Newsletter · 2026-09-10 · Ashutosh Agarwal

AMD Breaks Out as Nvidia Fades in the Accelerator Trade - AI Accelerators - Week of September 10, 2026

AI Accelerators issue 023 for the week of September 10, 2026: AMD broke out over 500 dollars on a raised long term forecast while Nvidia kept fading, the custom silicon trade fragmented in a way that is getting harder to model, Alphabet posted its first negative free cash flow, and CoreWeave argued from the Goldman conference that the unit economics are improving.

AI Accelerators

Week of September 10, 2026: AMD Breaks Out as Nvidia Fades in the Accelerator Trade

GPUs, Custom Silicon & Optics, Issue 023, Thursday, September 10, 2026 Research window: September 3–10, 2026. Prices are FactSet daily closes; the last completed session before publication was Wednesday 9/9 (Monday 9/7 was Labor Day, market closed).

Three weeks ago the story was simple: Nvidia blew out its quarter, then Marvell did, then Broadcom did, and each one sold off anyway. We called it the "sell the news" fade, and by Monday's issue it had a sequel, the day after Broadcom's print, the money that left the big designers didn't leave the group. It rotated down the stack into the parts suppliers (memory, optics) and up into software.

This week that rotation did something new and, if you own these names, important: it picked a new leader. AMD, the number-two GPU maker, and a stock this newsletter has not had a real reason to write about for six straight issues, gapped up over $500 and ran +9.1% in two sessions (Tuesday and Wednesday). Optics kept ripping (Lumentum +12.2% on the week). And the old king kept slipping: Nvidia fell -2.9%, drifting toward the bottom of its 52-week range of movers even as everything around it caught a bid.

Put plainly: for a year, "AI chips" meant "buy Nvidia." The trade right now is the opposite instinct, sell, or at least fade, the priced-for-perfection leader and spread the chips around the rest of the stack. Whether that is healthy rotation or the first crack, the smartest people on this week's podcasts genuinely disagree, and this issue is mostly about why they disagree, in their own words.

A quick, honest note on the podcasts themselves: the marquee finance-and-tech shows an analyst actually wants, BG2, All-In, Acquired, Odd Lots, Stratechery, Sharp Tech, The AI Daily Brief, ChipStock Investor, were silent on chips again this week. So the best investing content came from the CNBC/Bloomberg daily shows, a couple of excellent sector podcasts (The Circuit, Next in Tech), and a very loud bubble debate. I'll flag clearly where a claim is a pundit's number rather than a company's.

TL;DR

• AMD is the week's story. +9.1% over two sessions (to $521.10), a clean technical breakout over $500. The catalyst, per IBD's daily podcast: AMD's CFO told a conference the AI total addressable market could hit $3 trillion by 2030, up from a $2 trillion estimate "just in July." No new product; a bigger number and a chart breakout. After six quiet issues, AMD is back in the conversation.
• Nvidia faded -2.9% even as the group rallied. The leader keeps lagging the rotation.
• Custom silicon just got harder to model. On The Circuit, sector analysts flagged that Broadcom named a real new rival inside Google, MediaTek, and that customers now "mix and match" vendors (compute chip from one, networking from another, I/O from another). The clean "whoever wins the program gets it all" math is breaking down. They also declared the "bottleneck trade" over: "We've turned over every rock... you're going to have to be a lot pickier."
• The bubble debate got a real data point, not just vibes. Next in Tech noted Google reported its first-ever negative free cash flow and raised $85 billion of equity in March, the largest offering in history. Fresh bears (Jay Martin, David Rosenberg, Ed Zitron) all built the same "2008, not 1999" financing analogy.
• The best bull rebuttal came from Morgan Stanley's Dan Skelly on The Compound: 2027 AI capex heading to $1.5 trillion vs the Street's $1.2 trillion, and every data center built "embeds guaranteed purchases of servers and chips as far as the eye can see", because a GPU is a 3–5 year asset, not a 20-year one.
• From the Goldman Communacopia conference: CoreWeave's CEO conceded the industry has done a "horrendous" job explaining itself to communities, but said deal economics are improving, recent contracts carry 5–10 points more operating margin.
• Quiet corners: no AMD product deep-dive (MI400/MI450), nothing on TSMC/CoWoS packaging, and no standalone Google-TPU/Trainium/Maia/MTIA episode. Memory went quiet after last week, Micron's late-September report is the next test.

1) The lead: AMD finally shows up, and Nvidia keeps fading

For most of this newsletter's life, AMD has been a name we mention only to note that nobody is talking about it. That ended this week, not because AMD launched anything, but because of a number and a chart.

Here is the two-day tape (FactSet closes):

The AMD move got its clearest explanation on IBD's daily markets podcast (September 8), where the hosts walked through the chart in real time. On the price action:

"Up almost 6%... this was a gap up. It closed up near its high. Took out that round number of 500. So it kind of broke out of a little downtrend today." , Stock Market Today With IBD, Sept 8, 2026

And on why:

"With AMD in particular, the catalyst being the CFO at a conference saying the total addressable market could hit, what, $3 trillion by 2030, up from an estimate of $2 trillion just in July. So that's a pretty big increase there in the outlook in such a short period." , Stock Market Today With IBD, Sept 8, 2026

In plain terms: AMD's finance chief roughly doubled the size of the pie AMD is aiming at in the span of two months, a $3 trillion "total addressable market" (industry shorthand for the entire pool of revenue a company could theoretically go after) by 2030, and, per the hosts, a comment that "the AI super investment cycle [is] still at the very beginning."

Why it matters: AMD is the only credible alternative to Nvidia for merchant GPUs (chips sold to anyone, as opposed to the custom silicon Broadcom and Marvell design for one customer). When it breaks out on a raised long-term forecast, while Nvidia is falling, that is the market saying it wants exposure to the buildout but is done paying any price for the single leader. AMD now trades at roughly 90x trailing earnings with an $850 billion market cap; this is not a value stock catching up, it's momentum broadening to the next-most-obvious name.

Two cautions, both from the same IBD hosts who were describing the rally (this is pundit/technical commentary, not a company view):

"We're past that... rising tide lifting all boats period... the jury is still out on whether we can get back to the kind of market environment for AI that we were in in that April–May timeframe."

And they explicitly warned against chasing: "You don't want to be just kind of haphazardly buying and saying AMD is ready to run and ready to break out to new highs here." The move is real; the all-clear is not.

One gap worth naming: there was no dedicated AMD podcast this week, nothing on the MI400/MI450 roadmap, no operator interview, no deep-dive on where AMD is actually winning sockets. The stock moved on a headline TAM figure and a chart, not on a product catalyst. That's worth remembering if you're tempted to treat +9% as a fundamental re-rating.

2) The custom-silicon trade just got harder to model, and that's the point

If there was one genuinely new analytical idea this week, it came from The Circuit, Episode 191 (September 7), where two sector specialists dissected Broadcom's quarter after the dust settled. Their conclusion is a partial knock on the exact bull case (Broadcom and Marvell custom chips) that has led this newsletter for a month.

Point one: a real competitor now has a name, MediaTek. For quarters, Broadcom brushed off the idea that anyone could take its custom-chip business inside Google. This quarter it couldn't:

"Two quarters ago, they were basically denying competition existed within Google. And here they are now saying, yes, there's competition from MediaTek, but we're delivering faster... And that's a big step back... MediaTek is no slouch in the custom ASIC business." , The Circuit, Ep. 191, Sept 7, 2026

("ASIC" = application-specific integrated circuit, an AI chip built for one customer's exact workload rather than sold off the shelf.) MediaTek, per the hosts, won the "compute tiles", the core processing chunk, on a Google program that many investors had mentally assigned to Broadcom.

Point two, and the important one: the programs are fragmenting. It used to be that if you knew a hyperscaler was building a custom chip, you could hand the whole program to one supplier and multiply out a dollar value. Not anymore:

"Now there's so much mixing and matching happening... sure, you win the compute tile, that's great, probably the highest margins, but then you might use somebody else for networking. You might use somebody else for I/O... That makes it a lot harder to model." , The Circuit, Ep. 191, Sept 7, 2026

This is exactly what Marvell has been quietly telling investors, the hosts noted, that its Google and Amazon wins are in networking and attached products, not the headline compute chip. Different piece of the machine, different margin, different size.

Why it matters: the custom-silicon bull case (Broadcom's $115B FY27 / ~$230B FY28 guide inside a $400–430 billion total market by 2028) is still a big-growth story. But "harder to model" cuts both ways, the hosts think Broadcom beats those numbers, yet admit nobody can cleanly say how the pie splits when customers buy à la carte. That uncertainty is precisely why a stock can post a record guide and still trade sideways.

The hosts also flagged two things buy-siders keep circling:

• Margins. Broadcom's gross margin fell, and management's answer, "don't worry about gross margin, worry about operating margin", landed badly, especially after quarters of Hock Tan boasting about margin strength. (Fair economics, poor messaging, was the verdict.)
• Broadcom vs Nvidia isn't a war with one winner. One host built a spreadsheet comparing their operating leverage: "the result was a tie. They're both growing revenue very strongly, much faster than they're growing expenses."

And the framing that ties this whole issue together, the "bottleneck trade" is done:

"The bottleneck trade... that trade is fully played out. We've turned over every rock. We've searched under every couch cushion. And we know where all the bottlenecks are. So then what do we do now? I think we're in a much more tactical individual company analysis... you're going to have to be a lot pickier and choosier." , The Circuit, Ep. 191, Sept 7, 2026

That is the analytical version of the price tape: the easy "buy anything with a bottleneck" phase is over; from here it's stock-picking. One more useful anchor from the same episode, every hardware maker (Dell, HP, Broadcom, Nvidia) said the same thing this earnings season: growth is capped by supply, not demand. Michael Dell reportedly listed constraints all the way down to "tea glass," and the hosts' takeaway was that 2027 is the peak year of shortages, with real relief not arriving until 2028. "This industry is severely underbuilt."

3) The bubble debate finally has a hard number: Google's cash flow went negative

Every week features somebody yelling "bubble." What changed this week is that the bears stopped waving their arms and pointed at an actual line item.

On Next in Tech (September 8), equity research analyst Bren Daly laid it out:

"Google, for the first time in its history, reported negative free cash flow. This is a company that used to do $70 billion of free cash flow... And it's only going to get worse for the next two years. That's eight quarters. They're projected to be cash flow negative. This is a stunning reversal." , Bren Daly, Next in Tech, Sept 8, 2026

"Free cash flow" is the money left after a company pays for everything, including building things, historically Google's "play money." That it went negative, and that Google had to sell $85 billion of stock in March ("by far the largest equity offering in history... and they could have sold another 15 or 20"), is the cleanest evidence yet that even the richest buyers are now funding the AI buildout out of raised money, not spare cash. Daly's other blunt line on the AI labs themselves: "For every dollar they take [in] revenue, they're burning three" (his characterization of OpenAI; he noted Anthropic is more disciplined and "talking about perhaps getting to break even next year").

Three more bear voices this week sang the same tune, all pundits, all worth steel-manning, none to be taken as fact:

• Jay Martin (Commodity Culture, Sept 9) made the "2008, not 1999" case in the most detail. His claim: future compute orders owed by OpenAI and Anthropic to Oracle, Microsoft, Amazon and Google are "now over $2 trillion," while the hyperscalers' ~$750 billion of annual capex "is more than all the income of those tech companies... So every new shovel in the ground is debt financed." His analogy is the 2006 "2-and-28" subprime mortgage: the whole thing works only as long as the AI labs' valuations keep climbing so they can refinance. Crucially, he's not an AI bear, "I'm super bullish on AI", he's bearish the financing structure on top of it. (The $2 trillion and $750 billion figures are his; treat as his estimates, not company-confirmed.) A useful cross-read for commodity investors: he argues "60 cents of every dollar" of that capex flows into physical commodities, steel, copper, energy.
• David Rosenberg (Rosenberg Research, on Thoughtful Money, Sept 6), an economist rather than a stock picker, argued the buildout "is providing about half the growth in the economy," while non-tech business spending is "flat year-over-year." His warning is subtle and matters for multiples: "The bubble is not in the technology... The bubble is in our behavior." With real interest rates backing up, "basis point for basis point, the move in the P[/E] multiple is far more powerful for equity valuations than earnings growth", i.e. even if the chip companies keep beating, rising rates can compress what investors will pay for those earnings. His tells to watch: credit default swaps (a market gauge of default risk) "rising... especially for Oracle," and Caterpillar, a proxy for data-center equipment demand, "down more than 25%." His bar for the whole thing to pay off: revenues "balloon 50% a year for the next half decade."
• Ed Zitron (Better Offline, Sept 9) is the loudest and most polemical of the group, flag this as opinion, and colorful opinion at that. His substantive points: OpenAI and Anthropic carry "$1.3 trillion of compute commitments" on a take-or-pay basis (they owe the money whether or not they need the capacity), and OpenAI is "expected to spend $750 billion through 2030." He also relayed a real, checkable item: the Financial Times reported that bankers are seeking investment-grade credit ratings for OpenAI and Anthropic IPOs, a move he thinks would just let two loss-making companies pile on cheap debt.

A lighter but telling read from DHUnplugged (Sept 9): the hosts noted AMD, Intel and SK Hynix all rose "in sympathy" with hyperscaler demand, nuclear/small-reactor names (Oklo, SMR) and Bloom Energy popped, but also that 10- and 30-year Treasury yields hit their highest levels of the year, and market breadth is thin. Rosenberg's rate warning and the tape are pointing the same way.

Why it matters: the bear case has evolved from "valuations are silly" to something more precise and harder to dismiss, the buildout is now debt-financed, the biggest customers are unprofitable, and rising rates are the mechanism that could hit the stocks regardless of chip demand. You don't have to buy the doom to take the financing math seriously.

4) The steel-man bull: why "flat after a blowout" might be the healthiest thing here

The best rebuttal to all of the above came from Dan Skelly of Morgan Stanley Wealth Management on The Compound and Friends (September 4), a genuinely rigorous bull case, argued against two sharp skeptics (Josh Brown and Michael Batnick).

The demand argument first: "60% of 2Q earnings [growth]... came from AI infrastructure," and Morgan Stanley's number for 2027 AI capex is $1.5 trillion, versus the Street's ~$1.2 trillion, "the Street is still low." Why won't the hyperscalers cut? Two reasons: "a generational competitive risk" (nobody dares fall behind), and "the scaling laws continue to work" (bigger training runs keep producing better models, so there's a technical reason to keep spending).

Then the point Josh Brown called the best of the whole conversation, and the one most relevant to chip owners:

"When you build these data centers, you are embedding guaranteed purchases of servers and chips as far as the eye can see... it's not a 20-year asset. Like, you're going to need new chips all the time. And there's a story there for Nvidia, story there for Dell." , Josh Brown, The Compound and Friends, Sept 4, 2026

A car factory makes the same model for 20 years; a data center full of GPUs needs replacement silicon every 3–5 years. Even if new construction slows, the installed base becomes a recurring order stream for Nvidia and AMD, a structurally different, and more durable, demand picture than a one-time buildout.

On why the leaders sitting flat after record prints is good, not ominous:

"We don't have to have the rise and fall all take place inside of two weeks... the healthiest thing that has happened in terms of this cycle vis-à-vis 2000 is the rotation that's happening."

Skelly's read of Morgan Stanley's own hedge-fund client book: a lot of them "have not regrossed", they took profits in the first-half AI winners and have not piled back in at size, rotating instead into Mag 7, software and healthcare. Money on the sidelines is the "wall of worry" that lets a bull market keep climbing.

He was honest about the risks, which is what makes it a steel-man and not a pitch:

• Deployment, not demand, is the real constraint, and that creates double-ordering risk across "40 to 50 gigawatts of projected data center construction in the next 2 to 3 years." But he thinks "dark GPUs" are less likely than the "dark fiber" of 2000, because "a lot of those chips can be reverted back to cloud."
• How this differs from the dot-com bust: the spenders are "net cash" Mag 7 balance sheets, and the customers are "Amazon buying on behalf of its 30 million cloud customers" across every sector, not the newly-IPO'd Pets.com and eToys that propped up Cisco in 1999.
• The wildcard he named himself: the two biggest buyers in the whole ecosystem, OpenAI and Anthropic, are "not public yet, not profitable, not particularly transparent yet... and not proven through any sort of economic cycle. That's the wild card." On the circular-financing worry, he pointed to roughly "$320 billion in 2026" of investment-grade debt and off-balance-sheet vehicles, and noted only "Microsoft and J&J are the only two AAA-rated names left. Oracle's CDS spreads trade at very wide levels", i.e. the market is already pricing risk unevenly, as it should.

Why it matters: bull and bear this week are looking at the same facts, record capex, debt financing, private unprofitable anchor customers, and drawing opposite conclusions. The bull's strongest, least-crowded point is the recurring-replacement one: even a capex slowdown leaves a large installed base that must be re-chipped. The bear's strongest point is that the marginal dollar is now borrowed against the promises of two companies that lose billions. Both can be true; the question is timing.

5) From the Goldman conference: CoreWeave says the economics are getting better

The Goldman Sachs Communacopia + Technology conference ran September 8–9, the event we flagged last issue as the first management read after Broadcom's fade. The most useful accelerator-adjacent interview surfaced on Squawk on the Street (September 9), where CoreWeave CEO Michael Intrator (an operator, CoreWeave rents Nvidia GPUs to AI labs) sat with David Faber.

Two things stood out. First, an unusual admission that the industry has an image problem, prompted by the Treasury Secretary reportedly calling its public messaging "horrendous":

"I do not think that as an industry we have done a particularly good job of representing ourselves in terms of what the data centers' impact is on the communities." , Michael Intrator, CoreWeave CEO, Squawk on the Street, Sept 9, 2026

Second, and more relevant to whether the buildout pays, he insisted the unit economics are improving, not deteriorating, even as power and rent costs climb:

"The deal-level economics are excellent... we've added between 5% and 10% of operating income, operating margin, to the more recent deals... they're getting better in spite of the fact that the price of the inputs are increasing, because the amount people are willing to pay for them is increasing faster." , Michael Intrator, Sept 9, 2026

He pointed to CoreWeave's Q2 as an inflection, adjusted operating income of "$128 million from $21 million in the first quarter", while acknowledging the cost side is heavy (an extra $335M of rent, $87M of utilities and power, and sharply higher depreciation). He also said CoreWeave has "a gigawatt worth of pipeline for data centers outside of the U.S." to reach "communities that are less resistant", a concrete sign that the NIMBY pushback bears keep citing is starting to shape where the capacity actually gets built.

Why it matters: this is an operator, on the record, saying the thing pricing power should show: buyers are absorbing rising costs and then some. If that holds, it undercuts the "returns can't justify the spend" bear case at the level that matters most, the individual deal. If it reverses, it will show up here first.

6) Nvidia and the systems names: mostly last week's news, with a few new receipts

Nvidia's big item, the ~$12.9B Hugging Face acquisition, was Issue 022's #2 story, so treat this as follow-through color rather than fresh news:

• On The Six Five (September 7), analysts Patrick Moorhead and Daniel Newman framed the deal as Nvidia wanting to be "the front door for developers" (the logic: developers start at GitHub, then Hugging Face, then deploy, and Nvidia wants to own steps two and three). They expect AMD and Qualcomm to protest the antitrust review, and shrugged at the price ("86x, $150 million [in annual revenue]... it's like a blip").
• The same episode is a good scoreboard for the systems layer: Anthropic signed a $35B compute deal with Nvidia-backed Lambda, the third $30B+ compute commitment in three weeks, with Nvidia itself taking the lease (the building owned by former crypto miner Hut 8). Anthropic's revenue run-rate was pegged at "$65 billion at the end of July," up from ~$10B in 2025. Dell printed "$47 billion of revenue up 58%... 16.4 billion [in] AI servers... backlog... 95 billion"; HPE beat as well.
• Color only: on Moonshots (September 9), Jensen Huang claimed the industry has "reached AGI" and that OpenAI's next model trained on "more than 100,000... Blackwell GPUs." File under demand signal, not analysis.

The number that matters is still the tape: Nvidia -2.9% on the week. The most-owned, most-loved name in the complex is the one the rotation keeps leaving behind.

Where the valuations sit now

For context on why "sell the leader" has room to run (FactSet consensus, retrieved 9/10; fiscal years per each company's own calendar):

Note the tension: Nvidia trades at ~14x next year's expected earnings and is falling, while AMD trades at ~90x trailing earnings and is ripping. The market is not paying up for the cheapest, best-positioned name, it's paying up for the one with the freshest story. That is either healthy rotation or late-cycle behavior, depending on which podcast you believe.

What changed since Monday (Issue 022)

• The rotation matured and rotated up the stack. Issue 022's story was "only the suppliers got paid" (memory and optics up while Broadcom faded). This week the money climbed back into a large-cap GPU, but AMD, not Nvidia, while Nvidia itself joined the faders. The trade is now explicitly "own the buildout, just not the crowded leader."
• AMD ended a six-issue silence on a doubled TAM ($3T by 2030) and a technical breakout. New leadership name in the complex.
• The custom-silicon bull got its first real wrinkle: a named competitor (MediaTek) inside Google and program fragmentation that makes Broadcom/Marvell revenue genuinely harder to size, a partial offset to the story that led issues 019–022.
• The bubble debate got a hard fact (Google's negative free cash flow, $85B equity raise) plus a well-argued bull rebuttal (Skelly's $1.5T capex and embedded replacement demand). It's no longer just noise on either side.
• Hugging Face is now old news and recedes to color.

Next catalysts

• CPI, Friday 9/11, the rate/multiple pressure Rosenberg flagged runs straight through the inflation print.
• Micron (MU) reports late September, the cleanest test of the memory bull, which went quiet this week.
• Oracle earnings (mid-September), the "poster child" for AI debt; watch its CDS spreads, which multiple bears named.
• Marvell Investor Day, October 6, expect it to try to put clean numbers on exactly the fragmented custom-silicon opportunity The Circuit says is hard to model.
• Broadcom Q4 FY2026, December 10; Nvidia next print ~November 18.