Newsletter · 2026-09-21 · Ashutosh Agarwal

Chip Stocks Rebound After the Fed Hike as Credit Markets Take Center Stage - AI Accelerators - Week of September 21, 2026

AI Accelerators podcast synthesis for the week of September 21, 2026. The Fed hiked for the first time in three years and all ten tracked AI hardware names rebounded, led by Coherent, Astera Labs and Credo, while Brad Gerstner made the no-bubble case, Dan Niles flagged a pullback in top-user AI spend, credit investors became the key voices, and SambaNova and the Qualcomm-Amazon deal advanced the custom-silicon story.

AI Accelerators

Week of September 21, 2026: Chip Stocks Rebound After the Fed Hike as Credit Markets Take Center Stage

GPUs, Custom Silicon & Optics: a twice-weekly read on the hardware powering AI. Issue 026. Monday, September 21, 2026.

Three days before this issue, the whole AI-chip complex looked like it was cracking. The Federal Reserve was about to raise interest rates for the first time in three years, the 10-year Treasury yield had punched through 5% for the first time since 2007, and the founders of Anthropic and OpenAI were publicly asking the world to slow AI down. Our last issue led with that danger: every one of the ten names we track had fallen, and the most expensive, most cash-hungry ones fell hardest.

Then the Fed hiked, and the complex went straight back up. Every one of the ten names rose over the following days. And the exact names that had been crushed the week before led the rebound. One podcast summed up the whole week in its title: "AI slowdown lasted a day."

That is the story of Issue 026. Not a new direction, but a violent round-trip that tells you how this trade now works: it is being priced day by day off a bond yield and a Fed chair's tone, not off any company's fundamentals. Below, what the podcasts actually said: who's an insider and who's an opinion, with the numbers and the exact words.

TL;DR

• The Fed hiked, and it's now confirmed. New Fed chair Kevin Warsh raised rates 25 basis points on September 16 to a 3.75%–4.00% range, the first hike in three years, and a unanimous vote. He signaled at least one more this year. (A "basis point" is one-hundredth of a percentage point.) This week's podcasts, recorded after the decision, confirm it.
• Warsh named the chip shortage as an inflation driver. In plain terms, the Fed chair is now pointing at AI hardware demand as one reason prices are rising, a striking link between this newsletter's subject and monetary policy.
• The complex round-tripped. All ten names we track rose September 14→18, led by the optics and custom-silicon names that had been the biggest losers the week before: Coherent +19%, Astera Labs +18%, Credo +17%.
• The bull case got its loudest voice. Investor Brad Gerstner laid out a detailed "no bubble" argument on All-In, with Nvidia at "14 times next year's earnings" and semiconductors making up "70% of the Nasdaq's return."
• But real cracks showed up in the data. Investor Dan Niles flagged that the heaviest AI users actually cut their spending 10% from July to August, plus a Texas moratorium on new data centers.
• Credit is the new tell. A parade of bond investors (Apollo, Wellington, PGIM, Vanguard) now say the AI trade lives or dies in the debt market, and that hyperscalers will "borrow at any rate."
• Anthropic delayed its ~$2 trillion IPO for "safety" reasons, despite roughly $100 billion of annualized revenue.

1. The biggest development: the Fed hiked, and the chips ripped anyway

Start with what changed since Friday, because it's the whole point of this issue.

Our last issue (published Thursday, September 17) had to flag the September 16 Fed decision as unverified, since the podcasts recorded before it could only guess. This week's podcasts, recorded afterward, confirm it clearly: the Fed raised rates by 25 basis points to a 3.75%–4.00% range on September 16, its first hike in three years, in a unanimous vote, with a signal that at least one more is coming this year.

The new Fed chair, Kevin Warsh, was blunt about why. On Bloomberg's Instant Reaction: Fed Chair Kevin Warsh on Fed Policy (September 16), the anchors replayed his words:

"I would be hard pressed to describe broad financial conditions as restrictive… so we removed a dose of accommodation."

"Today's action starts to show that we're serious about this."

Translation: the Fed doesn't think money is tight enough yet, so expect more. The bond market heard it and yields jumped: the 10-year Treasury yield hit 5.04%, its highest since 2007, per FT commentator Robert Armstrong on Prof G Markets (September 17). (The 10-year yield is the interest rate the U.S. government pays to borrow for a decade; it's the benchmark that sets the cost of almost all long-term borrowing, including the debt funding AI data centers.)

Here's the part that matters directly for this newsletter. In the same Bloomberg segment, an analyst broke down the three things Warsh blamed for stubborn inflation: geopolitics and oil, tariffs, and, most importantly, "the chip shortage, which I would argue is the most important one to keep an eye on, because that's something that hasn't gone away." Read that again: the Fed chair is effectively naming AI hardware demand as an inflation driver. The buildout this newsletter covers has become large enough to move monetary policy.

And yet the complex rallied. Stocks fell 0.4% on decision day, then rebounded the next session and kept going. Why did chip stocks go up into a rate hike? The cleanest answer came from an operator-analyst on the Futurum Equities podcast (September 18), whose whole episode was titled "AI slowdown lasted a day":

"This rate hike was due to a good problem. Growth is healthy and employment is low… The data good or bad gave the confidence for a unanimous vote to raise rates, thinking that that will not cause damage. In a lot of ways, this was a bullish hike."

Why it matters: the market has decided a rate hike driven by a strong economy is fine for AI, because, as nearly every credit voice below argues, the buildout doesn't care about the cost of money. That's the bull's whole foundation. It's also exactly what makes the trade fragile: it now trades tick-for-tick off the 10-year yield and the Fed's tone, which is why one Fed meeting produced a down-week and an up-week back to back.

2. The prices: a clean V-shape, led by last week's losers

Here is the verified move, using end-of-day prices from FactSet. The window runs from the September 14 close to the September 18 close (the last completed session before this issue). All ten names rose, the mirror image of last week, when all ten fell.

The pattern is the tell. Last week's worst performers (the optics and custom-silicon "tail") became this week's best. That is not a fundamentals move; it's a de-risk-then-re-risk whipsaw. When the macro fear (5% yields, a hawkish Fed, "slow down AI" headlines) receded, the highest-beta, highest-multiple names snapped back the most.

For context, here's where the group sits now, and how far each still trades below its 52-week high. Note the enormous spread in valuation, from Micron at ~23x trailing earnings to Lumentum at ~155x:

(P/E = price-to-earnings, a rough gauge of how expensive a stock is versus its profits. "Consensus price-target upside" is how much higher, on average, Wall Street analysts think the stock can go.)

Why it matters: even after this bounce, five of the ten names are still 20%–43% below their highs. The rebound was sharp, but the group has not repaired the damage of the summer. The stocks that "won" the week are the ones the market is least sure about over the long run.

3. The bull case gets its clearest voice: Brad Gerstner

The most complete "this is not a bubble" argument of the week came from Brad Gerstner, founder of the investment firm Altimeter, on All-In (September 17), in an episode literally titled "No AI Bubble, Semis Eat the Nasdaq." Gerstner is a pundit here (a high-profile investor talking his book), but his framework is worth understanding because it's the one bulls are leaning on.

His core point: this rally is powered by real profits, not by investors simply paying higher prices.

"This is no bubble like it was in 2000. Nasdaq, S&P, SOX, Nvidia, all trading well below their average multiples… Earnings are up 26%… but the multiple on the Nasdaq and the S&P is actually down."

He put a specific number on Nvidia: "trading at 14 times next year's fully taxed gap earnings"; in other words, cheap for a company growing this fast. (Our FactSet snapshot shows Nvidia at ~38x trailing earnings; the gap between the two numbers is exactly the point: bulls value it on expected future profits, not past ones.)

Two lines capture where he thinks the money is flowing:

"Semiconductors are 70% of the Nasdaq's return… The makers of the tokens are making the money, and the buyers of the tokens are basically going along for the ride."

("Tokens" are the units of AI output; every word an AI writes is measured in tokens. His point: the companies selling the shovels, the chipmakers, are capturing the profits.) He showed that hyperscaler capital spending is almost dollar-for-dollar the free cash flow of the chip companies: the spending of the buyers becomes the earnings of the sellers.

His single most-watched metric is the revenue of the top AI labs:

"The collective run rate… of the top three labs, Anthropic, OpenAI, and SpaceX, is about $100 billion… I think they need to collectively get to at least $180 billion by the end of the year… Is Anthropic's monthly revenue, is OpenAI's monthly revenue going to be $4 billion or $8 billion? This is the single most important data point in the market today."

On capacity, he cited SemiAnalysis: about 19 gigawatts of compute added in 2026, and a forecast of 43 gigawatts next year ("as much as the total compute we have in the United States this year"), with over half of all U.S. compute controlled by just two labs by 2028. (A gigawatt is a billion watts; it's become the industry's shorthand for the size of a data-center buildout, because power, not chips, is often the limiting factor.) And he argued the market it all serves is enormous: "You only have to get to about 4% of that TAM, or $1.2 trillion, in order to pay for the CapEx. I don't think it's a TAM issue." (TAM = total addressable market, the whole revenue opportunity.)

His three risks, in his own order: regulation, power, and interest rates.

Why it matters: Gerstner's argument is the intellectual backbone of the rebound. If he's right that this is earnings-driven and cheap on forward numbers, the dips are buying opportunities. The bears below aren't arguing the earnings are fake today; they're arguing about the second derivative: whether the growth rate is quietly rolling over.

4. The debate, steel-manned: "speed bump," not "top," but the cracks are real

The most useful bear-ish voice wasn't a doomer. It was investor Dan Niles on RiskReversal (September 18), whose episode was titled "This Isn't the Top, It's a Speed Bump." He's still constructive:

"I'm in the speed bump category… I feel like we still have at least a year or two to run."

But he then laid out the most concrete evidence yet that demand may be softening at the edges, and this is the part bulls should not wave away:

• The heaviest users are pulling back. "The top 1% of users, they actually saw a 10% reduction in spend from July to August, if you look at the Ramp Index", the first monthly drop since ChatGPT launched. (The Ramp Index tracks anonymized corporate spending; the top 1% are the cutting-edge tech firms, i.e., the leading indicator.)
• Prices are collapsing. Cost per token is "down 50% since the end of May," though total tokens quadrupled, and the volume is shifting to cheap open-weight models (mostly Chinese). OpenAI cut its prices and "profitability actually got worse."
• Politics is turning. Texas ("a deeply red state… the second biggest data center [state]") put a moratorium on connecting new data centers to the grid. Niles's sharp point: bulls can't have it both ways. "If you're telling me power and land is what is dictating how much you can spend, then how is bipartisan pushback on data-center spend… you're going to have less spent. You can't have it both ways."

On the "slow down AI" chorus from Altman, Dario Amodei, and Musk, Niles was cynical in a useful way: three motives can all be true at once:

"You can be concerned about what this technology can do. And at the same time, you can be wanting to have a government-sanctioned duopoly… And then finally, you want to get a great valuation, so you want to cut back on your spending by saying, hey, we're doing this for safety reasons."

His bottom line captures the risk/reward better than any price target: "At this point, to me, the downside risk, if something occurs, is much greater than the upside risk."

The credit-market echo. Bank of America's Savita Subramanian (Bloomberg Surveillance, September 17) gave the bull-market-with-a-catch version. Her firm's whole thesis (that hyperscalers would "de-rate" as they got more capital-hungry) has played out: "they're getting more levered… less asset light, and they spend less on R&D and buybacks and more on physical CapEx, which is generally a reason to expect multiple compression. And we got that." But she waved off the 5% panic ("we've never been that worried about 5% as some kind of red line") because big companies locked in cheap long-term debt (today ~80% of S&P 500 debt is fixed-rate, versus half floating back in 2007). Her warning is about time and cost, not collapse: the build "might just take longer and cost a lot more."

The hard bear. For the genuinely negative view, Danielle DiMartino Booth of QI Research (The Julia La Roche Show, September 17) went at the growth rate: "on a second-derivative basis, the growth rate of investment in AI absolutely collapse[d]." If the "robust capital investment" Warsh praised has already peaked, she warned, "we could see the other side of the AI boom much more quickly than anybody's anticipating."

Why it matters: the debate has moved. A month ago it was "is there enough demand?" Now the bulls (Gerstner) and the bears (Niles, DiMartino Booth) agree demand is huge today; they disagree on whether the growth rate is quietly rolling over. The Ramp Index and the 50% price collapse are the first real data points on the bear side. Watch them.

5. The new tell: this is now a credit trade

If there was one theme running through more podcasts than any other this week, it wasn't a chip. It was the bond market. The people financing the buildout have become the most important voices in it.

Steve Eisman (of The Big Short fame) tied the two together directly on The Real Eisman Playbook (September 18). His argument: roughly $500 billion of AI-related debt raised this year is "crowding out" buyers of U.S. Treasuries (investors would rather buy AI bonds), and that's part of why the 10-year is above 5%. "For now, 5% on the 10-year does seem to be the demarcation line." He also delivered the sharpest read on the "slow down" theater, arguing Amodei and Altman are faking it because "the entire future of Anthropic and OpenAI depends on there not being any slowdown… OpenAI alone represents $300 billion of Oracle's $600 billion-plus backlog." His theory: they're inviting regulation to build a moat before a price war. He quoted Musk's own jab: "some crazy 4D chess to say… a 10% chance of annihilating humanity. But by the way, how much allocation would you like in our IPO?"

Harley Bassman, the inventor of the MOVE index (the bond market's fear gauge), gave the most memorable line on MacroVoices (September 17):

"These guys are going to borrow money at any rate. Six, eight, 10, 12. It doesn't matter. They're in an existential race… The Fed can't stop this."

But, and this matters for how the trade eventually ends, he thinks the bonds are safe even if the stocks aren't: "Meta, Google, Amazon, Microsoft, Oracle… they have plenty of cash flow… I think the bonds are solid. It doesn't mean their stock's not going to go down." His historical rhyme was Global Crossing, which laid fiber-optic cable in the 1990s expecting to charge ten times what it eventually got: most of the money gets built out, only a couple of players win.

Two more credit voices sharpened the picture:

• Mike Medeiros, macro strategist at Wellington (The Wall Street Skinny, September 18), is the one to take seriously on downside: he sees the S&P "vulnerable to a correction… of around 10%" into year-end. His key evidence that rates aren't actually restrictive yet: new investment-grade bond deals are "super well received… corporate spreads at the tights." And his historical warning: past capex booms (railroads, the 1920s, the internet) "peaked around year four. Next year will be year four for the AI CapEx cycle."
• PGIM's co-CIO (Credit Exchange with Lisa Lee, September 18) sized the spending: hyperscaler capex ~$400 billion (2025) → $763 billion (2026) → $1.2 trillion next year, and called the near-term numbers "already baked." He framed the local backlash as "fracking 2.0" (everyone wants the cheap energy, nobody wants it in their backyard) and, tellingly, said a slowdown would be good for bond investors even as it hurts growth.

Why it matters: the center of gravity of this trade has moved from the income statement to the balance sheet. The number that decides where these ten stocks go next is no longer a chip company's revenue; it's the 10-year yield and the spread on hyperscaler debt. If bonds stay "money good" and spreads stay tight, the bulls win. The first sign of trouble won't be a bad earnings report; it'll be a widening credit spread.

6. Custom silicon: a fresh challenger and a concrete deal

Underneath the macro noise, the actual hardware story kept moving, and this week it was operators, not pundits.

SambaNova's Anton McGonnell (a senior executive, an insider) laid out the clearest challenger pitch of the week on Super Data Science (September 18). His company builds an alternative to Nvidia's GPU that it calls an "RDU" (reconfigurable dataflow unit). The pitch is about money, not benchmarks:

"If a customer buys a cluster of SN50s and runs inference on these systems… [they] will get their money back within six months."

("Inference" is the running of a trained AI model to answer real queries, as opposed to "training," which is teaching it. Most day-to-day AI compute is inference.) His technical claim: Nvidia's newest systems wire together 72 chips in a rack, and "the overhead of the communication between the chips really exasperates their inefficiencies," while SambaNova's design "scale[s] linearly… twice as fast on two chips… four times as fast on four chips." The commercial kicker: SambaNova's racks are lightweight and air-cooled, so they can drop into "existing brownfield data centers today where other providers can't" (old telephone exchanges and CPU data centers), sidestepping the multi-year wait for new, power-hungry builds. That is a direct attack on the power bottleneck everyone else is stuck behind.

The week's biggest custom-silicon deal got operator-grade detail on The Six Five (September 14): the Qualcomm–Amazon partnership. Analysts Patrick Moorhead and Daniel Newman put numbers on it: roughly $60 billion over a decade, about $9 billion of committed near-term orders ($5 billion in Qualcomm's fiscal year starting around October 1, $15 billion by 2029), plus Amazon taking the right to buy 25 million Qualcomm shares. Moorhead's read: "the market is underplaying this" because Wall Street won't credit a deal it can't see in the next 18 months. Crucially, he argued it doesn't come at anyone's expense: "it does not need to be at the cost of Broadcom… Marvell, or Nvidia or AMD… the demand is so outlandishly large."

And the startup landscape got a survey on TechSurge (September 16): Groq (sold to Nvidia) and Cerebras (going public) as early ASIC efforts that "neither one really won a hyperscaler"; OpenAI's own "clean-sheet" chip; and long-shots like Tensordyne (using log math instead of matrix multiplication) and Etched. The sober takeaway: challenging Nvidia takes "hundreds of millions" of dollars and a hyperscaler willing to commit.

Why it matters: the custom-silicon threat to Nvidia keeps getting more concrete: a shipping challenger with a six-month payback pitch, and a $60 billion Amazon–Qualcomm deal. But the operators themselves keep insisting the pie is growing fast enough for everyone. For the book: this supports owning the "arms dealers" (Broadcom, Marvell) alongside Nvidia, not instead of it, and it's a reminder that Astera Labs and Credo, the connectivity names that led this week's bounce, sit right in the middle of this fight.

7. Memory: the round-trip name into a live earnings test

Memory was the best-covered corner of the hardware world this week, and it's the cleanest way to play "AI needs more than GPUs." Micron rose 9.9% on the window and, at ~23x trailing earnings, is by far the cheapest name in the group, with the second-highest analyst upside (+55%).

The pitch, from technical analyst Charles Moon on Schwab Network (September 17):

"AI can't grow without memory… It's the bottleneck… even as the optics could fall back… memory is going to be something that's going to be in high demand."

The bulls are getting aggressive on price. The MoneyFlows Show (September 17) cited a $1,575 target on Micron at ~6x forward earnings; Buy Hold Rant (September 17) went further: "Micron deserves $1,800–$2,000." And Futurum (September 18) framed the whole trade as a valuation absurdity: "Dell should [not] be trading at 20 times earnings and Micron should be trading at six times earnings. One of those two facts has to be wrong," pitching a pair trade of long Micron and Nvidia against short Intel and Dell.

Meanwhile the SK Hynix-to-US-fabs story is still everywhere and still one story: roughly ten different AI-news podcasts carried the identical Reuters report that SK Hynix is in early talks with Intel to make high-bandwidth memory in the U.S. for the first time (possibly leasing Intel's Ohio complex or forming a joint venture) under the threat of a Commerce Department 100% tariff on Korean chipmakers that don't build stateside. It's a real story, but it's one story amplified ten times; treat it accordingly.

Why it matters: Micron reports earnings later this month, and it's the rare name that's both cheap and central to the AI thesis. It's the single best near-term test of whether the "memory is the bottleneck" trade is real or just a narrative. If the print confirms the pricing power these podcasts are describing, the cheapest name in the group has the most room to run.

8. Names and catalysts in play

• Anthropic delayed its ~$2 trillion IPO "for safety" (Elon Musk Podcast, September 20), despite roughly $100 billion of annualized revenue and ~30% market share (concentrated in high-income households: 35% adoption above $150k, 18% below $60k). Our last issue had this IPO as pending and roughly a coin-flip before November; now it's explicitly pushed back. The company is spending $1 billion each with Accenture and Faculty on "embedded evaluators" to red-team its models during training. Read-through: this feeds directly into Eisman's and Niles's cynicism about the "slow down" narrative, and a delayed mega-IPO removes a catalyst the whole complex was watching.
• Oracle keeps funding its buildout with stock, not bonds (Telltales, September 20): a September 11 print of ~$19 billion revenue (+30%), but $126 billion of net debt against a $561 billion enterprise value, roughly a fifth of the company already borrowed. Larry Ellison called off a ~$7.5 billion share sale, Oracle locked a 433-megawatt power deal with RWE, and Congress subpoenaed Ellison over a $27 billion VA contract. The host's line: "Not one of those answers came from the bond market." Read-through: Oracle is the credit-market canary. Locking up power years ahead is not what a company that's about to slow down does.
• "Nobody is pulling back." On Squawk on the Street (September 17): "we have yet to hear anything that indicates anybody pulling back in any way from their plans to add compute." Generac jumped 19% on a $2.4 billion Amazon backup-generator deal (with warrants for up to $340 million of stock), with the picks-and-shovels trade widening beyond chips into power.
• The scale keeps getting bigger. Alphabet is guiding to ~$200 billion of data-center capex in 2026 with ~$800 billion of off-balance-sheet commitments through 2030 (The Investor's Podcast, September 17 & 20). EQT Infrastructure's CEO (Dakota Live!, September 16) put hyperscaler spending at $4–5 trillion by 2030, adding 15–20 gigawatts a year at $1–5 billion per gigawatt. Tom Lee and Dan Ives (The Compound and Friends, September 18) cited a 13-to-1 order-to-production ratio and 1,200–1,500 new data-center facilities over 5–6 years.

What changed vs Issue 025

• The Fed flag is resolved. Last issue we could only say the Fed was "priced ~90% for a hike" and had to flag the outcome as unverified. It's now confirmed: a 25bp hike on September 16, unanimous, first in three years, with more signaled. Warsh explicitly named the chip shortage as an inflation driver, a new, direct link between this sector and rates.
• The tape flipped. Issue 025: all ten names down 9/11→9/16, tail (Broadcom, Astera, Micron, Coherent) leading down. Issue 026: all ten up 9/14→9/18, the same tail (Coherent, Astera, Credo) leading up. A full round-trip in a week.
• The debate advanced from "if" to "how fast." Last issue's bear metric was Ruchir Sharma's "5% tripwire." This issue, the bears (Niles, DiMartino Booth) have moved to hard data: the Ramp Index showing top-1% users cutting spend 10%, and a 50% collapse in the price per token. The bull (Gerstner) doesn't dispute the level of demand; the fight is now about the growth rate.
• Operators stayed in the mix: SambaNova's McGonnell, plus operator-grade detail on Qualcomm–Amazon and Oracle. And credit voices (Apollo carried over, now Wellington, PGIM, Vanguard, Bassman) have become the dominant lens.
• Anthropic's IPO went from "pending" to "delayed."

Bottom line

The AI-chip complex just proved it will round-trip a Fed decision in a single week. The bulls got their confirmation that a strong-economy rate hike doesn't break the buildout; the bears got their first hard data that the growth rate may be softening at the top. Both are right at once, which is why the tape is so violent. For a book, the message is consistent with last issue, only sharper: this is now a credit trade wearing a semiconductor costume. Watch the 10-year yield, watch hyperscaler credit spreads, and watch two specific prints: Micron's earnings (the cheap, central name) and Oracle's funding (the leveraged canary). The stock reports will tell you about last quarter; the bond market will tell you about next year.