Newsletter · 2026-09-28 · Ashutosh Agarwal

The Debate Over Whether AI Chips Are Deployed or Warehoused - AI Accelerators: GPUs, Custom Silicon & Optics - Week of September 28, 2026

A synthesis of what podcasts and a fresh IPO filing said about AI accelerators for the week of September 21-28, 2026, centered on the debate over whether AI chips are deployed or warehoused (Ed Zitron versus Jensen Huang), Nscale's S-1 with Nvidia on every side, Micron's September 30 gross-margin guide, and a week when Astera Labs and Credo jumped about 20 percent while Broadcom and Coherent fell.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of September 28, 2026: The Debate Over Whether AI Chips Are Deployed or Warehoused

Podcast window: September 21–28, 2026. Prices: FactSet closes through Friday, September 25.

Last Thursday I wrote that the AI debate had moved off the balance sheet, into special-purpose vehicles and hidden leases.

This week it moved somewhere even more basic: the loading dock.

On Friday, Ed Zitron put out a monologue on Better Offline arguing that a huge share of the AI chips already sold are not actually running. His headline estimate:

"I estimate that approximately 50% of all AI hardware and chips that have been sold is being warehoused, and will take more than two years to fully ingest."

The same day, the Hard Fork feed replayed Ezra Klein's interview with Jensen Huang, where Nvidia's CEO made the opposite case in one sentence of arithmetic:

"$50 billion to build a one gigawatt data center, one gigawatt AI factory. And you can rent it for 40 to $50 billion per year."

Those two claims cannot both be right in their strong form. If Jensen is right, every gigawatt that gets plugged in pays for itself in about a year and the build-out is rational. If Zitron is right, a big chunk of Nvidia's recent revenue is a pre-order book for data centers that don't have power yet.

And the week handed us a real document to test it against. The Run the Numbers podcast walked through the S-1 (the registration filing a company makes before an IPO) of Nscale, a two-year-old UK "neocloud" (a company that rents out Nvidia GPUs by the hour). What it shows is Nvidia sitting on almost every side of the deal: supplier, investor, rent guarantor, and customer.

Meanwhile, the stocks did something they haven't done in two weeks: they split. After two straight weeks where all ten names in our coverage went up, this week two of them fell (Broadcom and Coherent), while Astera Labs and Credo each jumped about 20%. And Micron reports on Wednesday, September 30, after the close, which is the single best test we'll get all quarter of whether "memory is the bottleneck" is still true.

Let me be clear about where I stand. I think the demand is real. The operators who run these systems keep saying so, and the GPU rental data (more below) backs them up. But "demand is real" and "every chip sold is earning money today" are different statements. This week was about the gap between the two.

TL;DR

• The new question is deployment, not demand. Ed Zitron (Better Offline, Sept 25) estimates Microsoft has only about 2.2 million GPUs in service (roughly 2 gigawatts of 12 gigawatts total capacity) despite more than $265 billion of capital spending since 2022, and that about $106 billion of Microsoft's GPUs and related hardware sit unpowered or in warehouses. These are his estimates, not company figures.

• Jensen Huang's counter: a 1-gigawatt "AI factory" costs about $50 billion and rents for $40–50 billion a year, and Nvidia chips are becoming "an asset class, kind of like an airplane" (Hard Fork, Sept 25).

• The rental data sides with Jensen, for now. A strategist on Full Signal (Sept 24) says on-demand GPU availability is "on the floor" and even the oldest chips' rental rates have been stable for years. Silicon Data (The Business Brew, Sept 23) says rental income implies GPU resale values are "quite a bit higher" than accounting depreciation assumes, and CME is launching H100 and B200 rental futures.

• Nscale's S-1 is the week's hardest data. $103.4 billion of contracted value, about 85% from Microsoft and Anthropic, negative gross margin, and Nvidia as supplier, roughly $2.2 billion investor, $860 million rent guarantor, and $1.2 billion customer (Run the Numbers, Sept 24).

• Micron reports Wednesday. An analyst on Investing Experts (Sept 27) warns DRAM price growth is cooling (TrendForce: +13–18% this quarter vs. roughly +50–60% last quarter) and China's CXMT is scaling fast. The stock trades at about 6.8x next year's consensus earnings, with 46% upside to the average analyst target.

• Broadcom is the widest gap between price and targets. It's down 28.7% from its 52-week high with 51% upside to the average target. 7investing (Sept 25) makes the bull case: a $34.8 billion Q4 revenue guide (+93% year over year, as stated on the podcast).

• Breadth cracked. Astera Labs +20.2% and Credo +19.9% for the week; AMD +12.7% to within 1.3% of its high; Broadcom −1.3% and Coherent −6.8%. Nasdaq +2.1%.

• Optics coverage is still almost nonexistent. The two biggest winners, Astera Labs and Credo, got zero podcast coverage this week.

The scoreboard

Weekly change is Friday Sept 18 close to Friday Sept 25 close. "Since last issue" is Wednesday Sept 23 close to Friday Sept 25 close.

Nasdaq Composite: +2.06% on the week (26,522.55 to 27,068.72).

What the table says, in plain English:

• The "everything goes up" phase ended. Two names fell for the first time in three weeks. That doesn't mean the trade is over, but it means investors started choosing.

• The market is paying up for connectivity. Astera Labs and Credo make the chips and cables that move data between GPUs inside a server rack. Both rose about 20%, and Credo added almost 9% in just Thursday and Friday.

• AMD has run out of analyst targets. It closed at $630.63 against an average analyst target of $635.54. Put simply, Wall Street's price targets haven't caught up with the stock, or the stock has gotten ahead of the story. It also trades at about 109 times the last twelve months' earnings.

• The cheap names stayed cheap. Nvidia (+48% to target), Broadcom (+51%) and Micron (+46%) have the most room to their average analyst target. Micron trades at about 14.7x this fiscal year's consensus earnings per share ($73.83) and about 6.8x next year's ($159.42). (A "P/E" or price-to-earnings multiple is how many dollars you pay for each dollar of a company's annual profit.)

1. Where are the GPUs? The deployment debate

The bear case: chips sold, not switched on

Podcast: Better Offline, "Monologue: The AI Data Center Overbuild" (Sept 25, 2026) Speaker: Ed Zitron, host and tech critic (Cool Zone Media). Pundit, not an operator.

Zitron's argument is detailed, so it's worth walking through step by step.

Step 1: Microsoft. He says he previously reported that Microsoft had only 2.2 million GPUs in service, drawing about 1.993 gigawatts of power, with an installed cost of "somewhere in the region of $50 to $60 billion." He then cites Bloomberg reporting that only 2 gigawatts of Microsoft's 12 gigawatts of data-center capacity is specifically for AI. (A gigawatt is a billion watts, roughly the output of a large power plant. It has become the industry's unit for sizing AI data centers.)

"Despite having spent over $265 billion on capital expenditures since the beginning of 2022, Microsoft has only put about $50 billion worth of GPUs into service, and I estimate that as much as $106 billion worth of GPUs and associated hardware are now sitting either unpowered in data centers, or incomplete data centers, or in warehouses."

He also points out that Satya Nadella admitted last November to having chips "sitting in inventory that he couldn't plug in."

Step 2: the rest of the industry. Zitron goes looking in the "construction in progress" line on balance sheets. That's the accounting bucket where companies park assets they've paid for but haven't started using yet, like a half-built data center or uninstalled servers. Across Google, Meta, Oracle, Amazon, SpaceX, Tesla, CoreWeave, IREN, Core Scientific and Applied Digital, he adds up about $374 billion, with Google alone at around $122 billion. He says the number has grown every time it has been reported for 12 quarters. Including companies that don't disclose it, he guesses the true total is $400–500 billion.

Step 3: the conclusion for Nvidia.

"I think NVIDIA has made somewhere between $200 and $300 billion in revenue on stuff that's been sold probably 24 to 36 months in advance, and that any scarcity around AI compute is a result of most of the capacity being sold immediately to Anthropic and OpenAI."

He adds that "anywhere from half to 75%" of the big cloud companies' revenue backlogs come from Anthropic and OpenAI, and that OpenAI's spending made up 70% of Microsoft's AI revenue in fiscal 2026.

How much weight to give this. Zitron is a long-time AI skeptic, and the core numbers (50% warehoused, $106 billion unpowered at Microsoft) are his own estimates, not disclosed figures. The construction-in-progress totals are a real, checkable balance-sheet line, but that line also includes normal building work like shells, land and power equipment, not just idle GPUs. Still, it's the most specific version of the bear case we've heard. It also connects with a figure from last issue: TIP848 said Alphabet had about $122 billion in assets not yet in service, which matches Zitron's Google number.

Why it matters for the stocks. If chips are bought years before they're switched on, Nvidia's revenue runs ahead of real usage. The risk isn't that demand disappears. It's that buyers pause orders while they plug in what they already own. That kind of pause would hit Nvidia, Broadcom and the networking names first, and memory last, because memory is priced on shipments.

The bull case: every gigawatt pays for itself

Podcast: Hard Fork, "The Ezra Klein Show: Jensen Huang Thinks A.I. Alarmism Has Gone Too Far" (Sept 25, 2026) Speaker: Jensen Huang, founder and CEO of Nvidia. Operator.

Klein opens by noting Nvidia is now "$5.4 trillion in market cap," consistent with FactSet's $5.45 trillion as of Friday. Huang's pitch has three parts:

• Demand scale. With "multiple hundreds of billions of agents in addition to the humans using the computer," he argues computing needs could go "up by a billion times."

• Payback. A gigawatt AI factory costs about $50 billion and rents for $40–50 billion a year. That's CEO arithmetic, not audited economics, but it is the claim the whole build-out rests on.

• Fungibility. Nvidia's chips are "general purpose," so they can be used "from data processing to pre-training to post-training to eval to inference." Because of that, he says:

"People are talking about NVIDIA compute as an asset class, kind of like an airplane. So this is, and if we could do this, if this happens, then of course, the cost of capital for, um, funding NVIDIA AI factories will be the lowest."

He also points to "$500 billion of venture funding" flowing into startups that "all need compute."

Why the airplane line matters. It's a direct answer to last week's credit debate. Planes can be financed cheaply because they hold their value and can be leased to another airline. If GPUs work the same way, lenders will fund them at lower rates, which keeps the build-out going. If GPUs lose value fast, the financing gets expensive. That is exactly the question the next two podcasts tried to answer with data.

The tiebreaker: what GPUs actually rent for

Podcast: Full Signal, "ELITE Strategist: How to position for the NEXT AI trade" (Sept 24, 2026) Speaker: the strategist guest (not named in the transcript excerpt), who has tracked GPU availability with a colleague since 2023. Pundit/strategist with proprietary data.

His team checks cloud providers throughout the day to see whether they can actually get a GPU on demand, and reports it as a percentage (100% means always available, a loose market; 0% means never available, very tight). His read:

"Compute availability, GPU availability is on the floor."

He says availability leads rental prices: when it drops, rental rates stabilize and then rise. He notes H200 availability ticked up briefly last month as customers switched to newer Blackwell chips, but the key point for the depreciation debate is that "even the oldest GPUs rental rates are stable over a multi-year look back period." He also argues it's "going to take more than 25 basis points to stop this multi-trillion dollar data center build out." (A basis point is one hundredth of a percentage point, so 25 basis points is a quarter-point rate move.)

Podcast: The Business Brew, "Steve Hou - Data Over Narrative" (Sept 23, 2026) Speaker: Steve Hou, speaking for Silicon Data, a firm that publishes GPU rental price indices. Data provider.

Hou's point: most people value used GPUs using standard accounting depreciation (writing an asset down in equal slices over its life). Actual rental income tells a different story:

"The realized, you know, uh, residual value implied by the currently observable, you know, rental income is quite a bit higher."

His analogy: if you owned a Florida rental property and "rental income went up 50% in a year, and then went up again in the next year, you will reasonably adjust upward the market fair value for that home." He's careful to note that an installed GPU earning rent is worth more than a loose one "for sale on eBay," and that lenders offering guaranteed residual values carry the risk if rental rates fall.

He also confirmed that CME is launching two futures contracts on Silicon Data's indices: one for the H100 and one for the B200 (Nvidia's Blackwell chip) neocloud rental prices, settling against monthly averages. CME announced this in August, with listing targeted for October 5, pending regulatory approval. (A futures contract lets buyers and sellers lock in a price today for something delivered later. Here it's a month of GPU rental time.)

Podcast: The Most Interesting Thing in AI, "The Case Against the AI Bubble - Azeem Azhar with Nicholas Thompson" (Sept 23, 2026) Speaker: Azeem Azhar (Exponential View) with Nicholas Thompson. Pundit/analyst.

Azhar puts AI services spending outside China at $126 billion in the 12 months to July 2026, up from $110 billion in the 12 months to June. The figure is deduplicated, so a dollar paid to OpenAI that then goes to Microsoft counts once. On depreciation, the conversation relays an Amazon executive (named in the transcript as "Matt Garvey") saying "we're seeing chips in use six, six and a half years after we have bought them," against Amazon's five-to-five-and-a-half-year accounting schedule. Azhar's framing of the whole debate is the right one: can AI revenue "escape the gravity" of "the depreciation, the operating expenses, the interest expenses of the build out?"

My read on the deployment debate:

• Zitron's warehousing figure is an estimate, but the direction is plausible. Power and construction, not chips, are the constraint. Last issue we cited around 230 gigawatts of data-center grid applications against about 93 gigawatts of planned grid build-out.

• The rental data says something important: the GPUs that are switched on are earning more than the bears assume, and holding value longer. That supports Jensen's airplane argument and weakens the "GPUs are worthless in three years" bear case.

• Both can be true. Chips can be scarce in the rental market and stuck in warehouses, if the scarcity is caused by power and buildings rather than chip supply. That's the Zitron nuance worth keeping: the bottleneck is plugging chips in, not making them. That favors power, cooling and electrical names over more chip orders in the near term.

• What would change my mind: GPU rental rates or availability loosening. With the CME futures listing, that signal will soon be visible every day.

2. Nscale's S-1: vendor financing, now in a filing

Podcast: Run the Numbers, "The NScale IPO | A Neocloud S1 Breakdown" (Sept 24, 2026) Speaker: the host of Run the Numbers. Pundit reading a primary document.

For the last three issues, "circular financing" (Nvidia funding the customers who buy its chips) has been a podcast talking point. Nscale filed its S-1 on September 18 to list on the NYSE under the ticker NSCL, which puts actual numbers on it. The podcast's line:

"The house that Anthropic built and NVIDIA financed."

What's in the filing (as read on the podcast, and matching reporting on the S-1):

• Contracts: $103.4 billion of total contracted value, up from $38 billion in December. Microsoft (up to about $43.8 billion through 2033) and Anthropic (up to about $44.6 billion) make up roughly 85% of it.

• Capacity: 1.4 gigawatts active and contracted, with a stated "line of sight" to 10 gigawatts.

• Nvidia is on every side:

• makes every GPU Nscale deploys (bought through third-party suppliers);

• has invested about $2.2 billion, including $1 billion that closes in November;

• "guaranteed up to $860 million of rent on a Texas data center";

• rents $1.2 billion of GPU capacity back from Nscale as a customer.

"NVIDIA vouched for the rent on a building that would be filled with its own chips."

• Economics: gross margin negative; lost about a billion dollars while operating cash flow was a positive $1.7 billion, because deferred revenue (customer prepayments) jumped from $2 billion to $6.5 billion in six months.

• Financing stack: $24 billion committed to equipment spending, an $830 million revolver draw, $4.2 billion of new loans, $2.5 billion of Dell equipment leases, $3.1 billion of convertible notes, and $2.6 billion of lease liabilities across 17 sites.

• Valuation: reported target of about $35 billion, against $14.6 billion in March.

The host's verdict: "if you like Florida Swampland, I've got a negative gross margin Neocloud to sell you."

Why it matters. The fair version of this is that Nvidia's total exposure (about $2.2 billion invested plus an $860 million guarantee) is small compared with its revenue, and small next to the $24 billion Nscale is spending on equipment, much of which flows back to Nvidia through suppliers. The concern is concentration, not size. Nscale's contracts depend on two customers, and the podcast notes that "most of the rest depends on data centers, power plants, and GPUs that haven't been built or fully financed yet." According to reporting on the S-1, Anthropic keeps the right to cancel if delivery milestones are missed. Put that next to Zitron's claim that half to three-quarters of cloud backlogs trace back to Anthropic and OpenAI, and to All-In's claim (below) that about 60% of the world's new compute is being added for those two labs. The whole complex is increasingly a bet on two private companies' ability to pay. How the Nscale IPO prices will be a live read on how much appetite investors have for that bet.

3. Memory: Micron reports Wednesday

Podcast: Investing Experts, "What will Micron's gross margin guide be?" (Sept 27, 2026) Speaker: an analyst at Tech Contrarians (not named in the excerpt). Pundit/independent analyst.

This was the only dedicated Micron preview in the window, and it's cautious.

The setup:

"There's a lot of pressure on Micron into print because we've had a switch, I would say, in the memory sentiment intra-quarter."

• Prices are still rising, but more slowly. TrendForce now sees DRAM average selling prices up "only" 13–18% this quarter, versus "closer to 50 to 60 percent" last quarter. (DRAM is the standard working memory in servers and PCs. HBM, or high-bandwidth memory, is a stacked, premium form of DRAM that sits next to AI chips.)

• China is adding supply. CXMT, China's main DRAM maker, is expected to end the year at around 20,000 wafer starts, which the speaker frames as comparable to Micron's own monthly output.

• The number to watch is gross margin. Gross margin is the share of each sales dollar left after the direct cost of making the product. Micron guided to 86% last quarter after printing 84.6%, which the speaker says is "higher than NVIDIA's own peak margins."

"We don't think the market is going to be very forgiving if Micron's gross margins aren't guided to expand again into the first quarter."

• A contrarian point on HBM: the speaker argues the market's "biggest misunderstanding" is thinking memory makers rallied on HBM, when "it's actually the non-HBM side of the business that's really been carrying gross margins higher." He relays that Micron's CEO said non-HBM margins were higher than HBM margins two quarters ago. That matters because ordinary DRAM pricing is exactly what's cooling.

• A reminder from last quarter: after the June report, "stock soared, and then a day later, it began to, you know, trade downward for a while."

(The margin figures, CXMT wafer comparison and CEO remark are as stated on the podcast and not independently verified here.)

Podcast: CNBC's "Fast Money", "Market Resilience Despite Surge In Yields… And Options Action Ahead Of Micron Results 9/25/26" (Sept 25, 2026) Speakers: Fast Money panel. Pundits.

The panel's valuation framing lines up with the FactSet numbers: Micron trades at "call it seven times actually less next year's earnings" (FactSet consensus: about 6.8x), while AMD "looks like the multiple is still anticipating a lot of good news rather than pricing in the good news that's already here." They also noted that longer-term visibility has been part of Micron's rebound and that investor optimism "has waned about 20 percentage points just since the end of June."

How I'd frame Wednesday:

• Bull setup: the stock is cheap on next year's numbers (6.8x), 46% below the average target ($1,580.88), and still 13.8% below its high after rising 6.6% this week.

• Bear setup: the market is paying for margins at an all-time peak. If ordinary DRAM price growth is slowing and CXMT is ramping, the guide can disappoint even if the quarter beats.

• The read-through: a strong gross margin guide supports the whole "memory shortage lasts into 2027" idea (SK Hynix, Samsung, and HBM-hungry accelerator makers). A flat or down guide would be the first real crack in the memory story. Last issue's MoneyFlows case (HBM at 30% of DRAM capacity but 13% of output, "bandwidth out several years") is the bull case this print will test.

4. Custom silicon: Broadcom's widest gap, and a Marvell wrinkle

Podcast: The 7investing Podcast, "Why Broadcom Has Been My Best Buy for 3 Months in a Row" (Sept 25, 2026) Speakers: 7investing analysts. Pundits (long Broadcom; it's their top "best buy").

Broadcom is the name where price and fundamentals have diverged the most. It's down 28.7% from its high and fell again this week, yet analysts' average target implies 51% upside. The 7investing case:

• Acceleration, as stated on the podcast: revenue of $19 billion (+29% year over year), then $22.2 billion (+48%), with free cash flow (cash left after capital spending) reaching 46% of revenue. They also say a Q4 guide of $34.8 billion, which "would be 93% year over year growth." (Speaker-stated figures, not reconciled to filings here.)

• The size of the market: "$40 billion last year, $60 billion this year for each gigawatt of power capacity," and with Meta and SpaceX talking about 5-gigawatt campuses, "that's a $300 billion project, half of which is going to be going through the chips themselves." They put custom chips (ASICs, or application-specific integrated circuits, built for one customer's workload) at about 30% of more than $1 trillion in global chip spending in 2026.

• Why the stock has fallen, in their view: "A lot of people think it's because they don't have the sole supply of the TPUs that they're providing for Google." They say Alphabet has signed Marvell as a secondary supplier for its TPU processors, which they call "very smart, very wise of Google."

• Diversification: deals "with Anthropic, with OpenAI, with all of the other large hyperscalers that are worth $20 billion or more of backlog."

A co-host raised the fair pushback: what happens to Broadcom "if capital gets more expensive" and the big AI labs slow deployment of data centers "costing $60 billion plus now"?

Why it matters. Last week's rally was led by the beaten-down names, and Broadcom didn't join in. The market seems to be pricing Broadcom as the custom-chip supplier most exposed to share loss (Marvell getting a slice of the TPU) and to any capital-spending pause. With the stock at 36x trailing earnings, below Nvidia's 38.5x and well below Marvell's 79x, the "Broadcom is losing Google" story looks priced in, provided the $34.8 billion Q4 guide holds. Marvell's Investor Day is October 6. It's the next chance to hear, from the company, what the TPU second-source win is worth. Broadcom reports Q4 fiscal 2026 around December 10.

Meta's MTIA, briefly. On Chip Stock Investor Podcast, "Meta Muse: The AI Agent Era Has Arrived!" (Sept 25, 2026), the hosts (who hold a full Meta position) listed where Meta's roughly $89 billion of trailing-twelve-month capital spending is going: "the servers from NVIDIA, CapEx on Meta's own custom chip design like MTIA, working with Broadcom and others on those, working with TSMC to manufacture them." Their open question: can Meta keep operating margins "in the 30% or better range" as depreciation on all of that arrives?

5. Meta's Muse and the CPU angle

The consumer AI story of the week was Meta's Muse agent launch, which several podcasts said got a strong reception in its first week (the All-In episode title calls it "Meta's Muse Pop"). For our purposes, it matters because of the hardware it runs on.

Podcast: Dumb Money Live, "Meta's Muse Just Changed the Al Trade, Here's What We're Buying" (Sept 23, 2026) Speakers: Dumb Money Live hosts. Pundits (openly long and levered).

The hosts say that in April, Amazon struck a deal with Meta to power the virtual machines behind Muse with its Graviton CPUs (Amazon's in-house processors), plus Bedrock and AWS. One host sized the Graviton piece at "tens of millions," which is small. Their argument is that Muse's early success makes it much bigger: "guess who's writing a check to Amazon to cover CPUs and data center cost?" The Chip Stock Investor hosts flagged the other side: Amazon blocked Muse agents from shopping on its site, which could make agents a battleground between platforms.

The CPU read-through. The Investing Experts analyst noted that on Monday "Intel, AMD and ARM really fly on the sentiment of, you know, CPU demand is going to be much stronger." AI agents do a lot of general-purpose work (browsing, calling tools, handling files) that runs on ordinary CPUs, not just GPUs. That is part of why AMD, which sells both, rose 12.7% for the week. I'd treat the specific Graviton figures as unverified podcast claims, but the direction is new: agents make CPUs part of the AI trade again.

6. Token prices fell by half, and open models took over

Podcast: All-In, "Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails" (Sept 26, 2026) Speakers: Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg. Pundits/investors.

A "token" is the unit AI models are billed in, roughly a word fragment. Three claims from the hosts matter for chip demand:

• Prices halved: "the Frontier Lab Corporations, Anthropic and OpenAI, they both released models this week at 50% less token prices."

• Open models took share: "in the last 12 weeks alone, token use has flipped from 80-20 closed versus open to 80-20 open versus closed." Running an open model yourself can cost "less than 10 cents for a million tokens."

• Compute concentration: "over the next year or so, something like 60% of the worldwide compute that's being added is being added for these two companies," meaning Anthropic and OpenAI. They also cited Jane Street's publicly announced $19 billion in cloud capacity contracts.

Why it matters. Cheaper tokens usually mean more total usage, so this is good for chip volume. It is less clear that it's good for the labs' ability to pay for the compute they've signed up for, which is the Nscale and Zitron worry. The open-model shift also helps explain the Wall Street Skinny discussion below: Nvidia wants a foothold where the open-model developers are.

Related: The Wall Street Skinny, "How NVIDIA Is Secretly Coming for Anthropic & Open AI | Morgan Stanley's Head of US Thematic Research" (Sept 26, 2026). The hosts (Kristen and Jen, both former Morgan Stanley) framed Nvidia's pending acquisition of Hugging Face, the main hub where open-weight models are uploaded and downloaded (announced September 3), as a defensive move. The idea: as big customers build their own inference chips (Amazon's Trainium and Inferentia, plus startups like Etched), Nvidia wants to own the distribution point for everyone else.

"It is about ensuring this customer diversification... NVIDIA is basically trying to own the platform they are kind of buying the company at the heart of open weight AI."

(Guest Michelle Weaver, head of US thematic equity research at Morgan Stanley, spoke mainly about AI adoption across companies rather than about chips.)

7. Optics and networking: the winners nobody talked about

For the third issue in a row, the optics and connectivity names moved the most while getting almost no podcast coverage. Astera Labs (+20.2%) and Credo (+19.9%) led the whole group this week, and not a single podcast in the window discussed either one. Coherent fell 6.8% with no explanation on any podcast either.

What we did get was technical, and it points in a useful direction:

Podcast: Cisco Podcast Network, "Cisco Optics Podcast Ep66: Confronting the Physics of 400G per Lane - John Calvin, Part 3 of 8" (Sept 25, 2026) Speaker: John Calvin, strategic planner and datacom technology lead at Keysight. Industry engineer.

The industry is working on the next speed step for data links, 400 gigabits per second on each "lane" (each single signal path). Calvin's point is that the physical connectors, not the chips, are now the limit. Today's OSFP connectors (the standard plug-in sockets for optical modules) resonate at around 84–90 GHz, which garbles the signal. He sees "the end of the road" for current connector designs. Meanwhile, the optics side is refusing to change its signaling scheme:

"The optical community has already spoken very clearly here and saying that it's PAM-4 or it's the highway."

(PAM-4 is a method of packing two bits into each signal pulse. More complex schemes pack more data but pick up more noise, and in optics, "noise is actually the biggest force against you.")

Why it matters. Every speed step that strains copper and connectors strengthens the case for the retimers (chips that clean up and re-send a degraded signal) and active cables that Astera Labs and Credo sell, and eventually for co-packaged optics. That's the logic behind the price moves, even though no podcast made the stock-specific case this week.

Follow-up from last issue's Ayar Labs interview (The Circuit, Sept 21). Two additional details worth noting:

• Timeline: CEO Mark Wade expects co-packaged optics (putting the optical engine right next to the chip instead of in a plug-in module at the edge of the box) to show up from "the leading-edge GPU guys" in "2028, 2029," then spread to "the rest of the accelerator community" in "29, 2030, 2031." Until then, he says, everyone is looking for a "2027, 2028, you know, kind of stopgap solution." Those stopgaps are what the listed connectivity names sell today.

• Strategic backing: Ayar Labs' Series D had "strategic participation from, you know, NVIDIA, AMD, Intel, and TSMC," according to VP of capital strategy Jesse Leiter. Ayar Labs is private.

8. Macro backdrop, briefly

• Fast Money's September 25 episode was titled "Market Resilience Despite Surge In Yields." Tech held up. The panel's argument was that tech is "less reliant and less impacted by the narrative of higher inflation, higher energy prices."

• On The Julia La Roche Show, "#411 George Noble: Market Is Entering a Dangerous Phase" (Sept 22), veteran investor George Noble flagged the timing mismatch inside AI capital spending. When Microsoft buys chips, the revenue shows up immediately for the chipmakers, but "the depreciation gets amortized over a number of years." (Amortized means the cost is spread over several years rather than counted all at once.) Combined with "runaway government deficits" (he cites US debt "north of $40 trillion"), that flatters today's economy at tomorrow's expense.

• The Full Signal strategist said Fed futures went from "three cuts priced in" before "the Iran war" to a Fed that is now hiking. His view remains that it will take far more than a quarter-point to stop the build-out.

(Rate and yield levels here are as characterized on the podcasts, not independently checked this issue.)

The debates, steel-manned

1. Are AI chips earning money, or sitting in boxes?

• Bull (Huang, Full Signal, Silicon Data): A gigawatt rents for roughly what it costs to build, every year. GPU availability is "on the floor." Old chips still rent at stable prices and resale values beat accounting assumptions.
• Bear (Zitron): A large share of shipped chips aren't plugged in. Scarcity comes from power delays and two AI labs soaking up capacity, not broad demand. Nvidia's recent revenue is partly a two-to-three-year pre-order book.
• Where I land: Both are describing the same fact, a power and construction bottleneck, from opposite ends. That supports utilities, electrical equipment and memory more than new GPU orders in the near term.

2. Is Nvidia's financing of its own customers a problem?

• Bull: The dollar amounts are small next to Nvidia's revenue, and seeding neoclouds broadens the customer base beyond the hyperscalers building their own chips (the logic behind Hugging Face, too).
• Bear (Run the Numbers): Nvidia guaranteeing rent on buildings full of its own chips makes its reported demand look healthier than end demand really is, and Nscale's backlog rests on two customers.

3. Is Micron at peak margins?

• Bull: About 6.8x next year's earnings, 46% to the average target, and HBM capacity sold out for years (per last issue's MoneyFlows).
• Bear (Investing Experts): DRAM price growth is decelerating sharply, China's CXMT is ramping, and margins are already above Nvidia's peak. The guide has to go up again for the stock to work.

Operators vs. pundits

Operators and insiders (people who run the businesses):

• Jensen Huang, CEO, Nvidia: $50 billion per gigawatt, $40–50 billion a year in rent, compute as "an asset class, kind of like an airplane" (Hard Fork)
• Mark Wade, CEO, and Jesse Leiter, VP capital strategy, Ayar Labs (private): co-packaged optics in 2028–29, Nvidia/AMD/Intel/TSMC backing (The Circuit)
• John Calvin, Keysight: 400G-per-lane connector limits, "PAM-4 or it's the highway" (Cisco Optics Podcast)
• Steve Hou, Silicon Data: GPU residual values above accounting assumptions, CME H100/B200 futures (The Business Brew)
• Nscale's own S-1 disclosures, as read on Run the Numbers

Pundits and investors:

• Bears/skeptics: Ed Zitron (Better Offline), the Run the Numbers host on Nscale, George Noble (Julia La Roche), the Tech Contrarians analyst on Micron margins (Investing Experts)
• Bulls: 7investing (Broadcom), the Full Signal strategist (GPU availability), Dumb Money Live (Meta/Amazon), Chip Stock Investor (Meta, full position)
• Framers: Azeem Azhar (AI revenue vs. depreciation), the All-In hosts (token prices and open models), The Wall Street Skinny (Nvidia and Hugging Face)

Names in play

• NVDA: Jensen's payback claim versus Zitron's warehouse claim. The stock barely moved (+1.3%). 48% upside to average target at 38.5x trailing earnings. Next report around November 18.

• MU: reports Wednesday, Sept 30, after the close. The gross margin guide is the whole event.

• AVGO: biggest gap to target in the group (+51%). A TPU second source (Marvell) is the stated worry. A $34.8 billion Q4 guide (podcast-stated) is the answer.

• MRVL: the reported Google TPU secondary-supplier win is the upside story. Investor Day October 6.

• AMD: within 1.3% of its high, less than 1% below its average target. Agents and CPUs add a new leg to the story, but the stock has priced in a lot of good news already.

• ALAB / CRDO: +20% each with no podcast explanation. The connector-physics discussion is the strongest logic for them, but there's still no fundamental catalyst to match the move.

• COHR: −6.8% in a week when other optics names rose. Not discussed on any podcast.

• Private/IPO: Nscale (NSCL) IPO pricing will be a live read on appetite for neocloud risk.

Read-throughs

• Power and electrical equipment: if deployment is the bottleneck, the companies that energize data centers have the most pricing power right now.

• Neocloud credit: Nscale's financing stack (revolver, term loans, Dell leases, converts, Nvidia guarantee) is a template for how the next layer of AI capacity gets funded. Its IPO reception will affect CoreWeave-style peers.

• GPU rental futures: once CME's H100/B200 contracts list (targeted for October 5), anyone can watch GPU rental prices daily. A falling curve would be the earliest warning sign for the whole chip trade.

• Memory peers: Micron's guide will set the tone for SK Hynix and Samsung into their October reports.

What changed since Issue 027

• From balance sheet to loading dock. Last issue's question was how AI data centers are financed (off-balance-sheet vehicles, Goldman's $6 trillion capex estimate). This issue's question is whether the chips already bought are switched on.

• Vendor financing moved from podcast talk to a filing. Nscale's S-1 shows Nvidia as supplier, investor, rent guarantor and customer.

• Breadth broke. Issue 026 and 027 covered all-up weeks. This week Broadcom and Coherent fell while Astera Labs and Credo rose about 20%.

• AMD reached its analyst targets. Upside went from +3.0% to +0.8%.

• The Micron wait is almost over. Last issue it was a week away. It's now two days, and the only dedicated preview is cautious.

• New data tools. GPU availability tracking and soon-to-list rental futures give the depreciation debate hard numbers for the first time.

Negative space (what the podcasts didn't cover)

• Astera Labs and Credo: the week's two biggest winners, with zero dedicated coverage.

• TSMC and advanced packaging (CoWoS): no dedicated episode again.

• SK Hynix and Samsung: no specific commentary ahead of Micron's report.

• Marvell: the Google TPU second-source claim came from a Broadcom-focused podcast, not from Marvell or a Marvell-focused show.

• Coherent's drop: unexplained on any podcast.

• Hardware operators: outside Jensen Huang, no large listed chip company executive spoke on a podcast in the window.

Next catalysts

• Wed Sept 30 (after close): Micron Q4 FY2026 earnings. Watch the gross margin guide for next quarter.

• Mon Oct 5 (targeted): CME H100/B200 GPU rental futures start trading (pending regulatory approval).

• Tue Oct 6: Marvell Investor Day. Custom-silicon economics and the TPU second source.

• Nscale (NSCL) IPO pricing: date not yet set; target valuation reportedly about $35 billion.

• Next Fed decision: hike-versus-hold path, with podcasts pointing to December as the live meeting.

• Later: Nvidia around November 18; Marvell Q3 FY2027 December 1; Broadcom Q4 FY2026 around December 10.