Newsletter · 2026-08-29 · Ashutosh Agarwal

Nvidia Turns Financier as Big Tech's Off Balance Sheet AI Bill Hits 3 Trillion - Semiconductor Podcast Briefing - Week of August 29, 2026

Semiconductor Podcast Briefing for the week ending August 23, 2026. Podcast synthesis on Nvidia's roughly 500 billion dollar leasing vehicle with six private-equity firms, a Wall Street Journal count of about 3 trillion dollars of off-balance-sheet AI commitments across nine tech firms, Micron committing 10 billion dollars to a research lab, Marvell's 120 billion dollar Google deal, Analog Devices' first 4 billion dollar quarter, and strategists cutting chip exposure as the SOX breaks support.

Semiconductor Podcast Briefing

Week of August 29, 2026: Nvidia Turns Financier as Big Tech's Off Balance Sheet AI Bill Hits 3 Trillion

Week ending August 23, 2026.

This week the chip conversation on the podcasts stopped being about "how big is the demand" and started being about "who is paying for all of it." Nvidia is quietly turning itself into a lender. Big Tech's real AI commitments turn out to be roughly $3 trillion larger than their balance sheets show. Micron's CEO is spending $10 billion on a research lab because "memory is no longer a commodity." And underneath the AI-financing drama, the analog and memory chipmakers just put up some of the best numbers in their history, even as several strategists quietly cut their chip exposure and warn the group has "run too far, too fast."

TL;DR

• Nvidia is becoming a financier, not just a chipmaker. Its recently announced roughly $500 billion structure with six private-equity firms buys chips and leases them to AI labs like OpenAI and Anthropic, so the labs don't have to put the hardware on their own balance sheets. Broadcom is doing its own version, a debt deal now reported at $60 billion to as much as $90 to $100 billion.
• The real AI bill is about $3 trillion bigger than it looks. A Wall Street Journal analysis, dissected on one podcast, found nine top tech firms carry roughly $3 trillion of off-balance-sheet AI commitments: $1.9 trillion in purchase commitments plus $1.2 trillion in leases that haven't started. Alphabet alone shows $811 billion in purchase commitments as of June 30, nearly triple its roughly $600 billion of reported capex.
• Hyperscaler capex estimates keep leaping. The 2026 forecast has gone from $350 billion a year ago to about $1 trillion, with roughly $1.3 trillion pencilled in for 2027; one strategist put next year's number near $900 billion and said it is undershooting current demand, not building ahead of it.
• Two chip stocks now carry the market's earnings. Nvidia is expected to supply 18% of the entire S&P 500's 2026 earnings growth, and Micron another 14%, so 32% from two names.
• Micron is "tripling down on America." CEO Sanjay Mehrotra announced a $10 billion Micron Research Labs, calling memory "the strategic infrastructure of AI."
• Marvell landed a $120 billion Google deal. Google will spend up to $120 billion on Marvell's custom-silicon and networking products over six years, with warrants attached, potentially $20 to $25 billion a year by 2028, more than twice Marvell's current annual revenue.
• Analog Devices had its first-ever $4 billion quarter, up 40% year over year, with its communications segment (mostly AI data center) up 84%.
• The bears got louder. The chip index (SOX) broke a key support level after doubling in April and May; one strategist says semis "run too far, too fast," another has already cut exposure expecting "negative earnings" in 3 to 4 years, and a third called the AI setup a potential crash "five times worse than the dot-com bubble."

1. AI chip demand and hyperscaler capex (NVDA, AMD, AVGO, MRVL)

The single biggest storyline this week: Nvidia is using its balance sheet, not just its factory, to keep chips selling.

On Bloomberg Intelligence (Aug 21), Mandeep Singh, the firm's global head of tech research, explained the mechanics. Nvidia's roughly $500 billion arrangement with six private-equity players sets up a "special purpose vehicle," a separate company that buys the chips and then leases them out:

NVIDIA is using its balance sheet along with the private equity players to set up this special purpose vehicle, which will buy the chips and then lease it out to... users like OpenAI or other companies that want to use those chips.

The point, Singh said, is that fast-growing labs like Anthropic "don't have the balance sheet of the hyperscalers" to buy chips outright, so leasing keeps the purchases off their books. Broadcom is doing the same thing with its own $60 billion debt deal to fund chips for Anthropic. Importantly, Singh does not think Nvidia itself is at risk of running out of cash:

They're generating... like almost $100 billion plus in free cash flow every year... And with 75% gross margins, they will continue to do that in the foreseeable future. And they're not setting up... their own fabs. So the only investment they have is really in terms of backstopping a lot of these SPV structures.

How big is the spending? The numbers keep getting bigger. On Market Call (Aug 20), Nick Mersch, a portfolio manager at Purpose Investments, laid out how fast the estimates have moved:

At the beginning of last year, we thought we were going to spend $350 billion on CapEx for 2026. Now that number is looking like a trillion, going up to $1.3 trillion for next year... there's still a significant amount of demand that is not being met by the current supply.

On Excess Returns (Aug 22), Charles Schwab's chief investment strategist Liz Ann Sonders put next year's figure near $900 billion and drew the key contrast with the late-1990s telecom boom, since this build is chasing demand that already exists rather than betting on demand that might show up:

This time the build is associated with current demand, and it's undershooting current demand. But that doesn't mean it lasts forever.

She also flagged just how concentrated the market's earnings have become: Nvidia is expected to account for 18% of the S&P 500's 2026 earnings growth and Micron another 14%, so "32% for two stocks, two chip stocks."

The custom-chip wave: Marvell's $120 billion Google win. On The AI Investor Podcast (Aug 20), the hosts broke down a landmark deal: Google agreed to spend up to $120 billion on Marvell's products (custom inference chips, networking, storage controllers, memory interfaces) over the next six years, with Google receiving warrants to buy Marvell stock if it hits that spend. The context is that Google wants to diversify away from Broadcom on its TPU (its in-house AI chip) program:

Google is getting warrants to purchase Marvell on this. And they only invest if they spend 120 billion on Marvell's products across the next six years... by 2028, they're already going to be spending maybe 20 or 25 billion on Marvell. That's more than twice the size of the current company.

The hosts cited a Bank of America estimate for the total market for these custom chips: Broadcom $250 to $350 billion, MediaTek $100 to $150 billion, and Marvell $50 to $100 billion, so Broadcom still keeps the majority. They argued Broadcom's roughly 10% two-day sell-off on the news was an overreaction, since Broadcom controls scarce Taiwan Semiconductor capacity, packaging, and high-bandwidth memory allocation. Their bigger structural point: the AI networking market is expected to grow "from about 10 billion in 2023 to 245 billion by the end of the decade."

A note on the newer hardware economics. The same podcast explained why the buildout is accelerating: Nvidia's newest GB300 systems "produce 50 times more tokens than an H100," which makes the economics attractive enough that everyone wants to build as fast as possible, raising the genuine risk of an eventual overbuild in 2027.

2. Memory: HBM, DRAM and NAND (MU, SK Hynix, Samsung, Sandisk)

Memory was the feel-good story of the week, headlined by Micron's CEO. On Squawk on the Street (Aug 20), Jim Cramer interviewed Micron CEO Sanjay Mehrotra from the company's Boise headquarters, where Micron is committing $10 billion to a new "Micron Research Labs" for long-horizon memory research. Mehrotra's framing, that memory has graduated from cheap commodity to essential AI infrastructure, is worth quoting in full:

Memory is no longer a component in a system. Memory is the strategic infrastructure for AI. And it's no longer a commodity. It is a high value... Without memory, you cannot make AI smarter. You cannot make AI faster. You cannot scale up AI.

Cramer summed up the American-manufacturing angle: "Sanjay doubled down on America. Now, this way, he's actually tripling down." Micron holds 62,000-plus patents and, Mehrotra noted, kept investing in leading-edge research even through "the deep downturn of 2023."

The memory rebound is broad. On The AI Investor Podcast (Aug 20), the hosts highlighted two more data points that lit a fire under the group: SK Hynix announced a $29 billion share buyback (a vote of confidence in its own shares), and Sandisk, at an investor event, guided to 80% adjusted gross margins and 75% operating margins from 2026 to 2030, with long-term agreements now covering two-thirds of its output. Their takeaway is important for how to read the cycle: memory margins have historically peaked around 60% and rolled over, but this time companies are locking in mid-80% margins through multi-year contracts rather than volatile spot pricing, which, if it holds, makes the cycle far more durable than the old boom-bust pattern.

The cautionary counterpoint. Liz Ann Sonders on Excess Returns (Aug 22) used Samsung as a warning about how high the bar has gotten: Samsung beat the published analyst consensus "by a pretty handy margin" but still "underperformed the buy-side expectation," and the miss-versus-hope helped drive a roughly 40% drawdown in Korea's KOSPI index (Samsung and SK Hynix are about half that index). Her point: with chip stocks this loved, beating the official estimate is no longer enough.

3. Analog, auto and industrial semis (ADI, TXN and peers)

Analog Devices delivered the cleanest fundamental beat of the week. On Schwab Network (Aug 21), Rick Ducat walked through ADI's fiscal third quarter (reported Aug 19):

• First-ever $4 billion quarter: revenue of $4.02 billion, above the $3.92 billion estimate; adjusted earnings of $3.45 per share versus $3.33 expected.
• 40% revenue growth year over year, led by data center and industrial.
• Industrial (about half of revenue) up 53%; communications up 84%, with data center now the bulk of that segment; automotive up 16%; consumer up 6%.
• Gross margin of 72.5%, down 50 basis points from the prior quarter but up 330 basis points year over year.
• Guidance for the next quarter of $4.2 to $4.4 billion, above the $4.07 billion estimate; management said optical circuit switching revenue is "poised to approximately double this year" with similar growth targeted for 2027.
• ADI also closed a $1.5 billion acquisition of Empower Semiconductor on July 7 to push deeper into power delivery.

Ducat named the honest risks too: ADI exited the quarter with "record balance sheet inventory and elevated inventory at distributors," while distributor "channel weeks fell below its 6 to 7 week target." And he framed the sector-wide vulnerability plainly:

Global AI CapEx by tech giants and major hyperscalers is projected to exceed $700 billion to $1 trillion in 2026, so any slowdown in spending could lead to potential correction risks for share prices as ADI is trading at elevated price-to-earnings multiples.

He also noted the wider analog group is mixed, with competitor ON Semiconductor having "seen a big collapse recently," and that ADI (up roughly 52% over the past year) has lagged the broader semiconductor ETF (up roughly 94%), largely because memory names have run so hard.

4. Foundry and manufacturing: Taiwan risk, TSMC, Intel (TSM, INTC)

The foundry conversation this week was mostly about geopolitics. On The Peter McCormack Show (Aug 22), policy analyst Dmitri Alperovitch laid out just how concentrated chip production still is:

The estimates are about 90% of advanced chips go through Taiwan. Over 40% of all chips, period, go through Taiwan.

He explained that beyond the handful of "advanced" chips (the AI processor, the memory, the modem), a device like an iPad contains "about 150, 160 foundational chips," the unglamorous parts that everything from cars to refrigerators depends on. On the reshoring push, he credited the Trump administration's tariff pressure with getting TSMC to commit "$250 billion in Arizona, committing to build six fabs there over the next five years or so," which, if completed, would let the U.S. "produce half of the chips we need." But he was blunt that this does not remove the dependence on Taiwan:

Most of the expertise is still back home in Taiwan. The process engineers and so forth. So without them, you could have very, very expensive warehouses, basically... It's a fallacy to think that just because we've increased production in the U.S., we no longer need Taiwan.

He put a number on the stakes: a conflict or blockade of Taiwan would wipe out "$5 trillion of GDP in the first year," and he called that estimate conservative. He also recounted a 2023 conversation with then-Intel CEO Pat Gelsinger, who warned that if Intel didn't get its CHIPS Act money that quarter, "Intel may not survive"; the money came too late, and Intel only stabilized after the Trump administration took an equity stake.

On the market reaction to TSMC, Matt Maley of the Maley Market Strategy Podcast (Aug 21) flagged a worrying signal: TSMC reported "amazing earnings, amazing guidance" but had already fallen 12% before the report and still fell afterward. Because it wasn't overbought going in, he argued that decline reflects genuine demand worries, not just routine profit-taking.

5. China, export controls and the commoditization threat

The China angle this week was less about export bans and more about cheap Chinese AI models undercutting demand for expensive Western chips. Multiple hosts pointed to low-cost Chinese open-weight models (Kimi from Moonshot, DeepSeek, Alibaba's Qwen) as a real risk to the chip thesis, because if enterprises can run good-enough models cheaply, they buy less compute.

• Arthur Hayes, on Unchained (Aug 21), said the market will eventually punish overspending "when you have a Chinese competitor that has a model that's just as good and is one one hundredth of the price for the same thing."
• Matt Maley (Maley Market Strategy, Aug 21) framed the debate directly: cheaper Chinese models could either raise chip demand (more usage) or lower it (if they displace paid demand at OpenAI and Anthropic), and "this whole issue of the commoditization of this business is one that is becoming a bigger one."

Dmitri Alperovitch (The Peter McCormack Show, Aug 22) offered a contrarian read: he expects China's giveaway of open-weight models to be short-lived, because "even China cannot afford to spend hundreds of billions of dollars on chips and infrastructure to give it away." He predicts leading Chinese labs go "closed-weight" within a couple of years, just as Meta did in the U.S.

6. The debt-and-financing debate

This deserves its own section because it ran through nearly every podcast: how the AI buildout is being paid for, and what happens if the money stops.

The $3 trillion off-balance-sheet number. On The Banker Next Door (Aug 20), Dr. Joseph Bergquist walked through a Wall Street Journal analysis showing that nine top tech companies carry roughly $3 trillion of AI-related commitments that don't appear on their balance sheets: about $1.9 trillion in purchase commitments (chips and services locked in with suppliers) plus $1.2 trillion in leases that haven't started yet. The standout:

Alphabet's purchase commitments and contractual obligations have exploded and stood at $811 billion as of June 30th... it is hard to tell from its disclosures what precisely it intends to buy.

That $811 billion is up from $332 billion just three months earlier and is nearly triple Alphabet's roughly $600 billion of reported annual capex. He also detailed Meta's Hyperion data center in Louisiana ("the size of about 1,700 football fields"), where a joint venture owned 80% by Blue Owl Capital keeps the project and its $27 billion of construction debt off Meta's books. And he underscored the cash-flow strain: Alphabet, Amazon and Meta have all gone to negative free cash flow this year and are projecting 10 to 20% negative free cash flow into 2027.

"The gray swan." On Eurodollar University (Aug 21), Erik Snyder argued Nvidia's half-trillion-dollar financing pledge is itself a warning sign, proof that customers can't fund the chips directly. He compared it to Nortel and Lucent lending money to their own customers during the dot-com era, and explained the mechanism that puts Nvidia on the hook: a "residual value guarantee," where if leased chips lose value and a customer defaults, Nvidia promises to make the lender whole. His conclusion:

NVIDIA sort of becomes the gray swan because it's at the middle of so many of these Jenga pieces that are stacked on top of it.

He argued this is why "NVIDIA shares are trading at a discount to... their cash flow," because investors are treating it "more and more like a risk."

"The maths don't math." On The Enterprise AI Show (Aug 23), hosts Brian, Brandon and Aaron zeroed in on how completely Jensen Huang's message has flipped. For years he told customers that being one generation behind meant being "close to going out of business," implying short chip lifespans. Now, to keep the chips selling, he is recasting GPUs as long-lived, bond-like assets:

Jensen has essentially... done a 180 from you have to be one day ahead or you're useless to 10, 12 year horizons for AI and GPUs is perfectly fine.

They pointed to Nvidia's $105 billion backing of OpenAI's Ohio data center as the latest example, and drew the uncomfortable "Big Short" parallel of "taking a bunch of stuff that may be volatile [and] making it seem really stable."

The lifespan debate matters more than it sounds. On the bullish side, Nick Mersch (Market Call, Aug 20) argued the chips last far longer than accountants assume, since NeoCloud providers CoreWeave and Nebius are still renting Nvidia's 2020-era A100 chips and have contracted them out to 2029, a nine-year life against a six-year depreciation schedule (Michael Burry has argued they're only worth three years). If GPUs really do earn cash for nine years, the return math improves a lot.

The bears with a target. On The SharePickers Podcast (Aug 22), Justin Waite used CoreWeave to illustrate the "circular deals" risk: CoreWeave has a $50 billion market cap, roughly $100 billion enterprise value once debt is counted, and negative free cash flow of $13.66 billion over the past 12 months. Nvidia has put in over $2.1 billion of equity and granted preferred chip allocation, which CoreWeave uses to buy, and borrow against, more Nvidia chips. His warning:

This is fine at the moment, as long as there's massive demand for chips and AI. But what happens when demand plateaus?... Companies will default on their debt obligations. Chip prices will crash and bring the entire financial system down with it.

He argued a crash could be "three to five times bigger" than the dot-com bust because the companies are far larger and passive index funds are so heavily concentrated in them. On Unchained (Aug 21), Arthur Hayes went further, calling AI capex "a massive capital misallocation," saying "you're spending trillions of dollars to build and power these containers for a chip that has a maximum two-year life cycle" and predicting a financial reckoning by "mid-2028."

The measured middle. On Monetary Matters (Aug 20), Luke Gromen of Forest for the Trees Research noted hyperscaler bond issuance of about $500 billion for 2026 and Google's $800 billion in forward purchase commitments, and warned every one of the prior five great capex booms (canals, railroads, telecom and more) "ended in a bust, and this one will end in a bust too." But host Jack Farley pushed back hard against shorting the chips, arguing "the track has been laid for NVIDIA to make $300 billion in operating profits." Gromen's practical advice was to take profits into strength rather than bet against it. That "genuine technology, overbuilt anyway" framing, productivity miracle first and financial bust second, was the closest thing to a consensus across the week.

7. Cyclicality: has the chip cycle already peaked?

Several strategists think the correction has begun. On Something More with Chris Boyd (Aug 20), portfolio manager Brian (of Wealth Enhancement/AMR) said semiconductors and semi ETFs have driven all of his portfolio's outperformance this year, and that he is now cutting the position:

In the next three, four years, we're probably going to see negative earnings. And that's how these markets go. That's generally what we mean by a cyclical market.

His logic: as the cost of running AI ("token costs") falls and hardware supply catches up, compute demand will eventually moderate, echoing the dot-com fiber glut where too much capacity got built and the reckoning came later. He noted the multi-year "take-or-pay" contracts (customers must pay even if they don't need the chips) should keep the money flowing for a few more years, but flatly stated: "I think we already have entered [a bear market] in the semiconductor space."

The charts agree, according to Matt Maley (Maley Market Strategy, Aug 21): the SOX semiconductor index has broken below its key 12,000 support level after doubling in just April and May. Because the group doubled, he warned, it "could fall a lot further without doing any major damage to the charts."

They have run too far, too fast... more correction to go.

Nick Mersch (Market Call, Aug 20) was more balanced. He rates the SOXX semiconductor ETF a "hold" after its July momentum reversal (it had been up roughly 115 to 120% year-to-date at the peak), noting fundamentals actually improved even as prices fell, but he "wouldn't necessarily be adding" here. His nuance is worth keeping: memory chips are now "a little bit more commodity-based," so the real thing to watch is when margins start rolling over, since "you can have periods of glut."

Earnings reactions

Context on the AI labs (the demand behind the chips): podcasts cited Anthropic at roughly a $60 to $65 billion revenue run rate and reportedly profitable on a GAAP basis, versus OpenAI at about $40 billion (rising toward $48 billion) with a Q2 operating margin near negative 122% and an operating loss around $9.2 billion. Anthropic is reportedly preparing an IPO filing, with prediction markets putting roughly 82% odds on a listing this year (Chit Chat Stocks). These run-rates matter for chips because, as one host put it, if OpenAI or Anthropic demand slows, "it could ripple throughout the ecosystem."

What I'm watching next week

• Nvidia's earnings. Every podcast this week was effectively a preview, and the results themselves are the pending catalyst. Watch the reaction function, not just the beat: the same strategists warned that "buy the rumor, sell the news" has already hit TSMC, Broadcom, Micron and Samsung.
• The SOX 12,000 support level. Matt Maley flagged it as the line in the sand for whether the chip pullback stays a "rotation" or turns into something the whole market follows.
• Anthropic's IPO filing. Reportedly due around the end of August. It would be the first clean look at a frontier lab's real financials, and a direct read on whether the demand underpinning the chip buildout is as strong (and profitable) as advertised.
• Memory pricing and margins. With Micron, SK Hynix and Sandisk all pointing to durable 80%-plus margins locked in through long-term contracts, watch for the first sign of any margin roll-over, historically the tell that the memory cycle is peaking.
• The 30-year Treasury yield and hyperscaler debt issuance. With hyperscalers now funding capex through the bond market (roughly $500 billion of issuance in 2026), higher long-term rates directly raise the cost of the AI buildout. Several hosts tied rising yields straight back to AI borrowing.
• ADI's fiscal Q4 (guided to $4.2 to $4.4 billion) and any read-through to the rest of the analog group (TXN, MCHP, NXP, ON) on whether the industrial and auto recovery is broadening.
• Any further off-balance-sheet disclosures. After the $3 trillion figure, watch whether more of these purchase commitments and leases get pulled into the spotlight, and how the market reacts when they do.