Newsletter · 2026-10-08 · Ashutosh Agarwal

Marvell Bets $80 Billion on Connectivity - AI Accelerators: GPUs, Custom Silicon & Optics - Week of October 8, 2026

A synthesis of what podcasts said about AI accelerators for October 1 to October 8, 2026, centered on Marvell's $80 billion fiscal 2031 target and the connectivity thesis, SpaceX's reported $40 billion loan to buy Nvidia chips, Google's two-chip TPU strategy, rising GPU rental prices, and the debate over whether the memory cycle has broken.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of October 8, 2026: Marvell Bets $80 Billion on Connectivity

Issue 031, Thursday, October 8, 2026. Podcasts covered: October 1–8, 2026. Prices are closing prices through Tuesday, October 7, 2026.

On Monday, Marvell (MRVL) told investors it thinks it can turn about $9.5 billion of yearly sales into roughly $80 billion by fiscal 2031. Fiscal 2031 runs to early 2031, so think of it as calendar 2030.

The next morning, CEO Matt Murphy went on CNBC and explained the bet in one sentence. The first wave of AI money went to the chips that do the math. The second went to memory. The third, he says, goes to the wiring between them:

"So you've got a lot of installed compute, a lot of installed memory, but guess what? They all need to talk to each other through high-speed optical interconnections, switching networks. And what you don't want is any GPUs or memories sitting idle."

The stock jumped about 5.8% on Monday, then slipped 0.8% on Tuesday as the market pulled back from record highs.

The same day, a second story showed what is paying for all this. Reports said SpaceX is in talks to borrow $40 billion, specifically to buy Nvidia chips. Apollo would lead the debt and PIMCO may buy part of it.

So this week we saw both sides of the trade at once. Chip suppliers are giving more confident long-range forecasts than this industry has ever given. And the money paying for those chips comes more and more from bond markets, private lenders and the chipmakers' own balance sheets.

I think both are true together. Which one matters more for your portfolio depends on how far out you are looking.

TL;DR

• Marvell's investor day raised the bar for the whole custom-chip group. Murphy says the market for custom AI chips has gone from a $40 billion estimate in April 2024 to $235 billion by 2030. Marvell is targeting $30 billion of that (about 13%), plus a "scale-up optics" business that sells $0 this year and has had its forecast raised three times. Next year's revenue target went from $18 billion to $20 billion, "primarily because of connectivity." (Operator, Squawk on the Street)

• SpaceX's $40 billion Nvidia chip loan is the biggest single sign yet that AI chip buying is moving onto credit. The reported split is $10 billion through banks and $30 billion through private lenders led by Apollo. Apollo's John Cortese wouldn't discuss SpaceX, but said of chip financing in general: "I would expect more of it." (Bloomberg Tech)

• Google's infrastructure chief explained why Google now makes two TPUs instead of one. Amin Vahdat says inference (running trained models to answer users) could be "30, 40, 50, 60% of the market" over a chip's life. That is why Google split TPU 8 into an inference chip (8i) and a training chip (8t). He also said Google has to double its effective serving capacity about every six months. (Operator, Training Data)

• The credit worriers had a busy week. Ed Zitron says CoreWeave bonds issued at about 9% would cost 13–14% to issue today. Steve Eisman says $500 billion of AI debt this year is "crowding out" US Treasuries. Bloomberg's Sarah Fryer says her team tracks more than $1 trillion of "circular" chip-financing deals, mostly Nvidia's. (Pundits)

• Memory: the post-earnings argument has moved from "how good" to "how long." On The Circuit, Ben Bajarin says the market is pricing Micron as if there's a "crash in 29." Jay Goldberg's worry is that Chinese memory maker CXMT can build a fab in 18 months versus 2.5–3 years elsewhere. Micron rose 4.1% on Tuesday. Samsung's preliminary Q3 results landed after Tuesday's close and no podcast had covered them by the end of this window.

• Since the last issue (Oct 2 → Oct 7 close): Astera Labs +9.1% led, then Broadcom +6.0%, Marvell +4.6% and Arista +4.1%. Coherent was the only loser at −0.7%, with TSMC flat. The Nasdaq rose 1.3%. Nvidia closed at $237.47, 2.4% below its 52-week high, with a market value of about $5.75 trillion.

1. Marvell's investor day: the "connectivity gap" thesis, in the CEO's words

Source: Squawk on the Street: "9AM HOUR: Stocks Pull Back From Record Highs, Marvell CEO Exclusive, 'Halftime Report' Rings NYSE Opening Bell 10/7/26" (Oct 7). Speaker: Matt Murphy, Chairman and CEO of Marvell. Operator.

This is the most important interview of the week. Murphy laid out his growth plan in four steps, which he called legs. Here they are, with his own numbers:

Leg 1: Custom AI chips (what the industry calls XPUs).

• Murphy confirmed that Marvell put the custom-chip market at $40 billion in April 2024 and now puts it at $235 billion by 2030. "The custom number, it really ran on us for sure."

• Two years ago, he said, the debate was "can custom even materialize as a portion of the market?" Today that question is "clearly, clearly been solidified."

• Marvell's target is $30 billion of custom revenue in 2030. "Somebody pointed out in the meeting, look, if you just divide that by the $235, it's only 13% market share."

• The nearer number is more useful: "Our business in 2028 is going to be $12 billion plus in custom already." Getting from $12 billion to $30 billion in two years, he argued, "is not a stretch."

Leg 2: The optical links that connect racks to each other (scale-out).

More and more of the cables inside a data center have switched from copper to fiber over the past 15–20 years. Marvell makes the chips that turn light signals into electrical ones. Murphy said this leg is why he raised next year's number: "We took everybody up next year from $18 to $20 billion, primarily because of connectivity."

Leg 3: Optics inside the rack (scale-up). This is the newest leg, and probably the most interesting for anyone who owns optics stocks.

• Today, the links that tie GPUs together inside one rack are almost all copper. "It's all copper today. And it's going to move to optical." When? "Oh, it's going to start next year."

• Here is the forecast history, in his words: "It's $0 this year. And we said next year, actually, two quarters ago, I said it would be $300 million in revenue. I then said it was going to be material higher than that last quarter. And then it's gone up again. So this is ramping now."

• Asked about a fourth leg, he joked: "There's always one more leg."

On the $80 billion target and its range. Morgan Stanley reportedly asked whether Marvell needed to set the bar this high. Murphy pushed back on the idea that he is either cautious or aggressive: "I always look at most probable outcome judged and pragmatic. And we call it like we see it, and we give a range." He said Marvell gave ranges for every business so "investors can take a look, and they can make their own judgments. But we think the midpoint of $80 billion is very achievable."

On politics and data center moratoriums. Some local governments are pausing new data centers. Murphy said Marvell sees "no impact from any of these externalities right now on our market potential. And certainly not right now, even in the near term, on our revenue growth."

Insider buying. Jim Cramer pointed out that Murphy bought stock in the open market twice after earlier sell-offs. Murphy confirmed: "I bought twice in the 70s." The stock closed at $284.68 on Tuesday.

Why it matters.

• The $235 billion custom-chip figure is a market-size claim from someone with an obvious interest in it being large. But the $12 billion-plus for 2028 is a revenue commitment from Marvell's own order book. That is the number to hold him to.

• The scale-up optics comment matters well beyond Marvell. If copper inside the rack starts giving way to optics next year, the market for optical parts gets a whole new category. That would matter for Coherent (COHR), Lumentum (LITE) and the makers of retimer and cable chips, Astera Labs (ALAB) and Credo (CRDO). Credo's business is heavily tied to copper cables, so for it this is both a risk and an opportunity.

• The custom-chip numbers also support Broadcom (AVGO), the market leader in custom chips. If the market really reaches $235 billion and Marvell takes 13%, the bigger share has to go somewhere. Broadcom rose 3.7% on investor-day Monday and 6.0% across the week.

2. The analysts on Marvell: a model "blown out of the water," and one real risk

Source: Schwab Network: "MRVL 'Compelling' Risk-Reward: Morningstar's Bull Thesis in Earnings, AI Buildout" (Oct 7). Speakers: William Kerwin, Senior Equity Technology Analyst at Morningstar; Tom White, host of Schwab's Fast Market. Analyst / pundit.

• Kerwin raised Morningstar's fair value to $360 a share, from $300. "Simply put, the guidance that management unveiled yesterday blew our model coming in out of the water."

• He finds the targets believable because Marvell has a record of setting big targets and hitting them.

• His math: from about $8 billion in sales last year to an $80 billion midpoint in calendar 2030 is 10x in five years, or "upwards of 50 percent, nearing 60 percent annualized growth."

• Why can a chip company see five years ahead now? "We are seeing these multi-year far out spending commitments, trying to lock down deals, lock down supply." He thinks this is true across the AI supply chain, not something special to Marvell.

• The risk he flags is the AI labs. Asked how much depends on OpenAI and Anthropic, he said: "An enormous amount." He thinks the big cloud companies are showing a real return on AI. For the labs, "it's a little bit murkier," and "if you see even one of these large AI model deployers fall off or cut its spending, it can have a big ripple effect through the supply chain."

• Tom White added two details. Management raised fiscal 2028 targets by about $2 billion and put its total addressable market (all the sales it could possibly compete for) at $400 billion by 2030. He also noted that November options were pricing in about a plus or minus $50 move in the stock.

A cooler view of the valuation. On Saxo Market Call: "Europe wobbling again as US posts another all-time high." (Oct 7), the host called the $70–90 billion range "an aggressive forecast," and added that the stock "is largely priced to get there as well," at about a $250 billion market value on roughly $9.5 billion of current sales.

A hotter view. On Wall Street Unplugged: "Did Google just revive the nuclear trade?" (Oct 7), the hosts called chips "no longer cyclical. This is a secular growing industry", meaning growth driven by a long-term trend rather than the usual boom-and-bust. I'd treat that as cheerleading, not analysis, though the same hosts usefully reminded listeners that Marvell was long seen as "the redhead stepchild of the chip industry."

My read. At about $250 billion in market value and roughly 86 times trailing earnings, the $80 billion midpoint is already largely in the price. The upside case needs the top of the range, or the scale-up optics business to turn out much bigger than expected. The consensus price target of $342.83 is about 20% above Tuesday's close.

3. SpaceX borrows $40 billion to buy Nvidia chips, and Wall Street wants more deals like it

The story broke Tuesday and was all over the podcasts by Wednesday morning.

The deal structure: Bloomberg Tech: "SpaceX Eyes Borrowing for Nvidia Chips, Apple Plans Smart Home Push" (Oct 7):

• Bloomberg's chief Wall Street correspondent said the package is $10 billion syndicated through banks and $30 billion through private-market lenders led by Apollo, with PIMCO possibly involved. "It is very clear that the tech story today is the debt story."

• Bloomberg's count: about $360 billion raised this year for data centers, the chips inside them and the power to run them. That is roughly three times last year's pace.

• The warning sign: SpaceX's five-year credit default swaps hit their highest level since they started trading in June. A credit default swap is insurance against a borrower defaulting, so a higher price means more worry.

• Operator voice: John Cortese, co-head of corporate credit at Apollo, wouldn't discuss SpaceX. But he said: "It's not surprising that the largest companies are looking at this marketplace and saying, we'd like to find structured solutions for complicated build-outs... chip financing being one of them... I would expect more of it."

• The bullish credit view: Robert Schiffman of Bloomberg Intelligence called it "wildly positive" and said "CDS has barely budged and bonds have barely budged." He also said "demand for bonds like this are going to be through the roof."

Why SpaceX is buying: Squawk on the Street: "10AM Hour: Stocks Pull Back from Records, SpaceX Raising Money to Buy Nvidia Chips..." (Oct 7):

• CNBC's Seema Modi, citing a source: the financing is for "about $40 billion worth of NVIDIA GPUs" and it is "specifically NVIDIA chips, not any of the chips being designed by the hyperscalers." The talks are "preliminary, still in early stages." Her source also said: "There are no projections as to when the shortage will ease."

• Louis DePalma, William Blair analyst with an outperform rating on SpaceX: "NVIDIA does not give away its GB300s for free." He estimates SpaceX will spend about $50 billion on AI capex this year and "north of $100 billion next year." He added that Anthropic and Google come online as SpaceX compute customers in October. He also claimed SpaceX has achieved "a payback of less than a year." That last claim is unverified, and he has a buy rating on the stock.

• Earlier that morning, on the 9AM hour of Squawk on the Street, a host paraphrased Elon Musk's comment from SpaceX's last earnings call: "We're all about NVIDIA. They are the best-performing chips. Vera Rubin is what we are."

The arithmetic, with a caveat: The Best One Yet: "'Re-Poked' - Facebook's movie sequel. Crocs' jibbitz bars. SpaceX's debt doozy" (Oct 8):

• The hosts said Apollo takes a 1% fee and PIMCO earns 8% interest. They also said SpaceX plans to rent compute to Anthropic and Google "for a billion dollars a month." These are the hosts' own figures, not checked against filings.

• Their math on Nvidia: at a 67% profit margin, a $40 billion order is about $27 billion of profit for Nvidia. That is "more profit for NVIDIA than its entire first 20 years of existence." The math is rough: it applies gross margin to the whole order. But it gets the scale across.

Why it matters.

• This is Nvidia-specific demand, paid for with someone else's balance sheet. It comes one week after Nvidia's $150 billion buyback (Issue 029).

• Put the two together and Nvidia is both returning cash to shareholders and seeing customers borrow heavily to buy its products. The bull case is that demand is so strong that buyers will borrow at 8% to get chips. The bear case is the same sentence.

• Nvidia rose just 1.5% between our issues and closed 2.4% below its 52-week high. The consensus target of $334.45 is about 41% above Tuesday's close.

4. Broadcom's lending, and the case that the financing is the risk

Broadcom's deals with Anthropic came up again and again this week, on both sides of the argument.

What's been reported:

• Bloomberg Intelligence: "Micron Gives Bullish Forecast, Even as Pay Raises Crimp Margins" (Oct 1). Sarah Fryer, Bloomberg Managing Editor for Technology, described Broadcom agreeing to lend Anthropic "as much as $42 billion to lease its chips." "That is what's known as circular financing. We've tracked altogether over $1 trillion in these types of circular financing deals, largely driven by NVIDIA." Circular financing means a supplier lends money to a customer so the customer can buy the supplier's products.

• Wall Street Unplugged (Oct 7). The hosts said Broadcom is talking about raising $60 billion to lend to Anthropic, and that Anthropic is "said to become Broadcom's largest custom chip design customer by 2027." This is unverified. They compared the structure to AIG before 2008. That is their opinion, not reporting.

The bear case, in detail: Monetary Matters with Jack Farley: "Ed Zitron on Anthropic's IPO (S-1), AI Debt, and Counterparty Risk" (Oct 1). Speaker: Ed Zitron, writer and well-known AI skeptic. Pundit; several figures are his own estimates.

• CoreWeave: he says about $7 billion of CoreWeave bonds sold in 2025–26 at around 9% would cost 13–14% to issue at today's prices. "This is why they're doing convertible notes."

• Oracle: its bonds trade at roughly 7.1–7.5% for short maturities and 8.1–8.3% for long ones. He says debt tied to its New Mexico data center trades at 89–91 cents on the dollar, and Meta's Hyperion data center financing vehicle at about 89 cents.

• Broadcom: he describes a $35 billion loan package where Anthropic is liable for only about $5 billion if it goes bust (unverified). His summary: "there is counterparty risk here at scale." Counterparty risk is the risk that the other side of a deal can't pay.

• The link back to chip demand: Anthropic's $252 billion of non-cancellable compute agreements are "factored into analyst expectations for Google, Amazon, and Microsoft. What happens if they don't get paid?"

• On Nvidia vs Broadcom: Nvidia "is, by comparison, actually being a little smarter," because it mostly makes equity investments rather than taking on loans.

The macro version: Prof G Markets: "Steve Eisman: One Company Could Break The AI Boom" (Oct 2). Speaker: Steve Eisman, portfolio manager. Pundit.

• He calls AI "a $500 billion issuance this year" and says it's crowding out US Treasuries: "If there was no AI, none, the 10-year would be much lower."

• On who is weak: "I don't lose sleep over those companies," meaning Meta, Google and Microsoft. "I worry about Oracle, CoreWeave." Oracle is rated BBB-minus, "one level above junk." And: "If something bad ever happened to Anthropic or OpenAI, there'll be no place to hide."

Steel-manning both sides.

• Bull case. The lenders are the biggest, most careful credit firms in the world: Apollo, PIMCO and the investment-grade bond market. Hyperscaler credit is strong. And every podcast this week, bull or bear, agreed that compute is still scarce. Lending against a scarce asset that is earning money is normal finance, not a bubble. Robert Schiffman's point that CDS and bonds "barely budged" backs this up.

• Bear case. The weak links aren't Nvidia or Broadcom themselves. They are the two private AI labs that a growing share of the order book now depends on, and the neoclouds (GPU rental companies like CoreWeave) borrowing at 9% and up. Broadcom isn't just selling to Anthropic. It is lending to it. If the lab stumbles, Broadcom takes the hit twice: lost revenue and a bad loan.

• Where I land. For Broadcom specifically, this is the main reason the stock sits 24% below its 52-week high even as its custom-chip story keeps getting confirmed. The consensus target is about 42% above Tuesday's close, the widest gap in our group alongside Micron's. The Anthropic IPO filing will be the most useful document of the quarter for this debate.

5. Inside Google's TPU program: two chips, six-month doublings, and Gemini designing hardware

Source: Training Data: "Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI" (Oct 6). Speaker: Amin Vahdat, head of AI infrastructure at Google. Operator.

This was the most detailed operator view of custom silicon all week. It speaks directly to the Broadcom and TPU part of our coverage.

Why two TPUs this year.

• About two years ago, Google faced a choice: "in 2026, should we have two chips or one?" It chose two: 8i for inference and 8t for training.

• The deciding factor was how big inference would get: "By 26, we saw inference and serving really taking off. And so having a chip that would be significantly faster for serving that we thought might be 30, 40, 50, 60% of the market in its lifetime, started making a lot of sense for us."

• If inference had been only 2–5% of demand, a specialized chip would not be worth it, "even if that specialized chip is, let's say 2x faster."

• Each chip can still do the other's job. That matters because Google must predict demand "over a six year period, lifetime of the hardware." So Google plans on using each chip for about six years. That is useful context for the GPU depreciation debate covered in Issues 028–030.

The demand number. Asked whether Google has to roughly double its serving capacity every six months, Vahdat said yes, measured as the ability to generate tokens (the units of text AI models produce), not raw chip power. "You have to double the capability of that hardware to generate tokens every six months. And as much or more of that is going to come from software as it is from hardware." He said most efficiency gains come from improving the models. Hardware still delivers "2x or more year over year performance improvements."

How far ahead Google plans. "We are planning hardware two, three, four, five years in advance." TPU 9, 10 "and maybe some others are in concept execution." He also said: "We're using Gemini to design hardware for future Geminis."

TPU vs GPU. He was generous to Nvidia: "NVIDIA is an incredible whole systems company... They give you a really, really strong reference stack." On whether customers treat TPUs and GPUs as interchangeable: "There's for sure a better set of problems that are better suited for one or the other... GPUs are more general purpose than TPUs."

Why it matters.

• Google's capacity needs doubling every six months, and only about half of that comes from hardware. That is a demand curve for TPUs, and so for Broadcom, that does not depend on AI lab financing.

• The split into inference and training chips also backs Marvell's point that compute is becoming "more diverse," which makes the connections between chips more complex and more valuable.

6. GPU rental prices: up, but messy, and nobody agrees on the price

Source: Empire: "Compute Is A Trillion-Dollar Market Trading In Group Chats | Brett Harrison & Andrawes Bahou" (Oct 5). Speakers: Brett Harrison, who is building a compute derivatives exchange, and Andrawes Bahou, whose firm publishes H100/H200/B200/B300 compute price benchmarks on Bloomberg. Market-structure operators.

• Prices are rising across the board. "On average, the price of compute across every single chip type, every single data center grade GPU type has gone up." The reason: "a lot of demand and not enough supply that can make it on time."

• But hyperscaler margins on GPUs are shrinking. "If you look at the filings of the hyperscalers, the margins they make on their GPU compute have shrunk... that's an encouraging signal that this is indeed a commodity."

• Fast growers pay up. Asked whether a fast-growing startup overpays by 10–20% or 2–3x, he said: "probably two X up to two X... for them, the opportunity cost from revenue destruction beats... threshold unit economics."

• Where the price data comes from. "Most compute is not bought off of a website." His firm gets its data from neoclouds' invoicing systems. "We pay them a lot of money."

• Why futures matter. Harrison said a buyer needs one year of compute while "someone really wants to sell five years of compute. There's a whole mismatch of duration there." A futures market (contracts that lock in a price today for later delivery) would let someone else take on years two through five.

A second price-index view: "The Cognitive Revolution": "AI:AM: Was Trump-Xi Anything? ... AWS GPUs Cost 3X & AI Diagnoses Rare Diseases" (Oct 1). The guest, who runs a GPU price index and a GPU benchmarking service, said hyperscalers "charge regularly, consistently, at least two to three times, sometimes more, compared to a typical NeoCloud." He put that down to bundled software, compliance and sticky enterprise relationships. He also said the frontier-model token price index fell from about $4 to about $1.6 between late June and mid-September, a drop of more than 50%. Tokens got much cheaper even as GPU hours got more expensive.

The extreme end of willingness to pay: FYI - For Your Innovation: "Starship Reached Orbit. Here's the $1B-Revenue-Per-Launch Math | The Brainstorm 152" (Oct 7). The ARK hosts relayed a SemiAnalysis report: OpenAI's fastest service tier, running on Cerebras chips, is sold out, "mostly apparently to Jane Street," at about $200 billion per gigawatt. For comparison, "Anthropic is willing to pay $30 billion per gigawatt to rent capacity from SpaceX AI." Both figures are second-hand and unverified. Even so, the gap shows how much some buyers will pay for speed.

Why it matters. Rental prices for GPUs are the cleanest real-time read on whether chips are scarce or oversupplied. Everyone in this week's podcasts who actually measures prices said they're still rising. The rental business is also starting to look more like a commodity market. That is good for chip demand and bad for neocloud margins.

7. Memory after Micron: the debate shifts to 2028–29

Issue 030 covered Micron's results in detail. This week's podcasts moved past the quarter to the bigger question: is the old boom-bust memory cycle really over?

Source: The Circuit: "Has the Memory Cycle Broken? Plus: Nvidia & Broadcom's $40B+ Financing Playbook" (Oct 5). Speakers: Ben Bajarin (Creative Strategies) and Jay Goldberg (D2D Advisory). Industry analysts.

Bajarin's view: the floor has moved.

• Memory industry revenue used to peak at about $100 billion. Now "our model included, you could be at... between a $600 or $800 billion depending on ASP floor." ASP is average selling price.

• The market, he argues, "is assuming that crash comes in 29." He thinks that ignores ongoing growth in the amount of memory sold.

• His reasoning: memory chipmakers will keep older production lines running rather than scrap them, "a little bit more like TSMC... That's never been a dynamic for memory." He would not expect margins "to go back into the thirties" and sees a possible settling point in the "mid sixties, low seventies."

Goldberg's pushback: supply catches up sooner.

• "Prices are basically doubling this year," he said. Meanwhile, the amount of memory produced is growing only "low twenties for DRAM and mid twenties for NAND into 27, 28," and every maker is adding capacity.

• On China: "It takes two and a half, three years to build a memory fab in the U.S., Korea, Japan. It takes 18 months in China." He also cited reports that CXMT is pulling forward its capacity targets. "I think that leads to, we're going to have a lot of memory in 2028."

• His other worry is the PC and phone market. Memory makers are pouring capacity into high-end HBM (the stacked memory that sits next to AI chips), and he noted Tesla is reportedly cutting back on memory in its robots "largely because they just can't get enough memory."

Price action and the next event. On Stock Market Today With IBD: "Stocks Pare Losses As Yields Slash Gains; AbbVie, Apple, Micron In Focus" (Oct 7), IBD's Ed Carson pointed to Micron's 4.1% reversal higher on Tuesday after three down days. He flagged Samsung's preliminary Q3 operating profit, due after Tuesday's close, as something that "could influence that sector." No podcast in this window covered the Samsung numbers. That reaction comes Monday.

Why it matters. Micron trades at about 14 times trailing earnings, by far the cheapest in our group, and sits 47% below its consensus target. That discount is the market betting on Goldberg's timeline. Bajarin's case is that the next downturn will bottom much higher than past ones did, which the current price doesn't reflect.

8. AMD, neoclouds and the rest of the week

AMD: a fresh Wall Street target, and a sober take on market share.

• On Rob Black Show: "The 500 Rises To Record High" (Oct 6), Rob Black read out Citi's move on AMD: target raised from $575 to $800. The reason given was personal AI agents (Meta's Muse, the startup Instinct) increasing demand for regular processors (CPUs), "with AMD as the primary beneficiary." AMD rose 2.8% that day.

• On Market News with Rodney Lake: "Episode 104 | Can AMD Gain Ground on Nvidia?" (Oct 1), Rodney Lake of the GW Investment Institute gave rough market-share figures: "it was 95 and 5 and 90 and 10 now" (Nvidia vs AMD in data-center AI). He put consensus revenue at about $51 billion for 2026 and $88 billion for 2027. His key question: is AMD "going to be a real alternative? Or is it going to remain in this category that it's negotiating leverage" for buyers bargaining with Nvidia?

• AMD closed at $645.86, only 1.3% below its consensus target. That is the smallest gap in our group, so the stock already prices in most of what analysts expect.

Neoclouds: one investor calls a "localized bubble."

• On Chip Stock Investor Podcast: "Are AI Data Centers a Bubble? Navigating the Neo-Cloud Boom" (Oct 6), Nicholas Rossolillo split the large players from the rest. CoreWeave and Nebius are "almost in a class by themselves." Below them are former Bitcoin miners and companies that recently pivoted, including former shoemaker Allbirds and GoPro's move into optical transceivers.

• His argument: "As more capital gets deployed, the productivity of that capital starts to slow down." His fund trimmed its Nebius position and kept its Amazon and Alphabet holdings.

Co-packaged optics: the engineering view.

• On Cisco Podcast Network: "Cisco Optics Podcast Ep67: The Hidden Intelligence Inside Every Ethernet Link" (Oct 7), Keysight's John Calvin explained the next speed jump: 3.2-terabit transceivers with 400-gigabit lanes. Each lane now needs five separate signal-cleanup stages (equalizers), and each one adds delay and burns power.

• He says that power cost is what pushes the industry toward co-packaged optics and co-packaged copper, which put the connection right next to the chip. "It will happen. It's only a matter of time." Standards are the holdup: "every connector vendor out there wants their connector vendor to be the standard." This fits Murphy's timeline for scale-up optics. It also suggests the market will be split across vendors for a while.

Nvidia's revenue expectations. On Saxo Market Call (Oct 7), the host noted that consensus has Nvidia's revenue at about $411 billion for the fiscal year ending January 2027, up from $61 billion in the year to January 2024. Bloomberg estimates have it above $1 trillion by the year ending January 2030. "I'd be amazed if they get there, but that is what is built into expectations."

The key debates, steel-manned

Names in play

Prices are FactSet closing prices. The 52-week-high gap uses each stock's reported 52-week high. Targets are the FactSet consensus 12-month mean price target.

Read-throughs

• Marvell → optics and connectivity chips. If Murphy is right that in-rack links start moving from copper to optics next year, that is a new source of demand for COHR and LITE lasers and for ALAB's connectivity chips. It is a mixed signal for CRDO, whose active copper cables are the thing being replaced over time, though Credo also sells into optics.

• Marvell → Broadcom. A $235B custom-chip market with Marvell at 13% leaves the biggest share for Broadcom. The company with the most to gain from Murphy's market-size claim is his biggest competitor.

• SpaceX → Nvidia and TSMC. $40B of Nvidia-only orders, paid for with debt, adds to Nvidia's FY28 visibility. Every one of those chips goes through TSMC's advanced packaging (CoWoS) and SK Hynix's HBM, as The Best One Yet hosts noted.

• Google TPU split → Broadcom. Two TPU designs a year, plus capacity doubling every six months, means more custom-chip work and more chips for Google's custom-chip partner.

• Credit spreads → neoclouds. If Zitron's 13–14% estimate for new CoreWeave debt is close, smaller neoclouds face higher costs. That supports the "localized bubble" view (Rossolillo) more than any worry about hyperscaler chip demand.

What changed vs Issue 030

• New: Marvell set its five-year plan ($80B midpoint, $235B custom-chip market, scale-up optics forecast raised a third time). Issue 030 only listed the investor day as an upcoming event.

• New: SpaceX's $40B Nvidia chip financing. The financing debate moved from labs and neoclouds to one of the world's most valuable private-turned-public companies borrowing specifically for GPUs.

• New: A detailed operator view of Google's TPU (Vahdat). Earlier issues had to infer TPU demand from Anthropic's commitments.

• Moved on: The Micron debate shifted from the quarter (Issue 030) to the shape of the cycle in 2028–29.

• Repeated: Broadcom's lending to Anthropic. Issue 030 flagged the $42B figure as unverified. This week Bloomberg's tech editor discussed it on air. A larger $60B raise is now being reported but is still unverified.

• Leadership reversed: Optics led Issue 030 (COHR +17.1%, CRDO +12.2%, LITE +11.8%). This week Coherent was the only loser and Credo rose just 0.6%, while Astera Labs (the worst performer last issue at −1.6%) led at +9.1% and Broadcom gained 6.0%.

• Still quiet: No podcast covered Arista, Coherent, Astera Labs or Credo specifically, despite ALAB and ANET's moves. There was also no TSMC preview, no Samsung or SK Hynix results coverage, and no AMD MI450/Helios update. The CME GPU futures launch flagged for October 5 got no direct podcast coverage either. The Empire episode on compute derivatives is the nearest discussion.

Next catalysts: TSMC September sales and Q3 results (mid-October); Samsung and SK Hynix Q3 results (late October); Alphabet Q3 (TPU sales to outside customers); Anthropic's public S-1 (reportedly before Thanksgiving); final terms of SpaceX's $40B deal; Nvidia results (~Nov 18); Marvell Q3 FY27 (Dec 1); Broadcom Q4 (~Dec 10).