Newsletter · 2026-10-01 · Ashutosh Agarwal

Nvidia's $150 Billion Buyback and the High Bar for Micron - AI Accelerators: GPUs, Custom Silicon & Optics - Week of October 1, 2026

A synthesis of what podcasts said about AI accelerators for September 24 to October 1, 2026, centered on Nvidia's new $150 billion buyback and the circular-financing debate, its $180 billion of Anthropic hardware commitments, the high bar going into Micron's fiscal fourth-quarter report, rising GPU rental prices, fresh skepticism about AMD's Helios rack, and Raja Koduri's case that China is the real competitor.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of October 1, 2026: Nvidia's $150 Billion Vote of Confidence

Issue 029, Thursday, October 1, 2026. Podcasts covered: September 24 – October 1, 2026. Prices: FactSet closing prices, September 25 and September 30, 2026; average analyst targets and EPS estimates are FactSet consensus as of October 1, 2026. Unverified speaker claims are marked as such in the text.

On Monday morning, September 28, Nvidia (NVDA) did something that many people who have followed this cycle didn't see coming.

It authorized another $150 billion of share buybacks. A buyback is when a company uses its own cash to buy its stock back from the market, which leaves fewer shares and gives each remaining share a bigger slice of the profits. Nvidia says it expects to have bought back $235 billion of stock by fiscal 2028.

That came just four months after it added $80 billion to the program.

Bernstein's Stacy Rasgon, who rates the stock Outperform with a $400 target, did the math on CNBC's Squawk on the Street that morning:

"$235 billion by the end of fiscal 28 is over the next six quarters. It's like $40 billion a quarter, which is double what we were modeling."

Here is why I think this matters more than one big number.

For a month the podcasts have run one bear case after another. In Issue 027 it was debt moving off balance sheets. In Monday's Issue 028 it was Ed Zitron's claim that huge numbers of GPUs are sitting in warehouses without power. Both stories come down to one question: is the money going around in a circle, with Nvidia funding the customers who buy its chips?

The buyback is Nvidia's answer, given in cash rather than words. It says: we earn more than we can sensibly reinvest, and we're handing it back.

The stock moved a little but didn't take off. Nvidia rose +1.47% from the September 25 close to the September 30 close, ending at $228.38. The Nasdaq fell 0.77% over the same days. Most of the rest of the chip group fell too.

And on Wednesday night Micron (MU) reported its fiscal fourth quarter, the event the whole memory trade had been waiting for. The episodes below are all previews, recorded before the results, and they set up the bar very clearly. Micron's CEO Sanjay Mehrotra is booked on Squawk on the Street this morning at 9 AM ET. Monday's issue will cover the reaction in full.

So this issue is about two things: what Nvidia's cash is telling us, and how high the bar is for memory.

TL;DR

• Nvidia's $150B buyback reframed the "circular money" debate. Bernstein's Rasgon says the pace (~$40B a quarter) is double his model, yet still under 5% of the company's market value. On Bloomberg Tech, Bloomberg Intelligence credit analyst Robert Schiffman called the circular-financing worry "a great media buzz line." Nvidia also disclosed $180 billion of Anthropic hardware commitments.
• Nvidia's valuation has shrunk. The Exchange (CNBC) noted Nvidia's forward P/E (share price divided by expected earnings per share over the next year) has fallen from about 66x five years ago to 18x. Broken Pie Chart pointed out that Costco trades at 40.5x.
• Micron walked in with a high bar. Susquehanna's Mehdi Hosseini ($2,000 target) said on Squawk on the Street that memory, not compute, is now "the defining link." Options were pricing a move of roughly 6.5%–8% either way. The stock was trading at around 6–7x forward earnings, a price that only makes sense if investors think today's profits won't last.
• GPU scarcity data grew stronger. On Frictionless, P Equity Research cited CLSA's estimate of a 73% supply-demand gap for GPUs and custom chips lasting until 2030, plus B200/B300 rental prices going from $5 to $7–8 an hour.
• The challengers drew fire. On Monetary Matters, long-time Nvidia bull Ben Pouladian said AMD's Helios rack is "twice the size of a Blackwell system," "weighs like eight tons," and may face delays. That last part is hearsay and unverified. AMD fell 3.0% this week.
• A former AMD and Intel graphics chief says China is the real competitor. Raja Koduri (founder of Oxmiq) said on TechSurge that a gigawatt of AI data center costs the West $50–60 billion, while China is aiming for under $10 billion.
• Optics split. Credo (CRDO) fell 7.7% this week, the worst in our group. Lumentum (LITE) was the best performer (+3.1%) and got one investor's endorsement on The Exchange.

The scoreboard

FactSet closing prices. "This week" runs from the September 25 close to the September 30 close, which is the same span as the change since Issue 028. Micron's September 30 close came before its results.

What jumps out:

• Only three of ten names rose. This is the second issue in a row where the group didn't move together. In Issue 028, Credo and Astera Labs (ALAB) each gained about 20%. This week Credo gave back a big part of that.
• Nvidia beat the group. That's rare this year. On Squawk on the Street, one of the anchors noted the stock is up only about 30% over twelve months, "pretty much on par with, like, Archer Daniels Midland."
• Broadcom (AVGO) is still the cheapest name relative to Wall Street's targets (+52% to the average target), now tied with Micron (+50%) after Micron's pre-earnings drift.
• AMD is nearly at its analysts' average target (+3.9% upside). The market is pricing it for near-perfect execution, which is the title Futurum gave its episode last week.

1. Nvidia's $150 billion answer to the "circular money" question

What happened

On Monday, September 28, Nvidia announced the extra $150 billion buyback authorization, along with new software to keep AI agents in check ("OpenShell" and a monitoring tool). It also disclosed that its order book keeps growing.

CNBC's Seema Modi set the scene on The Exchange, "Nvidia's Security Push, Consumer Conundrum, and the Diesel Divide 9/28/26":

"$150 billion announced this morning, the largest in history, four months after adding $80 billion to its program. And just as the stock has lagged its peers in the month of September, its price-to-earnings ratio, by the way, has fallen from about 66 times forward earnings five years ago to 18 times."

She added that Nvidia is now invested in 13 public companies and over 200 private ones.

The sell-side view: big, but not that big

Stacy Rasgon of Bernstein, on Squawk on the Street, "10AM Hour: Nvidia's $150B Buyback, Mortgage Rate Outlook, Meta Pulls Back 9/28/26", had the most useful framing. The pace is aggressive: about $40 billion a quarter, double his model. But measured against free cash flow (the cash left after running the business and paying for investment), it's only:

"about 50% of my free cash flow across fiscal 27 and 28. So it's kind of in line maybe with what they've talked about."

And measured against a company worth over $5 trillion:

"In the context of a $5 trillion market cap, it's still not that big. And presumably, they're not going to be finished by the end of 28."

The anchor compared it with Apple's 2018 $100 billion buyback, which was about 11–12% of Apple's market value at the time. Nvidia's $235 billion is under 5%.

Rasgon also gave two reasons why the stock has lagged even as earnings sped up, and both are worth knowing:

• Investors have been playing the bottlenecks. For most of the year, he said, people bought memory, chip equipment, networking, power and more recently CPUs. Those offer "more torque," meaning bigger percentage moves in earnings, than a company Nvidia's size.
• Nvidia is simply too big to own more of. "It's 8% of the S&P or whatever it is. And for some of my clients, it's just hard for them to own more of it." He noted that Samsung and SK Hynix together make up about two-thirds of Korea's Kospi index, so in his view weights this large can work.

His conclusion: "if they're not ultimately successful, I don't think anybody else can be. So I still like it."

The credit view: Nvidia is "acting as a bank"

The sharpest reply to the circular-financing worry came from someone who reads balance sheets for a living. Robert Schiffman, a credit analyst at Bloomberg Intelligence, on Bloomberg Tech, "Nvidia Goes Big on Buybacks as SpaceX Reaches Orbit":

"circular financing is a great media buzz line to try to bash this space. And I just don't think it's working... They're making credit-style investments where they're both leasing facilities. They're providing residual value guarantees. They're providing support behind the value of their chips beyond their depreciable life. I just think what they're doing is they're just acting as a bank for everybody, which is not a bad thing."

A "residual value guarantee" is a promise that a chip will still be worth a set amount after a certain time, which makes lenders more willing to finance it. Schiffman's point is that Nvidia is lending out its strong balance sheet. Chipmakers are "collecting right now," and they're sharing that money with the builders, the hyperscalers (the giant cloud companies) and the neoclouds (newer GPU-rental companies). He expects that to flip "in the next couple of years" once those facilities are running and the builders turn free-cash-flow positive.

The show's host said Nvidia had just north of $99 billion in cash and short-term investments at the end of its fiscal Q2 FY27. Schiffman also pointed to a deal Nvidia just announced with large alternative asset managers to help raise $500 billion: "I'm surprised it wasn't a trillion."

Here's the skeptic's version, from Marketplace on Marketplace All-in-One, "Nvidia's record-breaking stock buyback":

"Companies are filling the dump trucks with money and sending it to NVIDIA to buy their chips. NVIDIA is getting dump trucks of money every day. That's why it can do stock buybacks."

The program noted that Nvidia "has committed hundreds of billions to the chip supply chain and invested in AI companies that turn around and spend their money on more NVIDIA chips." Futurum's Daniel Newman gave the other side: "When you have a balance sheet that looks like NVIDIA, you have the optionality to return to shareholders, and you have the optionality to invest in the ecosystem."

The $180 billion Anthropic number

Hidden among Monday's news was a disclosure that matters more for revenue visibility than the buyback does. On the same Bloomberg Tech episode, the host said Nvidia had disclosed that Anthropic alone has $180 billion of Nvidia hardware commitments across its different deployments.

Erica Clower, founder and CIO of Science and Technology Partners (a long-biased fund), read the buyback as a cash signal:

"NVIDIA is saying not only are we stepping in to buy more of our own shares and a vote of confidence about the future of the industry, but we also have the cash to do it."

She also gave a useful reminder of how much damage the group took below the surface: "we've seen a tremendous correction in technology shares since June. In many cases, there are companies, premier companies that are down 30, 40, 50%. So I think we're in the midst of a bottoming process."

The pundit math (handle with care)

On The Compound and Friends, "Bad feeling, weak internals, confidence collapse, Nvidia breaking out | WAYT?", Matt Cerminaro calculated that if the stock stays flat, $235 billion of buybacks through Q1 2028 would cut the share count by about 4.3%. Josh Brown then offered a much bigger, unverified estimate:

"This is a company that this fiscal year is going to do something like $96 or $98 billion in free cash flow. Next year, it could be $360 billion in free cash flow on $680 billion in revenue."

That's Brown's own extrapolation from capex plans. It is not company guidance, so treat it as an opinion. Cerminaro added that the timing was a statement in itself: Jensen "dropped this, like, two weeks after Dario came out and said to slow the pace," a reference to Anthropic CEO Dario Amodei's recent call to slow AI development.

Michael Sansaterra, CIO of Sylvan Capital Management, for whom Nvidia is the largest holding, gave the growth case on The Exchange. Free cash flow has grown "something like 190 percent for the last three years," with forecasts calling for 75% growth next year. On rates: "The cost of debt to NVIDIA is insignificant."

Why it matters

• The bear case is losing its best argument. "Circular money" works as a bear case only if Nvidia's cash is fake. A $40 billion-a-quarter buyback is hard to pay for with fake cash.
• But it doesn't settle the deployment question. Monday's debate was about whether chips are getting powered on, and the buyback says nothing about that. The data on GPU rental prices (Section 3) does.
• The valuation gap is now the trade. At $228.38 and FactSet consensus EPS of $9.27 this fiscal year and $15.75 next, Nvidia trades at about 24.6x this year and 14.5x next year. Analysts' average target implies 46% upside.

2. Micron: the bar going in

Micron reported after the bell on Wednesday, September 30. The episodes in this window were recorded before the numbers came out, and the figures were not checked against Micron's own release in time to publish them here. So here's what the market was asking for, so you can grade the results yourself.

The bull case: memory is the new bottleneck

Mehdi Hosseini, senior analyst at Susquehanna (Buy, $2,000 target), on Squawk on the Street, "10AM Hour: Contrarian Market Call, Micron Results Preview, New Eli Lilly Weight-Loss Data 9/30/26", made the boldest claim of the week:

"I make an argument that compute is actually becoming more commoditized. And it's memory that makes the difference... we expect Micron earnings to reach $30, $40 on a quarterly basis. And ultimately, Micron should be able to hit a $50 quarterly earning or $200 annualized."

His test for the print was not the quarter itself but whether margins can last. He expected management to talk up capex discipline and "how they're able to sustain operating margin in the 70%." Operating margin is the share of revenue left after running costs. He framed the whole AI trade around one question: "who can have a 70% margin profile in AI. Is that memory? Is that foundry? Is that customers like NVIDIA and AMD?"

The anchor added some context: the stock was up "nearly 300 percent so far this year" but has been flat since early June. It "popped 16 percent" after last quarter's report. And 95% of covering analysts rate it a Buy, with an average target implying 50% upside. FactSet's average target gives the same +50% from the September 30 close.

What the numbers looked like going in

Rodney Lake, director of the GW Investment Institute, gave the clearest walk-through of the numbers on Market News with Rodney Lake, "Episode 103 | Micron's AI Memory Boom and What Comes Next". These were figures he cited in mid-September:

• Market value: about $1 trillion, up from $62 billion in 2022.
• Trailing 12-month revenue: about $90 billion, versus $37 billion in fiscal 2025.
• Gross margin: 72% on the latest data. That compares with negative gross margins in the 2023 downturn and 23% in 2024.
• Projected revenue: about $129 billion for fiscal 2026 and $249 billion for fiscal 2027.
• Competition: He named China's CXMT, which listed this summer, as a fast-growing fourth player in high bandwidth memory (HBM, the stacked memory that sits next to AI chips).

Ed Carson of Investor's Business Daily, on Stock Market Today With IBD, "Market Bounces To Cap Strong Week For Growth; Microsoft, Amphenol, Bloom Energy, Micron In Focus", cited expectations of about 940% year-over-year earnings growth and 351% revenue growth for the quarter. He named the one outcome that would hurt:

"If they say we're going to have one more good quarter and then boy, oh boy, nobody, the pricing is going to go down... you could imagine that really negative reaction for Micron and the, you know, the broader memory space."

The traps hidden in the numbers

The most useful preview came from Futurum's analysts on Futurum Equities Podcast, "Meta is up big on Muse. AMD is priced for no mistakes. Micron is next. | Ep 58". They flagged an accounting detail that anyone reading this morning's headlines needs to know:

"Q4 has 14 weeks in it. Q1 only has 13 weeks in it. People aren't going to normalize their guidance and they're going to jump to conclusion. Sequential growth is lighter than expected."

In plain English: the quarter just reported had an extra week. So even a strong forecast for next quarter may look like slowing growth until you adjust for that week.

Their three other things to watch:

• Strategic customer agreements (SCAs). These are Micron's multi-year contracts, and the question is whether they make "the next downturn look different from previous cycles."
• HBM4 share. Can Micron hold about 20% of the market as the industry moves to the next generation?
• Capex. Spending plans show whether management expects a big wave of new supply in 2028 or sooner.

They also quoted consensus going in: $50.9 billion in revenue and $31.49 EPS. Stacy Rasgon gave his own short list on Squawk (9/28): "ASP upside and whether or not the long-term agreements that they put in have capped their ability to continue to take price." ASP is average selling price.

What options traders were betting

Options prices tell you how big a move the market expects. Three podcasts did the math:

• OVTLYR. Christopher Uhl said on How to Trade Stocks and Options Podcast with OVTLYR Live, "STOP Gambling on MU Earnings – Do This Instead" that options implied a move of about $69 either way (about 6.5%). He also reminded listeners that after Micron's June report the stock fell 41% from peak to trough.
• Broken Pie Chart. On Broken Pie Chart, "SpaceX Starship Reaches Orbit | S&P 500 Breadth Lowest Since Dot-Com | 59% of S&P 500 Stocks in Bear Markets...", the hosts estimated an at-the-money straddle (buying both a call and a put, which pays off if the stock moves a lot in either direction) would need a move over 8% to break even. They put Micron at 7.1x forward earnings.
• Saxo Market Call. On Saxo Market Call, "Burnham declares end of Thatcherism. Micron earnings on tap.", the host put Micron's quarterly operating income at "on the order of $33 billion... more than half of what NVIDIA does." Yet Micron is valued at "less than a quarter of NVIDIA's $5.x trillion."

Why the multiple is so low

Every preview came back to the same tension. Micron trades at a single-digit multiple because history says memory runs in cycles. The best explanation of why this cycle might be different, and why it probably still is a cycle, came from P Equity Research on Frictionless with Logan Jastremski, "AI spending can't grow forever with P Equity Research | EP 167":

"No, I still think it's a boom-bust cycle, but the, the bust side of it may just be softened a little bit because of the LTAs... In order for it to not be cyclical, I believe the spending has to just continue infinitely, which I don't see happening."

LTAs are long-term agreements: multi-year supply contracts with set volumes, prices and financial guarantees. Two new facts from that episode:

• Ten-year contracts. Citing Korea's Daishin Securities, the analyst said Samsung has started receiving requests for 10-year contracts, double the typical five. He called that "negative for both parties," because no hyperscaler can really forecast past two years.
• HBM share at SK Hynix. SK Hynix is expected to get about 35% of its DRAM revenue from HBM by 2030, roughly double this year's share.

The same episode had a striking range of estimates for memory's share of total AI capex:

UBS's figure for memory spending next year is about $900 billion. As the analyst put it, that's "basically more than the CapEx that is being spent by all the hyperscalers this year." The range is so wide that his honest conclusion was: "I don't think anyone knows how much is being spent on memory."

Why it matters

• The bar was very specific. Results beating $50.9 billion and $31.49 EPS were already expected. The swing factors were (1) next-quarter guidance, adjusted for the 14-week quarter, (2) whether LTAs cap pricing, and (3) capex.
• Micron fell 1.6% in the days before reporting and closed at $1,065.11. That was a muted lead-in compared with the 16% jump after the June quarter, which was followed by a 41% drawdown.
• We'll grade the print on Monday, using Mehrotra's comments on Squawk this morning and whatever the podcasts make of it over the weekend.

3. The scarcity data keeps coming in

Monday's issue asked whether GPUs are actually being used. This week brought more data, mostly suggesting they are.

Rental prices are going up, not down

On Frictionless (9/25), the P Equity Research analyst laid out the numbers:

"there was CLSA research that came out today saying the supply-demand gap of ASICs and GPUs is 73%, and they expect the shortage to last till 2030... the rental prices show that because 4 to 5-year-old chips are commanding a higher pricing than they did before."

• New chips: B200/B300 rental rates have gone from $5 to "$7 or $8" an hour.
• Used chips: H100s are still selling for about $25,000.
• Old chips: Nvidia's Volta, nearly ten years old, is still in use. A neocloud (the host thought Nebius) is reportedly running chips for 8–9 years.

That last point goes to the heart of Michael Burry's depreciation argument from last year: that hyperscalers are spreading GPU costs over too many years. If 8-to-10-year-old chips still earn rent, the accounting looks conservative, not aggressive.

The analyst was fair about it, though. He had spoken with an expert from AMD who "believes we have enough, and the bottleneck is in other areas like data center power, advanced packaging and memory, and memory being the big one." In other words, the chips may be plentiful. What's short is everything needed to switch them on.

a16z: "a model buster"

The a16z team took the broadest view on The a16z Show, "The $1 Trillion AI Buildout | State of Markets" (note they are venture investors with a stake in AI):

• Capex: Alphabet, Amazon, Meta, Microsoft and Oracle are spending about $780 billion in 2026, up from $416 billion in 2025, and "all expectations point to them spending over a trillion dollars annually from 2027."
• Backlog: Microsoft, Google and Amazon have about $1.7 trillion of combined cloud backlog, meaning signed customer commitments not yet delivered.
• Supply chain: "There are certain elements in the data center supply chain where you can't get access to materials or products until 2028."
• GPU spot prices: "any GPU that you can bring online is being priced at an attractive rate."

One host called compute demand "a model buster at this point." He noted that two weeks ago OpenAI "had to pause new subscriptions on their pro plans," its $2,000 service.

Why it matters

• Rental prices are the cleanest tiebreaker in the deployment debate. Rising rents for both new and decade-old chips fit a shortage story, not a glut. Warehoused, unused GPUs would push rental prices down. They're going up.
• The bottleneck is moving to power, packaging and memory. That favors Micron, TSMC's advanced packaging, and the power and optics suppliers over the GPU makers at the margin. It's the same "bottleneck trade" Rasgon blamed for Nvidia's lag.

4. Nvidia vs. everyone else: the challengers come under fire

The bull case for the full stack

The week's longest discussion of competition came from Ben Pouladian, a long-time Nvidia bull, on Monetary Matters with Jack Farley, "Token Bill Dwarfs CPU Bill | Ben Pouladian on Meta's Muse, Anthropic S1, and Why All Roads Still Lead to NVIDIA". He's openly an advocate, so read this as a well-argued bull case, not a neutral view.

His main point: buyers are moving from measuring chips to measuring factories. The new metric is cost per token (a token is a chunk of text an AI model reads or writes):

"What does it cost to extract a billion tokens per megawatt?... if there's other companies that are serving tokens at a higher cost, they won't be as profitable... if you have a higher token production cost, your margins are going to be less. And I think you're going to get hit with a discount on your valuation."

His evidence that the full stack wins: Elon Musk dropped Tesla's in-house Dojo chip for Nvidia's NVL72 Vera Rubin platform. "All roads and rockets lead to NVIDIA."

On AMD: "it weighs like eight tons"

Pouladian's most specific claims were about AMD's first full-rack system, Helios:

"when I saw the Helios rack, it's basically twice the size of a Blackwell system... If they're coming out with their first rack scale solution, Helios, and she said they're going to ramp at the end of this year into early next year, and already I'm hearing that there's potential delays in scaling it and manufacturing it, and the thing weighs like eight tons."

Treat the delay claim as unverified. It's secondhand ("I'm hearing"), and he is an open Nvidia advocate. But it's the sort of claim that matters for a stock sitting 3.9% below its average analyst target, with no room for a slip. He also said Nvidia's new agent-security sandbox "works on x86, ARM and NVIDIA chips, but not on AMD."

His wider view: AMD is "squeezed in the middle." Above it is Nvidia. Below it are the hyperscalers' own chips (OpenAI's "Jalapeno," Meta's MTIA, Microsoft's Maia) and inference startups. "Why would they try to buy an AMD product for inference, which might be just as good?"

On custom chips and Anthropic

Pouladian also brought up Anthropic's leaked S-1 filing:

"with the new S1 that leaked for Anthropic, they've committed over a half a trillion dollars over the next few years on compute contracts. I don't think that's Tranium. I think a lot of that is probably NVIDIA GPUs and Google TPUs."

He said early Claude training on Amazon's Trainium chips came with "a lot of outages." That's his characterization and unverified. Read it next to Nvidia's own disclosure of $180 billion in Anthropic commitments: if Anthropic has committed over $500 billion in total, Nvidia's share is roughly a third. The rest likely goes to Google's TPUs, which Broadcom builds, and to others.

The other side: the market sees more upside in challengers

Futurum's Daniel Newman summed up the market's logic on the Futurum Equities Podcast (Ep 58):

"the market is treating NVIDIA as the, as the leader in the incumbent that has everything to lose. The market is treating AMD as the entrant competitor that has everything to gain... Nobody has ever seen a company go from five to 25 trillion. There is no precedent for that."

Pouladian agreed on the arithmetic: "it's a lot easier to go from half a billion to one trillion instead of five trillion to 10 trillion." Newman also said Futurum raised its 2030 CPU market forecast from $110 billion to $250 billion between January and September, and that its estimate for cumulative capex through 2030 now tops $14 trillion. Both are Futurum's own figures.

Why it matters

• This week, the pressure moved from Nvidia to AMD. AMD fell 3.0% while Nvidia rose 1.5%. The talk on the podcasts went the same way.
• The Helios timeline is now the key thing to watch for AMD. Any confirmed delay in the year-end ramp would hit a stock priced for perfection.
• Broadcom gets a quiet boost from the Anthropic math. If roughly two-thirds of Anthropic's commitments are not with Nvidia, a large share likely flows to TPUs. Broadcom remains the name furthest below its analysts' targets.

5. Raja Koduri: the real competition is China

The week's most thought-provoking operator interview came from Raja Koduri. He has run graphics at AMD twice, led Apple's graphics architecture, and was chief architect of Intel's graphics group. He now runs Oxmiq, which he calls "a Broadcom for NeoCloud." Note that he is pitching his own startup. He spoke on TechSurge: Deep Tech Podcast, "Gaming Chip Pioneer Raja Koduri on China's AI Cost Advantage and Moving Beyond the GPU".

The capital math

"Every gigawatt today is a 50 to 60 billion dollar capital expense, out of which 45 billion dollars goes to these things called GPUs, some CPUs, some networks, some storage... we need over 400 gigawatts of new compute infrastructure by 2030... it's around 24 trillion dollars of capital needed, right? That's a lot of capital. I don't think we have that capital."

His near-term estimate is more modest: "about 100 plus gigawatts" over the next five to seven years, "15 gigawatts per year or something."

The China cost gap

"The target for China Inc. for gigawatt is to be less than 10 billion dollars. We are at 60... I thought they would be able to do that in the next three years. I see indications that they will be able to do it much sooner than that."

That's a 5–6x cost gap. His argument for why export controls on advanced chipmaking may not close it is the most important technical claim of the week:

"you can take TSMC 7 nanometer die, which is, I think, first launched in 2019... And attach it to DRAM... you 3D stack, hybrid bond it. Okay. You get 10x the bandwidth than the current HBM. And 10x token generation rate. That package will beat Vera Rubin."

That "beat Vera Rubin" claim is Koduri's own and unverified. The logic behind it is the same one driving the Micron bull case: for inference (running trained models to answer requests), the hard limit is moving data between memory and compute, not raw processing power. "The bottleneck has shifted quite significantly to how you deal with memory hierarchy."

He also said that for a given enterprise customer, a typical Nvidia-plus-InfiniBand data center could "easily cut 20 to 30 percent of that cost" by tuning the mix of compute, memory and connections to the workload.

Why it matters

• It backs memory's pricing power. If memory bandwidth decides inference performance, HBM and advanced packaging capture more value, and leading-edge chip process nodes capture less. That's good for Micron, SK Hynix and TSMC's CoWoS packaging, and a quiet challenge to the idea that only the newest process nodes matter.
• It's a real, long-term risk to the West's cost per token. If China delivers sub-$10 billion gigawatts, the "dollar per million token bar set by China" becomes the global price benchmark, as Koduri puts it.

6. Optics and networking: Lumentum's case, and big moves elsewhere

The one real data point came from Michael Sansaterra of Sylvan Capital on The Exchange (9/28), who pointed to Lumentum's position in optical circuit switches:

Copper is "being brought to the limit as far as processing is concerned," and "you're going to see more and more copper both inside circuits, outside networks" replaced by lasers, with Google championing the process and Lumentum "best positioned to replace that copper transfer with lasers."

Lumentum was the best performer in our group this week (+3.14%).

On the other side, Credo fell 7.67%, Coherent (COHR) fell 2.71%, and Astera Labs fell 2.37%, without a fundamental explanation from the podcasts. With Credo now 37% below its 52-week high, moves this big without a story usually reflect positioning, meaning funds piling into or out of a crowded trade, rather than news.

Erica Clower's comment on Bloomberg Tech that security is becoming "table stakes for other chip vendors, whether it's AMD or Marvell or Broadcom" was the only mention of Marvell (MRVL) or Broadcom beyond the Anthropic math. Marvell's Investor Day is on October 6.

7. The macro backdrop: yields and politics

Two risks came up in almost every market-wide podcast.

Rates. The 10-year Treasury yield was quoted at 5.25% on The Exchange (9/28), 5.27% on Bloomberg Tech (9/28) and 5.28% on Squawk on the Street, "11AM Hour: Schwab CEO, Counselor to the Treasury Secretary David Zervos & Expectations for Micron Earnings 9/30/26". On The Compound and Friends, Matt Cerminaro noted that 18 of the last 24 trading days saw a higher 10-year yield, which has happened only 11 times since 1965. Historically, he said, yields were lower a year later about 80% of the time, by about 90 basis points on average. A basis point is a hundredth of a percentage point. He played a clip of Stanford's Darrell Duffie on Odd Lots explaining that foreign central banks "have had all that they need and they're not buying more."

Politics. On Squawk on the Street (9/28), an anchor raised growing calls for a moratorium on data centers, "including from politicians in states that were once very open, like Governor Greg Abbott in Texas." His guest agreed that "regulation policy is the biggest near-term risk." He still said, "I don't see anything that can break supply and demand imbalances for stuff like Micron in the next several quarters... I think it's a 10-year cycle. You're three and a half years into a 10-year cycle."

The Saxo Market Call host raised a deeper worry: that the build-out is being run like "a Manhattan Project towards AGI." If artificial general intelligence never arrives, "that in itself is a warning to where market values are right now."

The key debates, steel-manned

1. Is Nvidia's buyback a sign of strength or a sign it has run out of ideas?

• Bull: Rasgon says $235 billion is only about half of free cash flow and under 5% of market value. Schiffman says Nvidia is "acting as a bank," lending its balance sheet so the build-out can happen. Return capital and invest; it can do both.
• Bear: A company that can't find better uses for $40 billion a quarter may be near the end of its fastest growth. And part of that cash comes from customers that Nvidia funds itself ("dump trucks of money").
• My read: The bull side has the numbers. The bear side has a real point about circularity, but it can't explain rising GPU rental prices.

2. Is memory still a cyclical business?

• Bull: Hosseini says memory is "the defining link," with lasting 70% operating margins thanks to more varied demand and capex discipline. LTAs and talk of 10-year contracts point to a more stable industry.
• Bear: P Equity Research says it's still boom and bust, because capex can't grow forever. The multiple of about 6–7x forward earnings says the market agrees.
• What decides it: Micron's capex guidance and whether its long-term agreements cap price increases. We'll see this morning and over the weekend.

3. Do challengers win on value, or does the full stack win on cost per token?

• Challengers: AMD's jump from about $200 billion to about $1 trillion in market value shows how much upside a challenger has. Hyperscaler chips save money on known workloads.
• Full stack: Pouladian argues that cost per token across power, networking, software and security is what counts, and that the Helios rack's size and possible delays show rack-scale design is hard to do.
• Missing evidence: Nobody has published an independent cost-per-token comparison. Until someone does, both sides are arguing from anecdotes.

Names in play

Read-throughs

• Buyback, then cost of capital. Nvidia says debt is cheap for it while the 10-year sits near 5.3%. That widens the gap between companies that fund the build-out from cash and companies that borrow for it. Expect more Nvidia guarantees and financing for neoclouds, along the lines of the Nscale S-1 terms covered on Monday.
• The Anthropic math favors TPUs. If more than $500 billion is committed and $180 billion is Nvidia's, Google's TPUs, built by Broadcom, are the obvious home for much of the rest. Broadcom's December quarterly report is where that should show up.
• Memory first, then packaging. Koduri's bandwidth argument and the a16z note that some supply-chain items are sold out until 2028 both point to TSMC's advanced packaging and substrates as the next bottleneck.
• Optics is still a positioning trade. Until a podcast gives a real fundamental reason, the 20% swings in Credo and Astera Labs, up one week and down the next, look like crowded trades rather than news.

What changed since Issue 028 (Monday, September 28)

• New: the buyback. Nvidia answered Monday's "where are the GPUs?" debate indirectly, with cash. The circularity argument is weaker; the deployment argument is unchanged.
• New: the $180 billion Anthropic disclosure, which adds to the Nscale S-1 picture of how concentrated demand is.
• Stronger: the shortage evidence. Monday had GPU availability "on the floor" (Full Signal) and high resale values (Silicon Data). Today adds CLSA's 73% gap, $7–8 an hour B200/B300 rents, and decade-old Volta chips still in use.
• Shifted: AMD sentiment. Monday's CPU rally gave way to Helios skepticism. AMD went from +2.6% over Issue 028's window to -3.0% this week.
• Reversed: optics momentum. Credo went from +19.9% (Issue 028 week) to -7.7%. Astera Labs went from +20.2% to -2.4%.
• Unchanged: Broadcom and Micron remain the two names furthest below their analyst targets (+52% and +50%).
• Pending: the Micron print. It came out last night. The podcast reaction is not out yet.

On the calendar: Micron CEO on Squawk (Oct 1, 9 AM ET); CME H100/B200 rental futures (targeted around Oct 5); Marvell Investor Day (Oct 6); Nscale IPO pricing; Nvidia Q3 FY27 (around Nov 18); Marvell Q3 FY27 (Dec 1); Broadcom Q4 (around Dec 10).