Newsletter · 2026-08-23 · Ashutosh Agarwal

Memory Steals the Show as Nvidia Earnings Loom - Weekly Semis & AI Infrastructure Podcast Recap - Week of August 23, 2026

How the semiconductor and AI infrastructure podcasts covered the memory boom, Nvidia's looming earnings, and the AI bubble debate for the week of August 17 to 23, 2026, with Micron's CEO calling memory strategic infrastructure and analysts split on whether the capex cycle can pay for itself.

Weekly Semis & AI Infrastructure Podcast Recap

Week of August 23, 2026: Memory Steals the Show as Nvidia Earnings Loom

This was a loud week for chips. Memory was the star of almost every show, Nvidia's earnings (due next week) hung over everything, and the "is this an AI bubble?" argument reached a fever pitch across finance podcasts. Below is what the week's episodes actually said, in plain language, with the numbers and the people behind each view.

Top of mind this week

1. The memory chip boom went mainstream. More podcasts talked about memory (the chips that store data, as opposed to the chips that do the math) than any other topic. A few concrete markers of how hot it got:

• JPMorgan raised its price target on Micron from $500 to $1,550, more than triple. Real Vision's Macro Mondays called the memory trade "probably the best performing trade over the past week or so," and argued the assumptions baked into memory stocks are "incredibly conservative still, which is in sharp contrast to the run-up to the year 2000" (Real Vision: Finance & Investing, Aug 17).
• SK Hynix (a Korean memory maker) reported revenue up ~250% and profit up ~500%, and Samsung announced a big shareholder return plan (dividend plus a roughly $29 billion buyback) (Bloomberg Tech, Aug 19; Behind the Ticker, Aug 23).
• CNBC's Jim Cramer got an exclusive with Micron CEO Sanjay Mehrotra at the company's Boise headquarters. Mehrotra's core message: "Memory is no longer a component in a system. Memory is the strategic infrastructure for AI. And it's no longer a commodity." He's putting $10 billion into a new "Micron Research Labs" for long-horizon memory research (Squawk on the Street, Aug 20).

2. Nvidia earnings are the next big event. Several shows framed the week as pre-Nvidia positioning. Chip stocks had sold off after a huge multi-year run, and hosts kept asking whether the print (due the week of Aug 25) would reset the whole AI trade (CNBC's "Fast Money", Aug 21).

3. The "hidden" cost of the buildout got a lot of scrutiny. A widely-cited Wall Street Journal analysis showing ~$3 trillion of AI-related off-balance-sheet commitments across nine big tech companies (far more than the ~$600 billion they've reported as capital spending) was picked apart on multiple podcasts, feeding the bubble debate below.

Dominant themes

Theme 1: Memory is the new bottleneck ("the mines, not the picks and shovels")

The clearest framing came from Howard Chan of Kurv Investment Management, who runs a memory-focused fund. His pitch, in his words: everyone calls chipmakers the "picks and shovels" of AI, but memory makers are actually "the mines", because supply is physically limited. His supporting points (Behind the Ticker, Aug 23):

• Just three companies make ~90% of the world's memory: Micron, Samsung, and SK Hynix.
• Memory used to be a commodity where the cheapest producer won. AI changed that. A specialized product called HBM (high-bandwidth memory, chips stacked on top of each other to move data faster to the GPU) has become a moat. The big three have sold out most of their HBM inventory through 2028.
• Why can't they just make more? Building a new fab (chip factory) costs $30-50 billion and takes years, and the essential lithography machines come from a single supplier, ASML in the Netherlands, which already has years-long waitlists. Chan calls it a "two-to-four-year thesis" and views the recent pullback as an entry point.
• A knock-on effect: because the big three are shifting capacity to high-margin HBM, ordinary memory is getting squeezed too, which is why Apple, and makers of MacBooks, iPhones and Xboxes, are raising prices.

The technical picture matched the investment case. At the Future of Memory & Storage Conference, analyst Jim Handy said DRAM contract prices are up 80-95%, DRAM revenue is set to rise "hundreds of percentage points" in 2026, and even without the HBM crunch, hyperscaler buying alone would have caused a shortage. Constraints may not ease until ~2028 (Grey Beards on Systems, Aug 22).

Theme 2: Foundry and equipment: the supply lag that could cut both ways

Ben Thompson (of Stratechery) gave the week's most nuanced take on the manufacturing side (Invest Like the Best, Aug 18):

• He compared memory to the shipping industry: prices spike when there's a shortage, everyone rushes to add capacity, then a glut crashes prices and washes people out. The difference this time is that memory consolidated to an oligopoly of three disciplined players who "learned the mistakes of the past."
• His warning for memory makers, long-term: by using their leverage now, they paint a target on their backs. He compared them to Iran and the Strait of Hormuz, the threat is most powerful when you don't use it, because using it makes everyone build around you. "The memory makers probably screw themselves in the long run by creating such a massive target on their back." He noted Apple is lobbying to buy Chinese memory, and that the number-one goal of algorithm work is "how can we use less memory."
• On TSMC (the Taiwanese company that makes the most advanced chips): he argued TSMC has been deliberately conservative on expansion, because a fab is a 30-year bet and overbuilding is its biggest fear. The result is that TSMC "offloaded risk onto the big tech companies", the risk now shows up as foregone revenue and profit for hyperscalers who can't get enough compute. "Risk doesn't disappear. It just gets handed off."

That conservatism is exactly why the equipment makers are booming. On the chip-tool side, Applied Materials (ticker AMAT), one of the "Fab Five" equipment makers (with ASML, Lam Research, Tokyo Electron and KLA) that form a genuine choke point, just did over $9.1 billion in quarterly revenue (up 25% year-over-year), with 50% gross margins, and guided to its first-ever $10 billion+ quarter. Heading into next year it becomes a $40 billion+ annual revenue business, and China is back in growth mode as its largest geography (~26% of systems/services). The host's caution: after the run-up, the stock is pricing in roughly a 24% annual earnings growth rate, "a high bar to clear" (Chip Stock Investor Podcast, Aug 17).

The bear version of the same lag: on This Time Is Different (Aug 18), the host argued it takes "three, four, five years" for foundries like Samsung or TSMC to add capacity, and by the time supply arrives, "today's demand has gone somewhere else." If the AI narrative shifts in that window, "that is going to completely obliterate the orders that people are banking on today."

Theme 3: Do GPUs really depreciate as fast as the bears assume?

A genuinely fresh data point came from Silicon Data, a startup that publishes daily GPU-rental price indices (think of it as an S&P 500 for renting chips). Founder's key finding (Equity, Aug 19):

• The rental rate on the A100, an "ancient" chip released in May 2020, has been steadily rising since late last year, not falling. His read: "That tells me inference demand is really strong."
• His analogy: it's like everyone still walking around using an iPhone 12 even though the iPhone 17 is out, a sign there aren't enough phones. "These things may not depreciate as quickly as people thought." This directly undercuts the popular bear argument that data centers are taking out loans against fast-depreciating collateral.
• Real Vision echoed it with a real example: CoreWeave said it re-rented five-to-six-year-old A100 GPUs for another three years, longer than the depreciation period usually assumed (Real Vision, Aug 17).

Theme 4: Inference is the trillion-dollar prize, and Nvidia's challengers smell blood

The action is shifting from training (building the model) to inference (running it for users). Sid Sheth, CEO of chip startup d-Matrix, laid out the economics (Eye On A.I., Aug 17):

• Inference will be "the largest part of AI compute... over a trillion dollar market in the next five years."
• A "premium token economy" is emerging for low-latency (fast-response) inference: users will pay $20 per million tokens for high interactivity versus $2 for standard, a 10x premium, and they're happily paying it. Serving that needs specialized chips with far more memory bandwidth than a general-purpose GPU.
• The named challengers to Nvidia in this niche: d-Matrix, Cerebras, Groq, and SambaNova.

The competitive threat got sharper elsewhere. Saxo flagged two Nvidia challengers in one breath: Cerebras (whole-wafer chips, new inference benchmarks) and Etched, a brand-new startup of twenty-something Harvard dropouts that got TSMC to make its first batch and landed Jane Street as a customer. Saxo's takeaway: "It is really hard to maintain a monopoly on technology... if you have a series of technologies all on par with one another, you wonder if the margins can hold up" (Saxo Market Call, Aug 19).

Theme 5: Optical/photonics: the wiring is becoming a bottleneck too

As chips get faster, the connections between them have to keep up, and copper wiring is running out of room. This week the optical-networking names had a rough tape but strong fundamentals (The MoneyFlows Show, Aug 20):

• Data speeds through fiber are exploding: 400 → 800 gigabit → 1.6 terabit, with Coherent already pushing 3.2 terabit. Scientists pushed 450 terabits/second through existing London fiber in a record test.
• Coherent (COHR) was highlighted as a favorite: backlog stretching beyond 2028, ~31x forward earnings, and Nvidia made a ~$2 billion strategic investment in it. Corning (GLW) is developing multi-core fiber (up to 19 cores). But note: Saxo pointed out Coherent and Lumentum were among the worst performers on the S&P 500 mid-week (down ~10-13% on a bad day), so the sub-sector is volatile (Saxo Market Call, Aug 19).

Theme 6: Custom silicon (ASICs): the hyperscalers build their own

Momentum for chips designed in-house by cloud giants: Marvell's partnership with Google on custom silicon and TPU (Google's own AI chip) expansion was flagged as a driver, alongside Amazon's Trainium/Graviton chips (Bloomberg Daybreak, Aug 21; Bloomberg Tech, Aug 19). This is the long-term counterweight to Nvidia: the biggest customers becoming competitors.

Theme 7: Geopolitics of the supply chain

• Taiwan's dominance is staggering: TSMC has over 70% of global chip sales and makes over 90% of the most advanced chips, and is building or expanding 13 fabs this year. A guest called it "path dependency", 50 years of ecosystem you can't easily copy (Taiwanology, Aug 18).
• A US–South Korea rift is a new memory risk. Trump moved to limit joint military drills with South Korea, and the theory floated on Macro Mondays is that it's tied to Korean HBM leaking to China via Malaysia. Samsung and SK Hynix run Chinese fabs that need US license renewals, a live risk to the memory trade (Real Vision, Aug 17).
• China's mature-node flood remains a longer-term overhang on cheaper, older chips (Taiwanology, Aug 18).

Key debates

Debate 1: Is the AI capital-spending boom a bubble?

This dominated the week. The camps:

Bears / skeptics:

• Steve Eisman (of "The Big Short" fame) focused on cash flow: Meta's quarterly free cash flow was just $785 million, Microsoft's ~$19 billion but down ~25% year-over-year, and Amazon was negative over 12 months. Google raised $85 billion in equity, its first big raise since the early 2000s. His conclusion: the buildout can't be funded from cash; it needs the bond and equity markets. He wouldn't short it (demand exceeds supply), but doubts hyperscalers can earn a decent return on this capital intensity (The Real Eisman Playbook, Aug 17).
• Former NY Fed President Bill Dudley: absolute capex will still grow in 2027, but the rate of increase will slow, which hurts suppliers' earnings growth, and hyperscalers face the challenge of justifying ~$2 trillion of revenue to support current investment. He expects a telecom/railroad-style overcapacity bust (Bloomberg Talks, Aug 20).
• Arthur Hayes called it massive capital misallocation into depreciating assets and forecast the bubble collapses by mid-2028 (Unchained, Aug 21).
• Organized Money: $6 trillion of AI capex can't be justified when the industry would need $2.5-3 trillion of revenue just to break even, and everyone's building the same commoditized technology (Organized Money, Aug 18).
• RenMac noted bubble signals across Korean, Taiwanese and semiconductor-index stocks, warning that historically ~90% of such bubbles fully round-trip their gains (RenMac, Aug 21).

Bulls / "this is real":

• Dave McClure & Aman Verjee argued AI infrastructure is a real, financeable asset class like railroads or the internet boom, with Goldman, Apollo, Brookfield and KKR investing on commercial terms as a positive signal, and GPUs financeable over 8-10 years like cars, not obsolete in 2-3 (Keen On America, Aug 17).
• Jack Kuykowski (Full Signal) is "broadly bullish" over the next 18-24 months on the semiconductor buildout, citing oligopoly pricing power (ASML's EUV monopoly, memory controlled by "two Korean families"). His candid nuance: he thinks the bears may well be right in three-to-five years on return-on-invested-capital: "$5 trillion of capex over five years... it's very hard to imagine a group of companies bringing in $5 trillion of revenue", just not yet. He noted off-balance-sheet commitments went from ~$1 trillion in Q1 to ~$1.5 trillion in Q2 (Google alone ~$800 billion), which is "bubblicious" but "extremely profitable to the semiconductor companies" (Full Signal, Aug 19).
• The Analytics Engineering Podcast ran five quantitative bubble checks and concluded it is not currently a bubble: economic strain under 1% of GDP, revenue doubling roughly every seven months, with only "funding quality" flashing yellow as capex now exceeds hyperscaler cash flow (The Analytics Engineering Podcast, Aug 21).

The middle ground:

• Mike Pyle argued the memory/chip windfall is real but not permanent, bottlenecks relax over "quarters to years", yet demand still substantially outstrips supply, implying a "pretty considerable" sustainable earnings path (Top Traders Unplugged, Aug 19).
• Liz Ann Sonders' panel (Excess Returns) put next year's hyperscaler capex near ~$900 billion, argued the build is currently undershooting real demand (not speculative), but flagged a coming "depreciation cliff" that will eventually dent earnings growth (Excess Returns, Aug 22).
• Ben Thompson summed up the honest position: "I believe in AI... the economic impact is going to be astronomical. You can believe all that and still be worried about... the bridge to this actually generating the level of returns necessary" (Invest Like the Best, Aug 18).

Debate 2: Can Nvidia keep its near-monopoly?

The bear structural point, sharply put on This Time Is Different (Aug 18): if you truly believe Cerebras can disrupt the GPU model, "you also shouldn't be owning Micron and Broadcom and SK Hynix... because that would send the share price of those companies down." In other words, the popular trade of owning both Nvidia's challengers and the memory/networking ecosystem that depends on the GPU status quo is internally contradictory.

Debate 3: Is the financing "circular"?

Multiple hosts zeroed in on Nvidia extending financing to customers (variously described as a ~$500-600 billion effort) so they can afford its chips, and on the idea of "collateralized chip obligations." This Time Is Different (Aug 18) drew an explicit line to the 2006-08 mortgage machine: "Now we're talking about collateralized chip obligations. This is getting absolutely freaking nuts." Bulls counter that vendor financing is normal (like a car maker's financing arm) and that external demand, not circular deals, is driving it (Buy Hold Rant, Aug 19).

Stocks discussed: bull vs. bear