Newsletter · · Ashutosh Agarwal

Chip Stocks Snap Back as the Buildout Numbers Go Up Not Down - Foundry & Chip Equipment Weekly - Week of August 9, 2026

For the week of August 3 to 9, 2026, the July chip meltdown was re-read as a leverage event rather than a demand event: the index rebounded more than 8% in a day, capex guides went up across all four big cloud buyers, and Intel landed a rare packaging win over TSMC.

Foundry & Chip Equipment Weekly

Week of August 9, 2026: Chip Stocks Snap Back as the Buildout Numbers Go Up Not Down

Last week ended with a trillion dollars of chip market value gone, a South Korean market in free-fall, and a marquee AI fund blown up and sold for parts. This week the same stocks came roaring back. On the recovery day, the main US chip index jumped more than 8% in a single session; SanDisk was up 26%, Micron 18%, AMD 13%. The bodies were still warm and the group had already reversed.

That whiplash is the whole story. And once you understand why it happened, the July panic looks less like a warning about the businesses and more like a warning about leverage.

Here's the tell. In a normal sell-off this violent, money runs out of the whole sector, investors dump the exchange-traded funds (the baskets that hold all the chip names at once) and head for the exits. This time the opposite happened. Even as the chip index had one of its worst months in two decades, chip ETFs were on track for their best month of inflows ever. The selling was concentrated in individual stocks, the specific names one forced seller had to liquidate, while everyone else was buying the dip with both hands. When the forced seller was done, the floor was in.

Underneath all of it, the demand signal that actually drives this newsletter, how much gear the chipmakers need to buy, didn't just hold. It got louder. And in a genuinely new twist, the most interesting foundry headline of the week wasn't about Taiwan. It was about Intel finally being ahead of TSMC on something.

TL;DR

  • The crash reversed almost as fast as it came. The chip index rebounded more than 8% in a day; the July meltdown is now widely read as a leverage-and-margin-call event, not a demand event. A prominent strategist notes chip multiples have already de-rated about 50% this year, "the market is well onto this."
  • The buildout numbers went up, not down. Roughly $580 billion of capital spending across the big cloud players over the past year, an estimated ~$900 billion in 2026, and a path some see toward $1.5 trillion in 2027. All four big spenders guided their spending higher and said they're still short of compute. One veteran investor says compute rental prices are actually rising 50-100%, and that Nvidia is trading at its lowest forward earnings multiple in ten years.
  • Intel got a rare win over TSMC, in packaging. Semiconductor analysts flagged that TSMC is now scrambling to build a packaging technology to match Intel's, and called an Apple-Intel foundry deal "inevitable." First time in a long time the headline ran that direction.
  • ASML's order book looks fine; the moat is the question. The Dutch monopoly booked €5.6 billion of new orders in mid-2026 on strong demand for its most advanced machines, even as a state-backed Chinese lithography effort spooked the stock. One big bank says investors are overreacting.
  • The Apple-Micron memory fight escalated in Washington. Apple is lobbying to be allowed to buy cheaper Chinese memory; this week Micron started lobbying to stop it. Analysts think Washington says no.
  • Quantum quietly became a supply-chain story again, a rare operator laid out the "picks-and-shovels" control-and-cryogenics layer, the math on how many qubits it takes to crack encryption collapsed further, and the White House put quantum on a presidential deadline.
  • The real bear case got sharper, and it's about demand, not China. New analyst work suggests roughly 70%+ of the big cloud players' AI revenue traces back to just two customers, OpenAI and Anthropic, a concentration-and-circular-financing worry that would, if it broke, hit everyone downstream.

What's new

The rebound, and what it says about the crash

Start with the mechanics, because they explain everything else. On The Compound and Friends (August 4), hosts Josh Brown and Michael Batnick, market commentators rather than operators, walked through a set of charts that made the point cleanly. When 88% of the roughly 3,000 largest US stocks' semiconductor names were in a bear market, the overall market was only about 3.5% off its high. As Brown put it, "that dynamic had never happened before." Normally a semiconductor washout means the whole market is falling apart. This time it didn't, "so investors... correctly looking past the blowup and they bought the ever-living [dip]."

The proof was in the flows. Citing Goldman Sachs data, they noted that while the chip index "has faced one of its worst months of performance in nearly two decades, semiconductor ETFs are on track to have their best month of fund flow since the inception of SMH," the largest chip ETF. Retail investors sold about $285 million of individual chip stocks (per Vanda Research, the biggest outflow since the April "Liberation Day" tariff scare) but poured money into the baskets. The selling of single names was concentrated; the buying of the sector was broad. Their verdict: "The momentum crash in July reset the bull market. It gave us a new foundation." Or, more memorably, "a healthy bull market takes out its own trash."

The trash, in this case, had a name: Leopold Aschenbrenner's fund, Situational Awareness. The clearest reconstruction came from two industry analysts, Patrick Moorhead (Moor Insights & Strategy) and Daniel Newman (Futurum), on The Six Five (August 3). "Ground zero was SK Hynix," Newman said, the Korean memory giant posted a roughly 500% jump in profit, but it came in "slightly off of expectation," and in a market this jumpy "good numbers weren't good enough." Then the leverage did the rest: Aschenbrenner was running about four times leveraged, a roughly $45 billion book sitting on around $175 billion of positions. Forced to sell into a falling market, he drove prices down; short-sellers piled on; and when Citadel bought the book, the same names "ran 30% a day" back up, "capitulation followed by short covering." Their one-line lesson: "watch the leverage."

Which brings us to the sober counterweight. On RiskReversal (August 7), Morgan Stanley's chief US equity strategist Mike Wilson made the case that the market has actually done its job. "Multiples have de-rated 30% for the hyperscalers... 15% for the S&P 500. Semiconductor stocks have de-rated by 50%. So that gives me comfort that the market is well onto this." He's not all-clear, he thinks interest rates are "headed to 5%," which could cost the broad market another 5-8%, and he won't rule out one more retest of the lows. But on the chip complex specifically, his read is that a 50% multiple reset is "a very healthy correction" that is "probably going to hold."

The demand didn't crack, it got tighter

If the crash was about money and leverage, the businesses told a very different story. Every big chip company that reported "blew past earnings," as Wallstreet Trapper put it bluntly on Trappin Tuesday's (August 2), Micron up 70-81% year over year, TSMC and ASML with "massive earnings," and got sold off anyway. His conclusion was about mood, not fundamentals: "the appetite for tech has changed, momentarily."

The spending that funds all of it went the right way. On the Chip Stock Investor Podcast (August 3), Nicholas Rossolillo tallied the big cloud players' capital spending, the money they lay out for data centers and the chips inside them, which is where the whole equipment industry gets paid. It runs to about $580 billion over the past twelve months across Amazon, Microsoft, Google, Meta, Oracle and Tesla, an estimated ~$900 billion in 2026 once SpaceX is included, and "possibly headed towards $1.5 trillion in 2027." His answer to his own headline, has anything fundamentally changed, was a flat "no."

The people closest to the plumbing said the same, more colorfully. On The Circuit (August 3), two long-time semiconductor analysts described a supply chain that simply cannot keep up: "There is no capacity. There is no room at the inn... I don't think the masses truly understand how big of a problem this is." All four big cloud buyers, they noted, said their spending was going to the high end or higher and that they remain "very constrained on compute," "which is what the semi complex wanted to hear."

"There is no capacity. There is no room at the inn." (The Circuit)

And from the investor who spent the week actively hunting for a reason to be bearish: Gavin Baker of Atreides Management, on Invest Like the Best (August 4). His mission in Silicon Valley, he said, was to "look for negative data points as hard as I could," and he mostly failed. The striking evidence: rental prices for Nvidia's chips are rising, not falling. A startup that rented a large cluster in the mid-$2-per-chip-per-hour range is bracing to pay "just under $4" seven months later, a 50-60% increase, and one cloud told him it expects "to pay 100% more" when its contract renews. "All the hyperscalers are under-earning," he argued. His punchline for anyone who thinks the group is expensive: "Nvidia is... at its lowest forward P/E of the last 10 years." It has only ever been cheaper at two prior panic bottoms.

Even the demand's rawest cheerleader made the equipment point for us. On All-In (August 8), the hosts relayed Elon Musk's first earnings call as a public SpaceX, $7.8 billion in quarterly revenue, up 92%, with a compute-rental business that more than tripled to $2.6 billion, and his plan to grow data-center power from 2 gigawatts toward "closer to 10" next year. The line that matters for this newsletter: Musk said the market is "constrained by whatever the bottleneck is... right now the bottleneck is memory," with memory supply maybe up 20% next year against demand up more than 200%. When the loudest builder in the world says memory is the thing he can't get enough of, that is a demand statement for the entire memory-and-packaging tool-chain.

Intel's quiet win: for once, ahead of TSMC

Here is the genuinely new foundry development, and it came from the analysts on The Circuit rather than from any company microphone. Two things surfaced at once. First, "it's inevitable that Apple uses Intel" as a manufacturing partner, "a matter of when and for which chips," with Intel described as "a huge beneficiary of EMIB," its advanced-packaging technology. (Packaging is the increasingly high-value step of stitching multiple chips together into one module; it's the current bottleneck of the AI supply chain.) Second, and more striking: "TSMC is exploring a new packaging technology [called CoPoS] to compete with Intel's EMIB." As the host put it, "for the first time in a long time, we have a headline saying Intel is ahead of TSMC on some technical aspect of manufacturing."

Keep the caveat the analysts themselves kept: the differences between these packaging methods are "pretty arcane," and TSMC is exploring an alternative mostly because it "can't make enough for their customers," not because it's technically behind across the board. But directionally, this is the first concrete evidence in weeks that Intel's foundry pitch, "you need a credible second source, and our technology is good enough," is landing. As they put it, "the quality of their technology is what makes them a viable second source."

The corporate scoreboard filled in the rest, via the Telltales weekend cashflow memo (August 2), worth flagging that this show is produced entirely with AI voices, so treat it as an aggregator, not a source with a reporter behind it. By its accounting, Intel posted its fastest revenue growth in 15 years, Q2 revenue up 25% to $16.1 billion (its best since 2011), a forward guide topping estimates on data-center strength, and a formal commitment to high-volume production of its next-generation 14A node in 2028, even as it announced 24,000 job cuts in the same stretch. Growth and a restructuring in one breath usually means the growth isn't coming from where the headcount is.

ASML: order book fine, moat still the debate

The same Telltales memo carried the cleanest ASML data point of the week, and it cuts against last week's panic. Yes, a Shanghai-based, state-backed Chinese company has begun mass-producing the kind of deep-ultraviolet immersion lithography machines ASML has largely had to itself (lithography being the process of printing circuit patterns onto silicon), reportedly after absorbing staff from a Huawei-backed startup, and ASML shares slid to their lowest level since early June. But: ASML "also posted €5.6 billion of net bookings in mid-2026 on strong demand for its high-NA EUV systems," its most advanced tools, and Bank of America "says investors are overreacting to the China threat." The memo's own framing is the right one: "the order book is fine, and the moat is the open question." Those are different problems on very different clocks, near-term orders are strong; the long-term monopoly is the thing now up for debate.

The Apple-Micron fight goes to Washington

Last week we flagged that Apple had gone to the White House to ask permission to buy cheaper Chinese memory. This week the counterpunch landed: on The Circuit, the analysts noted a fresh report that "Micron is [now] lobbying the US government to not let Apple do that." Their read is that this is "a political question... not going to be decided on a rational commercial basis," and that a waiver is unlikely, if anything, given the week's China-model tensions, "they're going to crack down even harder." The broader point stands from prior weeks: the memory shortage is flowing downstream into device makers' costs, and now into lobbying budgets. (Telltales separately noted Apple guided its September quarter below the Street and blamed supply constraints, the parts it can't get, not the demand it can't find.)

Quantum: the picks-and-shovels layer gets a face

After a genuinely strong quantum week last time (the D-Wave/AT&T deal, IBM buying a quantum lab), this week was quieter on deals but richer on the part that actually matters for a supply-chain investor: who makes the gear.

The operator voice was Yonatan Cohen, co-founder and CTO of Quantum Machines, on the Physics World Weekly Podcast (August 6). His company doesn't build qubits; it builds the control systems, the classical electronics that operate the fragile quantum hardware, sending the "very clean signals" that determine whether the machine works at all. It's the quantum equivalent of the tool-and-instrument layer, and it's already a real business: founded 2018, 300-plus people across 21 countries, 500-plus customers. His most telling point is that the industry is maturing from every company doing everything toward specialists at each layer, Quantum Machines demonstrated top-tier results running its own control and calibration on a Rigetti quantum chip, mixing vendors the way a PC uses parts from many suppliers. It has low-latency links into Nvidia GPUs, is now integrating with AMD chips, and just bought a Dutch software firm. Whoever wins the qubit race, this layer, control electronics, calibration, cabling, cryogenics, gets paid.

The investor's frame came from Barak Bussel of the deep-tech fund 7i, on The New Quantum Era (August 3). His headline number: the estimated hardware needed to run the algorithm that would break today's standard encryption has collapsed, from an original 20-30 million physical qubits, to about 500,000 in a recent Google/Caltech analysis, down to roughly 10,000 using a neutral-atom approach plus newer error-correction codes. He compared the moment to Intel's first microprocessor in 1971, a technology whose eventual scale nobody could see. And he was refreshingly honest about how hard the hardware is: leaning on a foundry like TSMC, taping out a chip and testing it eats 18 months, and one of his portfolio companies spent two years just solving how to physically plug a photonic chip into a system. Cabling and cryogenics, again, are the quiet gating items.

And the policy backdrop hardened. On Moonshots with Peter Diamandis (August 4), White House science-and-technology director Michael Kratsios said a presidential executive order now targets "a scientifically relevant quantum computer by the end of [the] term," with pharmaceuticals the use-case he's most excited about. In the same conversation he made the export-control read explicit: the 2019 controls on EUV lithography were "probably one of the most impactful export controls in the history of the United States," and America's chip lead over China "continues to increase year over year." He also let slip a nice supply-chain tell: Nvidia's Jensen Huang told him the order Nvidia placed with Corning for future glass "is the largest order in the history of Corning."

The debate

The bull case this week was overwhelming, and for once it came with receipts. The businesses beat. The spending that feeds the equipment makers rose across all four big buyers. The people closest to the supply chain say there is physically no spare capacity. A serious investor who went looking for cracks found rising, not falling, prices for compute, and pointed out that Nvidia has rarely been cheaper on forward earnings. A top strategist argued the 50% multiple reset in chips has already done the market's work. And the reversal itself, an 8%-in-a-day snapback the moment a forced seller was done, is textbook "it was the money, not the businesses."

The bear case did get voiced, and this is important: the strongest version this week was about demand, not China. On Better Offline (August 5, with a follow-up monologue August 7), the writer Ed Zitron, a committed AI skeptic and a journalist rather than an operator, pulled together analyst notes from Barclays, UBS and Wells Fargo estimating that roughly 70% or more of the AI revenue at Amazon, Microsoft and Google traces back to just two customers, OpenAI and Anthropic. UBS's figure: those two could be 44% of all Google Cloud revenue by 2027. His charge is circular financing, the cloud giants invest billions into the same two labs that then spend it renting the giants' compute, against a backdrop of ">$990 billion" of capital spending. Strip out this circle, he argues, and the underlying demand is far thinner than the capex implies. You don't have to buy the conclusion to see why it's the real bear case: it doesn't need China to be right; it just needs those two labs to stumble, at which point everything downstream, chips, tools, memory, packaging, feels it.

The China threat that detonated last week, notably, cooled from a fundamental scare toward a longer-dated worry. On Around The World With Yusko (August 5), investor Mark Yusko still called China's chip advances "a really, really big deal" and recapped CXMT's blockbuster IPO, up 5x on day one to a ~$500 billion market value, briefly China's most valuable listed company. But even he framed the read-through mostly as a memory-supply story ("this could be really bad for the memory companies"), and Bank of America, per the Telltales memo, is openly calling the equipment-stock reaction an overreaction. A useful reality check from the same Yusko episode, worth filing: only about 10% of announced data centers are actually under construction, because permits are hard and construction labor is tight, a reminder that the demand is real but the buildout is physically gated.

Read-throughs

Wafer-fab equipment sits on the durable side of this trade. The names that fell hardest on the China lithography headline, ASML, Lam Research, KLA, are the same ones whose customers just told the market they're short of capacity and spending more. ASML's €5.6 billion of fresh orders and a full industry chorus of "there is no room at the inn" are hard to square with a demand cliff. Applied Materials reports around August 13; it will be the first pure-play equipment operator to actually speak into this recovery, and worth watching closely.

Advanced packaging is where the action rotated this week. The Intel-versus-TSMC packaging skirmish (EMIB versus CoPoS) matters because packaging is the current bottleneck of the AI chip, and because it hands Intel its first "ahead of TSMC" headline in years. If the "Apple uses Intel is inevitable" thesis is right, it widens the customer base for Intel's foundry and adds a second credible home for advanced-packaging volume, a structural positive for the packaging and metrology tool-chain regardless of which foundry wins the socket.

Fabless customers remain the pressure valve, and the fight is now political. The Apple-versus-Micron lobbying war in Washington is the clearest sign the memory shortage has become a policy problem, not just a pricing one. Watch for any US decision on Chinese-memory waivers; a "no" (the analysts' base case) keeps Micron's pricing power intact, while a surprise "yes" would be a genuine negative for Micron and a signal that the shortage has cracked the political consensus.

Memory's durability is still the swing factor for the whole complex. Elon Musk calling memory the binding constraint, and SK Hynix's 500%-ish profit jump, argue the shortage is real and lasting. The bear counter, voiced on Smart Investing (August 1) by advisors Brent and Chase Wilsey, is worth keeping: those long-term memory contracts "may not be secure" if spot prices fall, because customers renegotiate when conditions change, and the stocks "falter when people start to feel the crisis is ending," not when it actually does. New memory plants plus the equipment to fill them could arrive in one to three years. That's the clock to watch on the memory-tool cycle.

Quantum is a picks-and-shovels story now, and that's where the investable read lives. The qubit companies grab headlines; the value is accreting in the layers Quantum Machines, cryogenics and precision-component suppliers occupy, the same "sell the gear, not the dream" logic that makes the chip-equipment names the quiet winners of the AI buildout. With a US presidential deadline now attached to quantum, the component-and-control supply chain gets a policy tailwind on top of a technical one.

What changed from prior weeks

The mood flipped completely. Two weeks ago the story was blowout earnings met with falling stocks; last week it was a trillion-dollar crash triggered by China and amplified by leverage; this week the same names snapped back 8-18% in a day and the demand data came in stronger, not weaker. The through-line across all three weeks is now clear: this has been a leverage-and-momentum unwind riding on top of a demand story that, by the companies' own numbers, keeps accelerating.

China faded from a fundamental scare to a longer-dated worry, still real, but reframed as a memory-supply and 2028-and-beyond monopoly question rather than a next-quarter demise, with at least one big bank calling the equipment sell-off an overreaction. And the foundry narrative finally moved on something other than Taiwan: Intel's packaging edge and the "Apple uses Intel" thesis are the first constructive, name-specific foundry developments in weeks.