Market Overview: August 25, 2026

Summary

• Shanghao Jin’s core call: back from vacation, he found “everyone was bearish,” but thinks he may have been among the earliest bears and is not pessimistic as a result—“The ghost stories being told in the market now are no different from what we were saying in June.” He had already sensed Anthropic’s ARR would fall when using the model in June, while current positioning and leverage are nowhere near late June or July levels, making a June-July-style liquidation less likely.
• Anthropic’s $6.5B ARR, versus the previously reported $7B July target, disappointed the market. He offered 2 explanations: the company may have standardized its reporting ahead of an IPO, creating inconsistent figures and “understating it somewhat”; or growth is genuinely slowing because Google, Amazon and Microsoft, each with roughly $20B in revenue—“that’s revenue; ARR should be at least 2x”—are siphoning off part of the enterprise demand from the 2 labs. Whether Anthropic can keep growing after that demand is stripped out depends on how fast the overall market is expanding.
• 2 earnings reports are the clearest evidence of what the actual chip buyers think: Alibaba’s annual report explicitly says that buying U.S. chips offshore pays back in 2.2 years, while Microsoft Corporate AI has grown from the low single-digit billions at the start of the year to a $20B year-end target against just $1.5B of costs, implying a 6-month payback on chips. “Read Microsoft’s and Alibaba’s financial statements carefully, and listen closely to what 梁文锋 says… semiconductor demand is nowhere near a peak.”
• AI is not a tool for programming; it is a tool for processing large volumes of data in parallel. Cross-border e-commerce inventory, tax, and traffic-acquisition data have never been fully exploited through training, and that is where enterprise AI demand lies. Programmers displaced may account for not 30% but 80%-90%, yet the programmer population could grow 10x; organizational redesign will also create huge numbers of agent “robot workers,” all consuming CPU, networking and storage.
• A second underappreciated positive is the depreciation cycle: V100s are still in use, while A100s have not fallen in price and are still rising. “Once the depreciation cycle of chips changes, it will completely change the earnings cycle of every CSP.” Combined with revenue growth, that means better hyperscaler cash flow next year, and he is “relatively more bullish” on the 12-month medium-term outlook.
• Long-end rates are a real macro problem, but not an acute one. U.S. Treasuries have no credit problem—“you need debt to become the reserve currency”—and the essence of rising yields is that nominal GDP is growing too fast. He cited a roughly 180bp gap in long-end rate premia that still needs to be closed: either it widens and produces very strong inflation expectations, or long-end yields keep falling and rising “like Sisyphus pushing the rock uphill,” implying continued valuation compression over the next 2-3 years. Liquidity, with GSFCI improving since July, and positioning do not present a near-term crisis.
• The trading takeaway: this is a “no-multiple” phase, but hedge funds’ 3-month momentum trade is long software and short semiconductors whose earnings growth could reach 50%-100%, while the long/short mix is materially different from June. Reverse covering could become a positive catalyst. Once cloud earnings reset expectations and the cash-flow impact becomes visible, “the market will come back.”

Deep dive

1. No New Ghost Stories for the Bears: Positioning and Leverage Make a Repeat of the Flush Unlikely

• He opened with the key observation: after 3 weeks off the air and a vacation, “I found that everyone was bearish.” He considers himself “possibly among the earliest bears,” so the current pessimistic narrative is no different from June’s; he had already felt that Anthropic’s ARR would decline when using the model at the time.
• The bears’ industry case is concentrated on money: cloud providers’ free cash flow, how capex is financed, and the fact that “anyone who comes out to raise financing gets crushed,” Intel included. There is no credit, debt or equity financing—and even less “money the market is actually spending on tokens.” Anthropic’s slowing growth is being treated as confirmation.
• But positioning had already started improving 2-3 weeks ago. There was no sign of large-scale net selling or momentum trading, no downward spiral in leveraged ETFs and no forced external deleveraging. Leverage was not as high as in May, June, late June or July, so even if the market falls sharply, “I don’t think” a June-July-style liquidation is likely to recur.

2. This Industrial Revolution Cannot Be Compared to 2000: The Cancer Cell Has Only Just Entered the Body

• The essence of 2000 was “money inside the system and money outside the system.” There were plenty of profitable internet companies within the system, and semiconductor orders already had more than a year of backlog before they were later cut. But the real economy was not transferring money from outside the system into it. Internet companies advertised to one another—“Was Amazon actually selling anything? It wasn’t”—and there was nothing like the large-scale advertising spend that later emerged with Google.
• His cancer analogy is worth preserving in full: AI replaces human labor, and “once you squeeze out all human work and drain all the nutrients, the system becomes exhausted.” Eventually even demand for coffee advertising disappears. But that is the end game. “We are not at the end game; this is the opening.” When the cancer cell has just entered the body, it extracts nutrients at extraordinary speed. Wage money, advertising money, bond money and stock-market money are all being pulled toward it now, and that process is already underway.

3. Open-Source Models Could Rewrite the Cloud Providers’ Role: From Nursemaid to the Main Enterprise AI Channel

• Under the old model, 80% of the profits went to OpenAI and Anthropic, with “the cloud providers playing nursemaid in the back”—raising equity and debt and passing backlog orders on to the labs. Open-source models are changing the game. Microsoft is deploying Corporate AI across the U.S. at scale, offering enterprises an end-to-end solution. The business was running at roughly $1B-$2B at the start of the year, with a $20B year-end target and just $1.5B in costs. “It really is a 6-month payback,” and it is easier to sell than Anthropic because it combines enterprise sales, cloud services and security.
• This enterprise AI business is not included in Anthropic’s ARR, nor is it AI-to-B services. A large portion is enterprise post-training demand. Anthropic’s own enterprise AI backlog is already full, but “it doesn’t have the compute to solve the problem.”
• Alibaba’s annual report provides another hard number: investing in U.S. compute and buying U.S. chips offshore pays back in 2.2 years. His conclusion is that the rational move for technology and AI companies “is to turn all their cash into chips.”

4. AI Is Not a Coding Tool but a Parallel Data-Processing Engine—Agents Are the Real Compute Demand

• He rejected Wall Street’s “coder headcount × 30% displacement rate = $300B market” calculation. The share of coders displaced “is not 30%; it could be 80%-90%,” but the programmer population will grow 10x. AI is replacing organizational structures: front end, back end, product and testing will be consolidated into “superhuman one-person units.” The process will also create huge numbers of agents—“robot workers”—all of which require CPU, networking and storage.
• The real use case is enterprise data. Consider cross-border e-commerce: when to stock inventory, how much tax to pay, how long shipping will take, and “whether you should buy traffic on Amazon.” “This data has not really been used well in any industry.” Enterprise AI outcomes depend on the model, compute, data and the harness—the way models are called and agents are built.
• His message to token skeptics was blunt: “If the number of tokens you use now is not enough to max out Anthropic’s Max plan every week, I think that person’s capabilities are almost certainly among those that will be displaced.” Anyone who thinks token demand is excessive “should look for the problem in themselves.”

5. 2 Readings of the $6.5B ARR—and the Depreciation Cycle That Changes Everything

• Anthropic’s $6.5B ARR has 2 possible explanations. First, the company may be standardizing its reporting ahead of an IPO, making the figure inconsistent with the previous “$7B in July” target and “understating it somewhat.” Second, the slowdown is indisputable—but Microsoft, which mainly uses OpenAI, along with Google and Amazon, each has roughly $20B in revenue and is siphoning off part of the enterprise demand previously captured by the 2 labs. “If demand is being siphoned off and you can still grow, you have to look at how fast the market is growing.”
• He had already concluded in May and June that Adobe and OpenAI’s results would deteriorate. After using the products, he found them expensive; ARR will eventually hit a ceiling, and tokens will ultimately lead either to a price war or new entrants. Semiconductor demand, however, has not fallen.
• The depreciation cycle has been the key variable under discussion for several weeks. V100s are still in use, and “A100s have not fallen in price; they are still in the process of becoming more expensive.” “Once the depreciation cycle of chips changes, it will completely change the earnings cycle of every CSP.” Add revenue growth, and cash flow should be better next year.
• His time frames are clear: in the short term, watch Anthropic, OpenAI and market positioning; over the next 12 months, watch hyperscaler cash flow—“I am actually relatively more bullish on the medium term over the next 12 months”; over the long term, watch dollar rates and “whether there is ultimately any human demand.”

6. Macro: The 180bp Sisyphus Problem and the Valuation-Compression Cycle

• Macro bears are also worried about rising rates, runaway inflation and a supposed U.S. dollar-debt crisis. He considers the concern that the 10-year yield is on a structural uptrend “valid,” but says it affects the valuation regime rather than immediate liquidity. The S&P 500 is trading at roughly 19-20x forward P/E, while forward P/E has continued to fall in 2026 despite strong earnings growth. Liquidity, measured by GSFCI—which incorporates credit spreads, FX and carry—has improved continuously since July. “From a liquidity perspective, there is no particularly serious problem.”
• He completely rejected the U.S. Treasury credit argument: “No matter how much debt it owes, the U.S. will not have any credit problem… because you need debt to become the reserve currency.” The essence of rising long-end yields is nominal growth: U.S. nominal GDP is growing too fast. He cited a roughly 180bp gap that must gradually be closed. The Treasury and Biden have sought to suppress the long end by issuing more T-bills and reducing long-bond supply; historically, short-term issuance had a roughly 20% cap. The gap either widens and produces strong inflation expectations, or long-end yields fall and then rise again “like Sisyphus pushing the rock uphill.” The process of closing it will push U.S. corporate valuation frameworks lower over the next 2-3 years. “If a company does not have enough growth, there is no way for it to command a valuation.”
• At Jackson Hole on the 28th, he expects the Fed to have few effective options. Neither a hike nor a cut would solve the problem, and the Fed cannot control long-end yields. The term premium signals that the market strongly dislikes Fed control of the long end, so long-end yields will continue to grind higher.
• Positioning supports the near-term view. Prime book margin, measured over a 1-year horizon, is around 48 and not particularly high. Net positioning is not elevated, suggesting that much of the activity is momentum trading. Hedge funds’ 3-month momentum trade is long software and short semiconductors, where earnings growth could be 50%-100%; over a 1-year horizon, they remain broadly long semiconductors. The mix of positions makes a June-style, across-the-board liquidation less likely, and shorting semiconductors is “a contrarian positive.” The conclusion: this is a valuation-compression cycle, and the market will keep debating whether to pay multiples—“this is a period when the market refuses to pay a multiple” (“不给你估值期”). Once cloud providers’ earnings open up expectations, “the market will come back.”