Newsletter · 2026-09-20 · Ashutosh Agarwal

AI Slowdown Scare Hits Chip Stocks as Bulls Call It a Speed Bump - Weekly Semis & AI Infrastructure Podcast Recap - Week of September 20, 2026

The Weekly Semis & AI Infrastructure Podcast Recap for the week of September 14 to 20, 2026: a synthesis of investor and operator podcasts on the AI slowdown scare that knocked chip stocks roughly 25% off their July highs, Jensen Huang calling the fears a hoax, a severe memory shortage, still-rising hyperscaler capex, and electricity emerging as the real bottleneck.

Weekly Semis & AI Infrastructure Podcast Recap

Week of September 20, 2026: AI Slowdown Scare Hits Chip Stocks as Bulls Call It a Speed Bump

Top of Mind This Week

The whole week bent around one scare and the loud pushback to it. Over the prior weekend, the heads of the biggest AI labs, Anthropic's Dario Amodei, OpenAI's Sam Altman and Elon Musk, publicly called for the industry to slow the pace of AI development. Chip stocks were hit hard, and the semiconductor group is now down roughly 25% from its July highs even though it is still up more than 60% for the year. The rest of the week was a parade of investors, chip executives and even President Trump arguing the sell-off was overdone. Jensen Huang of NVIDIA called the fears a "hoax," Microsoft's Satya Nadella and Meta's Mark Zuckerberg rejected the slowdown idea, and long-time tech investor Dan Niles summed up the bull camp in a single phrase: "This isn't the top, it's a speed bump." Underneath the noise, the hard facts stayed strikingly bullish: memory chips are in severe shortage, hyperscaler spending plans keep rising, and the real bottleneck everyone kept coming back to was electricity, not demand.

Dominant Themes

1. The "AI slowdown" scare and the counterattack. The single loudest story of the week. The lab founders' call to "pace" development spooked markets, but almost every investor podcast read it cynically. On The MoneyFlows Show, Jason Bodner argued the call to slow down "is ultimately a call for regulation, which benefits these massive AI companies" and does nothing to cut actual demand for chips. On The Six Five, analysts noted David Sacks went as far as saying an Anthropic safety report was serious enough that "we have to stop the IPO," a reminder that Anthropic and OpenAI are both weeks away from going public, giving them a reason to shape the narrative.

2. Memory is the tightest link in the chain. This was the most consistently bullish, number-heavy theme. On The MoneyFlows Show, the hosts cited Elon Musk saying on the SpaceX call that memory output is rising about 20% a year while demand is rising 200%, and Apple's Tim Cook calling the memory situation a "100-year flood." They explained why high-bandwidth memory (HBM, the stacked memory glued onto AI chips) makes it worse: making HBM eats about 30% of global DRAM capacity but yields only about 13% of normal output, "a crunch within a crunch." The vivid image of the week: Jensen Huang reportedly walked up to SK Hynix's booth at a conference and signed a memory wafer that read "please make more."

3. Hyperscaler spending keeps going up, not down. Despite the slowdown chatter, every capex estimate cited this week was enormous and rising. On FT Tech Tonic, Richard Kramer of Arete Research pegged big-tech capital spending at $700 billion this year rising to a trillion next year, plus another $380 billion of R&D, and said this investment is "single-handedly propping up the U.S. economy." Multiple shows put 2026 spending near $200 billion each at Alphabet and Amazon, roughly $150 billion at Microsoft, and $100-150 billion at Meta.

4. Electricity is the real bottleneck. More episodes covered power and the grid than any single chip topic. The recurring point: the constraint on AI is no longer chips or money, it is getting enough electricity fast enough. Estimates flew around: one energy executive said hyperscalers are adding 15-20 gigawatts of data-center power a year at $1-5 billion per gigawatt, while a former U.S. energy official said American projects face 7-10 year power delays even in friendly states like Texas. Not everyone agreed it is a crisis (see the debates below).

5. The bubble question: circular money and rising debt. A quieter but sharper theme. Valuation expert Aswath Damodaran laid out the math: the AI industry has built a roughly $2 trillion "factory" but is only selling about $250 billion a year of actual AI products and services, and would eventually need $8-10 trillion of revenue to justify the spend. Several hosts worried the money is increasingly circular (companies investing in each other and then renting compute back) and, more dangerously, increasingly funded by debt rather than stock.

6. Custom chips and cheaper inference chip away at NVIDIA's edge, at the margins. NVIDIA still dominates, but this week brought fresh evidence of competition: Qualcomm landed a landmark data-center deal with Amazon, and challengers like SambaNova, Cerebras and diffusion-model startup Inception all pitched being cheaper and faster for "inference" (running AI models, as opposed to training them).

Key Debates

• Is the AI slowdown real, or a self-serving story? Slowdown skeptics won the week. Dan Niles on the RiskReversal Pod argued the labs cannot actually slow down, "if you slow down and Google keeps going, you're screwed," but conceded a company about to IPO might ease off for a quarter or two "just to make your cash flow look bigger." Jason Bodner (The MoneyFlows Show) called the whole thing a regulatory play. On the other side, the fact that the founders themselves raised the alarm was enough to knock chips down hard and keep a bearish minority nervous.

• Bubble or the best setup in a generation? The clearest clash of the week. Brad Gerstner of Altimeter, on All-In, made the full bull case: "This is not about multiple expansion. This is an earnings-driven market expansion," noting NVIDIA trades at "14 times next year's fully taxed gap earnings. This is no bubble like it was in 2000," and that semiconductors have driven 70% of the Nasdaq's return this year. Directly opposite, Aswath Damodaran on BiggerPockets Money warned of a "big market delusion" where overconfident giants all overinvest into the same market: "there might be one winner and three losers, and collectively, you might end up losing money." His caution was measured, though: using his "3P test," he said $10 trillion of AI revenue is "possible," is "pushing the limits of what's plausible," and right now he'd "attach a low probability to it."

• Is NVIDIA propping up its own customers? Damodaran drew a sharp line between NVIDIA and everyone else: "NVIDIA wants this circus to continue... like financing the purchase of their own chips by companies that can't afford to pay for the chips themselves." The bulls counter that the demand is simply real: Gerstner showed that hyperscaler capex is flowing almost dollar-for-dollar into semiconductor company cash flow.

• Is the power shortage a true ceiling, or overstated? Genuine disagreement here. Vistra's CEO, on Energy Gang, argued the panic is overblown: near-term AI demand is more like 17-30 gigawatts by 2030 (not the 700+ some throw around), and U.S. grids run under 50% utilization for 99% of hours, so the real problem is slow interconnection queues, not a lack of generation. Against that, guests on POWERS, Clean Power Hour and World Nuclear News described a hard wall: 230 gigawatts of data-center grid applications over five years versus only about 93 gigawatts of new power plants planned (a 2.5x gap), and a single Texas site needing power equal to "7 nuclear power stations." The workaround everyone described is going off-grid with on-site gas turbines, exactly what xAI did to get its Colossus data center running in months.

• Debt versus equity, how dangerous is the financing? On Simply Put, Robin Wigglesworth warned the AI build is shifting dangerously from stock-funded to debt-funded, hyperscalers issuing about $500 billion of bonds this year plus roughly $1.5 trillion of off-balance-sheet lease commitments, and drew a parallel to the debt-fueled 1873 railway crash, noting debt-driven bubbles are far more destructive than equity-driven ones. Dan Niles offered the more relaxed version: circular financing "only going to matter when things start to slow," pointing out Cisco survived multiple 30%+ drawdowns in the 1990s.

• Does spending peak soon? On the Chip Stock Investor Podcast, Nicholas Rossolillo argued hyperscaler capex growth slows in 2027 and more sharply in 2028, calling fiscal 2027 the "likely peak," citing Oracle's jump to $28.5 billion of quarterly capex from $8.5 billion a year earlier. Most other guests saw the build running well into the 2030s.

Stocks Discussed With Bull/Bear Angle