Newsletter · 2026-08-01 · Ashutosh Agarwal

Record Chip Profits Meet a Selloff as Jensen Calls for a Ten Times Bigger Industry - Semiconductor Podcast Briefing - Week of August 1, 2026

A synthesis of what analysts, investors and executives said on podcasts about semiconductors for the week ending August 1, 2026, including SK Hynix's record quarter and 10 percent drop, Jensen Huang's call for a ten times larger industry, Intel's foundry lock-ins, TSMC's fifth straight record, and China's CXMT shock.

Semiconductor Podcast Briefing

Week ending August 1, 2026: Record Memory Profits Meet a Chip Crash

Episodes from July 22 to 31, 2026.

This was one of the strangest weeks the chip world has had in years. The companies at the center of the AI boom reported some of the best numbers in the history of the industry, and their stocks fell anyway. South Korea's stock market, which is essentially two memory-chip companies wearing a trench coat, dropped more than 30% from its peak. A Chinese memory start-up you have probably never heard of, CXMT, had the most explosive stock-market debut in the country's history. And Nvidia's Jensen Huang went on a tour of South Korea telling anyone who would listen that the entire semiconductor industry needs to get ten times bigger.

TL;DR

• SK Hynix printed the best quarter in its history and the stock still fell ~10%. Revenue roughly tripled (up ~250 to 354% year over year) to about $55 billion, operating profit jumped more than six-fold to roughly $42 billion, and margins topped 80%. It still missed Wall Street's sky-high estimates by about $1.7 billion, and the stock has now lost roughly half its value since its June peak. (The Rundown, Jul 29)
• Jensen Huang says the chip industry needs to be "10 times larger" within a decade and announced over $500 billion of business with Korea's SK Group. He also flagged that Nvidia is short of almost everything: high-bandwidth memory, LPDDR memory, land, power, even construction workers. (Bloomberg Talks, Jul 25)
• A "chip crash." The week's dominant narrative was a sharp sell-off in chip stocks. Lam Research went from +154% year-to-date at end-June to +53%; Intel from +278% to +128%. Two triggers: fear of Chinese competition (CXMT's memory IPO and a Chinese lithography machine) and renewed worry about Nvidia's "circular" financing of its own customers. (Facts vs Feelings, Jul 29; Morning Brew Daily, Jul 28)
• Apple's Tim Cook called the memory shortage a "100-year flood." Apple's stock fell after earnings purely because it cannot get enough memory chips; hyperscalers are outbidding everyone. (Squawk on the Street, Jul 31)
• TSMC had its fifth straight record quarter: revenue +33.7% year over year, earnings +77%, gross margin 67.7%, and it raised full-year guidance to ~40% revenue growth with $64 billion of capital spending. Nvidia is now its largest customer at 22% of revenue. (The Canadian Investor, Jul 23)
• Intel bucked the sell-off. Revenue beat, it raised capital spending from a planned ~$18 billion to $20 billion (Wall Street had modeled $15 billion), and, crucially, it locked in 10-plus foundry customers on 3-to-5-year price-and-volume contracts. (CNBC Fast Money, Jul 23)
• China arrived as a supply-side force. CXMT's IPO jumped ~466% in a single day, briefly making it the most valuable company ever listed in China, and a Chinese firm claimed a home-grown deep-ultraviolet lithography machine, knocking equipment makers ASML, KLA, Lam, and Applied Materials down 6 to 8%. (Limitless, Jul 30; Rob Black Show, Jul 27)

1. AI chip demand and hyperscaler spending

The single most quoted moment of the week was Nvidia CEO Jensen Huang's swing through South Korea, where he laid out both an enormous long-term vision and a blunt admission that Nvidia can't get enough parts. Speaking to Bloomberg's Ed Ludlow on Bloomberg Talks (Jul 25), Huang framed why he thinks demand is structural rather than a bubble, the world is shifting from computers built for people to computers built for AI itself:

So instead of just a billion people using computers, we're going to have 100 billion agents and billions of people using computers... My guess is that the semiconductor industry is probably going to have to be 10 times larger than it is today, over the next decade or so.

On the supply side, he was candid that this can only happen slowly:

Well, we don't have enough bits. We're constrained in HBM memories, LPDDR memories, we're constrained in just about every part of the supply chain. We're even constrained now with land and power and construction workers to set up the data centers... I think we have the ability as an industry to double each year, but we're going to have a hard time growing much faster than that.

He put hard numbers on the Korea partnerships: over $500 billion of business with SK Group, spanning Nvidia buying memory from SK Hynix and selling AI supercomputers to SK Telecom as it builds out 2 gigawatts of AI data-center capacity, plus a $1 billion investment in Korean cloud leader Naver.

The other big demand story was AMD's up-to-$5 billion investment in Anthropic, which will deploy up to 2 gigawatts of AMD's next-generation MI450 chips in "Helios" rack-scale systems, with the first gigawatt starting in 2027. On Bloomberg Tech (Jul 22), Franklin Equity portfolio manager Sarah Rahi (herself an Anthropic investor) framed it as the new normal, chipmakers now take equity stakes to lock in customers rather than simply selling chips:

We've heard, we've seen Anthropic do deals with Broadcom, Google, all the players. They need more compute... this ecosystem is thriving. We're seeing all the success here and we're still in the early innings.

She also made a useful distinction that recurs all season, the difference between the company spending the capital and the company receiving it:

That recipient is seeing it in their cloud revenues. We talk about the ROI of AI. The hyperscale, the cloud companies are already seeing that right away... over 60% cloud revenue growth last quarter for Alphabet.

The clearest illustration of that shift came on Facts vs Feelings (Jul 29), where Ryan Detrick and Sonu Varghese described what Varghese called the "chart of the year," using Bloomberg data, the free cash flow of the big five hyperscalers (Meta, Amazon, Google, Microsoft, Oracle) has been collapsing while the free cash flow of the semiconductor names (Nvidia, Micron, Broadcom, Applied Materials) has gone nearly vertical:

But Ryan, this quarter, the lines crossed. So free cash flow for the semis is now larger than free cash flows for the hyperscalers... for the first time.

In plain terms: one company's spending is another company's revenue, and right now the chipmakers are the ones cashing the checks. Varghese also flagged just how dependent the whole US economy has become on this spending: of the roughly 2% real GDP growth over the last five quarters, about 0.9 percentage points, roughly 45%, came from AI capital spending.

2. Memory pricing (HBM, DRAM, NAND), the story of the week

Memory was where the "record profits, crashing stocks" paradox was sharpest. A quick definition for newcomers: HBM (high-bandwidth memory) is the premium memory stacked next to AI chips; DRAM is the ordinary memory in your phone and PC; NAND is flash storage. SK Hynix is the leader in the premium HBM tier.

The numbers from SK Hynix's quarter were, in the words of The Rundown (Jul 29), "insane":

Last quarter, revenues more than tripled to roughly $55 billion. And operating profit jumped more than six-fold to roughly $42 billion... Margins topped 80% as the company continues to raise prices due to the shortage.

And yet the stock fell nearly 10%, because those results still missed estimates. On Limitless (Jul 30), hosts Ejaz and Josh explained the counterintuitive reason the best news was actually bad news, SK Hynix is so focused on HBM that it literally sold out its entire supply, which caps any future upside surprise:

They've fully sold out of their inventory for 2026... There is no more capability for them to sell more. And 27. There's no ability for them to sell more or sell it at a higher margin because it's already pre-sold. So therefore, you eliminate a lot of the upside surprises. And there really is only downside surprises possible.

They added a mechanical accelerant: a 2x leveraged single-stock ETF that began trading July 13 fed a cascade of forced selling, and JPMorgan reported that the leveraged-ETF liquidation was roughly 90% complete, a sign the drawdown may be near its bottom. They also noted the eye-watering figure that SK Hynix posted 118% net margins (more profit than revenue, thanks to investment gains).

For the bullish case on memory as a structurally changed industry, venture investor Steve Jang of Kindred Ventures made the argument on Squawk Pod (Jul 29). He argued HBM is a locked-in, high-barrier market that has broken out of its historic boom-bust cycle:

And SK Hynix is the 65% market share leader in HBMs... in physical AI, robotics, humanoids, LLMs and agents, you're going to see demand that is going to 10X from here just over the next three, four years.

On the threat from China's CXMT, Jang was dismissive on technical grounds, making HBM is far harder than making commodity DRAM:

They make commodity DRAM. They are unable to make the stack-to-die structure and architecture of an HBM... So it'll be many years before there's a fab for memory chips that will be able to compete effectively in the market. And we're not talking about one or two years. We're talking about three, four, five years to even be able to enter into the market.

Becky Quick pushed the classic bear question at him, that 80%-plus margins invite competition and that today's shortage may be met with tomorrow's glut. That tension is the whole memory debate in one exchange.

The demand side of the memory squeeze showed up vividly in Apple's results. On Squawk on the Street (Jul 31), Jim Cramer relayed Tim Cook's description of the situation, Apple's revenue was up 16% and demand was strong, but it simply cannot get chips:

The memory situation is difficult. We'd describe it as a 100-year flood. He's been in the business a long time. He's never seen anything like it.

Cramer's framing of why: "this is hyperscalers overpaying everybody." Apple, one of the largest buyers of components in the world, is getting outbid for memory by AI data centers, and no one is promising 2027 will be any easier.

3. Semiconductor equipment: ASML, Applied Materials, Lam, KLA

The equipment makers, the companies that sell the machines that make chips, had a rough week, and it was almost entirely about China. On the Rob Black Show (Jul 27), Black ran through the damage after news that a Chinese company (tied to the Xingjian/CXMT complex) claimed it could produce immersion deep-ultraviolet (DUV) lithography machines, the workhorse tools for making chips:

ASML... declined about 6.2%. KLA Corporation down 6%. Lam Research down 8.2%. And Applied Materials down 7%.

The bear read is obvious: if China can make its own chipmaking machines, the Western equipment oligopoly's pricing power erodes. But the bulls pushed back hard on scale. On Limitless (Jul 30), the hosts pointed out that the Chinese DUV machine is a prototype, and the production targets are tiny, five units this year, a target of ten next year, versus the market's fear of a flood:

But the market saw this news and were like, oh crap, China's about to flood the market with EUV machines... It makes no sense. It has a massive overreaction.

The one bright, company-specific equipment story came from Chip Stock Investor (Jul 24): Swedish fab-equipment maker Mycronic raised its full-year 2026 guidance by roughly 500 million Swedish krona (about $50 million) after its global technologies segment grew net sales 118% year over year. Host Nicholas Rossolillo praised the business but said it remains a "pass" for their portfolio on size grounds, a useful reminder that a great business and a great stock are not always the same thing.

4. Foundry and manufacturing: TSMC and Intel

TSMC delivered what may be its best-ever quarter, per The Canadian Investor (Jul 23): revenue up 33.7% year over year, earnings up 77%, gross margin 67.7%, its fifth straight record profit, and $9 billion of free cash flow, all while, unlike the debt-and-equity-hungry hyperscalers, staying entirely self-funded. It raised full-year guidance to ~40% revenue growth and lifted capital spending to $64 billion, saying capacity, not demand, is the bottleneck. The host highlighted the growing customer concentration:

NVIDIA eclipsed Apple as the company's largest client now making up 22% of revenue. 78% of TSMC's revenue comes from 10 companies.

He argued the concentration is less scary than it looks because those customers "really have nowhere else to go," and relayed TSMC's own line that it wins regardless of which AI company comes out on top: "it doesn't matter who wins this race, the AI race, they're all TSMC customers." Still, the stock dipped after the print, a theme all season, where even a perfect quarter isn't enough. Separately, Network Break (Jul 27) covered TSMC's plan to raise prices ~10% across the board starting in 2027 and pledge another $100 billion for US fabs (bringing its US commitment to $265 billion); the hosts judged the 10% hike "restrained" and "not predatory" given how dominant TSMC's position is.

Intel was the exception that stood out from the whole sell-off. On Bloomberg Surveillance (Jul 23) and CNBC Fast Money (Jul 23), CNBC's Christina Partsinevelos relayed detail from CFO Dave Zisner: Intel locked in more than 10 customers on 3-to-5-year long-term agreements covering both price and volume, committed to ramping its most advanced 14A process to high-volume production by 2028, and raised capital spending to $20 billion (Wall Street had modeled $15 billion), with next year "significantly higher." Her key insight on why Intel rose while everyone else fell:

The stock is trading higher after hours, and that matters because TSMC posted its best quarter in history and fell. ASML, TXN... raised outlooks and yet their stocks fell. Samsung print printed record profit and yet fell. But Intel locked in these long-term agreements... Intel now has visibility into demand.

Not everyone was impressed. On the same Fast Money, Gene Munster of Deepwater put Intel's AI exposure in perspective and questioned the valuation:

Their best performing segment, that data center and AI segment, is 1/13 the size of Nvidia's GPU business. It grew at 56%. Nvidia is going to grow their GPU business at 96% in the July quarter... I'm surprised that the stock's up.

He noted Intel trades near 94 to 100x forward earnings versus a 10-year average of about 25x, a rich price that only works "if you believe the growth." Intel also scored a symbolic foundry win: security-hardware maker Fortinet chose Intel's foundry (Intel 4 process) for a custom security chip, its first-ever chip partnership with Intel and a rare named, disclosable customer for a foundry business Intel is desperately trying to get off the ground (Network Break, Jul 27).

5. Analog, auto, and industrial chips

This corner of the market, the everyday chips from Texas Instruments, Analog Devices, Microchip, ON Semi, NXP, and STMicro, sat in the shadow of the AI-memory drama. The main data point: Texas Instruments raised its outlook and still saw its stock fall, which fits the season's pattern. As Christina Partsinevelos noted on Fast Money (Jul 23), when a company like TI says the cycle is improving on the back of autos, it does not carry the same multi-quarter demand visibility that memory makers and Intel are now offering, so investors were unimpressed. If you follow the analog names, the read-through is that AI is not yet the tide lifting these boats, and they remain tied to the slower auto and industrial cycle.

6. China, export controls, and tariffs

The clearest explanation of the week's China anxiety came from Morning Brew Daily (Jul 28). The hosts laid out the logic: US export controls stopped China from getting the best chips, so China has raced to build its own supply chain, first AI models, now the memory and the equipment underneath them:

But all of a sudden, like China has done in so many other industries, they have caught up and they're starting to make their own domestic supply chain. So CXMT is their global champion for memory... they're starting to do it cheaper because they're not getting the most high-end models from the United States.

They tied it to the US side of the fear, Nvidia's reported $250 billion financing guarantee for an OpenAI data center in Ohio, a deal that could swell past $500 billion including chips, which "reignited fears of the circular financing going on in the U.S. industry." Two directions, one bad day for chip stocks.

7. China indigenization: CXMT, SMIC, Huawei

The breakout name of the week was CXMT (ChangXin Memory Technologies). On Limitless (Jul 30), the hosts detailed a debut that was hard to believe:

They had themselves a public IPO... in which they traded up 466% in one day, which is instantly the most valuable China-listed company ever, which is more than Alibaba or Tencent. And they raised about $9 billion.

CXMT is the world's number-four DRAM maker, having gone from roughly 1% market share to nearly 10% in about four years, with the Chinese state firmly behind it. But the Limitless hosts made the crucial nuance for US investors: CXMT mostly matters for China's own memory-starved AI labs (Moonshot's Kimi K3, Zhipu's GLM), because China is largely cut off from both Nvidia's frontier chips and the Western memory that is being funneled to US buyers. Their bottom line was that the market is wrongly treating dynamic, multi-architecture memory like it treated commodity solar panels:

You can't build a memory company to subsidize the prices... one, the demand is so high. And two, there's so many different types... DRAM is kind of like tap water... HBM is that premium bottled stuff.

One more indigenization data point worth watching: on the Sinica Podcast (Jul 23), China analysts Samm Sacks and Paul Triolo discussed WAIC 2026, Xi Jinping's AI speech, and Huawei's "SuperPod" hardware, and noted that SMIC and CXMT are planning IPOs that will plow capital back into building domestic capacity. The financial specifics were thin, but the strategic direction, state-backed and capacity-first, is unmistakable.

8. Merger and acquisition chatter

There was little pure-semiconductor M&A this week. The deal energy in tech went instead into equity investments and financing structures: AMD into Anthropic, Nvidia's multi-hundred-billion-dollar backstops for OpenAI's data centers, and Samsung reportedly in talks to invest in French AI start-up Mistral at a ~$22.8 billion valuation (Bloomberg Tech, Jul 22). The largest named takeover chatter was in payments, not chips, private-equity firm Advent's ~$60/share bid for PayPal, which is outside our scope. Net: no major chip-company merger surfaced, but the "invest to lock in supply and customers" model is quietly reshaping the industry's cap tables.

9. Cyclicality: is this a peak?

This is the debate under everything. The bear case got its most articulate hearing from venture investor Rick Heitzmann on RiskReversal (Jul 23). His core worry is that the "compute is scarce" story is cracking, because hyperscalers like Meta and xAI now have so much capacity they are renting it out:

If I have Microsoft and I have XAI and I have Meta all telling me that they have excess compute and they want to rent it out there, I have an issue with the whole notion that compute is so constrained.

He layered on two more concerns: that AI chips may last longer than assumed (people "skip cycles" the way they now keep an iPhone for years, reducing replacement demand), and that Nvidia's customers are all racing to diversify away from it toward Google's TPUs, AMD, and Intel CPU clusters. On memory specifically, he called the twin US listings by SK Hynix ($26.5 billion) and Samsung a classic top signal, and reminded listeners how brutal this industry's history is:

These capital-intensive equipment chip memory infrastructure companies are incredibly cyclical... Cisco took 26 years to get back to their February 2000 pricing.

The bull rebuttal came from Dan Ives of Wedbush on The Pomp Podcast (Jul 27). He argues the ROI is showing up, that demand still outstrips supply "12 to 1" on his latest Asia trip, and that memory is the tightest link in the chain:

Memory is the biggest bottleneck because there's no obvious... memory data centers that we're going to build. You probably don't have equilibrium until 2028, 2029.

Interestingly, even Ives thinks the memory stocks (up ~1,200%) may have gotten "overdone on one side" and that the hyperscalers, not the memory makers, will be the biggest ultimate winners. And he insists the real bottleneck is not chips at all, but power: the US needs "3 to 4x more energy" within a few years, and data-center permitting fights (he singled out New York's moratorium as "so dangerous") are the true risk.

The clean summary from Facts vs Feelings: most of the big chip leaders were up 200 to 300% in the first half, so a pullback and rotation into other sectors may be healthy rather than the end of the boom, as long as the hyperscaler CEOs keep spending like they say they will.

10. Earnings reactions at a glance

What to watch next week

• Whether the memory sell-off has bottomed. JPMorgan's estimate that the leveraged-ETF unwind is ~90% complete suggests the forced selling in Korea may be nearly over. Watch the KOSPI and SK Hynix/Samsung for a base, and watch whether the "record profits, falling stocks" pattern reverses.
• Micron's read-through. With SK Hynix's HBM sold out through 2027 and DRAM pricing spiking, Micron is the most direct US beneficiary. Any Micron commentary on pricing, HBM allocation, or 2027 capacity will move the whole memory complex.
• The circular-financing debate. Nvidia's ~$250B OpenAI backstop and the web of chip-for-equity deals (AMD and Anthropic, Nvidia and SSI) are increasingly the bear case. Watch for more scrutiny of whether these arrangements are demand or financial engineering.
• China's supply-chain progress. CXMT's ramp, the reality behind the Chinese DUV lithography claim (is it really only ~5 machines this year?), and any SMIC or CXMT IPO news are the key indigenization signposts.
• AMD's Advancing AI event and MI450/Helios detail following the Anthropic deal, the clearest test of whether a credible Nvidia alternative is emerging at scale in 2027.
• Power and data-center permitting as the next bottleneck Dan Ives keeps pointing to, including political fights like New York's data-center moratorium.