Newsletter · 2026-09-19 · Ashutosh Agarwal

Memory Crunch Within a Crunch as Gerstner Takes On the AI Bubble Callers - Semiconductor Podcast Briefing - Week of September 19, 2026

Semiconductor Podcast Briefing for the week of September 12 to September 19, 2026: podcasts argued the AI slowdown selloff in memory was backwards, with HBM eating about 30% of DRAM capacity for 13% of output, SK Hynix entered early talks to make memory in Intel's Ohio fab under a 100% tariff threat, Brad Gerstner made the no-bubble case against Richard Kramer and Aswath Damodaran, and ON Semiconductor fell 9% despite an upbeat investor day.

Semiconductor Podcast Briefing

Week of September 19, 2026: Memory Crunch Within a Crunch as Gerstner Takes On the AI Bubble Callers

This was a week where the loudest chip story wasn't a chip at all, it was memory. A wave of "let's slow down AI" comments from the heads of Anthropic, OpenAI and others rattled the whole complex early in the week, memory names got hit hardest, and then the podcasts spent the next several days arguing the sell-off was exactly backwards. Underneath that noise: the Federal Reserve actually raised interest rates (chair Kevin Warsh's decision landed mid-week), which matters for chips because so much of the AI build-out is now paid for with borrowed money.

Here is what the week's podcasts actually said, with the numbers and the direct quotes.

TL;DR: Things That Mattered This Week

• Memory is the tightest link in the chain, and the podcasts think it's underpriced. Micron ($MU) is trading near $940 at a forward price-to-earnings ratio of about 5.9, with a consensus price target of $1,575 (roughly 70% higher) even as analysts keep raising estimates (2027 earnings per share now pegged at $156). One host summed up the physics: making high-bandwidth memory eats "about 30%" of global DRAM capacity but only yields "about 13%" of usable output, "a crunch within a crunch." (The MoneyFlows Show, Sep 17)
• SK Hynix is in early talks to make memory chips inside Intel's Ohio fab, the first time SK Hynix would manufacture memory in the US. It's driven by the HBM shortage and by a threat from Commerce Secretary Howard Lutnick of up to a 100% tariff on Korean chipmakers that don't build in America. Both stocks rose ~3% on the report. (The Rundown, Sep 16)
• Brad Gerstner (Altimeter) made the definitive "no bubble" case: NVIDIA revenue up 2x and hyperscaler capex up 2x this year, NVIDIA trading at "14 times next year's fully taxed" earnings, and "semiconductors are 70% of the NASDAQ's return." (All-In, Sep 17)
• The bears pushed back hard. Former Goldman analyst Richard Kramer called equity markets "deeply unmoored from traditional intrinsic valuation… It's all vibes," and NYU's Aswath Damodaran pegged the AI build-out at "$2 to $2.5 trillion… the largest leap into the unknown" in business history. (FT Tech Tonic and BiggerPockets Money, Sep 15–16)
• Oracle's record capex became the tell for the whole trade. Its quarterly capital spending hit roughly $28.5 billion, up from about $8.5 billion a year earlier, and Chip Stock Investor now expects hyperscaler capex growth to "slow down in 2027" and slow "even more significantly" in 2028. (Chip Stock Investor, Sep 14)
• NVIDIA told the market it plans to roughly double chip sales next year, and "no key AI players" have signaled pulling back compute plans. Post-Fed, NVIDIA, Broadcom, Arm and Intel all traded up 2–3%. (The KE Report, Sep 18; Squawk on the Street, Sep 17)
• ON Semiconductor used its investor day to insist there is no slowdown, but the stock still fell 9% and is down 40% over three months. (Closing Bell, Sep 16)

1. AI Chip Demand and Hyperscaler Capex

The single most-discussed chart of the week came from Brad Gerstner of Altimeter Capital, speaking on All-In (Sep 17). His core point is that this is an earnings-driven move, not a valuation bubble:

"We have NVIDIA revenue up 2x, hyperscaler CapEx up 2x, OpenAI and Anthropics valuation up 2x… This is not about multiple expansion. This is an earnings-driven market expansion… Look at NVIDIA, trading at 14 times next year's fully taxed gap earnings. This is no bubble like it was in 2000."

He put a striking number on how concentrated the rally is: "Semiconductors are 70% of the NASDAQ's return." And he flagged that hyperscaler capex is flowing almost dollar-for-dollar into chipmakers' cash flow: "Dell up 5x, [SK] Hynix up 9x in just 18 months." (All-In with Chamath, Jason, Sacks & Friedberg)

The catch, in Gerstner's own framing, is whether the customers of all that compute can actually generate the revenue to pay for it. He estimates the top AI labs are at roughly a $100 billion run-rate today and says the industry needs to climb from about "$200 [billion] to $450 to $800 or a trillion dollars" of "offtake" revenue over the next few years "just to keep up." He also thinks the widely-cited 43 gigawatts of new compute forecast for next year (from Dylan Patel's SemiAnalysis) is "too aggressive," and that the real number is "closer to 25 gigawatts." And he was explicit on rates: "I think it's now over 90% chance that we're going to have rate hikes tomorrow," a warning that the "hurdle rate for that money is going up."

The demand side stayed strong in the corporate commentary. On The KE Report (Sep 18), markets veteran Joel Elconin relayed that NVIDIA "came out this morning and said we're going to sell the double amount of chips in the following year. So there seems to be no slowdown," adding: "If NVIDIA can really double their chip sales in production next year, NVIDIA is cheap." (The KE Report)

The clearest read on the capex curve came from Nicholas and Kasey Rossolillo of Chip Stock Investor (Sep 14), reacting to Oracle's results. Oracle's quarterly capex hit a record of roughly $28.5 billion (versus about $8.5 billion a year earlier), driven by its OpenAI relationship. Their base case:

"We'll continue to have really high capital expenditures from the hyperscalers, but we'll at the very least see a slowdown in the growth rate of that CapEx in 2027 and an even more significant slowdown in 2028." (Chip Stock Investor Podcast)

The contrarian, cautious view came from Joe Albano of TechCache on Investing Experts (Sep 13). His worry is cash flow, not demand:

"We're reaching a point where CapEx is substantially outpacing the growth of the AI… we're now dipping into negative free cash flow for the first time in some of these companies' history."

He was especially pointed on Broadcom's model of backstopping customer financing with its own custom chips (XPUs). Because those chips are built for one customer's workload, he questioned whether they're really fungible collateral: "So I think it's a house of cards in the financing. And as long as nobody blows on it, it's going to be fine. But as soon as something rattles it or blows on it, it's going to happen." He also flagged that Broadcom's AI revenue is "almost doubling" its overall revenue growth, but with "compressing margins" because "memories cost more, chips are costing more… we're in a supply constraint environment." (Investing Experts)

Who pays for all this? On Bloomberg Talks (Sep 16), Apollo president Jim Zelter made the financing scale concrete. He noted Apollo is behind "the Intel financing, the Broadcom financing, the NVIDIA financing," and made a key point for chip investors: "the gross margin is highest away from the models", meaning suppliers like Broadcom "are doing very, very well… notwithstanding who's the winner of the LLM race." He also explained why even NVIDIA leans on debt markets: its top investors already hold "a $500 billion exposure" in the equity, so "the biggest companies in the globe are not going to have people provide that scale of debt." He expects the AI ecosystem to become "10% of the [investment-grade bond] market." (Bloomberg Talks)

A note on the "slow down AI" scare that kicked off the week: On Squawk on the Street (Sep 14), Byron Deeter of Bessemer Venture Partners (an Anthropic investor) argued the industry is coalescing around "pacing" and "inspectors and transparency" rather than a hard "pause," which he called "bad policy both geopolitically and economically." He noted the odds of an Anthropic IPO announcement before Nov 1 had fallen to 54% from about 75% a week earlier. (Squawk on the Street)

2. Memory: HBM, DRAM and NAND

Memory was the trade of the week, and The MoneyFlows Show (Sep 17) made the most detailed case that the sell-off was a gift. Hosts Jason Bodner and Lucas Downey argued the shortage is structural and worsening:

"I think it's much worse than Wall Street realizes. First of all, this call to slow down is ultimately a call for regulation, which benefits these massive AI companies. What it's not doing is cutting back the demand for AI."

The technical reason it's so tight: high-bandwidth memory (HBM) is made by stacking DRAM chips, and it's brutally capacity-hungry. In their words, HBM "has about 30% accounting for global DRAM and it only outputs about 13% of DRAM's capabilities… So you have a crunch within a crunch." They stacked up three vivid data points:

• "Elon Musk on the SpaceX call basically said memory output is rising 20% a year while demand is rising 200%."
• "Tim Cook described the memory situation as a 100-year flood."
• "NVIDIA CEO Jensen Huang at a recent convention went up to the SK Hynix booth and he signed a wafer that said, please make more."

Their top pick was Micron ($MU): last price "just under $940," a forward P/E "right at six," 2026 EPS "$73 and change," 2027 "$156," 2028 "$171," and a consensus price target of $1,575, implying 70% upside. They also flagged SanDisk ($SNDK) (P/E ~7, PEG 0.42, target $2,263, ~46% upside) for NAND storage, and the smaller Silicon Motion ($SIMO) for NAND controllers, which "absolutely destroyed earnings": Q2 revenue of $451 million versus a $403 million estimate, and a Q3 guide of $530 million versus the street's $430 million. (The MoneyFlows Show)

The dissenting note on memory came, again, from Joe Albano, who framed the same tightness as a cost problem for the chip buyers: "DRAM and NAND memory from Micron is now hitting peak headwinds… we're seeing compressing margins." Same shortage, opposite investment conclusion depending on whether you own the memory maker or the memory buyer. (Investing Experts)

3. Foundry and Manufacturing: The SK Hynix and Intel Story

The week's biggest manufacturing headline, covered on The Rundown (Sep 16): SK Hynix, the world's leading supplier of HBM that sits next to NVIDIA's GPUs, is in early talks to make memory chips in the US for the first time. Two structures are reportedly on the table: SK Hynix leasing part of Intel's massive Ohio fab, or a joint venture between Intel, SK Hynix and a hyperscaler that wants to lock in memory supply.

Why it matters for both sides:

• For Intel, whose foundry ambitions have struggled and whose Ohio complex has slipped to 2030, landing SK Hynix would be "a major customer" to help fill capacity.
• For SK Hynix, it's a way to defuse the tariff threat: Commerce Secretary Howard Lutnick "has threatened up to a 100% tariff on Korean chip makers who don't build in the U.S."

The host also captured just how tight memory has gotten with one line: "that's why the price of RAM costs more than rent these days." Both Intel and SK Hynix rose about 3% on the report, though "these talks are still early. Nothing has been finalized." (The Rundown)

4. Analog, Auto and Industrial Semis: ON Semiconductor's Investor Day

ON Semiconductor ($ON) held its investor day, covered live on Closing Bell (Sep 16), and the stock still fell 9% on the day (down 40% over three months). CEO Hassan El-Khoury pushed back on slowdown fears directly:

"I don't think a slowdown is anywhere on the cards, because everybody is doing the build-out… we're looking at [it], and we anchored on a 10% on top of whatever market you want to pick."

Key numbers from the day: a core addressable market of $213 billion, rising to $243 billion once the pending Synaptics acquisition closes (expected mid-2027); a quadrupled forecast for the AI data-center chip market to nearly $48 billion by 2030; and a gross-margin target of 53%, which El-Khoury called "a milestone, not a destination." He also defended ON's auto franchise after a quarter where rival STMicroelectronics grew EVs and industrial faster, pointing to "over 9% CAGR year on year over five years" in automotive content. (Closing Bell)

5. China, Export Controls and Tariffs

Two threads ran through the week. The first is the tariff stick aimed at Korean memory makers (the 100% threat behind the SK Hynix and Intel talks, above). The second is a broader wait-and-see on China tariffs: reporting this week indicated the US is holding off on new capacity-related tariff announcements until after a planned Xi–Trump summit, keeping chip-relevant trade policy in limbo. The unresolved question for the group is whether Washington uses tariffs to force more memory and logic capacity onshore, or leaves the current export-control regime in place.

6. China's Chip Self-Sufficiency: Huawei's Logic Folding

On the Telecoms.com Podcast (Sep 11–14 window), the panel dug into Huawei's new Kirin 9050 Pro chip, shipped in a new foldable phone and, by inference, made domestically in China. The significance is the workaround: because US pressure keeps China from buying ASML's EUV lithography machines (the Dutch company has "a monopoly on the piece of kit… you need" for the most advanced nodes), Huawei is leaning on a technique it calls "logic folding" (or "time scaling"), stacking chip components vertically "like a cityscape of skyscrapers" rather than shrinking them geometrically.

The panel was skeptical it's a true breakthrough: one cited analyst view was that "it's not really that innovative. It's more like a packaging solution to the problems they face," and that Intel and IBM have discussed similar stacking concepts. Still, the framing from China is that "US attempts to suppress Chinese technological development… is now being undermined." (Telecoms.com Podcast)

7. The Bubble and Cyclicality Debate

This was the real fight of the week: is the AI/chip build-out a durable super-cycle or a bubble? Gerstner (Section 1) is firmly in the "no bubble" camp. The counterweight came from two heavyweights.

On FT Tech Tonic (Sep 16), former top-rated Goldman analyst Richard Kramer (Arete Research) argued the investment is real but the valuations are not:

"There is right now a degree of consensual hallucination of what I call an extrapolationist fantasy… equity markets are deeply unmoored from traditional intrinsic valuation right now… It's all vibes."

His most useful line for chip investors: the trillion-dollar hyperscaler capex wave "is going to create a huge value transfer from the internet industry to the semiconductor industry." He put big tech's spending at roughly $1 trillion of capex and ~$380 billion of R&D next year. Fellow panelist Ophelia Brown (Blossom Capital) offered the insider counterpoint: "no participant there [in San Francisco] is thinking of this as a bubble." (FT Tech Tonic)

On BiggerPockets Money (Sep 15), NYU valuation professor Aswath Damodaran reframed the question by separating the "architecture" builders (NVIDIA, TSMC, the power and data-center suppliers) from the companies hoping to sell AI products:

"The actual amount being invested in the architecture… is in excess of $2 to $2.5 trillion. This is the largest buildup in business history for any new business… we've also built the largest factory in history without a sense of what that factory will produce… It's the largest leap into the unknown."

His key point for chip holders: the architecture suppliers have "already made money," and "NVIDIA has already made its money." The open question is whether the buyers of that compute can generate enough revenue to justify the spend. (BiggerPockets Money)

A cost-innovation wrinkle worth watching: on Squawk on the Street (Sep 14), SambaNova CEO Rodrigo Liang argued his company's chip for "agentic" AI can cut inference cost roughly 10x through lower power draw, letting it run in existing air-cooled data centers rather than requiring new liquid-cooled buildouts, with "cost per token" as the metric that matters. If cheaper inference chips gain share, it complicates the straight-line "more gigawatts forever" narrative. (Squawk on the Street)

8. M&A and Deal Chatter

• NVIDIA / Hugging Face (~$13 billion): referenced across several shows this week, mostly as background. The framing was that NVIDIA is buying the open-source model hub to support open-weight AI and, in some tellings, to "acquihire" talent, not a classic semiconductor consolidation deal.
• ON Semiconductor / Synaptics: the pending deal (adding ~$30 billion to ON's addressable market) remains on track to close mid-2027, per CEO Hassan El-Khoury on Closing Bell.
• SK Hynix / Intel: the potential Ohio-fab lease or joint venture (Section 3) is the week's most consequential structural chatter, even if it's early.

Earnings and Event Reactions

What I'm Watching Next Week

• Micron's earnings (expected around Sep 30). After a week of podcasts calling memory the tightest, most-underpriced link in AI, MU's print and its HBM/DRAM pricing commentary will be the direct test of that thesis.
• The Fed aftermath. With rates just raised into a build-out financed heavily by debt, watch whether the "higher for longer" backdrop starts to bite on data-center financing costs, a risk Gerstner, Zelter and Elconin all flagged.
• Any hard detail on SK Hynix and Intel. A confirmed lease or joint venture would be a genuine catalyst for Intel's foundry story and a signal on how the 100% Korean-chip tariff threat gets resolved.
• The US–China summit and tariff decisions. The paused capacity tariffs and the Korean-memory tariff threat both hinge on what comes out of Xi–Trump talks.
• The Anthropic IPO odds and the "pacing vs. pause" debate. IPO odds fell to ~54% this week; a delay or a real regulatory brake would hit sentiment across the AI-chip complex.
• Whether cheaper-inference challengers (SambaNova, and custom ASICs broadly) gain traction, the one narrative that could soften the "more gigawatts forever" demand story that props up the memory and GPU trades.