Newsletter · · Ashutosh Agarwal

Treasury Yields Top 5% as Custom Chip and Memory Stocks Lead Declines - AI Accelerators - Week of September 17, 2026

AI Accelerators podcast synthesis for the week of September 17, 2026. The 10-year Treasury yield hit the 5% tripwire as the Fed went in priced for a hike, custom-silicon and memory names led the chip complex lower, the Amazon-Qualcomm warrant got a $4 billion price tag, and operators from Micron, SambaNova and Apollo spoke directly.

AI Accelerators

Week of September 17, 2026: Treasury Yields Top 5% as Custom Chip and Memory Stocks Lead Declines

GPUs, Custom Silicon & Optics: a twice-weekly read on the hardware powering AI. Issue 025. Thursday, September 17, 2026. Podcast window: Sept 10–17.

Two Mondays ago, in Issue 024, I handed you a single number to watch. Ruchir Sharma had gone on the Financial Times' podcast and said the whole AI build-out has one tripwire: the day the 10-year US Treasury yield climbs "decisively above 5%," the companies pouring a trillion dollars a year into chips and data centers start losing the fight for money to the US government itself. At the time the yield was around 4.8%. It felt like a line on a distant horizon.

This week the line got hit.

On Prof G Markets this Monday, host Ed Elson opened the show with the market vitals almost as an aside: "the yield on 10-year treasuries topped 5% for the first time in three years." (For anyone who tuned out in economics class: the 10-year Treasury yield is the interest rate the US government pays to borrow for a decade, and it quietly sets the floor under almost every other borrowing rate in the country, including the rates hyperscalers and neoclouds pay to finance their GPUs.) And the Federal Reserve, which had spent months cutting or holding, walked into its two-day meeting this week with the market betting roughly 90% odds it would raise rates: the first hike of a new cycle, straight into the most capital-hungry construction boom in corporate history.

The chips did exactly what you'd expect. Every single one of the ten names I track fell over the four sessions from Sept 11 to Sept 16. Not a crash, but an orderly, rates-driven bleed. But the pattern underneath tells the story: the stocks that got hit hardest were the custom-silicon and memory names that need the most outside money to keep growing, while the plain-vanilla GPU names held up best.

So this issue is about what happens to the build-out when the cost of money finally bites. The good news, if you're tired of pure doom: this was also the richest week of operator commentary in a while: an actual Micron executive on the memory shortage, the CEO of a company trying to undercut Nvidia, and the president of Apollo explaining who's writing the checks. Let's get into it.

TL;DR

  • The tripwire from last issue got hit. Multiple podcasts put the 10-year Treasury yield at or above 5% this week (the exact level Ruchir Sharma called the danger line in Issue 024), and the Fed went into its meeting priced ~90% for a rate hike. The whole accelerator complex fell.
  • The fade broadened. Two weeks ago networking (Arista) held while memory and optics led down. This week all ten names I track were red, led lower by the custom-silicon and memory tail: Astera Labs -7.6%, Broadcom -6.2%, Micron -5.0%, Coherent -5.1%. The GPU names (AMD -0.7%, Nvidia -2.0%) held up best.
  • The custom-silicon warrant got real. Last issue I flagged a Vaughan Nelson manager's claim that hyperscalers are getting custom chips "basically free" via stock warrants. This week Steve Eisman put concrete numbers on the Amazon–Qualcomm deal: a warrant for 25 million Qualcomm shares at $161.26, already worth over $4 billion.
  • The operator drought ended. A Micron SVP walked through the "memory wall" and the company's five-fab build-out; the CEO of Nvidia-challenger SambaNova claimed 10x cheaper inference with six-month paybacks; and Apollo's Jim Zelter (who's financed Intel, Broadcom and Nvidia) dropped the line of the week: "the gross margin is highest away from the models."
  • A new named bear on Nvidia. Seaport's Jay Goldberg reiterated an underperform: "they're sold out... for this year, probably for next year. And once you're sold out, what's the upside?"
  • Fresh memory story: SK Hynix is reportedly in early talks to make memory chips in the US for the first time, possibly inside an Intel factory, as Washington threatens 100% tariffs on Korean chipmakers who don't build stateside.

The Big Picture: What Happens When Money Finally Costs Something

For two years the AI hardware trade ran on a simple, unspoken assumption: that capital was effectively free and infinite. You could raise it, borrow it, or print equity for it, and there was always someone to buy. This week the podcasts spent most of their oxygen on the possibility that the assumption is breaking.

Start with the mechanics, because a strategist laid them out cleanly. On Bloomberg Surveillance's "The Fed, Rate Hikes, and AI" (Sept 16), BNY's Alicia Levine described a Fed that feels boxed in:

"A Fed hike today, maybe the beginning of maybe one or two or three hikes... The Fed needs to raise rates because the market's telling the Fed it needs to raise rates. And the rhetoric has been that if we don't, we lose credibility."

Then she connected it directly to our world in one sentence every accelerator investor should tape to their monitor:

"We cannot pretend that hikes are good for multiples or good for marginal areas of the market that require funding."

That is the entire thesis of this issue in one line. "Marginal areas of the market that require funding" is a polite way of describing exactly the stocks in this newsletter: the ones trading at 50, 100, even 150 times earnings whose whole story depends on borrowing or issuing stock to build the next data center. Levine's own framing had the 10-year around 5%, the 30-year at 5.3%, and oil at $105 a barrel, what she called a return to a "pre-GFC" world of higher rates and real, physical inflation. And she made the striking point that AI has become so large it's now "one-third to one-half of US GDP this year" in terms of its contribution, in part because the chips being imported from Korea and Taiwan are so enormous they distort the national accounts.

Here's the part that keeps the bulls in the game, though: Levine still said keep the holdings. "You can't start picking sectors here. I think it's too volatile... what we know is that the investment will continue." Even the person warning you about the funding math isn't telling you to sell.

Who is actually writing the checks? The most useful voice of the week on that question was Jim Zelter, president of Apollo Global Management, on Bloomberg Talks (Sept 16). Apollo is one of the largest private-credit lenders on earth, and Zelter confirmed they are right in the middle of this, financing "the Intel financing, the Broadcom financing, the NVIDIA financing." His framing of the scale problem was genuinely clarifying. The build-out is so big, he argued, that no single pool of money can fund it:

"It's going to take any and all precincts, equity, public equity, private equity, private capital, investment grade debt, and everything else in between... I call it calling all precincts."

("Investment grade," or IG, just means debt from financially solid companies, the safest tier of corporate borrowing.) He said the AI ecosystem is on track to become 10% of the entire investment-grade bond market, a market that took decades to build, now being reshaped by an industry that barely existed four years ago. And he explained why even Nvidia, the most valuable company on the planet, can't simply borrow its way through:

"If you look at the largest company on the planet, NVIDIA, their four or five top-10 investors have upwards of anywhere from 2%, 3% of the equity to almost 9% of the equity. That's a $500 billion exposure. The biggest companies in the globe are not going to have people provide that scale of debt."

The punchline Zelter delivered (and this is the single most important sentence for anyone deciding which accelerator names to own) was about where the profit actually sits:

"It's interesting to see today the gross margin is highest away from the models."

Translation: the AI labs (OpenAI, Anthropic) burn cash and may or may not win. The place where the fat, durable profit margins actually live is upstream: in the chips, the memory, the networking, the picks and shovels. Whoever wins the model race, the toll is collected by the hardware. That's not a new idea, but hearing it from the man financing the whole thing carries weight.

Steve Eisman's audience made the same point more colorfully. On The Real Eisman Playbook's "AI Terminator Fears Grow & Rates Breach 4.9%" (Sept 11), Eisman read a subscriber note that framed the entire supply chain as a knife fight:

"In this knife fight, you want to be the one selling the knives. Upstream, semiconductor suppliers such as ASML, KLA, LAM Research, and Applied Materials, together with memory suppliers such as Micron and SK Hynix... benefit regardless of whether the winner is NVIDIA, a hyperscaler, an AI lab, or some architecture we haven't seen yet."

And the counter-argument, so I'm not just stacking bears. The most quotable bubble skeptic of the week was Richard Kramer of Arete Research, a former top-rated Goldman Sachs analyst, on FT Tech Tonic's "Is the AI boom a bubble?" (Sept 16). Kramer doesn't dispute the spending. He disputes the stories being told about it. Big tech, he said, will spend a trillion dollars on capex ("capex" = capital expenditure, the money companies spend building physical things like data centers) next year, up from about $700 billion this year, plus another $380 billion on research. And here's the number that should stop you: the combined cash flow of the megacaps is actually going down this year, to about $175 billion, "but $200 billion of that is showing up at NVIDIA. So this AI investment is kind of single-handedly propping up the U.S. economy." His phrase for the whole thing was a "value transfer from the internet industry to the semiconductor industry."

Kramer's warning wasn't about the building; it was about the valuations:

"On the valuation side, there is right now a degree of consensual hallucination of... an extrapolationist fantasy."

He pointed to Anthropic reportedly heading for an IPO citing a $30 trillion total addressable market ("coincidentally, also all the software applications in the world") and noted that when these pure-play AI companies finally file to go public, the interesting question won't be how much money they make but "who has the largest losses." The elegant bull rebuttal came from his fellow panelist, Blossom Capital's Ophelia Brown, who noted that in twenty trips to San Francisco this year she'd met no one who thinks it's a bubble, and that OpenAI reached a billion monthly users in under three years, a scale it took the iPhone thirteen years to reach.

So where does the funding-gap math actually land? The bears keep sharpening it. On Hedgeye (Sept 15), Mike Taylor ran the arithmetic: roughly $3 trillion of AI compute by the end of 2027 means about $400 billion a year in depreciation alone, on top of maybe a trillion for power, staff and the cost of capital, against roughly $100 billion in combined OpenAI and Anthropic revenue. Luke Kawa, on Full Signal (Sept 16), gave the cleanest one-liner on what's really being questioned:

"The bubble is expected earnings, the bubble is clearly not valuations."

His point: the stocks aren't insane on today's numbers; they're insane only if the future earnings everyone's penciling in don't show up. And the swing factor, he said, is whether credit markets keep their appetite to finance it all, a question he doesn't think gets resolved until roughly mid-2027. Which is a long time to hold your breath at a 5% 10-year.

Custom Silicon: The "Free Chip" Trade Gets a Price Tag

Issue 024's number-two theme was a fresh and provocative claim from Vaughan Nelson portfolio manager Adam Rich: that the custom-chip boom (hyperscalers designing their own accelerators instead of buying Nvidia's) is "the real bubble," and that the deals are structured so hyperscalers get these chips "basically free" through stock warrants. This week, that abstract claim got a hard, specific number attached to it.

On The Real Eisman Playbook (Sept 11), Steve Eisman walked through the Amazon–Qualcomm partnership, a deal to build custom AI chips and optical networking gear aimed squarely at competing with Nvidia. The eye-opener was the payment structure:

"As part of the deal, Qualcomm issued a warrant to Amazon, giving Amazon the right to buy 25 million Qualcomm shares at $161.26. Now, since Qualcomm's stock is over $175, the warrants are in the money and are worth over $4 billion."

(A "warrant" is a coupon that lets the holder buy stock later at a fixed price. Because Qualcomm's shares already trade above the $161.26 strike price, Amazon's coupon is "in the money," instantly worth billions.) Eisman admitted he was "a bit bewildered": not by the partnership, but by why Qualcomm would hand a customer $4 billion of equity upside when it doesn't need the cash. His subscribers offered two theories worth chewing on: first, that it's a "strategic customer acquisition cost: Qualcomm trades equity upside to guarantee a $60 billion revenue pipeline and secure AWS as a flagship anchor customer," a way for a company known for phone modems to buy instant credibility in data centers; and second, that Qualcomm's expertise in these XPUs (the industry's catch-all for custom accelerator chips) is unproven, so Amazon's own chip-design know-how may be the thing actually making the partnership work, which would explain why Amazon got paid.

This is exactly the dynamic Adam Rich flagged last issue, now visible in the open. When a chip vendor has to give a hyperscaler billions in stock to win the business, it tells you who holds the bargaining power, and it isn't the chip vendor. Network Break (Sept 14) added the technical shape of the deal: AWS deploying Qualcomm silicon for AI inference with optical links up to 1.6 terabits per second, and Amazon reciprocally giving Qualcomm access to its Bedrock AI platform for chip-design work. The host's reaction was the same skepticism Eisman voiced: this looks like a lot of favors changing hands.

The fresh challenger operator. If custom silicon is one threat to Nvidia, the startups are another, and this week gave us a live one. On Squawk on the Street (Sept 14), in a CNBC exclusive, SambaNova CEO Rodrigo Liang made his pitch. His whole company is built on a single metric: "cost per token served" (a "token" is a chunk of text an AI model processes; cost per token is essentially the unit cost of running AI). His claim:

"If we can find a way to actually reuse existing data centers, brownfield data centers that are air-cooled with the power already allocated to it... cost of inference is going to drop 10x."

("Brownfield" just means existing, already-built facilities, as opposed to new "greenfield" construction.) Liang's argument is that Nvidia's newest systems demand liquid cooling and enormous power, forcing expensive new builds, while SambaNova's lower-power chips can slot into data centers that already exist. He claimed a six-month payback on the hardware versus 12 to 24 months for competitors, and confirmed SambaNova raised a $1 billion round at an $11 billion valuation on July 8, with a second close "almost finished" as public-market investors pile in. He's got a multi-year partnership with Intel (whose CEO Lip-Bu Tan was SambaNova's chairman for years), and he pointed to customers like JP Morgan "repatriating" AI in-house, bringing it behind their own firewalls rather than renting it in the cloud.

Take the specific numbers with salt: they're a CEO selling his book. But the strategic read is real, and it's the one Nvidia bulls should worry about most: if inference (the everyday running of AI models, as opposed to the one-time cost of training them) really can be done 10x cheaper on cheaper hardware in existing buildings, a chunk of the trillion-dollar build-out thesis softens. Notably, that same panel featured Bessemer's Byron Deeter, an Anthropic investor, defending the industry's new "pacing" language as responsible rather than a slowdown, and Kalshi's odds of an Anthropic IPO announcement before Nov 1 had fallen to 54%, down from about 75% a week earlier.

And the bull's structural rebuttal, from TechSurge (Sept 16): Nvidia's moat isn't the GPU, it's the system (GPUs plus software plus networking plus cooling, sold as one integrated rack), and cloud buyers keep paying Nvidia's fat margins because they value speed-to-market over squeezing out cost. AMD is competing at that same system level now (its Helios rack), but the ASIC startups, the podcast noted, need to raise enormous sums just to stay in the race.

The Operator Voice Returns: Micron on the "Memory Wall"

Last issue I complained about an operator drought: plenty of pundits, no actual company insiders. This week fixed that, and the best of them was on a small engineering podcast most investors will never find.

On Embedded Insiders' "Hitting The Memory Wall" (Sept 10), Jeremy Werner, Senior Vice President and General Manager of Micron's Core Data Center Business Unit, gave a genuinely useful tour of why memory has become the choke point in AI. The core idea, the "memory wall," is that raw computing power has grown far faster than the ability to feed it data:

"The big change with integrating AI is that the bottleneck really shifts from computation to memory... the rate of flops in the case of a GPU... has grown at a much faster rate than the ability of memory to service data to the components. So that creates a widening gap that props up the wall."

His solution is Micron's HBM4, the newest generation of "high-bandwidth memory," the specialized chips that sit right next to a GPU and feed it data. Werner said HBM4 "doubled the bandwidth from the prior generation, or more than doubled, and brings down the energy meaningfully." He described a five-layer hierarchy of memory and storage in a modern inference system, and, when the host teased him about whether integrating memory directly onto the chip package locks customers into Micron, delivered the quote of the segment:

"No one has to use our memory, but usually the smartest people do."

On the shortage that's sent memory prices soaring, Werner was frank. Micron is building five fabs at once: two 600,000-square-foot plants in Boise, Idaho; the first of an eventual four in Clay, New York; plus expansions in Taiwan, Singapore and Hiroshima. But the constraint isn't demand or even money:

"There's just not enough of the supply chain, of the labor force, of the know-how to be able to build these things as fast as we'd like."

That's the operator confirming from the inside what the macro folks keep saying from the outside: the bottleneck is real, physical, and slow to fix. Which is bullish for anyone who already makes memory, and sets up this week's fresh memory story.

Memory Supply: SK Hynix Eyes American Fabs

Here's a new thread that broke this week and matters for Micron and the whole memory complex. On The Rundown (Sept 16), the report (sourced to Reuters) is that SK Hynix, the Korean company that is the leading supplier of HBM to Nvidia's GPUs, is in early talks with Intel to manufacture memory chips in the United States for the first time ever.

Two structures are being floated: SK Hynix leasing part of Intel's massive (and repeatedly delayed) Ohio complex, or a three-way joint venture between Intel, SK Hynix, and a hyperscaler that wants to lock in its own memory supply. The push behind it is political: Commerce Secretary Howard Lutnick has threatened up to 100% tariffs on Korean chipmakers who don't build in the US. The host's summary of the memory market was blunt: "the price of RAM costs more than rent these days."

Nothing is signed, and these talks are early. But the read-through is worth holding: if HBM production genuinely starts localizing to the US, driven by tariff threats and hyperscalers desperate to secure supply, it reshapes the competitive map for Micron (the one major HBM maker already US-based), Samsung and SK Hynix. This same story got syndicated across roughly ten AI-news podcasts this week, all repeating the same framing (three HBM makers globally, a "global bottleneck"), and the sheer echo tells you the market cares.

Names in Play

Credo (CRDO): the fastest grower nobody wants. The most name-specific work of the week came from Chip Stock Investor's "The Fastest-Growing Chip Stock Is Also the Most Hated" (Sept 15), where Nicholas Rossolillo dug into Credo, which makes active electrical cables and networking chips for AI data centers. The numbers are striking: revenue up 115% year-over-year last quarter (the fastest in its peer group) with a 25% operating margin, 17–18% free-cash-flow conversion, and $764 million of net cash. Management is guiding to 70–80% revenue growth for the new fiscal year. And yet the stock has sold off harder than almost anything else in the group. Rossolillo's reverse-DCF (a valuation method that works backward from today's price to ask what growth the market is assuming) found that at around $171 the market was only pricing in about 58% annual growth over three years, below the company's own guidance. He frames Credo and Astera Labs as the "new Baby Broadcoms." The catch he flagged himself: if the AI data-center build-out slows sharply, both Credo's growth and its newly-fattened margins are exposed. Note Credo closed Sept 16 at $161.49, now below even that $171 reference, and a remarkable 47.7% below its 52-week high despite the growth.

Oracle (ORCL): the numbers behind the drama. Last issue Oracle was all vibes, with Cramer taking his skepticism "off the table." This week Chip Stock Investor (Sept 14) put the hard capex numbers on the table, and they are staggering. Oracle's capital spending hit $28.5 billion in a single quarter, up from $8.5 billion a year earlier and $16.5 billion the prior quarter, the biggest in company history. Cloud infrastructure revenue rose 121% to $7.4 billion (mostly OpenAI), and multi-cloud database revenue was up 353%. But the balance sheet shows the strain: $125 billion of debt against $37 billion of cash, capex running at 1.61 times operating cash flow (you cannot spend more building things than your business generates in cash forever), and free cash flow still negative. The one reassuring signal: Oracle finished its $20 billion stock-issuance program and has now paid down debt two quarters running, management trying to calm the bond market. Rossolillo's base case is that capex growth peaks around fiscal 2027 and Oracle reaches free-cash-flow breakeven within a year. It's the clearest window we've had into what "financing the build-out" actually looks like on one company's books.

Nvidia (NVDA): a new named bear. Also on Bloomberg Surveillance (Sept 16), Seaport Research Partners' Jay Goldberg laid out his underperform rating, carefully not a short, just a bet that Nvidia lags the rest of the chip complex:

"It's gotten very big, and so it's harder and harder for it to move the needle. Expectations are always very high and it's increasingly hard to beat them. And on top of that, they're sold out... for this year, probably for next year. And once you're sold out, what's the upside?"

That last question is the whole bear case in seven words. Goldberg still called this "the strongest semiconductor upswing cycle any of us have ever seen"; his point isn't that Nvidia is bad, it's that being sold out two years forward caps how much good news is left to surprise on. He also noted Apple's new A20 chip is a 2-nanometer part from TSMC, while cautioning that Moore's Law (the old promise that chips double in density every couple of years) has "slowed down considerably."

The "slowdown" that isn't. Worth flagging against all the doom: on Prof G Markets (Sept 15), Charlie O'Neill of Base10 pushed back hard on the idea that the AI labs' new "pace the frontier" safety language means less spending. His read is the opposite: labs will divert up to 20% of their internal compute to safety and monitoring while still training ever-bigger models, which means "we might even see the labs be even more aggressive with compute build-outs and securing compute." If he's right, the demand side of the accelerator trade isn't softening at all; it's the financing side that's the question.

Operators vs. Pundits (Who's Actually Talking)

I keep these separate on purpose: a company insider with skin in the game is a different signal from a commentator with a microphone.

Operators / insiders (relayed or direct):

  • Jeremy Werner, SVP & GM Core Data Center, Micron: direct, on the memory wall, HBM4, and the five-fab build-out (Embedded Insiders).
  • Rodrigo Liang, co-founder & CEO, SambaNova: direct, claiming 10x cheaper inference and six-month paybacks (Squawk on the Street).
  • Jim Zelter, President, Apollo Global Management: direct, as the lender financing Intel, Broadcom and Nvidia (Bloomberg Talks).
  • Relayed only: the Amazon–Qualcomm warrant terms (via Eisman); Oracle's capex and balance sheet (via Chip Stock Investor's read of the earnings call).

Pundits / investors / analysts:

  • Bears & skeptics: Ruchir Sharma's 5% tripwire (carried over), Alicia Levine/BNY (hikes hurt funded sectors), Richard Kramer/Arete ("consensual hallucination"), Mike Taylor/Hedgeye (capex math), Jay Goldberg/Seaport (Nvidia underperform).
  • Bulls & counters: Ophelia Brown/Blossom (no bubble in SF), Charlie O'Neill/Base10 (labs will spend more), Byron Deeter/Bessemer ("pacing" is responsible), Nicholas Rossolillo/Chip Stock Investor (constructive into year-end).

What Changed Since Issue 024

  • The tripwire moved from theory to test. Two weeks ago the debate was "can it get financed?" with the 10-year around 4.8%. This week the 10-year hit Sharma's exact 5% danger line and the Fed went in priced to hike into it. The question stopped being hypothetical.
  • The fade broadened and rotated. In 024, networking (Arista) held while memory and small optics led down. This week every name fell, and the leadership of the decline shifted to the custom-silicon and memory tail (Astera Labs -7.6%, Broadcom -6.2%, Micron -5.0%, Coherent -5.1%) while AMD (-0.7%), Lumentum (-0.8%) and Credo (-0.9%) held up best. That AMD and Nvidia now fall less than the custom-silicon names is itself a read on which part of the trade the market is de-risking first, and it lines up neatly with last issue's "custom ASICs are the real bubble" thesis.
  • The "free chip" claim got a receipt. 024's abstract warrant argument became a concrete $4 billion coupon (Amazon–Qualcomm), plus a live challenger (SambaNova) putting numbers on the cost-cutting case.
  • Operators showed up. After an operator drought, we got Micron, SambaNova and Apollo speaking directly.
  • Oracle went from anecdote to spreadsheet, and Nvidia picked up a new named sell-side bear (Seaport).

What I'm Watching Next

  • The Fed's actual decision and the 10-year's path. The single most important input to this trade is no longer a company; it's a yield. If the 10-year holds above 5%, Levine's "marginal areas that require funding" get squeezed further.
  • Micron's next print (late September): the memory bull's real test, and the biggest US-based HBM maker.
  • Marvell's Investor Day (Oct 6): custom-silicon economics finally get numbers, right as the market is questioning the whole category.
  • The Anthropic IPO clock (odds now ~54% before Nov 1) and Oracle's credit spreads: the two cleanest real-time reads on whether the financing machine is still humming.
  • Any concrete move on SK Hynix–Intel: the first real sign of HBM production localizing to the US.