Newsletter · 2026-09-07 · Ashutosh Agarwal

Broadcom Blowout Fades and Suppliers Get Paid - AI Accelerators - Week of September 7, 2026

AI Accelerators for the week of September 7, 2026: Broadcom guided to doubling AI revenue two years running and more than 30 dollars of EPS by fiscal 2028, and the stock still fell about 2.5 percent, completing a three-name fade pattern, while memory, optics and connectivity suppliers ripped and Nvidia paid 12.9 billion dollars for Hugging Face.

AI Accelerators

Week of September 7, 2026: Broadcom Blowout Fades and Suppliers Get Paid

GPUs, Custom Silicon & Optics. Issue 022, Monday, September 7, 2026 (Labor Day, U.S. markets closed). Podcast window: August 31 to September 7, 2026.

Broadcom just did something remarkable, and the market barely blinked.

On Wednesday night (September 2), the company told investors it expects to roughly double its AI chip revenue next year, double it again the year after, and clear more than $30 of earnings per share by fiscal 2028, a number CEO Hock Tan says he has now "secured the supply" to hit. It named a roster of six customers, four of them "simply going to be huge," and laid out gigawatt-by-gigawatt build plans for Google, Anthropic, OpenAI and Meta. By any normal standard, it was one of the great guidance nights of this cycle.

And then the stock fell. Broadcom (AVGO) closed Thursday down 2.7%, and after a flat Friday it ended the short week down about 2.5% from where it sat before the print, at $357.90, versus a Wall Street average price target near $538. A "sensational" quarter, in the words of one podcaster, that the stock sold anyway.

That completes a pattern this newsletter has been tracing for a month. Nvidia guided way above expectations on August 26 and faded. Marvell beat and fell 10% on August 28. Now Broadcom, the biggest guide of the three, beats and fades too. Three custom-silicon champions, three "sell-the-news" reactions in a row. When the whole complex is priced for a flawless future, even a great result is only in line with a miracle.

But here's the twist, and it's the real story of the week. While the chip designers faded, the companies that supply them ripped. On Friday alone: Marvell +7.0%. Astera Labs +9.8%. Micron +6.1% to a new high. Coherent +6.6%. Lumentum +4.0%. As one podcast put it in its title: "Only the Servers Got Paid This Week. Every Other AI Beat Sold Off." Money isn't leaving AI hardware. It's rotating down the stack, out of the priced-for-perfection brains, and into the memory, the optics and the "picks and shovels" where the shortage is most acute.

TL;DR

• The fade is now a pattern, not a one-off. Nvidia, Marvell and now Broadcom all beat and faded. Broadcom guided to doubling AI revenue two years running and >$30 EPS by FY2028, and the stock still fell ~2.5% on the week. When 2027 is "locked in," the market is only paying for the uncertain 2028 upside, and it's discounting it hard.
• The rotation is the trade. The designers faded; the suppliers rallied Friday: MRVL +7%, ALAB +9.8%, MU +6.1% (new high), COHR +6.6%, LITE +4%. "Only the servers got paid." Software got paid too: Snowflake +24%.
• Nvidia bought Hugging Face for ~$12.9B (plus ~$1B to keep the staff). Jensen calls it worth "every single penny"; skeptics call it "chip demand insurance," paying 86x revenue to make sure open-source AI keeps defaulting to Nvidia chips.
• The bubble and credit debate got its named voices back. Paul Kedrosky: Nvidia is now "the single point of failure," and more than half of data-center spend is externally financed. Tyler Cowen: "the bears are asking the wrong questions." Both are worth steel-manning.
• Memory is the cleanest bull story. A structural shift to custom high-bandwidth memory means memory may stop being a commodity: pricing power up, not down. Micron hit a new high on Friday.

1) The lead: a blowout that faded, and what the market is actually pricing

Start with what Broadcom said, because it was genuinely enormous. On the Q3 call (covered last issue), Tan guided fiscal 2026 AI revenue to ~$58 billion, then ~$115 billion in FY2027 and ~$230 billion in FY2028, roughly $350 billion of AI revenue across the next two years, and said the company is "on target to exceed $30" of EPS in FY2028. He described "six customers, four of them simply going to be huge."

This week we got the market's verdict, and the podcasts to explain it.

The reaction, in numbers (FactSet): Broadcom closed at $367.24 the day of the print. It fell to $357.16 on Thursday (-2.7%), the first full session investors could trade the news, and finished Friday at $357.90, net -2.5% on the week. This was not a crash. It was a shrug at a triumph.

Why? The clearest explanation came from The AI Investor Podcast ("A Tale Of 3 Stocks, Credo, Broadcom and Snowflake," Sept 3), which walked through why three fast-growing companies all got very different receptions:

"We're seeing a difference from Wall Street here where they delivered this blowout guidance and Wall Street's essentially meeting it with a yawn. This is a very similar situation to Nvidia... Nvidia said next year we're going to deliver 70% growth, Wall Street was expecting 44%, and initially this was met with a yawn."

The host's diagnosis is worth understanding, because it's the mental model for the whole group right now. Two fears are doing the work:

1. "Peak earnings." As he put it, "2027 is largely accounted for," the orders are booked, the estimates are locked. That pushes the debate into 2028, which he calls the "zone of uncertainty," where "a lot less of the demand is as certain." Broadcom's monster number is a 2028 number. The market is paying for the sure thing and discounting the maybe.
2. Margin compression. As Broadcom scales its custom chips (XPUs, its term for a customer-designed AI accelerator), gross margin, the share of each sales dollar left after the direct cost of making the product, drifts down, because those chips carry a lot of bought-in memory. The fear is that when demand eventually cools, today's fat 70-something-percent margins mean-revert.

On the raw valuation, he argues the fade is almost absurd: Broadcom now trades at roughly 10.3x its fiscal-2028 earnings on ~100% growth into that year. Snowflake, by contrast, a software company growing ~36%, trades at 93x 2028 earnings and rose 24% on the week. The market is paying up for open-ended software growth and refusing to pay for hardware growth it can already see the end of.

He did flag the one genuine worry, and we'll come back to it in the debate section:

"If there is one fear about them, it is largely that they're having to use kind of financing techniques to be able to facilitate some of these sales to companies like OpenAI, Anthropic."

Bloomberg Intelligence ("Broadcom Slides After AI Chip Forecast Underwhelms Investors," Sept 3) took the glass-half-full read. Analyst Matthew Bloxham noted the striking customer detail, that Broadcom now expects Anthropic and OpenAI to become its most important customers, "eclipsing Google," and argued there's "substance behind the plans and capital raising." His core bull point is the diversification trade: the hyperscalers "don't want to be purely reliant on Nvidia," and they'll want "two or three" genuinely independent chip suppliers. But even he conceded the runway is "contingent on the current momentum continuing," which is exactly the fragility the stock is pricing.

And Best Stocks Now with Bill Gunderson (Sept 3) summed up the mood in a sentence: the earnings were "sensational," the stock sold off anyway, and he's sticking with a $700 price target on a 16%-growth, five-year view.

Why it matters: The lesson of the NVDA to MRVL to AVGO relay is that in this tape, the bar is the enemy. Beating consensus is not enough when consensus already assumes perfection two years out. For a book, the asymmetry has quietly flipped: the marquee names now need a clean beat and an out-year raise just to hold the line, while the read-through names below them have room to surprise. Which brings us to the rotation.

2) The rotation: only the suppliers got paid

Here is the single most actionable table of the week, the Friday (Sept 4) session, when the market told you exactly where it wants to be:

Source: FactSet daily closes. AVGO % is week-to-date from its pre-print close.

The framing that captures it comes from Futurum Equities (Ep. 55, Sept 2): "Only the servers got paid this week. Every other AI beat sold off." The chip designers are where expectations are highest and therefore hardest to beat. The chip suppliers, the memory, the lasers, the retimers and cables that every one of those designs must buy, are where the shortage still bites and the estimates still have slack.

CNBC's Fast Money (Sept 2) made the intra-group version of the same call. On Marvell, which is now +143% year-to-date: "That's where the hot money is going, because they're the one at the top right now." The logic: hyperscalers are moving work off Nvidia and onto their own custom chips, and both Marvell and Broadcom design those chips, so both "should be the winner." Marvell has won so far; Broadcom, the panelist said, "gets interesting on a sell-off." On Nvidia itself, another trader flagged the tension holding the whole group together:

"You have a company like Nvidia that is beating so aggressively and it just doesn't seem to have that margin degradation that a lot of folks were calling for with all this competition... there's a huge disconnect where that stock is trading and the numbers that they're putting up."

Nvidia, notably, did not fade. It ground to within a whisker of its 52-week high ($230.36 vs $236.54). The market's message: it will still pay for the undisputed leader with no margin cracks, and it will pay for the shortage-driven suppliers. What it won't pay for is a #2 custom-silicon story whose best growth is dated 2028.

Why it matters: This is a rotation within the winners, not a rejection of the theme. For positioning, it argues for owning the bottleneck (memory, optics) and the leader (Nvidia), and treating the custom-silicon designers as sources of funds on strength and buys on the fade, precisely how the Fast Money desk is playing it.

3) Nvidia buys Hugging Face: a $12.9B insurance policy?

The biggest single piece of news this week wasn't an earnings print, it was a deal. Nvidia agreed to acquire Hugging Face, the hub where the world's open-source AI models live, for $12.93 billion, plus a reported ~$1 billion to retain the staff.

We got it straight from the principals on Squawk Pod (Sept 3), where Jensen Huang and Hugging Face CEO Clem Delangue appeared together. Huang's pitch was that open-source AI is now half his business and accelerating:

"Cloud service providers represent only half of Nvidia's business... the other half of our business are really largely driven by open models. And so Nvidia is growing in both directions."

On the price, Hugging Face was valued at ~$4.5 billion in 2023, he was blunt:

"We paid $12.9 billion because, one, that's what it's worth. And two, open models mean so much to us... It's worth every single penny."

Delangue said the three founders and the whole team join Nvidia, that Hugging Face will run as an "independent, neutral platform," and that the ambition is to go from 18 million developers today to "100 million AI builders."

That "neutral platform" claim is exactly where the skeptics pounced. The sharpest analysis came from the Elon Musk Podcast (Sept 3), which did the math most bluntly: at ~$150 million of revenue, Nvidia is paying an 86x revenue multiple, which "isn't a traditional software purchase." Their read:

"What they're actually buying is chip demand insurance... a defensive maneuver to protect their core monopoly on the data center floor."

The mechanism is worth understanding in plain terms, because it's the whole point of the deal. When a developer writes an AI model, the code has to be translated into instructions a specific chip understands, and for a decade the default language for that translation has been Nvidia's CUDA (its proprietary software layer). Google's TPUs and Amazon's Trainium chips use different languages. The open-source world builds for CUDA. By owning the hub where millions of developers find, test and download models, the argument goes, Nvidia keeps the default path pointed at its own hardware, and stops an independent Hugging Face from ever making it "frictionless" to deploy on a cheaper Amazon or Google chip instead.

The counter, which the same podcast raised, is that this can backfire: "Software engineers view friction as a defect and they just route around it." Try to box developers in, and you may push them toward the very alternatives you were defending against.

The more benign read came from BofA's Vivek Arya on The Exchange (Sept 3): the deal gives Nvidia control of "the town square" for AI developers and a front-row seat to shape where silicon needs to go next. Not everyone was sold. On Fast Money (Sept 3), one panelist worried the acquisition could actually hurt Nvidia by dragging a pure hardware champion into a lower-margin software fight.

This lands alongside two other Nvidia ecosystem moves worth logging: a $3.5 billion investment in MediaTek (Bloomberg Intelligence, Aug 31, where Mandeep Singh pegged Nvidia at ~80% accelerator share and guiding to >$650B annual revenue next year), and AWS committing to another 2 million Nvidia GPUs for 2027-28 on top of a million booked five months ago (The Six Five, Aug 31, 6-7 gigawatts, $80-120B).

Why it matters: Every one of these deals, Hugging Face, MediaTek, the Anthropic and Lambda financing structures, is Nvidia spending balance sheet to defend demand rather than to sell more chips today. The bull sees a moat being widened. The bear sees a company that has to keep priming its own pump. Both are looking at the same $12.9 billion.

4) The debate, steel-manned: is Nvidia the moat, or the single point of failure?

With the podcasts back, the bubble and credit argument regained its named combatants. Both sides deserve a fair hearing.

The bear: Paul Kedrosky on RiskReversal Pod ("Nvidia Is Now the AI Bubble's Single Point of Failure," Sept 4). Kedrosky's case is the most complete bear framework we've heard this cycle, and it has two legs.

The first is financial. AI capital spending, he says, now accounts for "anywhere from 30 to 60% of GDP growth in the United States, which is bonkers," and, crucially, the way it's funded has changed:

"More than half of hyperscaler spending on data centers is now externally financed, whether it's through SPVs or off balance sheet."

His warning is that this is how every buildout ends badly: "Capital markets are very good at turning prime credits into subprime credits, they just give them enough time." He invokes the financial-crisis mantra IBGYBG ("I'll be gone, you'll be gone") to describe the bankers structuring the deals, and argues AI is uniquely dangerous because it sits at the intersection of the four forces behind history's biggest bubbles at once: "technology, credit, government policy and real estate." His most vivid example: the rise of "powered shells" (or "powered land"), speculative data centers with power and water secured but no customers and no GPUs, built on hope.

The second leg is the "single point of failure" itself, and it's a supply-chain argument, not just a valuation one. Competitors building inside the data center, he says, refuse to keep paying Nvidia's 70-80% margins:

"My margins are your margins... I'm sitting here with 70% and 80% margins. I don't just sit here and take it. I'm pissed off about it as someone building inside the stack. So how do I engineer around this?"

The escape route he watches is memory: Chinese manufacturers, he claims, plan to add Samsung-equivalent capacity within about 12 months, which could crush high-end memory pricing the way solar-panel prices collapsed, and he thinks that risk is already discounting into Nvidia's ~17.5x earnings.

Two fellow bears reinforced the credit angle: George Gammon on Rebel Capitalist News (Aug 31) called Nvidia a "shadow bank" with ~50% of its $96B revenue concentrated in just four customers (Microsoft, Meta, Alphabet, Tesla), and argues a 10-20% order cut could trigger a Cisco-1999-style collapse. And Dan Niles on Excess Returns (Sept 3) is "bearish into November, room to run after," with the hyperscalers' credit-default swaps as his tell.

The bull: Tyler Cowen on Prof G Markets ("The AI Bears Are Asking The Wrong Questions," Sept 4). Cowen doesn't dispute the risks, he reframes them. On demand, he's unworried: token usage (a token is roughly a word or word-fragment the AI processes) is "going crazy through the roof," so "I don't think there'll be a demand crisis. At some point, there may be an energy sector crisis." Power, not appetite, is the binding constraint. On the objection that AI labs lose money on every query, he concedes prices "probably" need to rise short-term, but notes token prices have fallen "more than 100x" since GPT-4, so efficiency is racing costs and may win. On the circular-financing worry, he offers the bull's frame directly:

"You can think of Nvidia as a kind of lender or buyer of last resort for the sector."

And on valuation, he leans on the wisdom of committed capital: private markets value Anthropic near $2 trillion and OpenAI above $1 trillion, "those are smart people with real money on the line... on average, we should believe them." His analogy isn't the dot-com bust; it's the 1920s auto boom, where most car companies died but cars were still one of the defining industries of the century. Gavin Baker on the a16z Show (Aug 31, "Why AI Demand Is Outrunning Compute Supply") and Sands Capital's Daniel Pilling on Pitch The PM (Sept 2, "Why the AI Boom Is Still Early") round out the bull bench.

Where they actually agree: Both sides now watch the same instrument, the credit market. Kedrosky watches SPV financing and CDS for the crack; Niles watches hyperscaler CDS; even the bull case concedes the debt-financed parts of the supply chain (data centers, not the cash-rich labs) are where "bad macro consequences" would start. If you take one operational cue from this issue, it's that the bond market, not the next earnings beat, is the tell everyone is now staring at.

5) Memory: the cleanest bull story, and it's getting structurally better

If the designers are priced for perfection and the bubble debate is unresolved, the one place operators and pundits broadly agree is memory, and Micron's new high on Friday (+6.1% to $1,016) says the market agrees too.

The clearest operator-grounded insight came from The Circuit (EP 189, Aug 31), reporting from the Hot Chips conference. The theme across every memory maker was a shift to custom high-bandwidth memory (HBM), memory whose base layer is tailored to a specific customer's chip, rather than an interchangeable commodity part. The hosts relayed the Micron CEO's framing that you should now think of custom base-die memory "more like custom ASICs, a unique design customized for that customer." The investment consequence is the important part:

"In a custom base-die world, memory doesn't go down. In fact, it probably goes up... and it's going to behoove one to two memory vendors per customer."

In plain English: the knock on memory has always been that it's a commodity, interchangeable, so prices crash in every cycle. If memory becomes custom, that changes. Customers get locked to one or two suppliers per chip design, fungibility drops, and pricing power rises. The catch the hosts flagged honestly: it also raises costs and complexity, and could mean the same GPU performs differently depending on whose memory is inside, a genuinely "weird world" for the buyers.

Fidelity portfolio manager Zach Schreiber, on Market Sense (Sept 1), gave the demand side with the best metaphor of the week. Think of the AI model as a brilliant student and memory as the books on their desk:

"For years the frontier labs focused on making the student smarter... The desk, meanwhile, didn't improve much. Every time the student writes a sentence, they have to get up, walk over to the library, grab a new book, and walk back. The walk is the limiting factor."

HBM is a bigger desk, and "only three companies in the world make it, they didn't build enough, and a new factory takes years." That's why Microsoft, Google and Amazon have raised memory budgets "by many, many billions of dollars this year." Schreiber also steel-manned his own bull case, which is the discipline that makes it credible: bottlenecks "invite creative thinking, new solutions, and ultimately overbuilding," after which "supernormal profits mean-revert." The whole memory call, he says, comes down to one judgment: is this shortage cyclical (temporary) or secular (structural)? The custom-HBM shift is the strongest argument yet that at least part of it is secular.

One cloud on the horizon, from TechLinked (Sept 1): China's CXMT beat Samsung and SK Hynix to mass production of LPDDR6 (a next-gen mobile memory). It's not HBM, but it's the same Chinese-capacity threat Kedrosky is worried about, showing up first in the commodity tiers.

Why it matters: Memory is the rare AI-hardware theme where the bull case is strengthening (custom dies mean pricing power) while the stocks still don't carry perfection-level multiples. It's the cleanest expression of the "own the bottleneck" rotation.

6) Read-throughs and names in play

• Optics and networking got the rotation treatment. Coherent (+6.6%), Lumentum (+4.0%), Astera Labs (+9.8%) and Arista (+2.9% Thu) all rallied, cashing the read-through from Broadcom's own call last week, where Tan said the company is "more than tripling" its laser capacity because "demand for lasers is far surpassing supply." When the biggest custom-silicon vendor says it can't get enough lasers, the laser makers are the trade.
• Credo (CRDO) is its own story, handle with care. Credo fell ~20% on Sept 2 on its own report, not Broadcom read-through. Per The AI Investor, it delivered 115% growth, guided to 80% full-year growth and carries 68% gross margins at just 14.6x calendar-2026 earnings, yet still got hit on the same peak-earnings and 2028-margin fear stalking the whole group. A high-growth name being punished at 14.6x forward is the fade pattern in miniature.
• Systems: Dell is the sleeper winner. On The Rundown (Sept 2), Dell's AI-server revenue doubled to $16.4B with a $95B backlog, CPU servers jumped 122%, and management raised 2026 revenue guidance from $167B to $192B and EPS to $25.50, with the stock +230% YTD at ~20x forward. The servers, quite literally, got paid.
• The constraint is physical, and it's power. Arm CEO Rene Haas, on No Priors (Sept 3), warned that "constrained wafer and memory capacity will throttle AI growth for 3-5 years," with data-center buildout the gating item. Moody's Talks (Sept 1) sized it: racks and servers are now two-thirds of AI infrastructure spend, and data-center construction is roughly $1 of every $4 of non-residential building in the U.S.

What changed since last issue

• Broadcom's reaction landed, and it faded. The print came after Wednesday's close, so the market had not traded it last time out. Now it has: -2.5% on the week, completing the NVDA to MRVL to AVGO fade pattern.
• The rotation is new and tradable. Last week was the print; this week is the market's answer, money down the stack into memory, optics and Marvell, and up the stack into software.
• Nvidia and Hugging Face is a brand-new development, with the deal confirmed Sept 3 direct from Jensen Huang.
• The bubble debate re-armed with named, tradable voices: Kedrosky, Cowen, Niles and Baker.

Next catalysts

• AVGO at the Goldman Communacopia + Tech Conference, Tuesday, Sept 8 (tomorrow). First chance for management to address the fade.
• Marvell Investor Day, Oct 6 (expect FY2029+ custom-silicon sizing).
• Micron reports, late September (the memory bull case gets tested).
• AVGO Q4 FY2026, Dec 10. Nvidia's next print, ~Nov 18.