Newsletter · 2026-07-16 · Ashutosh Agarwal

Memory Chips Whipsaw as the AI Bubble Debate Goes Mainstream - AI Accelerators - Week of July 16, 2026

AI Accelerators hardware newsletter for the week of July 16, 2026. SK Hynix's post-IPO whipsaw dragged the AI-bubble debate into the open, with Gavin Baker, Rory O'Driscoll and Dylan Patel making the bull, bear and 'can anyone catch Nvidia' cases in detail.

AI Accelerators

Week of July 16, 2026: Memory Chips Whipsaw as the AI Bubble Debate Goes Mainstream

GPUs, Custom Silicon & Optics: a twice-weekly read of what the best podcasts are actually saying about AI hardware. Issue 009. Research window: July 9–16, 2026.

Last issue we watched SK Hynix, the Korean company that makes most of the specialized memory chips inside Nvidia's AI systems, pull off the biggest US stock-market debut ever by a foreign company. This issue we watched what happened next, and it was violent. Within one trading day the stock had its worst-ever crash back home in Korea, then two days later it ripped 27% higher. Nothing about the actual business changed in that week. What changed was the mood.

That whipsaw is the story of the week, because it dragged a bigger question into the open: is this whole AI-hardware boom real, or is it a bubble? For months that debate lived on the fringes. This week it moved to center stage, and, usefully, some of the sharpest people in the business made the case on both sides, in detail, with numbers. Gavin Baker, a well-known technology investor, made the single best "this is not a bubble" argument I've heard in a while. A venture capitalist named Rory O'Driscoll laid out the math that scares the bears. And Dylan Patel, who runs the most-read chip-research shop in the industry, explained precisely why almost nobody can catch Nvidia, even when they build a genuinely good chip.

Meanwhile, the companies trying to build their own chips kept shipping. Meta's first in-house AI chip is reportedly going into production in September. The CEO of Cerebras said his order book is now $25 billion. Qualcomm told investors it wants $15 billion a year from AI data centers by 2029. And China, after months of refusing, quietly started letting its biggest AI companies import a limited number of Nvidia chips.

Let's get into it.

TL;DR

SK Hynix whipsawed. After its record $26.5 billion Nasdaq debut on Friday, the US shares fell more than 9% on Monday and the Korea-listed stock had its worst day ever (down ~15%, ~40% off its high), then the US shares rallied 27% by Thursday's close. Traders on CNBC's Fast Money called the name "priced truly for perfection," with "zero room for error." Jim Cramer insisted the sell-off was "not really related to the price of DRAMs" and that the story is still "secular," not a normal boom-bust cycle.
The bubble debate went mainstream, and got substantive. On The a16z Show, investor Gavin Baker argued flatly: "There are no dark GPUs." On SaaStr, VC Rory O'Driscoll countered with the scary math: roughly $688 billion of AI spending this year against ~$110 billion of revenue, "half a trillion dollars more than we're taking in," and payback that may not arrive until 2031–2032. Peter Schiff went further: "cash cows... have turned into cash vacuums."
"Can anyone catch Nvidia?" SemiAnalysis's Dylan Patel gave the clearest answer yet: to win you must be "5x better," and Nvidia's edge in networking, memory, manufacturing and time-to-market shrinks that 5x to nothing. Even AMD, he noted, "got to 2 nanometer before Nvidia... and yet they still lose."
Challenger silicon is actually shipping. Meta's first custom chip (codename Iris), built with Broadcom and TSMC, reportedly enters production in September. Cerebras CEO Andrew Feldman said the company has a $25 billion backlog. Qualcomm is targeting $15 billion of AI-data-center revenue by 2029, up from zero.
China cracked the door for Nvidia's H200. The Information reported Beijing will let Alibaba, ByteDance and DeepSeek buy a limited number of H200 chips: a demand-side reversal after months of blocking them.
Nvidia roadmap rumors rattled the small names. The Circuit flagged (unconfirmed) chatter about delays to Nvidia's next-generation parts. The takeaway: delays barely dent Nvidia's revenue, but they hit the smaller optics and networking suppliers far harder.
Negative space: AMD goes a sixth straight issue with no operator or roadmap content. Optics pure-plays (Coherent, Lumentum, Astera Labs, Credo, Arista) still produced no company voices, only pundits.

1. The Lead: a $27 Billion Memory Stock on a Rollercoaster

(Mostly pundit commentary + reporting; SK's chairman is the operator voice.)

Start with the facts, because they're wild on their own. SK Hynix priced its US listing last Friday (July 10) at $149 per share and raised $26.5 billion, the biggest-ever US market debut by a foreign company, as reported on Wall Street Breakfast (Seeking Alpha's Brian Stewart, July 10). SK Hynix makes high-bandwidth memory (HBM), the stacks of ultra-fast memory chips that sit right next to an AI processor and feed it data, and it controls a majority of that market. (Reporters this week put its share anywhere from 56% to 60%; SK's own figure is 57%.)

Then the ride began. By Monday, July 13, the US-listed shares were "tumbling more than 9%... just $3 above where they priced," and the Korea-listed stock "fell more than 15% overnight, their worst day on record," per CNBC's Fast Money (July 13). Other memory names went down with it: SanDisk fell almost 13%. Trader Tim Seymour pinned much of the blame on the home market rather than the business: "foreign flows are down $100 billion" this year and "the local Korean market is over-levered," with margin calls and leveraged funds amplifying every move. Technical analyst Katie Stockton noted the Korea shares were already "down nearly 40% off the high" (Bloomberg Intelligence, July 13).

On Squawk on the Street the same morning, Jim Cramer (July 13) made the point that matters for anyone holding the memory trade: the crash wasn't about chip prices. "The decline is not really related to the price of DRAMs, which is typically what it would be linked to," he said. His read is that this is a secular story (a lasting shift, not a normal cyclical boom that inevitably busts), "because you're getting longer-term agreements. That's what Micron's got. That's exactly what Sanjay Mehrotra, the CEO of Micron, said on my show." But even Cramer admitted the price action looks unhinged: the Korean market "seems to lack true pricing discovery... a mob that buys it, a mob that sells it," and he compared it to "Japan in 1988."

Then the whipsaw completed. By Thursday (July 14), the US shares "ramped 27% by the closing bell" and were back in positive territory versus the Friday debut, per Fast Money (July 14). CNBC's Oliver Rennick, reporting from the options exchange floor, added texture that a book should care about: options on the new stock launched that day with about 220,000 contracts traded, roughly evenly split between bullish and bearish bets, but "the top 10 trades by dollar amount were either neutral or bearish." A dozen leveraged funds also launched on the single stock the same day. Translation: this is now one of the most mechanically fragile, over-engineered trades in the market. As one trader put it, the shares are "priced truly for perfection" with "zero room for error": "an 8 percent miss reacts with a 30 percent stock fall."

Why it matters for a book: The SK Hynix debut was supposed to be the market's referendum on whether AI-memory demand is durable. Instead of a clean verdict, we got a two-sided brawl: euphoria, a record crash, and a violent bounce, all in four sessions, most of it driven by positioning (leverage, ETFs, a hard-to-arbitrage gap between the US and Korea listings) rather than fundamentals. The operator view hasn't wavered: SK chairman Chey Tae-won told Bloomberg Tech (July 10) and Squawk on the Street (July 10) that "the demand is enormous exponentially... I don't really see any shrink sign of the HBM," and that customers are asking him to double capacity again. But the stock has decoupled from the company, and that gap is now the risk. Watch Samsung: several podcasts expect it to follow SK Hynix with its own US listing, which would add supply of "memory paper" to a market already choking on it.

2. Is This a Bubble? The Best Arguments on Both Sides, in One Week

(All pundit, but unusually high-quality pundit.)

This is the debate that actually moved thinking this week, so let's steel-man both sides.

The Bull: "There Are No Dark GPUs"

The standout was investor Gavin Baker (of Atreides Management) on The a16z Show (July 14). His argument is the cleanest rebuttal to the bubble case I've heard, and it rests on a single historical analogy.

The 2000 tech crash, Baker argues, was really a telecom bust, and it was defined by "dark fiber": fiber-optic cable that companies laid in the ground but never lit up because there was no demand for it. "At the peak of the bubble, 97% of the fiber that had been laid in America was dark," he said. Then the line that's going to get quoted for months: "Contrast that with today. There are no dark GPUs." Every AI chip being built is being used: famously, the biggest engineering problem in a big training run is GPUs literally overheating from full utilization.

Baker's second point is about price. Cisco, the poster child of 2000, "peaked at 150 or 180 times trailing earnings. Nvidia's at more like 40 times." (These are Baker's figures.) And his third is about returns: "the return on invested capital of the biggest spenders on GPUs, who are all public... have seen, call it a 10-point increase in their ROICs." In plain English: the giant companies pouring money into AI chips are, so far, earning more on their overall capital, not less. "There's an interesting and open debate about whether or not it will continue to be positive... I personally think it will, but there's no debate that thus far the ROI on AI has been really positive." (Again, Baker's framing: treat the 10-point number as his estimate.)

The Bear: Half a Trillion Dollars a Year, and Payback in 2032

The most concrete counter came from VC Rory O'Driscoll of Scale Venture Partners on SaaStr (July 15). He was careful to call his own numbers "roughly right," so treat them as informed estimates, but the shape is what matters. The giant cloud companies are spending about $688 billion on AI this year, he said. What's coming out the other side in actual revenue? Around $110 billion, most of it (~$89 billion) from the two big model-makers, OpenAI and Anthropic. "The bottom line is we're spending half a trillion dollars more than we're taking in."

The kicker: "It takes until 2031 or 2032 before the revenue from those two companies surpasses the total CapEx." Five or six years of spending more than you make. His warning isn't that the technology fails, it's that the financing gets tested along the way: "at any point you could have a hiccup... someone wakes up and goes, 'Oh my God, what are we doing? We're spending half a trillion dollars, we're only making $200 million. Maybe we should slow down.'"

Peter Schiff took the same idea and turned the volume up on his show (July 15). His framing: Nvidia's profits "come at the expense of the hyperscaler's losses," and those big customers, having burned through cash, are now borrowing to keep buying: "these big cash cows have turned into cash vacuums." His sharpest point is about accounting: the chips are being depreciated (written down) over five or six years, but "the life cycle of this equipment is more like two to three years... that's more like an operating expense." If he's right, the reported profits of the AI buyers are flattered by stretching out the cost of gear that goes obsolete fast. Schiff, notably, is a long-time gold-and-doom voice, so weight accordingly, but the depreciation question is a real one that even bulls should be able to answer.

Why it matters: Notice that the bull and the bears aren't really arguing about whether AI is useful: everyone agrees it is. They're arguing about timing and financing: is the return on all this spending showing up fast enough to keep the money flowing before someone blinks? Baker says the returns are already here; O'Driscoll says the bill comes due in 2032; Schiff says the accounting is hiding the losses. For a hardware book, the practical read is that the risk has shifted from "will demand show up?" (it has) to "will the funding for demand hold together?", which is exactly why an over-levered memory stock crashing 15% in Korea rattled everyone.

3. "Can Anyone Catch Nvidia?" The Clearest Answer Yet

(Pundit, but from the industry's most-cited chip analyst.)

If you only listen to one thing from this week, make it Dylan Patel of SemiAnalysis on The a16z Show (July 15). It's the best plain-English explanation of why Nvidia is so hard to dislodge, and it doubles as a warning about the custom-chip programs everyone is excited about.

Patel's core idea: to beat Nvidia, you can't just match it, you have to be dramatically better, because Nvidia wins on everything around the chip. "NVIDIA's going to have better networking than you. They're going to have better HBM. They're going to have better process node. They're going to come to market faster... So you can't just do the same thing as NVIDIA. You have to really leap forward in some other way. You have to be like 5x better."

And here's the trap: even if your chip is 5x better on paper for one type of work, Nvidia's advantages in the supply chain (better deals with the factory, TSMC, and the memory maker, SK Hynix) quietly eat that lead. "The supply-chain stuff means that 5x actually turns into a 2.5x. And then NVIDIA can compress their margin a little bit if you're actually competitive. And then that 2.5x becomes like a 50% better," at which point it's not worth switching. His Exhibit A is the one rival that has done everything right: "Even AMD... got to 2 nanometer before NVIDIA. They had higher-density HBM. They use 3D stacking, all these things on supply chain that should be better than NVIDIA, and yet they still lose."

This is the bridge back to the custom-chip story. On the same show, Baker predicted that "most of those ASICs are gonna fail" (ASIC being a custom chip built for one company's specific needs) and that "in the next 3 years... a bunch of high-profile ASIC programs canceled." His view is that the real contest narrows to "Google and the TPU versus Nvidia," with AMD as the necessary "second source" that big buyers keep alive just so Nvidia has a competitor.

Patel also gave Nvidia a piece of advice worth flagging for its financial read-through: a new US tax rule lets buyers write off the entire cost of a GPU cluster in year one, which he estimated is worth roughly $10 billion a year to Meta alone and is "massive" across the big cloud companies. That tax break is a direct subsidy to the AI buildout, a reason spending may run hotter for longer than the skeptics expect. (Patel's estimate; not a company figure.)

Why it matters: These two episodes reframe the whole "competition" narrative that drives so many hardware trades. The bull case for the custom-chip suppliers (Broadcom, Marvell) and for AMD assumes buyers will diversify away from Nvidia. Patel's argument is that the physics and economics make that far harder than the headlines suggest, which is bullish Nvidia, cautious on the "Nvidia killer" names, and a reason to demand real evidence (shipments, not press releases) from every ASIC program before believing it.

4. Challenger Silicon Is Actually Shipping (This Is Where the Operators Are)

(Cerebras CEO = operator; the rest is reporting + pundit.)

The pushback to Patel's skepticism is that the challengers are no longer just talking. This week produced real, datable milestones.

Cerebras: the operator voice of the week. On All-In (July 10), Cerebras founder and CEO Andrew Feldman (whose company builds a giant, unconventional AI chip and is about to go public) said the quiet part out loud about demand: "We have a $25 billion backlog. And we are not alone in that." He described a global building spree ("Kazakhstan, Tajikistan... Georgia, Armenia") where customers "are trying to capture yesterday's demand" because "the demand is way outstripping our ability to build data centers and fill them with hardware." On why every big cloud company is designing its own chip, his answer echoes Patel and Baker but from the inside: "Nobody likes being dependent," the hyperscalers learned that lesson being hooked on Intel's processors for years. His bottom line: "there is so much demand right now that there is no silicon that will go unused." (That's a company CEO talking his own book, but it's the clearest operator demand signal of the week, and it rhymes with Baker's "no dark GPUs.")

Meta's "Iris" chip goes to production in September. Two podcasts converged on this. Motley Fool Hidden Gems (July 9) reported Meta will start producing its own AI chip in September, with Broadcom and TSMC as designer and manufacturer ("the essential model that Apple uses to design its own iPhone chips") to handle the easier AI jobs while still buying Nvidia and AMD for the hard ones. The Rundown (July 11) added the codename, Iris, citing an internal memo reviewed by Reuters (treat the codename and timing as reported, not company-confirmed). Motley Fool's Matt Frankel made the read-through that matters: if it works, it's a "bigger needle-mover for Broadcom than for Meta." One more detail worth noting because it cuts against the "everyone's dumping Nvidia and AMD" narrative: the same episode reminded listeners that Meta "signed the biggest AI deal in history to buy AMD chips for the hardest AI jobs." So Meta is doing both, building its own for easy work and buying more merchant chips for hard work.

Qualcomm quietly becomes a data-center chip company. On Chip Stock Investor (July 14), the hosts walked through Qualcomm's investor day: a target of $15 billion in annual AI-data-center revenue by 2029, up from zero in 2024. Two things stand out. First, Qualcomm bought a company called Alpha Wave for the networking technology it needs to move data around a data center, the exact "networking" edge Patel says Nvidia dominates. Second, Qualcomm's design stacks a computing chip underneath a memory chip to "sidestep the need for high-bandwidth memory," a clever way to dodge the very memory shortage that's making SK Hynix so valuable, aimed at the cheaper "inference" side of AI (running finished models rather than training new ones). If it works, it's a small threat to Nvidia but a genuine one to the HBM story. (These are company targets from an investor day, unproven.)

Why it matters: This is the strongest evidence against the "no one can catch Nvidia" thesis, and the honest read is that both things are true at once. The challengers are landing real milestones (Meta in production in September, Cerebras' $25 billion book, Qualcomm's 2029 target), but every one of them is aimed at the easier, cheaper AI work, not the frontier training that Nvidia owns. The name that keeps winning either way is Broadcom, which gets paid to help Meta, Google and others build their custom chips regardless of whose brand is on the box.

5. China Cracks the Door for the H200

(Reporting: The Information's exclusive.)

A genuine policy reversal, and a demand-side one. On The Information's TITV (July 9), reporters Chenner Liu and Jing Yang broke that Beijing is preparing to let China's biggest AI companies (Alibaba, ByteDance and DeepSeek) buy a limited number of Nvidia's H200 chips. (The H200 is a powerful-but-not-newest Nvidia AI chip; the current generation is called Blackwell, and the next is Rubin.)

The nuance is the whole story. "The approvals are not completely open-ended" (companies have to justify how many they need and why) and "this is also not a full reopening." The government's message, as the reporters framed it: "use NVIDIA chips whenever you must, but use Chinese chips when Beijing thinks they are good enough." In practice, that means Nvidia for training new models (where its software and reliability still can't be matched) and homegrown chips from Huawei and others for inference. Why the change after months of blocking? A real shortage: "Chinese AI companies now need more compute... they just don't have enough high-end chips," made worse by the US crackdown on chip smuggling.

Why it matters: This is a small, deliberately capped positive for Nvidia's China revenue, not a floodgate. But it's a notable shift from last issue's story, when China was framed as building its way off Nvidia with vertical, homegrown silicon. Both are true: China wants independence long-term but is pragmatic enough to buy Nvidia now because its own training chips aren't ready. It also lines up with a striking aside from Patel on the a16z show, that some Chinese provinces have banned the weaker H20 chip as "not efficient enough," because unlike the US, China has effectively "infinite" power and would rather run fewer, more-powerful chips. The constraint in China is chips; the constraint in America is electricity.

6. Nvidia Roadmap Rumors, and Why the Small Names Feel It More

(Pundit / unconfirmed rumor, flagged as such.)

The Circuit devoted its EP 183 (July 14) partly to a swirl of supply-chain rumors about Nvidia's future products. To be clear: Nvidia officially denies any of this. The company's public posture, per the hosts, is "no delays," "our transition to co-package optics remains on time," "800-volt data center remains on time." The unconfirmed chatter they relayed: that a top-end future part ("Rubin Ultra") is being "scaled down from four die to two," that a rack system ("Kyber") is delayed, and that Nvidia's move to co-packaged optics (a way of building the light-based data links directly into the chip package) is slipping. Treat all of that as rumor, not fact.

The genuinely useful insight is the hosts' framing of who gets hurt if any delay is real: not Nvidia. "Even if there's a delay... it doesn't change how much money they're going to make," because demand so far outstrips supply. But it matters enormously "for the bottleneck-trade people lower down that chain, in the optical or the PCB [circuit board]" world. Those suppliers have few customers and lumpy revenue, so one pushed-out quarter can mean "suddenly a third of your revenue is missing." And optical, specifically, "always gets delayed."

Why it matters: This is the read-through to watch as the real buildout ramps. A rumor that reads as "bad for Nvidia" is often, on inspection, bad for the smaller optics, circuit-board and connector names that depend on Nvidia's exact launch timing. The lesson for a book: don't reflexively short Nvidia on a delay headline, but do stress-test the tiny, single-customer suppliers whose whole year hinges on one product shipping on schedule.

Names in Play

SK Hynix: the week's rollercoaster; now a highly mechanical, leverage-driven trade "priced for perfection." Watch Samsung for a copycat US listing that adds supply.
Nvidia (NVDA): the bull/bear debate's center of gravity. Baker: ~40x earnings, not a bubble. Patel: near-impossible to dislodge. Rumors of roadmap slips; company denies. A year-one tax write-off on GPU clusters is a quiet tailwind to demand.
Broadcom (AVGO): the pick-and-shovel winner of the custom-chip wave: paid to co-design Meta's "Iris," plus Google's TPU and others. The one name that wins whether or not the "Nvidia killers" do.
Meta (META): first in-house chip ("Iris") reportedly to production in September; also the buyer of "the biggest AI deal in history" for AMD chips. Doing both build and buy.
Cerebras: pre-IPO; CEO claims a $25 billion backlog and better-than-Moore's-Law performance gains. Operator demand signal, talk-your-book caveat applies.
Qualcomm (QCOM): targeting $15B of AI-data-center revenue by 2029 (from zero); a design that sidesteps HBM for inference. Unproven but a name to add to the watchlist.
AMD (AMD): sixth straight issue without an operator/roadmap appearance; surfaces only through rivals ("the second source"; "got to 2nm and still loses") and Meta's big buy order.
Micron (MU): repeatedly cited as the "long-term contract" counterweight to SK Hynix's spot exposure; down ~22% from recent highs.
Monolithic Power (MPWR): flagged as a beneficiary of the shift to 800-volt data centers (power management), per Baird's Ted Mortonson (Bloomberg Surveillance, July 15).
Optics (Coherent, Lumentum, Astera Labs, Credo, Arista): pundits still like the "secular change" story; no company voices this week; most exposed to any Nvidia timing slip.

Read-Throughs

The risk has moved from demand to financing. Every serious voice, bull and bear, agrees demand is real. The debate is whether the funding holds up before returns arrive (Baker: already positive; O'Driscoll: payback in 2031–32; Schiff: borrowing to buy gear that obsoletes in 2–3 years). An over-levered memory stock crashing 15% in Korea is the first tremor of that exact fear.
"Nvidia killer" stories deserve a higher bar. Patel's "you must be 5x better, and the supply chain eats your lead" argument is the sharpest reason to treat every custom-chip headline as a claim to be proven by shipments, not press releases.
Broadcom is the through-line. If custom chips win, Broadcom gets paid to build them. If they mostly fail (Baker's view), Broadcom still collected the design fees. It's the lowest-beta way to play the custom-silicon theme.
Delays hurt the little guys. A Nvidia roadmap slip is a rounding error for Nvidia and a potential quarter-killer for a single-customer optics or circuit-board supplier. Position accordingly.
China is a capped positive, not a floodgate. The H200 approvals are deliberately small and training-only. Don't model a China re-opening; model a controlled trickle.

What Changed Since Issue 008 (July 13)

The lead evolved from the debut to the aftermath. Issue 008 led with SK Hynix's record debut and the chairman's "this is not a cycle" case. Issue 009 leads with the whipsaw that followed (a record crash then a 27% bounce), which reframed the "durability vote" as a positioning-driven brawl, not a clean fundamental verdict.
The bubble question moved to center stage, with real arguments. Last issue had a named peak-earnings bear (Susquehanna's Hosseini) on memory specifically. This issue has a full, two-sided referendum on the whole AI-hardware trade: Baker's "no dark GPUs" bull case vs. O'Driscoll's "$688B in, $110B out, payback in 2032" math and Schiff's "cash vacuums."
New this week: the definitive "can anyone catch Nvidia" framework (Patel's "5x better," AMD "still loses"), which also recasts the custom-chip names more skeptically than Issue 008's incumbent-coup framing.
Custom silicon shifted from incumbents to shipping milestones. Issue 008 emphasized Broadcom-Apple and China's vertical silicon. Issue 009 has concrete, datable milestones: Meta's "Iris" to production in September, Cerebras' $25B backlog (operator), Qualcomm's $15B-by-2029 target.
China flipped from supply-side to demand-side. Issue 008: China building its own ASICs. Issue 009: China importing limited H200s (Alibaba, ByteDance, DeepSeek): a pragmatic reversal.
New: Nvidia roadmap-rumor risk and its downstream "bottleneck trade" read-through (The Circuit EP183), which Issue 008 didn't carry.
Negative space escalates. AMD is now a sixth straight issue with no operator/roadmap content. Optics pure-plays remain pundit-only. The Nvidia roadmap story is, again, told by analysts and rumor, not by Nvidia on record.