Newsletter · 2026-07-20 · Ashutosh Agarwal

Chip Stocks Break Into Bear Market Territory as TSMC Roars - AI Accelerators - Week of July 20, 2026

AI accelerator podcast intelligence for the week of July 13 to 20, 2026. The semiconductor index fell into a bear market on a cheap Chinese model as bears led by Jim Chanos pressed the AI capex math, while TSMC's blowout print and operators building flat out made the bull case.

AI Accelerators

Week of July 20, 2026: Chip Stocks Break Into Bear Market Territory as TSMC Roars

GPUs, Custom Silicon & Optics, a twice-weekly read on the hardware powering AI. Issue 010, Monday, July 20, 2026, podcasts from July 13 to 20.

Two weeks ago the story was one stock, SK Hynix, swinging wildly on its Wall Street debut. This week the swing spread to the whole neighborhood. The main index of chip stocks fell into what traders call a "bear market" (down more than 20% from its recent high), and the podcasts filled up with people arguing about whether the great AI hardware build-out is a once-in-a-generation boom or a slow-motion accident.

Here's the twist that makes this issue worth your time: while the stock market was busy panicking, the companies at the very bottom of the AI supply chain, the factory that makes almost every AI chip, and the people actually running the data centers, were saying the exact opposite of the panic. Demand, they insist, still can't be met. So this issue is really about a single question: when the market and the operators disagree this loudly, who do you believe?

A note on how we use the word "operator." When we say operator, we mean someone who actually runs the business in question, a CEO, an engineer, a factory boss, not a fund manager or a TV pundit talking about it from the outside. That distinction matters more this week than usual, because the loudest voices (the bubble-callers) are almost all outsiders, while the calmest voices (still building as fast as they can) are the insiders.

TL;DR

The chip trade broke. The main semiconductor index slid into bear-market territory on July 17, triggered partly by a cheap new Chinese AI model. Even Bloomberg's own analyst called it a pricing scare, not a fundamentals scare, "the fundamentals haven't really changed."
The bubble debate got heavyweight. Famed short-seller Jim Chanos laid out the single scariest number of the week: the return hyperscalers earn on each new dollar of spending has fallen from ~40% eighteen months ago to ~20% now, heading toward 10%. Others (Ed Zitron, Peter Boockvar, Marco Papic, Fred Hickey) piled on.
But the bulls have a fresh, hard fact: TSMC. The world's most important chip factory reported profit up 77%, guided higher, and added another $100 billion to its US build-out (total now $265 billion) because "demand is still outstripping supply." The base of the whole pyramid is expanding, not shrinking.
The operators are unbothered. OpenAI's compute chief: "We can't build fast enough." Dell's AI/energy lead is racing to cool 250-kilowatt racks. A data-center operator just signed a $1.1 billion (up to $2.5 billion) deal with chip-maker Cerebras. None of them sound like people who think the party's over.
AMD finally got its close-up, a full strategy breakdown for the first time in weeks (still from a pundit, not the company): don't try to beat Nvidia, become the "indispensable second platform."
Optics broke a long silence. After many issues of near-total absence, the networking-chip names Astera Labs and Credo were finally named with real design wins, inside Amazon's and AMD's newest chips.
China is the pressure everywhere, a cheaper model that spooked the market, a cheaper memory-chip rival (CXMT) about to go public, and Google using its own money to push its TPU chips onto Nvidia's customers.

The lead: the AI hardware trade cracked, and the trigger came from China

The single cleanest description of the week came from Bloomberg's markets podcast, bluntly titled "Chip Stocks Sink Into Bear Market Territory as AI Rally Fizzles" (July 17). The spark: a Chinese startup called Moonshot released a new model (Kimi) that appeared to match the best American systems at a fraction of the price. That was enough to knock the whole chip index into a bear market.

But here's the important nuance, and it came from Bloomberg Intelligence's own analyst Mandeep Singh, who refused to call it a crisis. Asked if this was "another DeepSeek moment" (a reference to the January 2025 Chinese-model scare) or something bigger, he answered "neither." His read: "It's really a function of pricing", the new model is cheap and "good enough," which is what rattled people, not any breakthrough that changes who actually leads. And crucially: "I think the fundamentals haven't really changed in a big way, which I think will be validated when these companies report their earnings and guide for CapEx, which I expect numbers to go up." (CapEx just means capital spending, the money the big tech firms pour into chips, buildings and power.)

In other words: the machinery is fine; the mood broke. That framing, a positioning and confidence wobble rather than a demand collapse, is the fault line running through this entire issue.

One vivid sign of the mood shift: Apple, which has spent the AI boom sitting on its hands, briefly overtook Nvidia as the world's most valuable company. As Bloomberg's Charlie Pellett noted, Apple is up about 22.5% this year while Nvidia is up just 9.9%. Apple's own tech reporter Mark Gurman called it "a little bit of a wake-up call" for Nvidia, and pointed out why investors suddenly like Apple's restraint: because Apple runs more AI directly on your phone, "their capex is slim compared to the competition." When not spending becomes the applauded strategy, you know sentiment has turned.

The bubble referendum: the bears finally landed real punches

For months the "AI is a bubble" case has been mostly noise from perma-bears. This week it got serious, specific, and numbers-driven.

The scariest chart of the week, Jim Chanos. The legendary short-seller (famous for calling the Enron fraud) laid out his case on the RiskReversal Pod (July 17). His killer metric is "return on incremental invested capital", plain English: for every new dollar the big cloud companies spend, how much profit does it throw off? His answer, and these are his figures: "It's gone from, as a group, 40% a year and a half ago to about 20% today. And if the spend keeps up at this kind of rate, it's going to be moving toward 10%." At that point, he argues, the bosses of Google, Amazon, Microsoft and Meta have to ask a brutal question: "Does this next trillion dollars make sense if you're only going to earn 50 billion on it? We can earn that on treasuries." His timing call: "We're going to get there some point in the next 12 months."

Chanos's deeper worry is a mismatch of time. Companies are signing 20-year data-center commitments based on today's sky-high rental prices, which he says only last "a year or two." "People are committing long-term capital projects based on near-term spot pricing," he said, the same mistake that bankrupted 19th-century railroads and shale drillers. And his valuation rule of thumb is worth taping to your monitor: "No one that is dependent upon NVIDIA to exist should trade at a higher valuation than NVIDIA itself. And yet many, many companies do." He also punctured the "this is just like the dot-com fiber build-out" comparison, but in the bearish direction: the two big failed dot-com industries (local phone networks and fiber) together spent about $100 billion over five years; today, he notes, individual companies are raising more than that alone.

The OpenAI-can't-pay-its-bills case, Ed Zitron. On Tom Bilyeu's Impact Theory (July 14), tech critic Ed Zitron made the most alarming operator-level claim: by his reading of audited financials (reported by the Financial Times), "OpenAI burned $20.9 billion in 2025," with costs rising in lockstep with revenue and "no proof that they can improve their margins." His hardware-specific jab: "No amount of specialist silicon or supposed Vera Rubens", a swipe at Nvidia's next-generation chip, "will bring these costs down." His nightmare scenario is Oracle, which he says is building 7.1 gigawatts of capacity for a single customer that would need roughly $75 billion a year in compute revenue to pay for it, "OpenAI can't afford that." (These are Zitron's figures and projections, not confirmed by the companies.) Host Tom Bilyeu is actually an AI optimist and pushed back on the doom, but even he conceded the mismatch between revenue and debt is "terrifying."

The memory-is-a-classic-top case, Peter Boockvar. On Wealthion (July 15), the veteran investor drew the sharpest line of all: "The bubble is clearly in the CapEx side, not necessarily the technology." His target is memory-chip maker Micron, now earning an eye-popping 85% gross margin. His warning is a lesson in how cyclical stocks work: "Never tell me what a PE ratio is on a cyclical stock and tell me it's cheap." A dirt-cheap price-to-earnings ratio on a boom-and-bust company usually means earnings are at a peak about to roll over, not a bargain. And he flagged a threat almost nobody was talking about: while everyone fixated on SK Hynix, a Chinese memory maker called CXMT, the world's fourth-largest DRAM producer, was about to go public in Hong Kong, with a strategy of grabbing market share by underpricing everyone. As he put it, channeling Jeff Bezos: "Your profit margin is my opportunity... gross margin is my market share."

The "it's already slowing" case, Marco Papic. On the Stansberry Investor Hour (July 14), the strategist made a subtle but powerful point: AI spending is still growing, but the growth is decelerating. Capex went "from $200 to $400, from $400 to $800, and then to $900 [billion]. That's a slowdown." Since that spending has been the market's main engine, even a slowdown in the rate of increase is a headwind. His hard ceiling: "You cannot actually build $3 trillion worth of data centers... you cannot get the picks and shovels and real estate and land and electricity to do it." (Fred Hickey made the parallel "A.I.-Mageddon" case on Thoughtful Money the same day, also zeroing in on Micron's 85% margins as a cyclical peak.)

The bulls' answer

The bulls' response boils down to: the spenders are not stupid, and the base of the pyramid is still expanding.

Larry Fink and Kevin Warsh. On Wall Street Unplugged (July 16), the host relayed that BlackRock CEO Larry Fink, who runs the world's largest money manager, is "not worried about an AI bubble at all" and instead worried about "investing and building as fast as possible," adding that today's leverage is "nowhere near" 2008. Fink's build-cost rule of thumb: roughly $50–60 billion to stand up one gigawatt of AI data center, all-in. Former Fed governor Kevin Warsh, testifying to the Senate, offered the bull's ultimate backstop: even if AI underdelivers, "you're still going to have power. You're still going to have the infrastructure built that can be used for something else", and that spending on real assets is "better than just having the financialization of share buybacks and increasing dividends."
Brian Nowak of Morgan Stanley. On Power Lunch (July 16), he raised his forecast for big-tech spending to about $1.2 trillion next year and $1.4 trillion by 2028. His rebuttal to the bears is exactly the opposite of Chanos: "They are doing the math. They are only spending ensuring they're going to get ROIC" (return on invested capital). He argued the market punishing the chip and memory makers right now is a normal handoff, early in a tech cycle you make money on the hardware, then the returns move "up the stack" to software and internet companies.
Hugh Gimber (JPMorgan Asset Management) on Squawk Pod (July 17) split the difference: capex estimates have already climbed from $500B to $700B this year, but chip valuations now depend on the spenders proving they can keep pushing even higher, and he wants to see returns coming from outside the tech bubble before he's convinced.

Why it matters for a book: the bear case is now specific enough to trade around. If Chanos is right that the return-on-spending number keeps falling, the pressure point is late 2026 into 2027, and the most vulnerable names are the ones "dependent on Nvidia to exist" trading richer than Nvidia, the neoclouds and second-tier data-center plays. If Nowak is right, this sell-off is the buying window before earnings season proves the spending is disciplined. The next few weeks of hyperscaler results are the referee.

TSMC: the one fresh, hard fact, and it's bullish

Amid all the opinion, one company reported actual numbers, and they were spectacular. TSMC, the Taiwanese factory that manufactures essentially every advanced AI chip, for Nvidia, AMD, Google, Apple and the rest, is the closest thing this whole debate has to ground truth.

Per Schwab Network (July 16), TSMC beat on both revenue and profit, guided higher, and, the headline, announced another $100 billion of investment in its Arizona factories, bringing its total US commitment to $265 billion. The reason, per host Tom White: "demand is still outstripping supply." Its gross margin hit 67.7%, and it guided next-quarter revenue to $44.6–45.8 billion, about 12% growth in a single quarter. The Rundown (July 16) put the profit growth at 77% year over year.

Think about what that means. The company with the best view of real, paid-for chip orders 12–24 months out is raising its build plans by $100 billion. That is very hard to square with an imminent bust.

And yet, telling detail, TSMC's own stock still fell that day ("not enough for traders," as the episode title put it). The two nitpicks: management guided gross margins to be flat rather than up, and investors are nervous about the one thing TSMC can't control, "when are we going to see a top" in its customers' spending. Even the best fundamental news of the week couldn't fight the mood. That's the whole issue in miniature.

The operators: still building as fast as they physically can

If you want the sharpest contrast with the bubble panic, listen to the people actually running the build-out. Three genuine operators showed up this week, and none of them sounded worried about too much capacity, they're worried about too little.

OpenAI's compute chief: "We can't build fast enough." The most valuable interview of the window was Sachin Katti, OpenAI's Chief Compute Officer, on The MAD Podcast (July 16). His blunt summary of demand: "Demand far outstrips compute supply today. So anything we can bring online, we consume immediately. So there's no compute that is going waste for us." His actual fear is the opposite of the bears': "My worry is more on the downside of we are not able to actually build all the compute we want." He also gave rare detail on OpenAI's own custom chip, codenamed Jalapeno, designed with Broadcom in nine months from design to finished blueprint, "probably the fastest I've seen in my career", by a team that previously built Google's TPU chips. The metric that chip is built to maximize is "tokens per watt", how much AI output you get per unit of electricity, because power, not chips, is now the binding constraint. His list of real bottlenecks is a shopping list for the picks-and-shovels trade: gas turbines, transformers, electricians, and permitting, none of which the chip industry can conjure quickly. One more nugget with market implications: he said inference (running finished models for users) is now "perhaps even the majority" of compute, a tailwind for the cheaper, efficiency-focused chips that target inference.

Dell: the physics of the build-out. On SunCast (July 16), Dell's David Holmes gave the single most concrete stat of the week on how fast the hardware is changing: the power packed into a single server rack has gone "from 8 kilowatt racks eight years ago to 250 kilowatt racks today." That is a more than 30-fold jump, and it's why "everything we learned out of the last four decades of operating data centers isn't super helpful." Dell's answer is aggressive liquid cooling, including next-generation "Vera Rubin systems, fully liquid-cooled fanless systems" (Vera Rubin is Nvidia's coming chip generation), all aimed, again, at squeezing out more "tokens per watt." Read-through: the shift to liquid cooling and grid-flexible data centers is an operator reality now, not a slide-deck promise, which supports the cooling and power-equipment suppliers even if the chip stocks wobble.

A data-center operator just bet $1.1 billion on Cerebras. On Wall Street Unplugged (July 17), DigiPower X CEO Michel Amar described a 10-year deal to host chips for AI-chip startup Cerebras worth $1.1 billion, scalable to $2.5 billion, signed in May, first phase live by December. His demand read from the trenches: "60% of these projects are being delayed" for lack of power, which he argues hands a huge advantage to anyone who already has power and buildings ready. As a rough gauge of how hot data-center assets are, he cited Blackstone paying $27 million per megawatt a month ago (his figure). Two cautions for readers: this is a small company talking its own book, and his claim that Cerebras must guarantee 750 megawatts of power for OpenAI is his characterization, not a confirmed contract. Still, the signal, real money changing hands for capacity, right in the middle of the sell-off, is the point.

The operators' collective message: the demand is real, the constraint is physical (power, transformers, cooling, electricians), and nobody in the building is slowing down. Whether that demand earns its keep is exactly what the bears dispute, but the "no one wants this" version of the bubble story isn't what the operators are living.

AMD finally gets its close-up (still no operator, though)

For roughly six issues we've flagged the strange silence around AMD, no executive interviews, no roadmap talk, just the occasional mention by a rival. This week that silence broke, at least partway: veteran analyst Dave Vellante devoted an entire episode of Breaking Analysis (July 18) to AMD's strategy. It's still a pundit, not the company, so the operator gap technically stands, but it's the most substantive AMD framework we've had in weeks, and it's timely with AMD's "Advancing AI" event happening this coming week.

His core thesis is clarifying: AMD's job is not to beat Nvidia. "The objective is not to beat NVIDIA... Instead, AMD's challenge is to define what success looks like in a market where NVIDIA remains number one." The goal is to become "the indispensable second platform", the credible backup that every big customer wants so they aren't 100% dependent on one supplier. The playbook, in his telling: build the core (AMD's Epic processors, Instinct AI chips, and ROCm software, its answer to Nvidia's dominant CUDA software), buy the missing pieces (the $49 billion Xilinx deal, the Pensando networking acquisition, and ZT Systems for rack-scale assembly), and push open industry standards to win over customers wary of being locked in. The memorable line: "AMD is trying to compress 20 years of ecosystem development into five years of capital allocation."

The math that makes it interesting: Vellante argues AMD doesn't need much. Grab even a "mid-single-digit share", call it 6%, of the AI-chip market, keep its ~55% share of the x86 server-processor market, and that alone creates "enormous value." But he's clear-eyed on the risk: "NVIDIA is not Intel. And Jensen is in founder mode", meaning the usual playbook of beating a sleepy incumbent (which worked against Intel) won't work against a hungry, founder-led Nvidia whose software moat (CUDA) is "one of the strongest soft moats in technology history." A big chunk of the bull case rests on AMD's reported deal to build six gigawatts of capacity for OpenAI over four to five years. And note the shared risk with everyone else: he pegs hyperscaler capex at $700 billion this year and $900 billion next, "if that investment gets dialed down, it could affect AMD's trajectory."

Why it matters: AMD is the purest way to play the "customers demand a second source" theme that Vellante, Chanos (obliquely) and the operators all point to. But it's also the higher-beta bet, if the capex bears are right, a name valued for future AI success gets hit harder than Nvidia, whose current profits are already enormous.

Optics: the long silence finally cracks

Regular readers know optics and AI-networking pure-plays have been this newsletter's most stubborn blank space, issue after issue of zero operator coverage. This week the names finally surfaced, even if the voices were still investors rather than executives.

On Shooting the Bull (July 16), a portfolio manager who's held Astera Labs (ALAB) since its $40 IPO gave the specific, actionable detail we've been missing: the two connectivity-chip designers Astera Labs and Credo (CRDO) have landed slots in "the Trainium for Amazon and then the MI500 for AMD, which is pretty much like two of the best programs they could be in." In plain terms, these two small companies make the chips that shuttle data around inside AI servers, and they've won places inside Amazon's and AMD's newest AI chips, arguably the two best non-Nvidia programs to be attached to. Stock Market Today With IBD (July 15) added that Astera now carries roughly a $62 billion market value.

The same investor also captured this market's maddening character perfectly: "The stock goes up 300 percent, they report a great result, the stock goes down like 20 percent after because it was, quote unquote, priced in." That's the whole mood, good news isn't good enough when expectations are stretched. He was frank that both Astera and Credo had "got a little overvalued recently" before the July 16 reversal knocked them down.

So: optics is no longer a total blank, but it remains pundit territory, still no executive from Astera, Credo, Coherent, Lumentum or Arista on a podcast this window. We'll keep flagging it until an operator actually shows up.

Memory after the crash

The memory names were the epicenter of the sell-off, down 20% to 35% across the board, which drew the bottom-fishers out. On The MoneyFlows Show (July 16), the hosts summarized the odd spectacle: "the hottest part of the market is in a bear market," even though the largest foreign IPO in US history (SK Hynix, which raised $26.5 billion at seven times oversubscribed) had just happened. Their key market-share figures: SK Hynix holds about 56% of the high-bandwidth-memory market (the stacked, super-fast memory that sits next to AI chips), with Micron having climbed to about 26% as a fast follower. (The show also flashed some SK Hynix revenue projections that looked internally inconsistent, so we're leaving those out rather than repeat a number we can't stand behind.)

The bull-versus-bear tension here is really about one word: commodity. The bears (Boockvar, Hickey) say memory is a classic boom-bust commodity at a cyclical peak, with Chinese rivals like CXMT about to flood in and crush margins. The bulls counter that high-bandwidth memory isn't a commodity at all, as the Shooting the Bull investor argued, it's a fourth-generation, hard-to-make product with only three companies on Earth able to produce it (a "triopoly"), high barriers, and long qualification cycles. He also flagged the mechanic that's changed the game: Micron's new long-term "take-or-pay" contracts deliver upfront, unrestricted cash the company can plow straight back into new factories. Both things can be true, durable near-term because supply is genuinely tight through 2027, but still cyclical whenever the Chinese capacity and the next down-leg arrive.

China is the pressure behind everything

Step back and almost every thread this week runs through China. A cheaper Chinese model (Moonshot's Kimi) triggered the sell-off. A cheaper Chinese memory chip maker (CXMT) is the margin threat that worries Boockvar. And on the competitive front, the most interesting reporting came from The Information's TITV (July 17), where editor Amir Afradi (a reporter, not an operator) described "the Nvidia GPU crunch" and Google's aggressive push to get its own TPU chips into rivals' hands. Google's pitch, per the reporting: TPUs change less generation-to-generation than Nvidia's chips, so customers get "fewer surprises" when installing them. Google is even offering to financially backstop loans so other companies can buy TPUs, and has a joint venture with Blackstone to run a TPU cloud. Afradi's bottom line on the one company Nvidia truly fears: "The one single competitor that Jensen is paying close attention to, it's Google and TPUs."

One important clarification buried in that same conversation, because it cuts against the bubble narrative: those lender "backstops" that Nvidia and Google are offering are "not a reflection of demand at all. This is just a reflection of financing", a way for young companies without credit histories to borrow for $50–60 billion data centers. Worth remembering when you see a scary headline about circular AI financing.

Names in play

TSMC, the fundamental anchor of the week: profit +77%, +$100B US investment ($265B total), demand still outstripping supply. If you believe one data point over the mood, it's this one.
Nvidia (NVDA), de-rated as sentiment soured; briefly passed by Apple as most valuable company. The bears' rule ("nothing dependent on Nvidia should trade richer than Nvidia") is really a warning about everyone else.
Micron (MU) / SK Hynix / SanDisk / Seagate, memory names down 20–35%; the commodity-versus-triopoly debate is the whole argument. Watch CXMT's Hong Kong listing as the China-margin catalyst.
AMD, the "indispensable second platform"; Advancing AI event this coming week (new Venice CPU expected). Higher-beta way to play the second-source theme; OpenAI 6 GW deal is the swing factor.
Astera Labs (ALAB) / Credo (CRDO), networking-chip pure-plays finally named, with design wins inside Amazon Trainium and AMD MI500. Richly valued, freshly knocked down.
Broadcom (AVGO), surfaced mainly as OpenAI's manufacturing partner on the Jalapeno custom chip (via Katti); the quiet beneficiary of the custom-silicon wave.
Oracle, the bears' favorite pressure point (Chanos: worst returns; Zitron: 7.1 GW for one customer). Not a chipmaker, but the clearest place a hardware-financing crack would show first.
Cerebras (private), visible only through a customer (DigiPower X's $1.1B deal), but that's a real demand signal for challenger silicon.

What changed since Issue 009 (July 16)

The lead evolved from a single-stock whipsaw to a sector-wide break. Issue 009 was about SK Hynix's wild post-IPO swings as a positioning brawl. Issue 010's story is bigger: the entire chip index fell into a bear market, and the AI-bubble question moved from a debate to a market event.
The bubble debate went from two voices to a full panel, and got more specific. Issue 009 had Gavin Baker (bull) versus Rory O'Driscoll and Peter Schiff (bears). Issue 010 adds a heavyweight bear bench (Chanos's 40%→20%→10% returns math, Zitron's OpenAI burn, Boockvar's cyclical-peak case, Papic's deceleration) against a serious bull bench (Fink, Warsh, Nowak, Graham, Gimber). The debate is now tradable, with late-2026/2027 as the pressure window.
A brand-new bull anchor: TSMC's actual numbers. TSMC didn't feature in Issue 009 at all. This week its blowout print and +$100B build-out are the strongest fresh evidence the base of the stack is still expanding.
Genuinely new operators. OpenAI's compute chief (Katti), Dell (Holmes) and DigiPower X (Amar) all appeared for the first time, the demand-side counterweight to the pundit panic.
AMD's silence partly broke. After ~6 issues with no dedicated coverage, a full strategy deep-dive (Vellante) finally ran, though still a pundit, so the operator gap technically continues.
Optics cracked open. From near-total absence to named design wins (Astera/Credo in Trainium and MI500), still pundit-sourced, but the names are finally on the board.
China flipped from supply-side to systemic pressure. Issue 009's China story was H200 import permission (demand-side). This week China is the trigger (Kimi model), the margin threat (CXMT), and the competitive wedge (Google TPUs) all at once.
Not repeated this window: the Nvidia roadmap-rumor thread (The Circuit EP183) and the a16z Baker/Patel interviews were already covered in Issue 009 and are not re-run here.