Newsletter · 2026-09-27 · Ashutosh Agarwal

Meta Muse Sparks a Chip Rally as Power and Debt Become the Real Limits - Weekly Semis & AI Infrastructure Podcast Recap - Week of September 27, 2026

A synthesis of what podcasts and operators said about semiconductors and AI infrastructure for the week of September 21-27, 2026. Meta's Muse launch flipped an 'AI doom' selloff into a broad chip rally, but the deeper conversation moved to the real constraints: a looming power shortfall, a build-out increasingly funded by debt, and Micron's earnings as the next test of memory pricing power.

Weekly Semis & AI Infrastructure Podcast Recap

Week of September 27, 2026: Meta Muse Sparks a Chip Rally as Power and Debt Become the Real Limits

Top of mind this week

A week ago, the chip trade was licking its wounds. Semiconductors had fallen about 30% from their June peak, and "AI doom" headlines were everywhere.

Then Meta ($META) launched Muse, a consumer AI agent that can book trips, order groceries and negotiate bills for you. The chip stocks took off. On Monday, September 21, AMD ($AMD) crossed a $1 trillion market value for the first time, up 10% in a day. Intel ($INTC) rose 12%, Arm ($ARM) 17% and NVIDIA ($NVDA) 2.3%, as Ryan Detrick listed on Facts vs Feelings.

But the most interesting podcasts this week were about what happens underneath the rally: who pays for all of this, where the electricity comes from, and whether the debt piling up behind the data centers can ever be repaid. NVIDIA CEO Jensen Huang and economist Paul Kedrosky gave two very different answers.

Dominant themes

1. Meta's Muse turned "AI is a bubble" into "AI needs more chips" in a week

The mood swing was fast. On Facts vs Feelings (recorded Sep 22), Ryan Detrick noted that Meta rose 11% in one day after an analyst upgrade, its best day in more than a year. Muse was downloaded more than 900,000 times in its first six days.

Why did chip stocks move with it? Agents that actually do things for people use far more computing power than a chatbot that just answers questions. Co-host Sonu Varghese made the point that Meta was the company most people expected to cut spending first. Instead:

"They're going to go for guns blazing. Yeah. They need even more."

Daniel Newman of the Futurum Group said on the Futurum Equities Podcast that he went straight to his modeling team after the Muse launch: "we've got to remodel the CPU trade." (A CPU is the general-purpose processor in every computer, as opposed to the GPU, the specialized chip that does most AI math.) He believes AMD's CPUs are "leading the charge" inside Meta for Muse. Futurum's own forecasts show how fast these numbers are moving:

• CPUs: 2030 market forecast raised from $110 billion to $250 billion a year, revised on September 18.
• Custom AI chips: raised from roughly $150 billion to $385 billion.
• Total AI capital spending: cumulative through 2030 now over $14 trillion.

On the Chip Stock Investor Podcast, the hosts gave the cost side. Meta spent $89 billion on capital investment in the last 12 months, and that spending has "completely eaten up free cash flow." Part of it goes to NVIDIA servers. Part goes to MTIA, Meta's own custom chip, which it designs with Broadcom ($AVGO) and has built by TSMC ($TSM). Their question: can Meta keep operating margins at 30% or higher once the depreciation from all those servers hits the income statement? (Depreciation is the accounting charge that spreads the cost of equipment over its useful life.)

2. Electricity, not chips, is now the real bottleneck

Three very different podcasts said the same thing: the limit on AI is power.

Michelle Weaver, Morgan Stanley's Head of US Thematic Research, told The Wall Street Skinny that her team forecasts a 57-gigawatt power shortfall for data centers through 2028. For scale, the hosts noted that all of New York City at peak uses about 10 gigawatts, so that is roughly six New York Cities. Even after counting "time to power" fixes like Bloom Energy fuel cells, natural gas turbines and former bitcoin miners' sites, Weaver said:

"We still end up with around a 30 to 40 percent shortfall on the power needed."

The NScale IPO filing, broken down on Run the Numbers, says it directly: "Access to power and its delivery costs have emerged as the primary gating factor for AI capacity expansion." That is why even Microsoft rents computing power from "neoclouds," the newer companies that do nothing but rent out GPUs. As the host put it, a neocloud "is selling speed to power as much as it's selling compute."

Renen Hallak, CEO of VAST Data (valued at $30 billion), explained on The MAD Podcast how much the hardware has changed: "The old stack, you had 10 kilowatt racks. Today you need 500 kilowatt racks." A rack is one cabinet of servers. That is 50 times the power in the same footprint.

Demand is still climbing. On a Facts vs Feelings live social hour, one panelist cited a report that Anthropic's computing capacity goes from 1.5 gigawatts last year to 5 gigawatts this year to 10 gigawatts next year: "There's no end to the spending. And one person's spending is another person's revenue."

3. The money behind the build-out is increasingly borrowed

This was the week's deepest theme, and Paul Kedrosky of SK Ventures drove it in two separate interviews.

On Better Offline with Ed Zitron, Kedrosky laid out how AI spending now spills into the wider economy:

• GDP: AI capital spending has made up between 30% and 70% of US GDP growth, quarter after quarter, for 12 to 18 months.
• Debt: Total AI-related debt is "approaching in excess of a trillion dollars." More than 60% of AI financing is now debt, "up from something like 15% to 20% a year ago." It is now the largest piece of both the investment-grade bond market and the high-yield (junk) bond market.
• Treasuries: Hyperscaler bonds are now competing with US government bonds for buyers, which he says helped push the 10-year Treasury yield above 5%. The 10-year is up about 1 percentage point in three to six months.
• Trade: According to the World Trade Organization, about 19% of global goods trade in the last three quarters was AI-related, mostly NVIDIA GPUs. Almost 55% of global trade growth came from AI.

His warning is about what happens when rates rise. Every increase in the 10-year raises the "hurdle rate," the return a data center must earn to justify its cost. At the most stressed end of the market, projects need 11% to 12% returns:

"There's no economics that does that."

He puts the breaking point "within six to 12 months on the outside."

On Between Two COO's, he gave the timeline. Much of this debt is five-year money, especially what he calls the "classic Oracle-style five-year GPU-backed lease." That means a wall of loans comes due in 2029-2030. "Will it all get refinanced? No freaking way." He compares it to the 2006 mortgage boom, when a burst of loans all reset at about the same time.

The NScale filing shows how these deals look up close. The company has $103.4 billion in contracts, but only $2.6 billion (about 2%) is live today. Microsoft signed for up to $43.8 billion and Anthropic for up to $44.6 billion. Gross margin in the first half of 2026 was negative 159% after depreciation. And NVIDIA shows up everywhere in the deal. It has invested about $2.2 billion, guaranteed up to $860 million of rent on a Texas data center, and rents back $1.2 billion of GPU capacity from NScale. As the host summed up, NVIDIA is NScale's "supplier, investor, guarantor, and customer."

Two other stories fit the pattern. On Computer Talk with TAB, the hosts reported that Oracle ($ORCL) sent a "force majeure" notice to developers of its $200 billion Project Jupiter data center. A force majeure notice is a legal step that protects a company from paying if events outside its control stop a project. And on The Business Brew, Steve Hou of Silicon Data explained that the CME is launching the first futures contracts on GPU rental prices. They will be based on Silicon Data's indexes for NVIDIA's H100 and B200 chips at neoclouds, settled against the monthly average rental price. The point is to let lenders and operators hedge a question Bill Brewster called central to the debate: "What's the residual value of these GPUs?" In other words, what is a used GPU actually worth after a few years?

4. The "interconnect wall": the next bottleneck inside the data center

On The Circuit, Ben Bajarin and Jay Goldberg interviewed Mark Wade, CEO of Ayar Labs, a private company building optical (light-based) connections for AI chips.

The problem Wade describes: AI systems link many GPUs together so they act like one giant chip. Today those links are copper wires. With copper, "the faster you want to go... the shorter distance you can go." He calls this the bandwidth-distance tradeoff. His claim is bold:

"Large-scale AI data centers are limited in their revenue scaling by the performance of the interconnect in their compute systems."

The fix is "co-packaged optics": putting tiny lasers and light detectors right inside the chip package. Ayar's optical engine is built on TSMC's 3-nanometer process. Key numbers and names:

• Timing: Wade expects the leading GPU makers to adopt it in 2028-2029, and the rest of the chip world in 2029-2031.
• Market size: "tens of billion, probably approaching $100 billion."
• Backers: Multiple rounds from NVIDIA, AMD, Intel and TSMC. Jesse Leiter, VP of Capital Strategy, said the Series E was led by Neuberger Berman, with Singapore's GIC, Qatar's QIA, Sequoia and ARK participating. The company has raised $650 million so far this year.
• Path to hyperscaler chips: Partnerships with MediaTek (which works on Google's TPU inference chips), Alchip (which works with Amazon) and GUC. The goal is for optical engines to be bought "just like how I integrate HBM," the stacked memory that sits next to every AI chip today.

Wade was open about how hard this is. Goldberg's one-word association for optical networking: "pain." Wade's answer: "you have to enjoy the process of pain and suffering."

5. Memory: pricing power is still intact, and Micron reports next week

Memory chips have been one of the hottest corners of the market, and the podcasts this week showed why.

On the AppleInsider Podcast, the hosts described Samsung asking smartphone makers for a $0.50 increase on a price of $1.50 per gigabit of memory. Other makers refused to negotiate that early, but Apple ($AAPL) agreed to $2. By the host's rough math (which he said he did not have in front of him), that takes memory from a low single-digit share (he first said "something like 8%") to about 23% of the iPhone's bill of materials, meaning the total cost of its parts. Tim Cook reportedly called it "a 100-year flood." Because Apple usually sets the terms, "everyone else is going to have to agree."

VAST Data's Hallak described the same squeeze in storage. One AI cloud customer said it would need about 500 petabytes over three years, then came back a week later wanting "an extra two exabytes on top." (An exabyte is 1,000 petabytes.) "Sometimes it scares me," Hallak said.

On Futurum Equities, Shai Bolor previewed Micron's ($MU) earnings next week. He had a warning about the headline: Micron's fourth quarter had 14 weeks and its first quarter has only 13, so sequential growth will look weaker than it really is. What he is watching:

• Progress on "strategic customer agreements," the long-term supply deals that could make the next downturn milder than past ones.
• Whether Micron keeps roughly 20% share as the industry moves to HBM4, the next generation of high-bandwidth memory.
• Capital spending, as a signal of whether new supply arrives in 2028 or sooner.

His bottom line: "I haven't heard one plausible evidence that memory pricing power is eroding, not one." He noted the stock trades at about five times earnings.

6. Watching demand in real time: GPU availability and rental prices

Warren Pies of 3Fourteen Research told Phil Rosen on Full Signal that his firm has tracked GPU availability since 2023. Throughout the day, they try to rent GPUs on demand from neoclouds and cloud providers and record how often they succeed. Zero means you can't get one at all; 100% means the market is loose. They build this into an index across NVIDIA's A100, H100, H200 and B200 chips.

His finding: availability leads rental prices. "If your on-demand providers are renting a Blackwell for $5 an hour, and there's no availability, those prices will start to creep up." He considers both better real-time signals than earnings, which "are always backward looking."

Pies thinks semiconductors are "ideal leadership" for this bull market. They are now almost 20% of total US stock market value. He sees semis and megacap tech together pushing the S&P 500 to about 8,000 "over the next couple of months." His blunt framing: "there's no world where this market does well if A.I. is a bust."

Key debates

1. Is AI spending a productive investment, or a debt-fueled bubble?

• Jensen Huang (NVIDIA), on Hard Fork / The Ezra Klein Show: productive. "$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." He argues NVIDIA's chips are "fungible" (any AI lab can use them for any task) and "durable" (software updates keep old chips useful). So NVIDIA computing power can become "an asset class, kind of like an airplane." An airplane starts as a passenger jet and ends its life as a cargo plane. If lenders see GPUs that way, "the cost of capital for funding NVIDIA AI factories will be the lowest," because the chips work as collateral. He expects demand for computing to rise "by a billion times" as hundreds of billions of AI agents join the billion human computer users.
• Paul Kedrosky, on Better Offline and Between Two COO's: a bubble. He says AI tokens (the units of text AI models produce) are getting 60% to 80% cheaper each year on a quality-adjusted basis. So "if I'm seeing a 70% year over year price decline in tokens... to stand still, I need a roughly 500% year over year growth." He dismisses the bullish margin math as "earnings before bad things." His view: "There's never been an episode where we've had this spending on this scale or it hasn't ended in a major recession, if not a depression." He goes further. If you remove AI-driven spending and war-driven energy costs from inflation, he calculates the US economy is actually deflating by about 0.25%. That would mean the Fed's recent rate hike is a policy mistake, similar to the one before the 1929 crash.
• Daniel Newman (Futurum): buy the fear. He called the recent "AI doom" wave a manufactured story and argued "every sell-off has led to a rally... there's alpha in skepticism." Alpha means returns above the market.

2. AMD vs. NVIDIA: who is the better stock from here?

• Daniel Newman on Futurum Equities described how the market sees it: "The market is treating NVIDIA as... the incumbent that has everything to lose. The market is treating AMD as the entrant competitor that has everything to gain." Investors can picture AMD growing 5x to reach NVIDIA's size. They struggle to picture NVIDIA going from $5 trillion to $20 trillion. He added that NVIDIA's story is already remarkable, in round numbers: doubling revenue again, 75% margins, "a trillion dollars in revenue over six quarters."
• Shai Bolor on the same podcast pushed back on AMD's price. AMD trades at 40 times its expected 2027 earnings, while NVIDIA trades at 15 times. He also flagged dilution (new shares that shrink existing owners' stakes). He said "I want to say it's like 160 million" AMD shares are held as warrants by Meta and OpenAI at about a penny each, so future earnings per share could be much lower than today's share count suggests.
• Newman added the execution risk: Helios is AMD's first full-rack system to take on NVIDIA, and it isn't generating revenue yet. "The price of the stock is almost like zero execution risk at this point. And that's where I'd be a little scary."

3. Memory stocks: still going up, or already past the peak?

• Bull (Shai Bolor, Futurum): No sign that memory pricing power is fading. Micron at about 5x earnings could re-rate above 10x if long-term customer agreements make it less of a boom-and-bust business. The AppleInsider discussion of Apple accepting a 33% memory price increase supports this.
• Skeptic (panelist on the Facts vs Feelings social hour): "You just saw every classic bubble sign and they moved about 30, 40% off the highs." His bull case is "sideways action for like a year plus," the way NVIDIA moved sideways for a year in 2023 before its next leg up. "I'm not a buyer of the semiconductors based on that."

4. Who wins the shift to inference?

(Inference is running a finished AI model to answer requests. Training is building the model in the first place.)

• Morgan Stanley's view, via The Wall Street Skinny: As more computing moves to inference, hyperscalers, frontier labs and startups like Etched are building their own inference chips. That threatens NVIDIA's share. Co-host Kristen argued NVIDIA's purchase of Hugging Face, the main hub where open-weight AI models are shared, is its answer: "NVIDIA is basically trying to own the platform." It gets data on who downloads what, and more developers stay on its CUDA software.
• Jensen Huang: NVIDIA's general-purpose design runs "from data processing to pre-training to post-training to eval to inference," so every AI lab stays on it.
• Paul Kedrosky: Once inference becomes a cheap commodity, it is "really just an energy problem," and China is adding roughly three times as much power capacity per year as the US. On who wins: "it ain't going to be open AI and Anthropic."

5. Should America sell its best chips to China?

• Jensen Huang: Export limits deprive the US of a market more than they deprive China of chips. "Are we depriving them a chip for their industry, or are we depriving the United States a market to compete in?" He noted Vera Rubin, NVIDIA's newest chip, goes to US frontier labs first anyway, and he is fine if the government makes that a rule.
• Michelle Weaver (Morgan Stanley): "The US is still very much winning on global share of compute, China is still very much winning on power." She doesn't expect much more American computing power to go to China. China is making its own chips but "just not able to keep pace." Meanwhile, Bloomberg Tech reported Alibaba unveiled an accelerator aimed at frontier training and a long-term plan for a 5 to 10 trillion parameter model, a direct challenge to NVIDIA inside China.

6. Are neocloud contracts worth what they say?

• Bull (the filing's framing): NScale's contracts are "real," and customers prepay about 23% up front. Its deferred revenue (cash collected for services not yet delivered) jumped from $2 billion to $6.5 billion in six months.
• Bear (the Run the Numbers host): The Anthropic contract, signed August 25, lets Anthropic walk away from a tranche if it's delivered late or if uptime falls short. And NScale has "no long-term contracts" guaranteeing its own chip supply, nor "binding commitments for any of the financings required" to build the Anthropic campus. "The single largest contract in the backlog runs through a campus that isn't built on power that isn't generating yet, funded by money that hasn't been raised."

Stocks discussed with bull/bear angle