Market Overview: June 30, 2026

Summary

• Core call on the regime shift: stocks are driving the macro, not the macro driving stocks. In the past, indices followed macro conditions, while a single company influencing the macro in reverse was “very, very difficult.” Now the relationship has flipped: “the equity market is the macro, and models are the equity market.” Anthropic’s $60B-plus figure is ARR, not revenue, but it has “essentially driven the entire market”; many traditional macro indicators have lost importance, while liquidity remains critical.
• The quantitative warning is flashing red: CTA faces asymmetric risk over the next month. A 2-standard-deviation selloff would force $100B of selling, while an upside move would trigger only $30B-plus of buying; whole-market negative gamma is above 74M. The index has been hugging CTA’s short-term flip-to-short level for 2-3 weeks. VIX is not elevated, but single-stock vol in the S&P and Nasdaq is “very, very high”—the quant data alone makes it feel like “sitting on dynamite, with plenty of kindling underneath ready to burn you.”
• His offsetting argument is that selling pressure is immaterial relative to the model-side numbers. The 4 major cloud providers plus Oracle have enormous conviction: they are taking the money saved in the past and going all in today, playing a high-stakes poker game where exiting midway means losing everything; the total scale is above 1T. Storage will take 400B-500B from CSPs next year, while the 3 storage companies will generate roughly 1T in combined revenue at gross margins above 85%—even a 100B liquidation would be bought back.
• Liquidity verdict: not tight, but no longer a flood. eSLR relief would put 400B-500B of capital into banks, theoretically supporting 4T-5T of leverage, but the people willing to take risk left banking long ago after 2008, and the money will be paid out as dividends. With no QE, further leverage-ratio cuts are also unlikely; Kevin Walsh (likely Warsh) wants to “return interest rates to the market.” April’s “liquidity flood, with semiconductors all going up 2x” will be difficult to repeat. The market will keep asking, wave by wave, “where does the money come from?”
• The playbook as the field narrows is to buy “fuzzy beauty”—names in shortage but not fully priced in. After 800V, domestic power semiconductors are in shortage across the board. In SiC demand, what sounds like Vera Rubin in the audio (“vanguard rubin”) is 20x Blackwell, while Blackwell is roughly 10x H100. SiC is a “very boring, very old business,” but its current setup really resembles how AI looked then. Storage and optics, where expectations are already full, warrant more caution. Hynix at 5x PE captures the valuation dynamic: “at 7x you need faith; at 3x you only need not to have faith—the earnings don’t change.”
• Price increases are shifting from stimulative to destructive, with China the biggest casualty. Apple’s across-the-board price hikes are a lagging signal but a warning shot: “when memory prices rise, you can’t even sell refrigerators.” Consumer electronics account for roughly half of Guangdong’s GDP; next year’s 1T in storage revenue could bring 3T in knock-on damage, like rubber in the auto era taxing the entire value chain. Consumer electronics will be hit first; “I think EVs will be next.” The market has not fully priced this in.
• The core model-side call: extending Anthropic/OpenAI’s market-expected revenue growth through the end of next year is impossible, no matter how good they are. The chance that the company rendered in the audio as “Adobe Open AI” (likely Anthropic/OpenAI) reaches 400B in 2 years “may be zero—absolutely impossible.” Models face social constraints as they move through human society. At the same time, the market may be underestimating the challenge that open-source models such as DeepSeek and Zhipu pose to closed models; anti-distillation defenses “may not hold.” The endpoint will arrive, but whether capex can be justified by model revenue is a question the boomerang will keep coming back to.

Deep dive

1. Stocks Driving the Macro: Models Are the Equity Market

• The opening call set the tone for this “regime shift.” In the past, US equities rose on liquidity plus corporate earnings, dividends and buybacks; equity, particularly the indices, followed the macro, while a single company driving the macro in reverse was “very, very difficult.” Now the causal chain has flipped: these companies need to raise capital at scale and convert that financing into semiconductor earnings—“the equity market is the macro, and models are the equity market.” Anthropic’s $60B-plus figure is ARR rather than revenue, but it has “essentially driven the entire market.”
• He acknowledged that the setup is distorted, but said it is hard to know when it will stop and that it “may still need quite a long time.” Once the narrative reaches a certain scale, “the importance of these narratives has already started to exceed that of the macro.” Liquidity still matters; most other macro indicators have moved to the second tier.

2. Quant Data Says We Are Sitting on Dynamite, but Model-Side Scale Overwhelms Selling Pressure

• CTA faces asymmetric risk over the next month: a 2-standard-deviation move lower would force $100B of selling, while an upside move would bring only $30B-plus of buying. Negative gamma is above 74M, and the market has been trading right on top of CTA’s short-term flip-to-short level for 2-3 weeks. VIX is not high, but single-stock vol across the S&P and Nasdaq is “very, very high”—the quant indicators make it feel like “sitting on dynamite, with plenty of kindling underneath ready to burn you.”
• The offset comes from the scale of the earnings opportunity. The 4 major cloud providers plus Oracle have “very, very high conviction—they are taking the money they saved in the past and going all in today. All 4 are betting, and they are playing a high-stakes poker game; if you exit midway, it’s over”—with total spending above 1T.
• The storage arithmetic is his evidence. Storage will take 400B-500B from CSPs next year; the 3 storage companies will generate about 1T in combined revenue, with all 3 above 85% gross margin. “Just think about how much money they are making.” Selling pressure such as a Pension Fund Rebalance “is not particularly important in front of the model-side numbers”; in an extreme case, even a 100B monthly drawdown would attract enough buyers to take it back up.

3. eSLR Releases Capital, but Risk-Taking Capacity Is Missing

• In response to listeners’ questions last week about bank eSLR and this year’s liquidity, he said relief could add roughly 400B-500B to banks’ capital base and theoretically unlock 4T-5T in purchasing power. April was “definitely a liquidity flood”; with semiconductor valuations having been compressed for too long, basically the entire semiconductor complex went up 2x.
• But the mechanism is not sustainable because the people willing to take risk have left banking since 2008. “The culture of being willing to throw a hamburger in a colleague’s face and tell them they made a bad trade is completely different now.” Bank-to-bank liquidity could indeed provide ETF leverage, and 3x ETFs could push the market higher, but the banks’ own risk-bearing capacity has not recovered. Banks would pay out the capital as dividends, returning liquidity to equilibrium. The US is deregulating while Europe is still adding regulation; China’s banking system around 2015 looked completely different from today. “Those people were arrested or left. Who would do this job now?”
• There is no incremental liquidity from the Fed either: no QE, no Treasury-discounting arrangement like during the SVB period, and further cuts to leverage ratios “should also be unlikely.” Kevin Walsh (likely Warsh) wants the Fed to intervene less in markets and “return interest rates to the market.” The conclusion is that liquidity is not tight but not loose, while the market still needs substantial financing. “Where does the money come from?” will be asked in waves, answered, and followed by a rebound.

4. The Field Narrows; Valuation Becomes a Question of Faith

• When liquidity is abundant, all stocks rise together, as in April. When it is less plentiful, “the road narrows, and stocks take turns rallying within specific sectors.” Hynix is the obvious example: at 5x PE, “from 5x to 7x you need faith; from 5x to 3x you only need not to have faith, but earnings are unchanged.” Whether investors believe comes down to 2028-29, which in turn depends on whether the model side can capture revenue.
• Hyperscaler valuation works the same way. The market values them on next year’s earnings, but next year’s earnings depend on the capex planned for the year after. The current expectation is 1T; next year the market will be looking at 1.5T for the year after that, followed by 2T the year after. The financing question will keep cycling.

5. The “Fuzzy Beauty”: Power Semis in Shortage but Not Priced In, Like Optics 2 Years Ago

• He divides semiconductor names into 2 groups. “Fuzzy beauty” means names that are in shortage but have not been fully priced by the market. After 800V, domestic power semiconductors are in shortage across the board. In SiC demand, what sounds like Vera Rubin in the audio (“vanguard rubin”) is 20x Blackwell; what sounds like Feynman (“Framen”) is 5-7x Vera Rubin (“Varuban” in the audio); and Blackwell is roughly 10x H100.
• The team’s analogy is that optics also looked “very boring” 2 years ago, and AOI was “a very boring industry” when they first looked at it. SiC today “really does look like AI looked then.” MLCCs are also rising, while the market remains “quite naive,” finding the price increases in power semiconductors almost laughable. By contrast, expectations in storage and optics are already full and deserve some caution; the market is not awash in money lifting every stock.

6. The Destructive Side of Price Increases: 3x the Knock-On Damage, With China Hurt Most

• Apple raised prices across its product lineup and hit memory twice, but he sees the news as lagging information. Price increases themselves do not affect semiconductor demand, and consumer demand will not fall; the FT report that Apple went to ChangXin for memory chips “definitely will not affect” the revenue expectations of the 3 memory makers. The reason for the order transfers is precisely that CSPs are paying too much.
• The real warning is that AI’s impact is changing from stimulative to destructive. In the Nvidia era, AI absorbed N2/N3/N5 capacity and built up the entire ecosystem. Once it reaches a certain scale, the impact turns destructive. The pain can be quantified: “when your memory sticks are getting more expensive, you can’t even sell refrigerators.” Roughly half of Guangdong’s GDP comes from consumer electronics; next year’s 1T in storage revenue could create “3T in knock-on damage,” like rubber during the auto era—outside the industry’s core chain, but effectively taxing the entire value chain.
• “Who is actually hurt the most? China.” Open-source models are catching up, but China’s consumer electronics sector is taking a major hit as it competes with AI for resources. “I think EVs will be next.” The market has not fully priced this part of the damage.

7. The Model Side: The Open-Source Challenge Is Underestimated, While Closed-Source Revenue Expectations Do Not Work

• He pushed back on the prevailing narrative that smarter models simply serve users who can afford them. Model performance is a combination of compute, the model, Hynix and engineering—it is “not just the model itself.” GPT-5.6 is “clearly the best” today and has been made available to only a few dozen software companies, but the idea that keeping China from distilling it will preserve the lead indefinitely “may not be that easy.” “You may not be able to stop it; they can still distill you. Once the new thing has been distilled, a lot of your faith will be affected.”
• His personal view, with the caveat that he is not a model-industry practitioner, is that DeepSeek and Zhipu represent “open source challenging closed models.” The MoE debate and engineering advances are helping the entire open-source community. In previous software cycles, “open-source models ultimately do not lag too far behind; generally speaking, they lead.” He personally spends $1,000-plus a day using “cloud” (likely Claude), but “saying that deploying DeepSeek is far behind Claude is hard to sustain as a statement over the long term.”
• His strongest quantitative rejection is aimed at the revenue curve. If Anthropic/OpenAI sustain the growth rates currently expected by the market through the end of next year, “even if you are very good, it is impossible.” If one extrapolates the technology in a straight line and assumes the company rendered in the audio as “Adobe Open AI” (likely Anthropic/OpenAI) reaches 400B in 2 years, “I think the probability may be zero—absolutely impossible.” Models have to work through human society: “you fire this person, and that person may still be the leader’s relative”; not firing them creates an economic cost. He also joked that OpenAI’s new model names sound like crypto tokens—“SOL, Terra, Luna.” He does not question the endpoint: “the final result will definitely be reached, but the process in between will not be this smooth.” Whether capex can be justified by model revenue is the open question; “the boomerang will keep coming back and forth.”