Newsletter · 2026-10-04 · Ashutosh Agarwal

Nvidia's Record Buyback and Micron's Blowout Quarter Meet Anthropic's S-1 Debt Worries - Weekly Semis & AI Infrastructure Podcast Recap - Week of October 4, 2026

Weekly Semis & AI Infrastructure Podcast Recap for the week of October 4, 2026. Podcast synthesis on Nvidia's record $150 billion buyback and all-time high, Micron's $54.2 billion quarter at an 87% gross margin and the memory supercycle debate, the agent-driven CPU rally lifting AMD, Intel and Arm, hyperscaler capex heading past $1 trillion a year and increasingly funded with debt, and Anthropic's leaked S-1 with $518 billion in compute commitments that puts Broadcom, CoreWeave and the cloud providers in the counterparty-risk spotlight.

Weekly Semis & AI Infrastructure Podcast Recap

Week of October 4, 2026: Nvidia's Record Buyback and Micron's Blowout Quarter Meet Anthropic's S-1 Debt Worries

Nvidia puts its cash where its mouth is, Micron proves memory is the new bottleneck, and Anthropic's leaked IPO filing gives the AI skeptics their best ammunition yet.

Top of mind this week

This was a week where the money got very real, in both directions.

On Monday, Nvidia ($NVDA) added $150 billion to its share buyback, taking its total authorization to $235 billion to be spent by fiscal 2028. By Friday the stock had hit an all-time high, with a market value above $5.5 trillion. In between, Micron ($MU) reported a quarter so strong it would have been unthinkable two years ago: $54.2 billion in revenue and an 87% gross margin.

But the same week brought Anthropic's leaked IPO prospectus (the "S-1," the document a company files before going public). It shows $518 billion in future computing commitments, most of which cannot be cancelled. That set off the sharpest debate on the podcasts: is the AI build-out a self-reinforcing boom that still has years to run, or a chain of circular financing that ends with someone left holding the debt?

The chip companies themselves have never looked healthier. What worried people this week was the balance sheets of their customers, and their customers' customers.

Dominant themes

1. Nvidia's record buyback reads as a vote of confidence in years of demand

Nvidia's new $150 billion buyback is the largest in corporate history. Most podcast guests took it as a signal that management sees a long runway of demand, not as a sign the company has run out of ideas.

On Bloomberg Intelligence, analyst Jonathan Bloxham said the size reflects confidence in "a healthy forward pipeline rather than undervaluation." He pointed out that Nvidia expects to generate roughly $200 billion in cash flow this year and about $330 billion next year. In his view, the biggest limit on AI spending is now supply (chip factory capacity, energy, how fast data centers can be built), not demand. Bloomberg Intelligence

On Squawk on the Street, Bernstein's Stacy Rasgon kept his outperform rating and $400 price target. He noted Nvidia's 80%+ revenue growth and 100%+ profit growth, even though the stock had lagged other AI names by about 30% over the past 12 months. In his words, Nvidia is "the ones creating the market that everybody else is playing." Squawk on the Street In an earlier hour, DA Davidson's Gil Luria ($300 target, buy rating) said the stock trades at about 18 times earnings and could head toward a $10 trillion valuation within a few years. Squawk on the Street

Alongside the buyback, Nvidia launched an "Open Agent Safety Platform": software (OpenShell and Sentry) that runs on its own CPUs and BlueField DPUs to fence in AI agents that misbehave. More than 100 partners signed up, including Cisco, CrowdStrike, Dell, JPMorgan, SAP and Anthropic. OpenAI was notably not on the public list. Squawk Pod Jensen Huang's framing on CNBC was blunt:

"We hope it's an engineering problem. I believe it's an engineering problem. I know it's an engineering problem... If it's not an engineering problem, it's not solvable."

By Friday, Nvidia hit an intraday record. Closing Bell noted it would need about another 7% to reach $6 trillion, and that the Philadelphia semiconductor index led the market that day. Closing Bell On Fast Money, Morgan Stanley was reported to have reinstated Nvidia as its top semiconductor pick. CNBC's "Fast Money"

2. Micron's quarter: memory is now the scarcest part of the AI machine

Micron's fiscal fourth quarter was the semis story of the week. On Closing Bell, the numbers were laid out against expectations: earnings per share of $33.42 versus $31.61 expected, revenue of $54.23 billion versus $51.0 billion, and gross margin of 87% versus 86.3%. Next quarter's margin guidance came in a touch light, at about 86.3% versus 86.6% expected. Closing Bell The company guided next quarter's revenue to about $61.5 billion. Futurum Equities Podcast

The important part was not the beat. It was how Micron is trying to make the boom last. CEO Sanjay Mehrotra told Squawk on the Street that the company now has 26 strategic customer agreements (up from 16 in August) running through 2031. He also said the slight margin "miss" came from a decision to raise employee incentive pay, and that Micron plans $250 billion of investment across Idaho, New York and Virginia. Squawk on the Street On The Rundown, the CEO was quoted saying 75% of 2027 supply is already committed, with some contracts running five years and including price floors and customer deposits. The Rundown

The Futurum Equities hosts went deepest on where the growth actually came from. The beat was almost all price, not volume. DRAM revenue rose 27% from the prior quarter on only mid-single-digit growth in shipments. NAND revenue rose 42% on about 10% volume growth. Long-term agreements now cover 35% of revenue through 2030. Futurum Equities Podcast One host said the memory shortage is now so severe that it is backing up into the rest of the chain:

"I was talking to some insiders this week and literally there are accelerator companies giving wafers back to TSMC right now because they can't supply. They can't secure the memory. Which by the way is extremely bullish for Nvidia because Nvidia has secured memory more effectively than any other company."

The same host predicted Micron, with $70 billion of cash now on its balance sheet, will become "the NVIDIA of the next couple of years" as an investor in its own ecosystem. And yet the stock barely moved. Bloomberg Intelligence's Jake Silverman said it fell 1.9% because expectations were already priced in, and that meaningful new chip factory capacity does not arrive until 2028-2029. Bloomberg Intelligence

3. AI agents are creating a second chip boom, this time in ordinary processors

A quieter theme that came up across several podcasts: AI "agents" (software that carries out multi-step tasks on its own, like Meta's new Muse assistant) need far more general-purpose processors (CPUs) than chatbots ever did. That is reviving demand for chips from AMD, Intel and Arm.

On The Six Five, Futurum's Daniel Newman said his team had raised its 2030 forecast for AI-related CPU revenue from $110 billion to $250 billion on September 18, before Meta's Muse launch, and now expects to "probably double that again." The hosts also noted that AMD crossed a $1 trillion market value this week. The Six Five with Patrick Moorhead and Daniel Newman

CoreWeave's CEO made a similar point on Squawk on the Street. He said the company brought its first customer, Cognition, into production on Nvidia's new Vera Rubin systems, and predicted a big jump in CPU demand from the "agentic explosion" that would help AMD, Arm, Intel and Nvidia. Squawk on the Street ThursdAI reported Cognition seeing a 4.8x improvement in how many agents it could run on the same hardware. ThursdAI

4. Hyperscaler spending keeps climbing, and is now being paid for with debt

On The a16z Show, the team put numbers on the scale. The five largest cloud and AI spenders (Alphabet, Amazon, Meta, Microsoft and Oracle) are on track to spend about $780 billion on capital investment in 2026, up from $416 billion in 2025. Every forecast points to more than $1 trillion a year from 2027. Their argument against calling it a bubble: the stock market is up about 20% while valuation multiples are down about 20%, so the gains are coming from earnings, not hype. Memory companies trade at "six times, seven times forward earnings." The a16z Show

The new part is how it is being paid for. On Unhedged, the FT team cited Goldman Sachs: these companies have borrowed about $200 billion in AI-related debt so far this year, with about $1 trillion expected through 2030. Investors are starting to ask for higher interest rates, less because they doubt the companies and more because spending plans keep getting revised up every quarter. Unhedged The Financial Exchange Show noted that for the first time, 2026 spending is projected to exceed these companies' operating cash flow, just as the 10-year Treasury yield jumped about 50 basis points (half a percentage point) in September. The Financial Exchange Show Bloomberg Surveillance went further, citing projections that long-dated data center bond issuance could rival Treasury issuance by 2027. Bloomberg Surveillance

The Canadian Investor added the long view: those five companies spent $71 billion in fiscal 2019 and $586 billion over the last twelve months, roughly 8x. Meanwhile their return on invested capital is falling across all five. The Canadian Investor

5. Anthropic's S-1 puts a price tag on AI lab risk, and on the chip companies tied to it

The leaked Anthropic filing ran through dozens of podcasts. The core numbers: $518 billion in compute commitments, of which $413 billion are non-cancellable. That includes $252 billion across Microsoft, Google and Amazon, and $161.2 billion in Broadcom TPU lease obligations. Hard Fork reported about $7 billion spent on compute in 2025 and a $42 billion loss (about $34 billion of it from accounting on convertible notes), with two unnamed customers making up nearly 25% of revenue. Hard Fork

For semis investors, the point was where that money lands. Bloomberg Daybreak reported Broadcom ($AVGO) leading a $60 billion financing group for AI chip infrastructure that benefits Anthropic. Bloomberg Daybreak: US Edition The Elon Musk Podcast listed other commitments including $45 billion with Nscale and $35 billion with Lambda. Elon Musk Podcast See Key debates below for how the bulls and bears read it.

6. The real bottleneck is still places to plug chips in

Chase Lochmiller, CEO of Crusoe (which builds AI data centers, including the Stargate site in Abilene, Texas), gave the most concrete picture of the supply crunch on 20VC:

"Where it's actually manifesting is there are not places to plug in GPUs. So that's ultimately the supply constraint."

When Crusoe set out to build its first two Abilene buildings (a little over 200 megawatts) in one year, the next-fastest bid was two and a half years. One part alone, the power distribution center that steps down medium-voltage power, had a 100-week lead time. He also pushed back on the popular worry that GPUs wear out economically in three years. Crusoe depreciates over six years, and he said that rental prices for Nvidia's three-year-old Hopper chips are now higher than when they were new. Right now, he said, managed GPU clusters are "incredibly high margin" because of "a massive shortage of supply." The Twenty Minute VC (20VC)

Power showed up everywhere else too. Energy and grid podcasts this week cited 230 gigawatts of data center grid connection requests pending in the US Clean Power Hour, and 5-15 year waits to connect in the UK. Wake Up to Money One contractor on The Construction Corner said medium-voltage cable lead times jumped from 30 weeks to 52. The Construction Corner

Key debates

Is Nvidia's lead safe, or are cheaper chips catching up?

• Ben Pouladian (Monetary Matters, bull on Nvidia): He argued that "all roads and rockets lead to NVIDIA." As AI matures, the winning metric will be the cost of producing tokens (the units of text an AI model reads and writes) per megawatt of power, the way oil companies are judged on cost per barrel. Nvidia's full system wins on that: GPUs, networking, software and now agent security all tuned together. He was skeptical of AMD's GPUs. AMD's Helios rack is "basically twice the size of a Blackwell system," weighs about eight tons, and he is "already hearing that there's potential delays in scaling it." He also said Amazon's Trainium chips "sort of kind of fell behind," and argued that most of Anthropic's half-trillion in commitments are likely Nvidia GPUs and Google TPUs, not Trainium. Monetary Matters with Jack Farley
• Jack Farley (host, pushing back): He is overweight Nvidia himself, but noted that AMD and "almost every other semiconductor stock" have beaten Nvidia this year. Google and Amazon, he said, are "not walking around being like, oh man, I wish I spent a hundred billion dollars more" on Nvidia instead of their own chips.
• Raja Koduri (TechSurge, says the battleground is moving): The former graphics chief at AMD and Intel argued that for inference (running a trained model, as opposed to training it), the bottleneck is now memory, not raw processing. He said an old 7-nanometer TSMC chip stacked directly with DRAM could get "10x the bandwidth than the current HBM and 10x token generation rate. That package will beat Vera Rubin." His bigger warning was about China: a gigawatt of AI capacity costs $50-60 billion today (about $45 billion of it for GPUs and related gear), while "the target for China Inc. for gigawatt is to be less than 10 billion dollars." That is a five-to-six times cost gap. He also estimated the world needs over 400 gigawatts of new compute by 2030, about $24 trillion of capital: "I don't think we have that capital." TechSurge: Deep Tech Podcast
• Walter Goodwin (No Priors, Fractile CEO): He made a related point. Nvidia, Google's TPU, Meta's MTIA, Microsoft's Maia and OpenAI's new Jalapeño chip are all "relatively identikit": the same HBM memory, the same kind of math engines, the same TSMC packaging. Fractile is betting on about 25x more memory bandwidth per chip. He noted that a single Nvidia system contains "anywhere between kind of six and nine custom chips." No Priors

Is the memory boom different this time?

• Bulls: Melius's Ben Reitz (buy, $2,200 target) argued the stock doubles from here. In his view HBM margins can stay in the mid-to-high 80s rather than falling to the 60s, and buybacks of 10-15% a year will support the shares. Squawk on the Street Susquehanna's Mehdi Hosseini (buy rating) said memory, not compute, will be "the defining value driver" of AI infrastructure. Squawk on the Street BofA's Vivek Arya raised his AI market forecast from $1.8 trillion to $2.2 trillion by the end of the decade. He cited demand growing more than 100% a year against capacity growing 40-50%, and noted memory is about 50% of a data center's bill of materials. Closing Bell On Schwab Network, Ross Gerber argued Micron "is no longer cyclical." Schwab Network
• Skeptics: Morningstar's Dave Sekera on The Morning Filter rates Micron two stars and says it trades at a 27% premium to fair value. His view: new production arriving in 2028 should satisfy demand "even [including] the AI buildup boom," and once supply catches up, "prices are naturally going to start coming down." If management can't convince investors the shortage lasts past 2028, "I wouldn't be surprised to see this one roll over and gap to the downside." The Morning Filter On Saxo Market Call, the host noted that Micron's operating income is now more than half of Nvidia's, yet it trades at less than a quarter of Nvidia's value, because the market expects the cycle to turn. Saxo Market Call
• Middle ground: BofA's Arya also said investors are in "fool me once" mode after the leveraged June rally in semis unwound. He likes that, because long-term money can now take a calmer look, "but it's not going to happen right away." Closing Bell

Is AI financing becoming circular, and who carries the risk?

• Ed Zitron (Monetary Matters, bear): He called Anthropic "a dog... not a good business" and argued the risk spreads well beyond the lab. The $252 billion in non-cancellable cloud commitments "are factored into analyst expectations for Google, Amazon, and Microsoft. What happens if they don't get paid?" He singled out Broadcom: in one $35 billion loan arrangement, he said, Anthropic is only responsible for $5 billion if it goes insolvent. He contrasted that with Nvidia, which mostly makes equity investments (CoreWeave, xAI, Lancium) rather than taking on debt, and is "actually being a little smarter." He also said CoreWeave's bonds, repriced at today's levels, imply borrowing costs of 12-13%. Paraphrasing Dario Amodei's own comment from February, he said that if Anthropic misjudges how much compute it buys, it goes bankrupt. Monetary Matters with Jack Farley
• Dave Collum (BTC Sessions, bear): He called Nvidia's buyback at about 30x earnings "late 90s dot-com" behavior, and described a "vendor financing" loop in which customers use the chips themselves as collateral to borrow more. BTC Sessions Jesse Felder (Soar Financially) said debt markets are already tightening, with Nvidia credit default swap spreads up 100% and CoreWeave facing junk-level yields. Soar Financially
• The other side: Crusoe's Lochmiller said GPU rental contracts are typically "take or pay" (the customer pays whether or not it uses the capacity), and that he doesn't see weak consumer demand as the first crack: "we're honestly just scratching the surface." The Twenty Minute VC (20VC) On Merryn Talks Money, Stephen Yu said Nvidia at 16-17x earnings has no valuation bubble, only "earnings bubble" risk, and that this is unlikely to break in the next 18 months given how long customer commitments run. Merryn Talks Money On Empire, the hosts disagreed about whether Anthropic's flat revenue chart reflects weak demand or simply a lack of compute. One said Anthropic "keeps telling me to stop using it." Empire

Does the spending ever earn a return?

• Bearish math: Odd Lots cited an economist's estimate that justifying current AI spending would need revenue equal to about 9% of GDP from a handful of companies. The guests also cited Goldman's view that about half of recent earnings growth is driven by hyperscaler spending itself. Odd Lots Empire quoted payback periods of up to 35 years on $50-60 billion-per-gigawatt builds at current rental rates. Empire
• Bullish reply: Gene Munster on Fast Money predicted hyperscaler spending growth will top 60% next year versus the Street's 40%, driven by agents that today reach less than 1% of potential users. CNBC's "Fast Money" An investor on Squawk on the Street called it "a 10-year cycle. You're three and a half years into a 10-year cycle." He added: "We'll still be short power. We'll still be short compute." Squawk on the Street

Stocks discussed with bull/bear angle