Summary

This episode’s core call: U.S. equities have entered a new cycle. Hyperscalers are being forced out of their asset-light model of “collecting global rents, then returning capital through buybacks and dividends” and into factory mode—monetizing wealth at the top on a massive scale and converting it into cash to feed semiconductors. CapEx will total $750B this year and $1T next year; “these companies actually have no cash left.” Wall Street has the cash math right, but it has missed that the 5 players in the Texas wager have already bet too much to fold.
The low-P/E bubble’s “protective shield” is still intact. Visible semiconductor names fell about 10% Friday and the Nasdaq fell 4.x%, yet the VIX only reached the 20s versus 40-60 last year, while memory stocks recovered quickly Monday. A 30% pullback can actually make the stocks look cheap: one investment-bank report says HBM still has to reach RMB53/GB and keep rising, while Huaqiangbei spot memory saw 128GB products effectively sell out and prices for 128GB and below rise more than 30%. The shield breaks only when both incremental gains and quality in Anthropic/OpenAI models come under doubt: “if the models fail, the whole thing is a mirage.”
Trade 1: Equity exposure belongs in semiconductors. Of the $1T in spending, “even Huang can take less than half,” with the rest cascading down the supplier chain; outside semiconductors, almost everything depends on financing and is therefore less safe. Within semis, favor mature-node names with capacity already in place and utilization that can be monetized directly, such as 800V power-control CMOS chips, rather than companies that still need to finance capacity expansion.
Trade 2: Buy Amazon CDS as a hedge. Amazon has the weakest cash flow, while Oracle’s CDS is already at 500 points and Amazon’s is only 30-40 points; he entered at 30-something points. Pairing $1M of stock with $10M of CDS carries roughly -3 points a year, while in a panic the CDS can rise by “dozens upon dozens of times.” This is not a bet on bankruptcy, but on spreads widening as the company continues to finance growth, giving investors the confidence to hold semiconductor exposure.
He has shifted from being very bullish on Oracle to “a little hesitant.” CapEx execution will make the earnings data look strong, but the cloud backlog is coming from model companies with combined revenue of roughly $100B this year and no profits yet; the gap between model revenue and CapEx “keeps getting wider,” while the market is pricing it as though it is narrowing. That is the market’s central artery; it could be realized in 3 years, but “it absolutely will not be smooth.”
Model quality is visibly deteriorating. The same pipeline, run through the 4.8 model he describes as similar to Grok, produced materially worse PPTs within 1.5 months, but June results may not be weak because “when it gets dumber, it uses many times more tokens.” Meta and others are still treating token consumption as a KPI—“this is metaphysics”—replaying the cloud-spending script in which companies use everything until costs explode and then cut back. That is the model companies’ Achilles’ heel.
Friday’s selloff had 3 causes, none of them fundamental. Bokang beat sell-side expectations without beating buy-side expectations; Nvidia’s Val Rubin “uses less HBM” story was really an optimization response to HBM shortages, not weaker demand; and macro still matters, with an overly strong payrolls report making a Kevin Warsh rate cut harder. Leveraged ETFs amplified the move: Korea-linked products grew 4.9x to $65B and U.S. products exceed $180B; another 1-1.5% decline could flip short-term CTAs into sellers of $5B-$10B.

Deep dive

1. Models Get Dumber, but June Results May Hold—the Real Vulnerability Comes Later

The opening visual demo ran the same pipeline he had modified himself through the 4.8 model he describes as similar to Grok, generating a PPT and comparing it with the version from 1.5 months earlier. “The model’s quality has visibly deteriorated.” As he put it: “This is just writing a PPT; when you’re writing code, the gap in bugs could be much larger.”
Would Adobe’s results decline this month? His answer was probably not, but only because it is a matter of timing: “Because it got dumber, it used many times more tokens”—usage has instead exploded. He expects this month’s results may not be especially weak.
The real risk is that enterprises have not yet reviewed the return on their spending. Meta and others still require employees to hit token-use targets “as if it were a KPI”—“this is really a metaphysical thing.” His analogy is the cloud cycle: “When everyone was building cloud, everyone had to use it. Eventually they realized something was wrong, costs had exploded, and they needed to impose controls.” The protective shield breaks only when both incremental model gains and quality are questioned; that morning he had seen a report that a company had reduced API usage because Anthropic had “dumbed down quite a bit,” though he could not remember which company.

2. Wealth Is Not Cash: Hyperscalers’ Texas Wager

Borrowing from 达里欧’s framework: wealth is not cash. Hyperscalers with $4T-$5T in market value have plenty of wealth but no cash now—their combined operating cash flow is only roughly $300B-$400B, against $750B of investment this year and $1T next year. Cloud revenue also needs to be discounted: invest $100B in model companies and receive $10B a year in cloud purchases in return—“that money doesn’t count.”
The Texas poker game has evolved. 2-3 years ago, “if you didn’t participate in AI, the whole company might be eliminated”; a few tens of billions was just a small opening bet, and the bubble started in a healthy way. Now the bets are too large to fold. Wall Street is not stupid to be bearish on AI—it has done the cash math correctly, but it has not understood how large the bets are. These 5 companies are at the technological frontier, believe in their cards and are being forced to call one another.
The immediate consequence is a new valuation regime. As long as the companies keep matching CapEx commitments, turning market value from wealth into cash requires issuing stock and debt, which weighs on share prices. That explains why the Hyperscalers have underperformed both the indexes and semiconductors for more than half a year.

3. U.S. Equities Enter a New Cycle: Monetizing Wealth at the Top to Feed Semiconductors

His A-share comparison is that the A-share market is a financing platform that “is constantly being bled”; over the past 10 years, only Moutai rose because it is a company that earns cash rather than wealth. Xiaomi’s low-margin, scale-first strategy is structurally bad for the stock market: it shipped roughly 180M phones last year but is expected to ship only 60M this year—“a complete mess.” The U.S. bull market was built on high margins, an asset-light model and buybacks and dividends, but AI has changed that model; large-scale capital flows now look “just like China before.”
The financing queue is already forming. Google has issued roughly $80B of equity, Meta is reportedly looking to raise $25B, and debt and equity are running in parallel; the funding gap is $300B this year and $500B next year. This week, SpaceX is set to pull liquidity through a $75B-scale IPO; Anthropic has reportedly filed its materials, though he said he had only heard that, and OpenAI is also looking to raise. Microsoft’s CDS used to trade nearly at zero, “like a bank”—that was their old state.
The more than $1T of monetized wealth ultimately becomes semiconductor earnings. “Semiconductors have earnings, but not necessarily valuation.” Some investors want to put Micron, Samsung and SK hynix on 15x P/E; “I think that’s insane.” Commodity companies should normally trade at compressed multiples and receive growth valuations only temporarily, when the model narrative convinces the market that revenue and profits can catch up with CapEx. Inside the industry, optimism has reached manic levels: orders are fully booked, lithography costs have doubled, and “1 year of revenue can build an entire fab.”

4. The Low-P/E Bubble’s Shield: Every Dip Finds a Bid

Last Friday, every semiconductor name in plain sight fell about 10% and the Nasdaq fell 4.x%, but the VIX only reached the 20s, versus a peak of roughly 60 last year and 40-something early this year. The lack of panic was also reflected in light gamma positioning among put sellers; memory led a rapid recovery Monday.
That is the defining feature of the first low-P/E bubble: every pullback makes the stocks look cheaper. One investment-bank report says HBM still needs to reach RMB53/GB and rise further. At Huaqiangbei, spot memory in the 35G category rose 75% in a week, 128G products were almost entirely sold out, and prices for 128G and below rose more than 30%. “The more you understand the industry and the more you know the companies’ details, the more willing you are to rush in,” because after factoring in next year’s capacity additions and shortage-driven price increases, 2027-2028 forward P/E does not look high, nor does the implied valuation. “Don’t assume that once this kind of pullback happens, you’ll get a 2000-style bubble collapse. That probably won’t happen.”
The key flow valve is roughly 1-1.5% lower from here: short-term CTAs would flip and sell about $5B-$10B. Monday’s rebound was large but “actually fairly weak.” With the SpaceX IPO draining liquidity, if the flows do not appear this week, CTAs may emerge next week as well.

5. Friday’s Drop Was Deleveraging

First, Bokang’s results beat sell-side expectations but not the buy side’s. The market quickly translated that into a different narrative: once Google and its peers enter factory mode, they start counting costs—“You make me such high gross margins every year—hey, bring them down a notch.” Buyers have no bargaining power at the choke points—optics, memory and MLCCs—but they retain some leverage over the turnkey contractors that design systems for them. That is the problem with Bokang’s guidance.
Second, the news that Nvidia’s Val Rubin uses less HBM frightened the market, but he sees “nothing particularly major” in it. HBM shortages are limiting NVL72 shipments; inference needs less HBM than training, so optimized allocation can support more shipments. Or the company simply cannot buy enough and has to ship a lower configuration—just as the 512GB Mac Studio is now unavailable. “It doesn’t indicate that demand has fallen.”
Third, macro still matters, and it collided with leverage. Payrolls were too strong; the WSJ headline he quoted read, “Kevin Warsh’s job just got a lot more complicated.” Korea-linked leveraged ETFs grew 4.9x to $65B, while U.S. products exceed $180B. They roll in as prices rise and rebalance by selling as prices fall; combined with substantial leverage in Korean memory positions, Friday became a full-scale deleveraging event.

6. Trade Implementation: More Semis, Buy Amazon CDS, and a Hesitation on Oracle

The core allocation logic is that Equity exposure should sit in semiconductors. Of the $1T in spending, “even Huang can take less than half,” with the rest distributed down through every supplier; memory is a major link in the chain. “Outside semiconductors, everything is financing-dependent—especially the large companies—so none of it is particularly safe.” Within semis, favor companies with capacity already in place and only utilization to monetize, such as mature-node CMOS power-control chips, where demand below 800V is “very, very large,” rather than companies still facing long-cycle investment in wafer fabs.
His own hedge is “a very important trade: buy CDS.” The thesis is not bankruptcy; with financing continuing and supply expanding, spreads should widen. In a real crisis, “it would be perfectly normal to price Amazon like Oracle.” Amazon has the weakest cash flow, while Oracle’s CDS has blown out to 500 points and Amazon’s is only 30-40 points; he entered at 30-something points. A $1M stock position paired with $10M of CDS carries roughly -3 points a year, while in an extreme market the CDS “can rise by dozens upon dozens of times.” He cited how cheap Lehman’s CDS was in 08, but stressed, “I don’t think a problem that severe will happen.” The insurance is meant to “make you more willing to hold semiconductor exposure.”
He also made a public self-correction: he had previously been very bullish on Oracle, but “now I’m a little hesitant.” CapEx execution will make earnings data look strong over the next few quarters, but the orders come from 2 model companies with combined revenue of roughly $100B this year and no profits of their own; after the good results, they still have to raise financing, and “financing always pushes prices down.” The gap between model-layer revenue and CapEx—$1T of CapEx requires at least $200B-$300B of earnings to be sustainable—“keeps getting wider,” while the market is pricing it as though it is narrowing. “That is the lifeline and the main line of the market.” It may be realized in 3 years, but it will absolutely not be a smooth ride.
This week’s macro focus is CPI. Core CPI will probably not be too bad, with the last report distorted by delayed rent data; headline CPI will be hurt by oil prices, but the pass-through to core should be slow. A Kevin Warsh rate cut will be difficult, but he expects Warsh “not to be as hawkish as the market expects”; the market has priced no cut this year and possibly a hike, and he does not agree with the hike scenario. The longer-term point is that AI has tied together U.S. strategy, Trump, the GOP and the major cloud companies. In a U.S.-China AI cold war, “the chief paymaster has to keep funding it.”

Why the AI Boom Is Just Getting Started Gavin Baker interviews SpaceX CFO Bret Johnsen at Mission Control