Global Capital Markets Analysis — February 25

Summary

• The speaker does not believe this pullback marks the start of a major selloff, but sees tail risks as far from exhausted, so he has not bought the dip and is keeping total exposure at about 60%. Over the past 2 weeks, hedge-fund selling of tech stocks reached roughly the 1st percentile, or more than 2 standard deviations from normal. Yet Magnificent Seven exposure continues to decline, mutual funds have almost no cash, hedge-fund leverage remains high, and the market is still “over-stretched”; upside is limited while downside could be deep, so investors should keep cash on hand.
• Capital is rotating from Magnificent Seven into small-cap tech, SaaS and AI applications, but falling prices do not mean commercialization will arrive immediately. Even with DeepSeek lowering costs and GPT pricing falling by about 97%, enterprise transformation, actual usage, productivity gains and monetization will “build slowly, then explode” （慢慢累积然后最后爆发）; against high expectations and a high base, software earnings in Q1 and Q2 will “very likely disappoint.”
• The VIX briefly rose to around 20, but that may not be enough to trigger an immediate gamma squeeze capable of pushing the S&P lower. With the S&P trading around 6000, CTAs are expected to sell about $10B-$11B over 1 week and roughly $27B over 1 month. But 90%-92% of companies are in their buyback windows, and February is one of the 3 most active buyback months of the year, leaving short-term liquidity relatively “balanced.”
• The real tail risk is not a technical selloff over the next few days but policy beginning to change the direction of economic growth. Tariffs on China could rise 10% initially and another 10% later, while DOGE’s aggressive government-spending cuts have not been fully priced in. Michigan’s 5-year inflation expectations triggered volatility after edging only slightly higher, underscoring the market’s extreme sensitivity to the mix of policy, inflation and growth.
• Nvidia’s next report could deliver a “double beat,” but strong results would instead offer a window for strategic retreat. The speaker expects Q4 revenue of about $40B and Q1 guidance of roughly $42B-$42.5B, with actual revenue potentially reaching $44B. The real problem is how far growth can run in 2026 and the difficulty of sustaining gross margins near 73%; exiting around midyear or after the Q1 report is his explicit choice.
• The speaker places Germany’s rightward electoral victory and the US DOGE push in the broader context of an ideological shift across the US and Europe, whose economic costs may arrive before any long-term efficiency gains. Musk-backed AfD won a large share of the vote, while Trump urged Musk to “get more aggressive.” His analogy is that cutting government is like “removing the barnacles from a turtle”: potentially more effective over the long term, but painful in the short term.

Deep dive

1. Positioning remains too crowded; this pullback is not enough to complete deleveraging

• Chinese stocks gave back their gains quickly after a strong rally the previous week, with Alibaba among the more volatile names; US equities were similarly choppy. The speaker’s core view is not that a major selloff has begun, but that tail risks remain unexhausted, so he rejects treating a short-term decline as an automatic buying-the-dip opportunity.

• In data through February 22, hedge funds sold tech stocks aggressively for 2 consecutive weeks, with the intensity more than 2 standard deviations from normal, or around the 1st percentile. On a stock basis, however, the mutual-fund indicator was around 1.3, suggesting they “basically have no cash left,” while hedge-fund leverage has continued to rise since 2013.

• His actual moves came earlier than the previously stated “start retreating in March”: he had already begun trimming software and TMT names that had run up over the previous 2 weeks. Exposure remains around 60%. AI is genuinely “an Industrial Revolution-scale thing”; quality companies can remain core holdings, but trading positions should continue to be reduced regardless of earnings.

2. The AI applications expectations curve is running ahead of real monetization

• Since the middle of last year, exposure to Magnificent Seven has continued to decline, while Prime Book has rotated into small-cap tech, SaaS and application-layer companies. The thesis is that these names have greater upside beta: once products such as Copilot genuinely deploy AI, valuations and earnings could rise rapidly, prompting capital to front-run the “third stage” of AI investment. Infrastructure, cloud and chips were the second stage.

• The US economy held up reasonably well in Q4, and small software companies may not have reported poor results, which only raises the hurdle for the next 2 quarters. The speaker expects Q1 and Q2 to “very likely disappoint” because the market has treated the arrival of a new technology and the resulting productivity gains as almost immediate, whereas real-world adoption must accumulate gradually before it can potentially explode exponentially.

• DeepSeek’s lower costs and the roughly 97% drop in GPT pricing solve only the cost of the model. Enterprise workflow redesign, usage scale, productivity gains and ultimate monetization will not arrive in sync. “The market is always a comparison between expectations and reality”; before applications truly break out, expectations must come first, and the moment expectations peak is precisely when investors should retreat.

3. Short-term liquidity has offsets; directional economic deterioration is the real risk

• With the S&P trading flat or slightly below 6000, CTA models project roughly $10B-$11B of selling over 1 week and about $27B over 1 month. At the same time, 90%-92% of companies are in their buyback windows, and February is one of the 3 most active buyback months of the year. Together, the 2 forces should leave liquidity broadly “pretty flat” over the next few weeks.

• A gamma squeeze occurs when options sellers close out positions and dealers potentially shift from long gamma to short gamma, turning hedging from buying declines and selling rallies into chasing both directions. Outstanding gamma is currently insufficient, and a large number of options have recently expired, unlike the extreme phase last August when gamma was around $7B.

• A VIX move toward 20 would attract fresh trend-following sellers; if it reached 30 or 40, many traders would regard that as “free money, 白送钱.” The speaker’s view is that current positioning does not support an easy move to 30 or 40, nor does the technical setup point to an especially large near-term risk.

• More concerning was the market’s sharp reaction when Michigan’s 5-year inflation expectations edged only slightly higher. Compared with that day’s rumor that China had another new virus, the response is more revealing: once policy causes the economic fundamentals to change direction, volatility will be far greater than the immediate gamma risk. Unless one is day trading, the right move is to watch and wait.

4. Nvidia could deliver a double beat, but gross margin and 2026 growth are now hard constraints

• The speaker is relatively bullish on Nvidia’s next report: earnings should be fine and could deliver a double beat. Q4 revenue is expected at about $40B, Q1 guidance could be $42B-$42.5B, and actual revenue may reach $44B. Demand for China’s H20 is also increasing; whereas China previously did almost no training and focused mainly on inference, many customers are now beginning to experiment with training.

• The market’s biggest blind spot is margin. Many fund managers bullish on Nvidia appear to believe that gross margins near 73% can be sustained, but he calls that “unimaginable.” Gross margin could fall below 73% in 2025, even below 70%, and will be even less sustainable at current levels in 2026.

• The first hard constraint is the growth comparison base: capex investment was enormous in 2024 and 2025, so how much growth remains in 2026 needs to be reassessed. Citing Microsoft’s apparent cancellation of a data-center project in New York, he argues that even achieving growth in the 60%-plus range this year will be difficult.

• The second hard constraint is that hardware iterations are slower than software iterations. Even if DeepSeek has not changed the entire landscape, nobody knows whether the next generation of models will favor inference or large-cluster training. The fact that DeepSeek did not create a risk this time does not mean the next one will not, while the current valuation already discounts a fully loaded growth outlook.

5. Tariffs and DOGE will inflict pain first; the Western rightward turn is widening policy uncertainty

• Tariffs on China will rise 10% initially and another 10% later, alongside DOGE’s deficit-cutting drive, but the speaker believes the market has not fully priced the combination. If government layoffs follow Musk’s Twitter playbook—cutting until something breaks or becomes unable to operate, then hiring people back—the impact on the public sector will be more complicated.

• Government spending cannot be judged solely by profitability: USAID affects US influence overseas, health care affects social stability, and even inefficient spending is included in GDP. Cuts may improve long-term efficiency, but they could also resemble “removing the barnacles from a turtle”; the economy may first endure significant pain in Q1 and Q2, while Q1 earnings will not begin reporting until early Q2.

• In the same post, Trump first said “Elon is doing a great job,” then said he wanted Musk to “get more aggressive.” The speaker places this alongside the rightward victory in Germany’s election and the large number of votes won by Musk-backed AfD, arguing that the US and Europe are experiencing a conflict between left- and right-wing ideologies, shifting from exporting universal values toward a more closed process of self-reinvention.