Newsletter · 2026-09-05 · Ashutosh Agarwal

Nvidia Guides to 70 Percent Growth While Memory Sells Out to 2028 - Semiconductor Podcast Briefing - Week of September 5, 2026

The Semiconductor Podcast Briefing for the week of September 5, 2026, covering podcasts from August 26 to September 4: Nvidia's $96.2 billion quarter came with a first-ever full-year guide of roughly 70 percent growth, Broadcom's 221 percent AI ramp still knocked the stock down, and DRAM prices up more than 200 percent with no new supply until 2028 became the story under the story.

Semiconductor Podcast Briefing

Week of September 5, 2026: Nvidia Guides to 70 Percent Growth While Memory Sells Out to 2028

This was the biggest week of the chip calendar. Nvidia reported a blowout quarter and did something it had never done before, give a full year of guidance, and the number was so large it reframed the whole debate. Broadcom then reported an equally strong AI print a week later and the stock fell. Memory chips are, in the words of one veteran analyst, an "8 out of 10 on the crazy scale, heading to a 10." And underneath all of it, the "is this a bubble?" argument got louder and more specific on both sides.

Everything below comes from podcasts published in the last week (August 26 to September 4). Where a speaker is named, the role and the date they said it are given.

TL;DR: Things That Mattered This Week

• Nvidia's quarter was enormous, and the guide was the real shock. Revenue of $96.2 billion, up 106% year-over-year, with data-center revenue of $89 billion (up 117%), roughly 92% of the company. For next fiscal year it guided to ~70% revenue growth versus a Wall Street consensus of ~45%. The stock jumped 8.7% the next day and added a reported $442 billion of market value in a single session (Bloomberg Daybreak: US Edition, Aug 27; The Rundown, Aug 27).
• The headline capex number went vertical. Jensen Huang told investors he expects the top five hyperscalers to spend $1.3 trillion on AI next year, up from ~$800 billion this year and ~$500 billion in 2025, a 60% jump (Motley Fool Hidden Gems Investing, Aug 28). Morgan Stanley's wealth desk says one strategist's 2027 number is even higher, at $1.5 trillion versus the Street's $1.2 trillion (The Compound and Friends, Sep 4).
• Memory is the tightest it has ever been. DRAM prices are up more than 200% year-over-year, "historically unprecedented", and there is essentially no unsold supply left, with no meaningful new capacity until 2028 (The Economics Show, Sep 4).
• Broadcom reported a monster AI quarter, and the stock dropped ~3%. AI chip revenue was $16.7 billion, up 221% year-over-year, and the CEO sees a path to $230 billion of semiconductor revenue by 2028, but near-term margin and guidance detail disappointed (Schwab Network, Sep 2; Bloomberg Intelligence, Sep 3).
• Nvidia is spending like a strategic acquirer. It agreed to buy open-source AI hub Hugging Face for ~$12.9 billion (plus a $1 billion retention pool), roughly 86 times revenue on a company doing about $150 million a year, and separately hired ~100 engineers out of Poolside (This Week in Startups, Aug 28; Elon Musk Podcast, Sep 3).
• The bubble debate hardened. Bears zeroed in on Nvidia's customer concentration and the "circular financing" web; bulls argued the spenders today are far higher quality than the dot-com era. Economist Tyler Cowen flatly said "the AI bears are asking the wrong questions" (Prof G Markets, Sep 4), while writer Ed Zitron laid out the most detailed bear case of the week (The Compound and Friends, Aug 28).
• Washington wants the fabs onshore. Commerce Secretary Howard Lutnick said the goal is to move US chip production from ~1% to 40–50% of domestic consumption, backed by $1.2 trillion in commitments including TSMC's $265 billion and Micron's $250 billion, and that Samsung and SK Hynix are "on notice" (Balance of Power, Sep 2).

1. AI Chip Demand and the Hyperscaler Capex Number

The single most-repeated figure of the week was Jensen Huang's capex forecast. Jon Quast of Motley Fool Hidden Gems Investing framed the quarter as "as much as a macroeconomic pulse check as anything":

"Jensen Huang's saying $1.3 trillion is what he expects to be spent next year just by the top five. That's a 60% year-over-year jump."

That's up from roughly $500 billion in 2025 and about $800 billion this year (Motley Fool Hidden Gems Investing, Aug 28). His co-host Matt Frankel put the scale of the spenders in context, and offered the bull's favorite rebuttal to the "how can they afford this?" worry:

"Over the past four quarters, Meta, Alphabet, Amazon, Microsoft, and Oracle... generated almost $700 billion in operating cash flow. So maybe with that context, that big CapEx number isn't really as scary as it seems."

The point about who is actually paying Nvidia recurred all week. On Morning Brew Daily (Aug 27), the hosts noted Meta, Amazon, and Google are each guiding to roughly $200 billion of AI capex this year, most of it flowing to Nvidia, and that Nvidia is providing about $500 billion of vendor financing through outside financial firms to help smaller, less creditworthy customers afford its chips.

CNBC's Arjun Karpal, ahead of the print, pegged hyperscaler capex at "north of $730 billion this year" and flagged that Nvidia's next chip, Vera Rubin, is expected to begin shipping this fall, with investors watching the customer transition from Blackwell to Rubin (Morning Call, Aug 26).

Morgan Stanley's Dan Skelly made the most detailed bull argument for why the spending sticks, and added a nuance worth holding onto: the constraint may be physical, not financial.

"60% of 2Q earnings... came from AI infrastructure... His numbers for AI capex in '27 are $1.5 trillion. The Street is at like $1.2 [trillion]... every quarter, you can bank on the numbers going higher... it's a generational competitive risk among the U.S. players."

He also cautioned that "under the hood, in reality, it's more about deployment constraints", that the industry is ordering servers and chips ahead of "40 gigawatts, 50 gigawatts of projected data center construction," raising a real risk of double-ordering in the supply chain (The Compound and Friends, Sep 4).

2. Memory Pricing: The Standout Story Under the AI Headlines

If Nvidia was the loudest story, memory was the most striking. On The Economics Show (Sep 4), Dan Kim, chief strategy officer at TechInsights and former chief economist at the US Commerce Department's CHIPS Program Office, was asked how crazy the memory market is on a scale of 1 to 10:

"Right now it's an 8. Next quarter it'll be a 9, and the quarter after that it'll be a 10. It'll stay at a 10 until the end of 2027 at least."

The numbers behind that:

"From last year to this year, we're anticipating an over 200% price increase of DRAM. That is historically unprecedented... there is no new supply coming online in significant volume until at least 2028."

A few concrete consequences Kim laid out:

• There is essentially nothing left to buy. "Effectively there is no supply out there that hasn't already been spoken for... practically most of the supply has been sold out." That's true for high-bandwidth memory (HBM, the DRAM paired with AI accelerators) and increasingly for ordinary memory too.
• Price signals have stopped working normally. Even buyers willing to pay the quoted price are being told they can only get a fraction of what they want, and it's now hitting the low end of the market too, "automotives, medical device manufacturers, and others."
• This reaches your phone. Memory is already among the three highest-cost components in a smartphone (alongside the main processor and NAND storage). Kim said the memory share of a phone's bill of materials could rise from a typical 15–25% toward "30% and 40% and even 50%."
• DRAM is now a three-firm market, Samsung, SK Hynix, and Micron, with Chinese entrants YMTC (in NAND) and CXMT (in DRAM) the only realistic new competitors, "because there is almost no end to the state support that they're receiving."

Kim's explanation of why supply is short is a clean cautionary tale about cyclicality: in the 2023 down-cycle, memory makers rationally cut production and delayed buildouts just as the AI demand "tsunami" was forming, and new fabs take "somewhere between 2 to 6 years" to bring online.

This also showed up inside the Nvidia story. On Buy Hold Rant (Aug 26), the hosts noted Nvidia's memory purchase commitments have ballooned from about $119 billion to $279 billion, a direct read on how much memory the AI buildout is locking up.

3. Semiconductor Capital Equipment (WFE)

The week's one concrete equipment forecast came from The MoneyFlows Show (Ep. 40, Sep 3), where Jason Bodner and Lucas Downey pointed to the equipment names as the downstream beneficiaries of the Nvidia and memory boom, citing industry wafer-fab-equipment spending of roughly $190 billion in 2027, about 20% growth, and flagging Applied Materials, Lam Research and KLA specifically.

The logic is straightforward: the memory shortage described above only ends when Samsung, SK Hynix, Micron and the Chinese makers finish the fabs they are all racing to build, and every one of those fabs is an equipment order.

4. Foundry and Manufacturing: Policy Did the Talking

Commerce Secretary Howard Lutnick laid out the administration's onshoring push on Balance of Power (Sep 2):

"When the Trump administration walked in the door... we had like 1% production of semiconductors. And our goal is now 40%. And if Intel can be successful, which it's clearly on the way to be, we can get to 50% domestic."

He tied it directly to the tariff structure, "if you build here, there's no tariff", and cited $1.2 trillion in commitments already, including TSMC at $265 billion and Micron at $250 billion. On Samsung and SK Hynix: "You are on notice that they are going to build here." He conceded the fabs can't be finished in two years, "that's breaking the sound barrier", but wants "shovels in the ground." He repeated the same figures on Squawk Pod (Sep 2).

5. Analog, Auto and Industrial Semis

The clearest read-through for the analog, auto and industrial names came indirectly, through memory. Dan Kim's warning is that allocation is now squeezing "automotives, medical device manufacturers, and others" at the low end of the market (The Economics Show, Sep 4), a cost and availability headwind for exactly these end markets. The auto-adjacent podcast conversation this week was almost entirely about tariffs on finished vehicles and parts rather than the chips inside them.

6. China, Export Controls and Tariffs

Two threads mattered for chips specifically:

• China is essentially a zero in Nvidia's numbers. CFO Colette Kress said Hopper sales into China were less than 1% of data-center revenue in the quarter (versus zero the prior quarter), and Bloomberg Intelligence's Kunjan Sabhani said consensus still carries $0 for China, so any reopening is upside, but bounded: "even if they can open up and start shipping initial samples... that's about 10, 20 or so billion dollars per year... I don't think that's going to move the needle a lot" given the size of the business now (Bloomberg Businessweek, Aug 26).
• Tariffs largely missed the AI supply chain. On InvestTalk (Sep 4), the hosts noted tariffs meant to curb imports failed to target data-center gear, with record capital-goods imports of $140.3 billion in semiconductors and computers surging on the AI buildout, arguing the policy was "planned for an economy in 2018, not what's going on in 2026." Supply chains have shifted from China toward Vietnam, Mexico, Taiwan and South Korea.

7. Earnings Reactions

Three big chip prints landed inside the window. The reactions were as interesting as the numbers.

On Nvidia's one soft spot, gross margin and memory costs. Bloomberg Intelligence's Kunjan Sabhani flagged a "slight miss on the gross margin" and explained where Nvidia's newly announced 15%+ price increases (starting January 2027) actually go:

"Our estimate would be it's all going to offset the rise in memory prices... NVIDIA purchases the memory from the memory suppliers and has a pass-through model... most of this is going for that input rise from wafer, so to TSMC, and the rest to the memory makers."

In other words, Nvidia's price hike is largely a conduit: the beneficiaries are Micron, Samsung and SK Hynix, not the server makers like Dell (Bloomberg Businessweek, Aug 26). Sabhani also captured why a great print didn't move the stock much: "We are in that bubble gap timing-wise this quarter and next quarter where everyone's trying to figure out what happens in [20]28."

Why Broadcom sold off on a great quarter got the clearest explanation from Bloomberg Intelligence's Matthew Bloxham (Bloomberg Intelligence, Sep 3):

"What we heard from them was that they expect both Anthropic and OpenAI to become their most important customers, eclipsing Google... it kind of shows you that there's substance behind the plans and capital raising that these companies are undertaking."

The strategic story is hyperscaler diversification away from Nvidia, but The Information's TITV (Sep 3) added the wrinkle that some of Broadcom's Google TPU work is shifting (with MediaTek picking up training chips), and Google's TPU units are forecast to grow from 9 million in 2027 to 15–16 million in 2028. On valuation, The AI Investor Podcast (Sep 3) noted Broadcom trades around 10.3x fiscal-2028 earnings on ~100% forecast growth, cheap if the double-and-double-again AI revenue path holds.

8. Mergers, Deals and Financing

An unusually busy week, dominated by Nvidia:

• Nvidia buys Hugging Face for ~$12.9 billion (plus a ~$1 billion retention pool). Widely described as "the GitHub of AI," with ~18 million users and ~3 million models. The eye-watering part is the multiple, roughly 86x revenue on ~$150 million of annual revenue, which nearly every host read as a defensive move to own the open-source AI distribution layer and protect Nvidia's GPU position (Elon Musk Podcast, Sep 3; Motley Fool Hidden Gems Investing, Aug 31; Brew Markets, Sep 3). On This Week in Startups (Aug 28), Jason Calacanis argued Jensen is positioning Nvidia as "the U.S. open-source champion."
• Nvidia acqui-hires Poolside, roughly 100 engineers from the AI coding lab for a reported $6 billion plus a $1 billion licensing fee (This Week in Startups, Aug 28).
• Nvidia's MediaTek move was framed as a "power play against Broadcom" in the inference market (The Morning Market Briefing, Sep 1).
• Broadcom's giant financing deal. Telltales (Aug 30) reported Broadcom is in talks on a $70–80 billion chip-financing deal ($45B senior, $35B junior) with Blackstone and Apollo to support AI companies including Anthropic, with Broadcom providing credit guarantees rather than loans, a guarantee that would exceed the company's ~$65 billion of total debt.
• Nokia is buying an NXP fab in Arizona and has absorbed Infinera, part of a vertical push into optical data-center communications (Chip Stock Investor Podcast, Sep 1).

A recurring theme tying the deals together: the "circular financing" web. Motley Fool's Matt Frankel described Nvidia as sitting at the center of financing partnerships with "Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR, to raise over $500 billion of third-party capital", and was candid about the accounting concern: "If I pay Jason $100 to teach me something and he gives me the $100 to teach him something, did we each really make $100?" (Motley Fool Hidden Gems Investing, Aug 28).

9. Cyclicality and the "Is This a Bubble?" Debate

This was the philosophical center of the week, and it split cleanly.

The bears. The most detailed case came from writer Ed Zitron on The Compound and Friends (Aug 28). His core argument is that demand is dangerously concentrated in two unprofitable startups:

"The hyperscalers have now become financially dependent on the growth of OpenAI and Anthropic... $440 billion of cloud revenue across Google, Amazon, and Microsoft coming just from OpenAI and Anthropic, two unprofitable startups who need to constantly raise money."

On Nvidia's own customer base, he cited: "16% of their latest quarter revenue was one customer. 44% of their first half of fiscal 2027 was three customers. Five customers make up 70% of their accounts." And on Microsoft specifically: "Fiscal 26... $34.33 billion of AI revenue. $24.1 billion of that is OpenAI," which he argued leaves Microsoft's non-OpenAI AI business as "a single-digit billion" business against "$260-plus billion in capex." He also flagged a funding wall: OpenAI's last round was "$122 billion... only $12 billion of that came from venture capital."

Other bears in the week:

• Rebel Capitalist drew the Cisco analogy explicitly, Nvidia today as Cisco in 1999, "both experienced 10x gains followed by 80–90% declines", and floated a scenario of Nvidia falling ~90% (Rebel Capitalist News, Aug 31). A separate episode framed Nvidia as evolving into a "shadow bank" via off-balance-sheet guarantees (Rebel Capitalist News, Aug 31).
• RenMac called it an "earnings bubble," comparing hyperscaler vendor financing to Nortel and pointing to widening credit spreads as a sign "bond investors have figured it out" (RenMac, Aug 28).
• Jay Goldberg (Seaport Research Partners) said there are "definitely some signs of bubble percolating," with circular financing "creating demand that might not exist independently" (The Credit Edge by Bloomberg Intelligence, Aug 27).
• Dan Niles is "bearish into November, room to run after," and is watching hyperscaler credit-default swaps as his early-warning gauge (Excess Returns, Sep 3).

The bulls. Economist Tyler Cowen rejected the framing entirely (Prof G Markets, Sep 4):

"I think it's the wrong discussion. I don't like the word bubble. Were automobiles a bubble in the 1920s? ... A bubble to begin with is something like the South Sea Bubble or the Dutch tulip bulbs that just made no sense. This is not that."

On the same numbers the bears cite, "$2.5 trillion in CapEx so far, about revenues of $150 billion", Cowen acknowledged "a correction wouldn't surprise me," but argued the differences are decisive: "revenue growth for the current AI companies looks much better than what we saw before the dot-com bubble burst... debt plays a much less significant role in this sector." His candid worst-case: if the weakest players fail with too much debt, "as messes go, I don't even think it would be close to the worst we've seen. Not close, say, to 2008."

Morgan Stanley's Dan Skelly made the "quality of spenders" version of the bull case, contrasting today's buyers with the dot-com era (The Compound and Friends, Sep 4):

"In 1999, in order for Cisco to hit its growth targets... they were reliant upon selling to a customer that had gone public a week prior... The customers today are Amazon buying on behalf of its 30 million cloud customers who represent every sector of the economy."

His other memorable point: the "dark fiber" of 2000 stayed dark for a decade, but AI chips are more fungible. "A lot of those chips can be reverted back to cloud. And so do we have dark GPUs the way we had dark fiber? Maybe not as likely." He did flag the honest asterisk, though: "the two biggest players in the ecosystem on the buy side [OpenAI, Anthropic] are not public yet... not profitable, not particularly transparent... not proven through any sort of economic cycle."

And Aaron Katz, CEO of ClickHouse, called the bubble narrative "a widely held belief that is wrong," arguing the cycle is "accelerating at an unprecedented pace" (The Twenty Minute VC, Aug 31).

10. China Indigenization: SMIC, CXMT, Huawei

One genuinely notable milestone landed: CXMT reached mass production of LPDDR6 memory ahead of Samsung and SK Hynix, at speeds of 10.7–12.8 gigabits per second (TechLinked, Sep 1), a signal that China's domestic DRAM effort is closing the gap faster than expected in some product categories. Separately, CXMT has sued the Pentagon over its inclusion on a Chinese-military-company blacklist, arguing its chips are for civilian and commercial use (Mac OS Ken, Sep 1). Dan Kim's broader framing applies here: Chinese memory makers are the only plausible new entrants precisely because of near-unlimited state support (The Economics Show, Sep 4).

What I'm Watching Next Week

• Micron's print and any HBM/DRAM contract commentary. With DRAM up 200%+ and supply sold out to 2028, Micron is the cleanest public read on how much of this pricing is being locked into long-term agreements, and how much flows to the bottom line.
• The Vera Rubin ramp. Rubin enters full production in Nvidia's fiscal fourth quarter; watch for confirmation of the Blackwell-to-Rubin transition, the promised inference cost-per-token improvement (management has cited 60–70%), and whether the memory-driven gross-margin dip (guided ~74%) stabilizes.
• The January 2027 price increases. Nvidia's 15%+ server price hike starts in January; the question is how cleanly it passes through and whether any party in the chain (Dell, HPE, Supermicro, or Nvidia) ends up absorbing cost.
• Credit signals. Several bears are now watching hyperscaler and Oracle CDS spreads and the wave of AI-related investment-grade debt and SPVs (one estimate: ~$320 billion in 2026). Widening spreads would be the first crack the skeptics are looking for.
• Semicap equipment orders. If the memory shortage is real and lasts to 2028, the fab buildouts in South Korea, Idaho and China should show up as WFE orders for AMAT, LRCX, KLA and ASML, a lagging but high-conviction tell.
• China and export-control headlines. Any actual resumption of Nvidia shipments to China is upside to a consensus that currently assumes zero, and the CXMT blacklist litigation is worth tracking as a barometer of US-China chip tension.