisitions (including Celestial AI). (Chip Stock Investor Podcast)

The rotation within the chip trade. On Power Lunch (Jul 9), Jay Hatfield of Infrastructure Capital Advisors recommended rotating from momentum names into value, specifically backing Broadcom over Marvell, Broadcom "trading at 0.6 PEG ratio" (price/earnings-to-growth; below 1 is considered cheap) with potential upside toward $600, versus Marvell above 1.0. (Power Lunch) On Power Lunch (Jul 6), Dan Ives argued "chips are going to continue to lead this market higher," citing "12 to 1" supply-demand in Asia, and noted Astera Labs, AMD, and Qualcomm each up ~6% that day. (Power Lunch)

A structural threat to Nvidia's inference lead. On Limitless (Jul 7), the hosts argued inference (running AI models, versus training them) has grown "from one-third of AI compute two years ago to two-thirds today," heading "to 80% by year-end," while Nvidia's GPUs run at only "30–40% utilization" on inference. They pitched the startup Etched's new ASIC as hitting "80–90% inference utilization" using "75% less power," the kind of specialization that custom-chip challengers are chasing. (Limitless: An AI Podcast)

2. Memory pricing (HBM, DRAM, NAND, MU, SK Hynix, Samsung)

This was the richest topic of the week, and the numbers were striking.

On Limitless (Jul 9), Ejaz and Josh laid out the memory monopoly in detail. Samsung is "the number two largest provider of... high bandwidth memory," and SK Hynix "own[s] about 60% of the entire HBM market." Their explanation for why prices can rise so far: "one gigabyte of HBM consumes the factory capacity of roughly four gigabytes of regular DRAM. So every wafer that becomes AI memory is a wafer that never becomes laptop or phone memory," which is why Apple just raised device prices across the board. The margin math they cited: "in a grocery store for every $100 of chips sold, they collect about three... Apple... is about 30%. Samsung is 52 and SK Hynix is 72." And the pricing trajectory: "In Q1, they hiked their prices up 90%... In Q2, it went up a further 50% to 60%... Samsung [is] going to hike it up another 20% in Q3." They also noted Samsung's memory workers are getting bonuses "worth six times their annual salary." (Limitless: An AI Podcast)

On The Rundown (Jul 7), Zaid Admani cited the market-wide moves: "Bloomberg says that DRAM prices rose more than 40% last quarter, while NAND prices jumped more than 50%." (The Rundown)

Stephanie Link (Money Rehab, Jul 6) put the same dynamic on Micron: "DRAM, their ASPs, average selling price to their customers, was up 60% year over year... And in NAND, pricing was up 80%. Pricing power is king... they are minting money." She added Micron "signed 16 license agreements this past quarter valued at $100 billion... $22 billion of it in cash," contracted revenue through 2028. (Money Rehab)

On Mac OS Ken (Jul 6), the host relayed a telling detail on how far margins have swung: Micron Chief Business Officer Sumit Sadana blamed the current shortage on customers (implicitly Apple) who "took advantage to pay rock-bottom prices" during the last downturn when "Micron's gross profits went negative," even as "Micron's fiscal third quarter gross margin more than doubled to almost 85% from 39% for the same quarter a year ago." He also noted IDC analyst Nabila Popal called the shortage "an absolute existential crisis for companies such as smaller Android phone manufacturers or local players... making devices below $100," who "won't be able to get the memory because memory suppliers are only answering calls of big players." (Mac OS Ken)

Is it durable? On Bloomberg Intelligence (Jul 8), Mandeep Singh made the bull case for memory over the AI hardware names being sold off: "if there's any sector that has the best pricing power right now, it's memory... companies with pricing power, with margin expansion, with probably the best margins they ever had." (Bloomberg Intelligence)

The dissent came from the same house that owns the stocks. On Chip Stock Investor (Jul 4), the hosts warned that over the next "3–5 years," easing supply will lower these "sky-high profit margins" as all three majors add capacity, and that Chinese competitor CXMT plans to "raise capital via IPO to expand manufacturing," pressuring pricing. (Chip Stock Investor Podcast) On Investing Unscripted (Jul 8), Nick Rossolillo made the structural point that memory "is a commodity without sticky moats unlike CPUs/GPUs which ship with proprietary software instruction sets," a reminder that this may still be a cyclical, not a permanent, re-rating. (Investing Unscripted)

SK Hynix's blockbuster US listing was the week's marquee event. On Bloomberg Tech (Jul 9), the mechanics were spelled out: SK Hynix planned to price at "$149 per American depository receipt... a 3% premium" to its Korean close, with each ADR equal to "one-tenth of one share," in a deal "over seven times oversubscribed... north of $160 billion in demand," "a record-setting listing for a foreign issuer here in the US, so outpacing Alibaba but falling short of Saudi Aramco's domestic listing." Bloomberg's Ian King called it "a rags-to-riches story... revenue tripling to more than $200 billion this year at an 80% gross margin," for a company that "came very close to being bought by its competitor, Micron." Franklin Templeton's Katrina Dudley said that while you can build a seven-point AI-capex bear case, "it's the bull case that really should be winning here... we think there's ROI that supports these investments." (Bloomberg Tech) On Rich Habits (Jul 9), speakers noted SK Hynix "controls over 50% of the global HBM market and secured 70% of NVIDIA's next-generation orders," with demand outrunning capacity "for three years" per its head of sales, but warned a shift "from shortage to surplus in late 2026/early 2027" could compress the 30–80% price hikes. (Rich Habits Podcast)

Micron as the "cleaner" way to play it came up on Futurum Equities (Jul 8), which argued that with SK Hynix's listing 3x oversubscribed, US-listed Micron is the cleaner entry point from here on valuation. (Futurum Equities Podcast)

3. Semiconductor capital equipment / WFE (ASML, AMAT, LRCX, KLAC)

Equipment got less dedicated airtime, but the signal was consistently bullish.

On the Rob Black Show (Jul 9), the host relayed comments from Applied Materials CEO Gary Dickerson that "the AI chip boom is far from over," with "the semiconductor industry [having] tremendous visibility into the next 24 months for demand." His pitch for the tool makers as a lower-risk way to play AI: "Rather than trying to pick the next AI winner, companies that supply picks and shovels like Applied Materials, LAM Research, KLA Tencor, ASM Lithography... could continue benefiting no matter which AI models ultimately come out on top." He did flag Nvidia's ~80% margins as a future risk "when those margins start slipping back to 40 percent norms," but framed that as "down the road." (Rob Black Show)

Stephanie Link (Money Rehab, Jul 6) singled out semi-cap equipment as long-term winners regardless of which chipmaker wins: "Lam Research, Applied Materials, KLA... they are winners over the long term, 'cause it doesn't matter who wins." Her tactical advice was to wait for pullbacks: "20% below these levels, if these stocks fall, that's where I think you wanna start a position." (Money Rehab) In Closing Bell's "Chips Trade Back On" (Jul 9), Lam Research was highlighted as one of the semi-cap names leading the rebound after a report that Meta's capex is heading higher, and Applied Materials' CEO was cited pointing to "better long term visibility past 2027." (Closing Bell)

Note on coverage: No single equipment-earnings-focused podcast surfaced this week, these were references inside broader market shows. ASML and the WFE names report over the coming week, which should generate more direct commentary (see "What I'm Watching").

4. Foundry & manufacturing (TSM, INTC, GFS)

The most substantive foundry discussion came from The Circuit (Jul 6), which flagged Samsung's foundry comeback. The hosts, building a foundry model, argued Samsung is "officially back in the conversation compared to two years ago" at leading-edge nodes (3nm, 2nm and below) alongside TSMC and Intel. Samsung's differentiator: "they're leveraging memory for logic... they are working on stacking memory on logic," and "it's easy to get foundry business when you have memory, which is the scarcest thing in the world right now." One host admitted, "I've been very vocally skeptical about Samsung's ability to stay in the market, and I was wrong. They... have turned it around," crediting a new leadership team. On Intel, the read-through was that packaging (specifically EMIB) "got them in the door," and customers are "increasingly interested in [Intel's] logic process as well." Both Samsung and Intel, the hosts stressed, still need a marquee external anchor customer, with AMD, Qualcomm, and a rumored Google TPU variant floated for Samsung. They also flagged Infineon opening a new ~€9–10 billion, 300mm fab in Dresden (funded partly by the EU Chips Act). (The Circuit)

On the smaller end, The MoneyFlows Show (Jul 2) pitched GlobalFoundries (GFS) as a silicon-photonics pure play, citing GFS guidance from a TD Cowen conference of "silicon photonics revenue... in 2026, $400 million... a $1 billion run rate in 2028... [and] 2030, $2 billion." It paired that with Tower Semiconductor (TSEM), which reported "$1.3 billion of contracted silicon photonics revenue for 2027" and targets "$2.8 billion in annual revenue and $750 million in net profit in 2028." (The MoneyFlows Show)

Note on coverage: TSMC itself was mostly referenced indirectly this week (as Nvidia's and Samsung's manufacturer, and as capacity-constrained). Its earnings next week should change that.

5. Analog / auto / industrial semis (TXN, ADI, MCHP, ON, NXPI, STM)

This was the thinnest topic, but one podcast covered the auto angle in real depth. On Automotive News Daily Drive (Jul 2), SBD Automotive consulting director Alex Euler explained why the DRAM shortage hits automakers especially hard. The problem is structural: cars use last-generation memory locked in by "long sourcing cycles... designed and specified five, sometimes even ten years ago," and the memory suppliers "are going to go where the margins and the volume are." His blunt assessment of the auto industry's position: "If you look at automotive, the volume isn't there... it puts them in a difficult position of where to dedicate that fab time," with automakers "a little bit of an orphan in that entire process," squeezed between "legacy computing platforms and next-generation architectures." Tier-one suppliers, the show argued, "may be the ones taking the hardest hit." (Automotive News Daily Drive)

Note on coverage: No podcast this week featured a named analyst giving a fresh price target on TXN, ADI, MCHP, ON, NXPI, or STM specifically, the analog names remain under-discussed relative to memory and AI accelerators.

6. China / export controls / tariffs

The standout was ChinaTalk (Jul 7), a deep dive into the specialty chemicals and gases that fabs depend on. A materials expert described China's "Big Fund" as having invested "$120 billion" to build near-complete domestic self-sufficiency, and, crucially, a supply chain the West now depends on. On the cleaning gas NF3 ("the janitor of the semiconductor world"): China's domestic need is "about 8,000 tons," yet "one producer in one province... [is] making 55,000 tons a year," so those producers are "busy recruiting international sales... trying to sell their NF3 internationally into Taiwan, into South Korea, into the US." The dependency is stark: "Taiwan is 100% reliant on Chinese supply chains today... If the Chinese government decided to put an export restriction on NF3, then the Taiwanese fabs would shut down." Underlying it all: "China's got 70% of the world's fluorospar" (the feedstock for the fluorine gas family). The one Chinese vulnerability he named was helium, where "the geology" leaves China a net importer. (ChinaTalk)

On VoxTalks Economics (Jul 8), CEPR President Beatrice Weder di Mauro made the counterintuitive case that US export controls accelerate Chinese innovation: "US restrictions on exports of advanced semiconductors... China, against whom these restrictions were directed, has seen a bigger incentive to innovate. So think about DeepSeek... Or think about Huawei. The lack of access to advanced chips means that they are experimenting with stacking chips or linking a large number of chips together. This type of innovation is not conducted in the West because firms are focusing more on shrinking the size of transistors... the restrictions... could result in greater innovation and potentially even leapfrogging." (VoxTalks Economics)

On the demand side, several shows tracked Apple reportedly seeking permission to buy memory from blacklisted Chinese firms (CXMT and YMTC) to ease its own shortage. And on Squawk on the Street (Jul 7), a CNBC feature on China's Hesai illustrated the price gap driving the export-control fight: a Chinese automotive LiDAR unit "costs about $200 a unit," while a comparable US competitor is "50 percent more," Hesai is on a US blacklist and "appealing... in federal court" while still shipping into Nvidia's autonomous-vehicle platform and US robotics. (Squawk on the Street)

A related policy wrinkle from Mac OS Ken (Jul 6): Micron is lobbying Congress for the "MATCH Act," which would tighten allies' (e.g., ASML's) equipment exports to China, giving it "about a 30% chance of being enacted," even as the industry group SEMI urged the administration to "stay out of the memory thing" and instead offer consumer tax credits to offset rising device prices. (Mac OS Ken)

7. Earnings & post-earnings reactions

Samsung's print dominated. On Bloomberg Intelligence (Jul 7), BTech anchor Ed Ludlow captured why record numbers still sank the stock: "revenue more than doubled, profit 19-fold. And what's so interesting about it is fundamentally... nothing in the memory markets changed from when you and I went to bed last night... there is still tight supply pricing of DRAM and NAND... and Samsung has no line of sight to when that improves. It's really a stock story where... Samsung was a stock that was up 150% going into it. And I think everyone's just taking a pause." His caution on margins: "Historically, memory is a very cyclical market, boom and bust... in those periods where the memory names have had very high margins, they've never been maintained." (Bloomberg Intelligence) Marketplace (Jul 7) added that Samsung beat "by only a small margin rather than significantly," and that investors had already bought ahead of the print, so it "fell 7% despite an 1800% year-over-year profit increase," per Wedbush's Matt Bryson and reporter Kelly Wells. (Marketplace All-in-One)

See the table below for the week's ticker-by-ticker reactions.

8. M&A & partnerships

The dominant deal thread was Apple extending and expanding its partnership with Broadcom to 2031. On Bloomberg Businessweek (Jul 6), Mark Gurman reported the new leg is about ASIC chips: "What is a big piece of machinery that you need that has a single purpose? An AI server. It's all about AI." The chip, codenamed Baltra, deploys in "2027, 2028" and will have "four times the power of the M5 Ultra." Gurman's broader point on Apple's shifting position: "I believe Apple will be spending considerably more on AI, but nowhere near what these hyperscalers... are spending." And on Apple's lost leverage during the memory shortage: "Apple just does not have that clout or credibility that they once had. They're just one of the guys now. They're not at the front of the line." (Bloomberg Businessweek) On Market Maker (Jul 9), Piers Clemmensen framed it as a "$30 billion custom chip deal" that "locks in Apple's largest customer for another decade" and highlighted Broadcom's lead in FBAR RF filters that improve iPhone battery life. (Market Maker) The Rundown (Jul 6) called Broadcom "the arms dealer of the AI era," supplying custom chips to "Google, Apple, Meta, and OpenAI." (The Rundown)

Meta-AMD was the other big commercial event. On Motley Fool Hidden Gems (Jul 9), a host said Meta "signed the biggest AI deal in history to buy AMD chips for the hardest AI jobs," while building its own chips for easier workloads, adding it "could actually increase the company's CapEx needs" and might be "a bigger needle mover for Broadcom than... for Meta." (Motley Fool Hidden Gems)

Custom silicon as a Nvidia-hedge kept coming up. On AI Update (Jul 2), the host detailed that "Anthropic is in talks with Samsung to create a custom AI chip," while "OpenAI... unveiled their own custom chip... called a Jalapeño... building that with Broadcom and... claiming that it's going to offer better performance per watt." The strategic logic: NVIDIA "allocate[s] to different companies" limited supply, so building your own "eases that constraint... [and gives] better negotiating power." Anthropic's project is "definitely in the early stages... a multi-year development timeline." Samsung is a natural partner since it "already manufactures NVIDIA's chips" and is "building a joint AI chip factory in South Korea with NVIDIA." (AI Update) On Bloomberg Intelligence (Jul 8), Mandeep Singh cautioned these are "2, 3 years out," "there is nothing... over the next 6 months" that competes with Nvidia, while noting Google's TPUs (now "8 versions") "inspired all these companies," since "the Anthropic models, which are the best-in-class models, were trained on Google TPUs," and Microsoft's own chip "doesn't have the same customer preference." (Bloomberg Intelligence)

9. Cyclicality, inventory & the peak/trough debate

This was the through-line under everything: is memory being re-rated permanently, or is this a classic boom that ends in a glut?

The most detailed discussion came on Closing Bell's "Chips Trade Back On" (Jul 9). Wedbush's Matt Bryson made the structural bull case: "memory is a commodity. Back when [SanDisk] came public, there was too much production... Now we're in a situation where there is too much demand, not enough supply, and there really aren't new fabs coming on until late 27." His key new-regime argument was about contract length: "typically contracts between the memory vendors and their customers, they last for three months at most... Now we're seeing these five-year customer agreements roll out where... the penalties for abandoning these agreements are relatively stark. And so for the first time, it looks like the memory vendors have visibility into longer-term cash flow." On why the stocks still wobble: "whenever it looks like there's weakness... memory acts worse because if demand falls off, then you've got this potential pressure on pricing." He added Micron said memory contracted at current pricing "is going to be 50% of their sales." (Closing Bell)

The clearest cautionary framing came from The Rundown (Jul 7): Samsung's "$70 billion in CapEx spending this year... could mean that the memory supply shortage today could turn into a supply glut in the near future, a cyclical pattern the memory industry has [repeated]." (The Rundown) Bloomberg's Ian King echoed it: "this is an industry that's never got it right over the long term... what everybody is focusing on is how much gets spent when." (Bloomberg Tech)

The valuation-history skeptics were vocal. On The Dividend Cafe (Jul 3), the host noted the semiconductor index rose "237% over 14 months with 90% of gains in the last six months, mirroring the parabolic rise into February 2000," with technology now "36% of the S&P 500 (approximately 50% when including communication services)," similar to 1999. (The Dividend Cafe) On Jim Cramer's Squawk on the Street (Jul 8), the concern was capital supply, not chip demand: "there's going to be a lot of offerings and not enough money, and that is going to be when I'll turn bearish. Right now there's a lot of offerings, not enough money, and I am turning bearish... The one that I really am scared about is the $29 billion SK Hynix." He also explained why memory pricing has held: Micron CEO "Sanjay Mehrotra just said... we have agreements... that are going to prevent them from having the typical down [cycle]," so the DRAM makers can no longer be "pit... against each other." He noted the KOSPI was "down 22% from the June 18 high," at "the lowest [forward PE] since October of '08." (Squawk on the Street)

Morgan Stanley's Stephen Byrd, on Thoughts on the Market (Jul 7), anchored the demand-side bull case: "I'm very confident that the demand for compute is going to exceed the supply," invoking Jevons' paradox (cheaper compute drives more usage, not less). His token-economics point: a typical enterprise AI use case yields "$55" of benefit for "a few dollars of token costs" (at roughly "$5 per million tokens"), so token-cost fears are overdone. (Thoughts on the Market)

And the "melt-up not meltdown" framing came from Erik ("YWR") on Monetary Matters (Jul 2), who argued this cycle may lack "the ferocity and the euphoria... of 1999" but "plays out longer," even suggesting his "S&P 10,000" call "could be quite conservative" as earnings beat expectations and he lowers his assumed multiple "to 25 [x]." (Monetary Matters with Jack Farley)

10. China indigenization (SMIC, CXMT, Huawei) and the read-through to US names

The week's most eye-catching indigenization story was on The Generative AI Meetup (Jul 7): Meituan, "the Chinese DoorDash," released "Longcat," a "1.6 trillion parameter" open-source model (48 billion active parameters, 1-million-token context) that the hosts say was "trained state-of-the-art... [with] no American-made GPUs," using an estimated "40,000 or 50,000 Huawei" Ascend chips. Their economic argument for why this scales despite weaker chips: even if a Huawei chip is "half as good... China has a lot of money... they could just... buy twice as many. And then now they're equal to NVIDIA capability," and if the Huawei chip is "30 percent of the cost," Huawei can still run "20 percent margins" versus Nvidia's "70 or 80 percent." Their bottom line: "this shows that they're able to train a frontier model with completely Chinese software, Chinese hardware," and "I don't think this is reflected in the market yet." (The Generative AI Meetup Podcast)

That connects directly to the chemicals dependency (Section 6) and to the leapfrogging thesis from VoxTalks, together they paint a picture of a Chinese chip ecosystem that is both increasingly self-sufficient upstream (materials, gases) and increasingly capable downstream (Huawei silicon, frontier models). For US names, the read-through cuts both ways: it is a long-term competitive threat to Nvidia's China revenue and a reminder (per Jensen Huang's own argument, referenced across several shows) that export controls may be strengthening the very domestic supply chain they were meant to contain. On the flip side, CXMT's planned capacity expansion (Section 2) is the specific memory-supply risk that could eventually break today's pricing.

Note on coverage: No dedicated SMIC or CXMT deep-dive surfaced this week; the indigenization discussion was carried by the Meituan/Huawei model story, the ChinaTalk chemicals episode, and the VoxTalks macro conversation.

What I'm Watching Next Week

1. TSMC Q2 earnings. Flagged on the VRA Investing Podcast (Jul 9) as coming "next week," with the analyst noting TSM "produces nearly all of NVIDIA's chips and Samsung makes Blackwell chips for NVIDIA." This is the first hard read on foundry demand and 2026 capex intentions. (VRA Investing Podcast)
2. ASML earnings and any AMAT/LRCX/KLAC commentary. The equipment names are the cleanest tell on whether the "multiple years" of visibility Applied Materials' CEO described translates into real orders, and whether memory-fab spend is finally accelerating.
3. SK Hynix's first days of US trading. Watch whether the ADR holds its ~3% premium to the Korean line or widens "closer to double digits," and whether the KOSPI stabilizes after its ~22% drawdown from the June 18 high.
4. The pace of Micron's $250B and Samsung's ~$70B capex deployment. As Bloomberg's Ian King put it, the whole cyclical debate hinges on "how much gets spent when," a fast flood of supply is the bear trigger for late 2027.
5. Q3 memory price hikes. Whether Samsung's reported plan for a further ~20% HBM increase actually lands will test the "pricing power is durable" thesis.
6. Hyperscaler Q2 earnings (late July). Amazon, Microsoft, Meta, and Alphabet capex guides are the single biggest swing factor. As Cramer put it, the question is whether they "slow" or merely "combine" spending, and whether the Nvidia-as-source-of-funds selling reverses.
7. Policy: the MATCH Act and Apple's CXMT/YMTC request. A House vote on tighter equipment-export rules (~30% odds) and any US decision on Apple buying blacklisted Chinese memory would both move the China read-through.
8. Marvell and the custom-silicon names into the next print, given how much growth the reverse-DCF skeptics say is already in the price.