Market Overview — August 4, 2026

Summary

• Momentum selling is nearing a near-term floor. That is the core call this week: Goldman, the most bullish of the sell-side desks, says the market has “basically sold through,” and the speaker’s view is similar. JPMorgan and Bank of America, however, flag prime-book positioning still at 95%. Leveraged ETF exposure has fallen from $470B to $300B, but “there’s still $300B left; if it comes out, it will be scary.” Bottom line: look for a near-term floor first and revisit it in another week or two; the short-term setup remains manageable.
• Situational Awareness’s forced liquidation is a major backdrop to this selloff. It had $45B of AUM in early July, including roughly $5-10B of Anthropic private-equity shares, and used 4x leverage to build approximately $120B of positions. The book was concentrated in longs across semiconductors and new cloud names—BE, Intel, SanDisk, “Nebulance” [likely Nebius], and “Irene” [likely IREN]—against software shorts; 85%-90% of the equity book has been wiped out. One person close to the trade had previously worked at OpenAI, while a spouse reported directly to Dario at Anthropic: “They were playing with the map open.” But the position was too large. Once enthusiasm faded, it became a large piece of meat, and everyone came to the market to smash it and kill it.
• Thursday’s rebound was driven by short covering, not Microsoft’s earnings. Citadel took the final $16B of the position, with Goldman providing the financing and intermediating the transfer. Shorts kept hammering the market, expecting to break the buyer, only to discover that “the monster was at full health—and on their own team.” Microsoft’s positive cash flow and improvement in Amazon’s negative cash flow were only marginal factors. Shorts have not fully covered: more than $2B in short positions remain on CME, making shorting during earnings season particularly risky.
• CTA selling pressure cannot break semiconductors. CTAs could sell as much as $150-180B at the global-equity level, but half of that is outside the US and less than one-third of US exposure is in semiconductors. The AI basket’s one-week volatility is 121 versus just 66-67 for the cap-weighted Nasdaq; the inverse seesaw means VIX would need to rise to 45-50 or higher to take semiconductors and the broad market down together. Neither route looks viable, except perhaps in an extreme negative-gamma scenario.
• Rate-hike pricing has gone too far, and the curve should steepen. Markets price a 66% chance of a September hike and 2 hikes by next March, but core inflation breaks down into 70-80bp from tariffs, 20-50bp from AI-driven price increases, and roughly 10bp from a war with Iran—all supply shocks for which rate hikes are ineffective. Strip out AI, and “the US is in a severe recession”; AI is like a cancer draining the economy’s nourishment. The speaker expects the front end to reprice lower with economic data, while Treasury supply puts a floor under the long end. He somehow expects Friday’s payrolls to undershoot, making “bad news is good news” a meaningful possibility.
• The yen intervention is a coordinated US-Japan operation. Japan must sell dollars and Treasuries to intervene, pushing up long-end Treasury yields until the other side starts to feel the pain. Japan cannot directly defend 160 and has to fight a guerrilla war—buying time for room and making shorts afraid to press the trade—with another rate hike possible in September. Korea faces a similar dynamic.
• AI’s ultimate funding test runs through 3 lines. The first is equity—Anthropic still needs to raise capital and eventually list to fund capex; the second is debt, including CDS; the third is how many people ultimately pay for AI. Capex is approaching $1T against combined Anthropic and OpenAI AR of more than $10B, perhaps $20B by year-end, versus roughly $40B needed. Unlike the 2000 internet bubble, end users are paying and cash returns quickly. The real bubble problem may only arrive with a genuine recession: “I’m happy to use AI for customer service, but the people buying things have been put out of work by your AI.”

Deep dive

1. Situational Awareness Blow-Up: Playing with the Map Open Still Didn’t Avoid the Hunt

• In early July, the fund had $45B of AUM, including roughly $5-10B of Anthropic private-equity shares. Its equity book was approximately $35B and is “basically all gone,” with 85%-90% wiped out. The entire position was levered 4x—“How large do you think that makes the position? $120B”—and was concentrated almost entirely in longs across semiconductors and new cloud names: BE, Intel, SanDisk, “Nebulance” [likely Nebius], and “Irene” [likely IREN], against software shorts.
• The speaker stressed that this was not a quantitative momentum strategy but a directional bet on AI. People close to the trade had previously worked at OpenAI, while a spouse reported directly to Dario at Anthropic. “They had a very clear read on what AI’s next phase would require and where the maximum profits would be extracted. It was like playing with the map open.”
• The liquidation mechanism was straightforward: the market knew the position was too large, so “when they come for your book, nobody takes the other side.” Even an Intel stock reporting strong earnings could be flushed to $60 with ease. “People don’t look at the fundamentals; they wait for you to die and then buy.” The speaker’s Blossoms analogy had 阿宝 long and 强哥 short: “The market is a game of competing positions, not an emotion.”

2. The Truth Behind Thursday’s Rebound: Short Covering, With Microsoft Earnings Only Marginal

• Citadel took the final $16B of Situational Awareness’s position on Thursday. Goldman lent the money and then intermediated the transfer to Citadel, “a bit like the Qilin Club taking over 阿宝’s position.” Shorts kept hammering the market, expecting to break the buyer, only to find that “the monster was at full health—and on their own team”; they could no longer unwind their own positions. Most of Thursday was short covering.
• The speaker downgraded Microsoft and Amazon’s earnings impact: positive cash flow at Microsoft and an improvement in Amazon’s negative cash flow merely improved expectations. “The impact is marginal and second-order—not that large.”
• Shorts have not finished covering. CME showed approximately $6.4B sold against 4-point-something billion bought, leaving more than $2B in shorts. “In the current earnings season, the risk to shorts is pretty high.” The quantitative read is that if net selling remains unchanged for 2 consecutive days while market volatility falls, selling pressure is weakening and support is forming. “If that selling force disappears, that is definitely a long signal.”

3. Is Momentum Finished Selling? A Near-Term Floor, With 95% Positioning Still Hanging Overhead

• Sell-side strategists disagree: Goldman is the most bullish and believes the market has basically sold through, a view close to the speaker’s. JPMorgan and Bank of America, by contrast, say prime-book positioning is still at 95%. Leveraged ETF exposure has fallen from $470B to $300B, but “there’s still $300B left; if it comes out, it will be scary.” Gross leverage remains in the 3-point-something range and is still high by standards going back to 2016.
• Deleveraging by systematic funds is lagged. The Gorman [likely Goldman] momentum index, viewed on a one-month-forward basis, still points to gradual selling, though the volume should be smaller than before. The speaker’s conclusion is that this is likely a short-term endpoint; cloud names could benefit as funding needs decline, and semiconductors would benefit if clouds do. Whether the deeper deleveraging Bank of America expects is still ahead will be clearer in another week or two.

4. CTA Selling Hits the Broad Market, Not the Other Side of the Seesaw—Semiconductors

• CTA exposure could generate $150-170B of selling when the index falls 2 sigma; Goldman’s estimate is $180B. But that is global equity exposure: only half is in the US, and less than one-third of US exposure is in semiconductors. The impact on individual stocks is therefore far smaller than the liquidation of momentum positions.
• The seesaw logic is the strongest part of the argument. One-week volatility for the AI momentum basket is 121, versus 66-67 for the Nasdaq; index volatility has recently been extremely low. Selling the broad market means selling megacaps such as Microsoft. To take both ends of the seesaw down, the negative correlation must turn positive—“you would need VIX to jump to 45 or 50-plus”—or AI-basket volatility would need to fall from 120, which itself would signal stocks rising. Both routes are unrealistic; negative gamma could be the exception in an extreme scenario.

5. Rate-Hike Pricing Has Gone Too Far: Supply Shocks Make Hikes Ineffective; Curve Set to Steepen

• Warsh’s silence was read as excessive dovishness on inflation, pushing the long end higher. Markets price a 66% probability of a September hike and 2 hikes by next March. The speaker said, “I think this pricing has gone a bit too far”—not that the 10-year at 4.75 should move lower, but that the market is overestimating the likelihood that Warsh will hike.
• The core CPI argument has 3 parts: 70-80bp from tariffs, a one-time effect that should fade; 20bp from AI-driven price increases, through memory and software such as Copilot bundled into Microsoft 365, with a worst case of 50bp; and roughly 10bp if Iran is actually attacked. All are supply shocks. “Can a rate hike make Samsung, SK hynix, SanDisk, and Kioxia produce more memory? It can’t.” Higher rates work on services and rents, not on this supply constraint.
• The more aggressive view is that, excluding AI-driven price increases, “the US is in a severe recession, an industry-wide recession.” AI is “like a cancer,” draining the liquidity and demand the system can generate while the underlying economy fails to grow. That is also why the speaker believes memory cannot be repriced higher again: higher prices directly hit core CPI and push up token prices. “It should be token deflation, not token inflation. The whole world would turn against you.”
• The rate conclusion is simple: “Steepen—the curve steepens. I think that is very clear.” The front end could reprice lower as economic data soften; the long end has a floor because the US must issue a great deal of debt. “Every time yields are pushed down, they slowly move back up.” The key markers are Jackson Hole and Friday’s payrolls. The speaker somehow expects a payrolls miss, making “bad news is good news” a meaningful possibility.

6. Yen Guerrilla Warfare and AI’s 3 Funding Lines

• The yen intervention is a coordinated US-Japan operation. Japan has to sell dollars and US Treasuries, pushing up Treasury yields until the other side starts to feel uncomfortable. Japan “doesn’t have the ability to defend 160 directly” and can only fight a guerrilla war: “You never know where the next strike will come from, so shorts don’t dare keep selling. Buy time for room.” If that fails, another rate hike may come in September. Korea is facing a similar situation.
• The complete funding framework has “3 things to remember”: debt, including CDS; equity, because Anthropic still needs to raise capital and eventually list to fund capex; and the longest-term question—how many people ultimately pay for AI. Capex is approaching $1T, while Anthropic and OpenAI together generate more than $10B of AR. Even $20B by year-end would not be enough; the requirement is $40B.
• The fundamental difference from the 2000 internet bubble is that end users are paying and cash comes back quickly. Semiconductor orders were also full in 2000, but “there was no money from the ultimate users.” Today Microsoft and Amazon sell tokens, and the business pays back in 6-8 months: “Why should I care about your interest rate?” The real failure point is a recession: “I’m happy to use AI for customer service, but the people buying things have been put out of work by your AI.” That is when the real bubble problem arrives.
• The original audio phrase “跟也 talk 这个下来” could not be verified as referring to a specific entity or meaning, and was not attributed to DeepSeek.

Verification Notes

• The original audio phrase “跟也 talk 这个下来” could not be verified as referring to a specific entity or meaning, and was not attributed to DeepSeek.
• The identities or referents of “Gorman” and “白川村” in the original audio are unclear; no specific institution or US-side counterpart was inferred from them.