Market Overview — October 27, 2026
Summary
- Shanghao Jin attributes most of the AI selloff to a liquidity shock in rates and credit markets. His chain of causation runs from energy volatility to higher rate volatility, bond buyers pulling back, wider CCC spreads, and then pressure on neoclouds, semiconductors, and AI stocks. He says investors are treating every drop as a reason to turn more bearish, reading falling prices as evidence of deteriorating demand and creating a “downside bubble.”
- He sees little chance of an AI funding gap this year or next; the real debt test comes in late 2027 to 2028. He puts neocloud financing at about $230B this year and $420B next year, and says hyperscalers still have roughly $500B of funding capacity. A gap could emerge in 2028, but “that may not happen” if revenue and cash flow grow enough to support financing.
- In his view, lower token prices hurt the residual value of older GPUs but do not overturn the case that GPUs and compute remain scarce. He says a two- to three-year payback is still assured and cards can be used for five years; investors simply should not assume they will still be worth 95% of their original price after five years. The market demands roughly 60% IRRs without assigning valuations to match, which he sees as a pricing disconnect; his conclusion is that “compute is not actually going to be oversupplied” (“算力其实不会过剩”).
- He reads Anthropic’s slower revenue growth as a supply constraint, not a lack of customer demand. He cites Anthropic’s S-1, which attributes the slowdown in its revenue outlook to insufficient GPU supply: customer demand is strong, but the company cannot guarantee it will secure enough capacity. He says he has heard ARR reaccelerated in recent weeks, and views positive operating cash flow as a basis for continued investment in compute; his spoken figures for cash on hand were inconsistent.
- Simultaneous fundraising intensified the market’s liquidity drain, while stable rates and lighter issuance after early December offer a path to relief. Anthropic, OpenAI, Meta, and others raising at the same time amounted to “taking a few more buckets out of an already fairly dry well” (“比较枯的井里面又打掉了几桶水”). He has seen convexity hedging ease and safe-haven bond buying emerge, but says it is still impossible to know how the war will unfold or how long the shock will last.
- Low positioning and depressed valuations leave room for a sharp reversal on even a small positive catalyst. He says the long-short ratio is down to 1.5, near levels seen during the tariff shock, and semiconductor forward P/Es are below where they were when GPT launched. Software earnings have indeed slowed, while semiconductor and AI earnings are still growing without decelerating. He admits he misjudged how well stocks could withstand a liquidity shock, and says, “I’m not sure when the real rebound will come.”
- Continued product progress and room for enterprise adoption are the core reasons he has revised his June–July view and remains bullish. Code generation is only the start: review, cybersecurity, and other bottlenecks still need to be addressed through an “AI industrialization” approach. He sees the most aggressive compute buyers securing funding and willing to lock in supply at $50 per watt as a positive for semiconductor fundamentals; next, he is watching cloud earnings, improving liquidity, and shifts in investor flows after the midterm elections.
Deep dive
1. The market is reading a rates shock as an AI demand crisis
Shanghao Jin invokes Lloyd Blankfein’s comments on bubbles in rising and falling markets, made when Jin joined Goldman Sachs in 2009, to describe today’s broad pessimism toward semiconductors and AI. He traces the initial shock to energy-driven rate volatility: real-money investors such as insurers and overseas investors care about total returns, not just spreads, and sharp swings in rates make bond buyers pull back.
Bond markets had been under pressure since early October: French government bonds and high-yield credit were already struggling, while investment-grade credit had not yet shown problems. Stress then spread to lower-rated debt. CCC issuers, which account for about 10%, generally generate just enough revenue to cover interest and principal payments. Neoclouds are in the same pool of capital and have also seen spreads widen, but credit quality still varies: CoreWeave is in considerably worse shape than Nebius, so the two should not be lumped together.
His supporting evidence includes French government bond yields exceeding 5% for the first time since 2002, with spreads reaching 168 basis points. Muni bonds may also be having “one of the worst quarters in decades.” Credit markets have not fully healed, so widening neocloud credit spreads are not, on their own, proof that AI fundamentals have deteriorated.
Foreign exchange, Chinese equities, and Latin American markets have also been hit, another reason he rejects a purely AI-driven explanation. South Korea can be said to have been directly affected by AI, he says, but not all of Latin America’s weakness can be attributed to it. AI had rallied sharply beforehand and fell harder in a liquidity shock; that does not mean AI was the source of the shock.
2. Financing needs have to be assessed against funding capacity and cash flow
On the funding question, Jin estimates neocloud financing at about $230B this year and $420B next year, but argues that the absolute figures are not enough on their own. The real questions are whether the market has capacity to absorb the issuance, whether the companies can make money, and whether bond funds are approaching their allocation limits for the sector.
He estimates hyperscalers have about $500B of financing capacity left. Even if they raised aggressively and did not refinance maturing debt, he says, they would not hit the limit in 2027 at current financing levels. Profitability determines whether debt can be repaid; borrowing cannot be assessed by looking at interest expense alone, without considering project returns.
His timing call is that “it’s unlikely to happen this year, and unlikely next year either.” A large share of the financing comes in 2027, while the main refinancing wall does not arrive until 2028. The real test is in late 2027 to 2028, by which point revenue and cash flow may have more time to prove themselves—and the funding gap may never materialize.
3. Cheaper tokens do not stop GPUs from generating strong returns
Jin accepts that lower token prices will reduce the residual value of older GPUs: it is reasonable to assume a card will not sell for 95% of today’s price five years from now. But the cards will still produce tokens. The market has not valued neoclouds on the assumption that they can sell equipment for 95% of today’s price in five years, so a lower residual value does not mean the cards are worthless. He says payback in two to three years is still assured, and the cards can be used for five years.
Returns and valuation have to be judged on the same basis. When liquidity is particularly tight, investors expect a 60% IRR from companies like CoreWeave; Jin says the market is not assigning valuations that reflect it. If investors really believed in a 60% IRR, “these companies should be at 100x P/E.” He also cites Labs’ potential to pay back in ten months by building in-house, and says DeepSeek, despite cheaper tokens, can still generate an 80% ROI.
The key to his demand thesis is that Microsoft, Amazon, and other cloud providers are looking at customer demand on their own platforms, not just betting on how much Anthropic or OpenAI can earn. He believes that demand is already enough to absorb the compute available and that the shortage will last much longer than the market is pricing in. “Compute is not actually going to be oversupplied” (“算力其实不会过剩”), he says.
4. Anthropic’s slowdown is a supply constraint
Jin cites Anthropic’s S-1, which attributes the slowdown in its revenue outlook to a shortage of GPUs: the company has substantial customer demand but cannot guarantee it will secure enough supply. He also says he has “heard” ARR reaccelerated in recent weeks, and that the company shifted some training compute to serve customers; those are his accounts of the situation.
Google is another example of tight supply. Jin says the company is rolling out a personalized search agent, similar to Muse but simpler, and needs to free up compute for it. At the same time, even Google’s training teams are finding it difficult to rent compute on weekends, suggesting new products and training are competing for capacity.
Based on the disclosures, he says Anthropic is already generating positive operating cash flow and that accounting losses are mainly driven by convertible-debt accounting. His point is that the business is profitable operationally. His spoken cash figures were inconsistent: he first said Anthropic had raised $14B and had $12B on its balance sheet, then said it had more than $100B in cash.
In his argument, as long as a company can train models and sell tokens while generating positive cash flow, it has the capacity to keep investing in compute. The market demands that ARR continue to rise linearly, he says, but does not value Anthropic for exponential growth; he puts its valuation at $1.5T. That high bar ultimately reflects a market short of capital.
5. Concurrent fundraising drained liquidity; easing rates offer a path to recovery
Jin says Goldman’s FCI began weakening in early October, followed by several large fundraises at once. Anthropic filed an S-1 and is looking to raise a large amount; OpenAI is also seeking funding. Meta plans to issue non-voting shares, allowing it to retain control while continuing to invest in AI. Some SpaceX shares have also come out of lockup, adding to the pressure on the market’s ability to absorb supply.
He describes the combination as “taking a few more buckets out of an already fairly dry well” (“比较枯的井里面又打掉了几桶水”). He also sees OpenAI launching model 6.1 at the same time as Anthropic’s fundraising as part of the companies’ head-to-head competition. The near-term drain on capital does not mean these companies cannot raise money or that AI demand has weakened.
He traces the shock to the U.S.–Iran war, oil-price volatility, and the Fed’s reluctance to manage long-term rates. As rates rapidly rose to 5.3%, convexity hedging added pressure, but that pressure had eased two weeks ago. Last week, when equities fell hardest, clear safe-haven buying appeared in bonds and yields moved lower.
Stable rates would bring credit buyers back, while issuance should fall to a trough after early December, easing supply pressure. Jin argues that if tight financial conditions hurt the real economy for long enough, policymakers will inevitably intervene. But no one can predict the war: continued fighting could push rates back up, and whether the shock lasts until the election or longer is “hard to say.”
6. Positioning and valuations already reflect deep pessimism
Jin says hedge funds’ exposure to AI infrastructure, along with their market-weight-adjusted net exposure to semiconductors, has fallen sharply since early 2026. U.S. equity allocations among fundamental funds have returned to early-2025 levels, and the long-short ratio has dropped to 1.5, near its level during the tariff shock. Positioning is “very, very light.”
Valuations have also compressed: by his comparison, forward P/E ratios are below tariff-period levels, and semiconductor multiples are even below where they were when GPT launched. Earlier in the year, when software came under scrutiny, earnings really were slowing. This time, semiconductor and AI earnings are still growing without decelerating, yet the sectors face similar doubts and selling. That is why, he says, the two episodes are not directly comparable.
He admits he misread the situation two or three weeks ago—or perhaps a month ago. He thought progress in models and cloud-provider earnings could support stocks, and at first equities did not follow bonds lower. Only later did he realize the shock was delayed: “When the wave came crashing down hardest, nothing could survive.”
On a reversal, he thinks that if there is a spark, the market could catch fire, but he is cautious on timing. Money will come back as liquidity improves, but he does not know when risk appetite will turn. Implied volatility remains high, though it has stopped rising; uncertainty about the war is also keeping investors from selling volatility.
7. Product adoption and cash flow have changed his medium-term view
Jin says the shorts raising funding and growth concerns today were already discussing them in June and July. What changed was his experience of using the models. Enterprise effectiveness and adoption still have a long way to go, while progress in models, agents, Muse, and other products continues to open up new applications. AI has not yet reached the maturity of established internet products.
He contrasts AI with Taobao: mature products are hard to improve in ways that lift adoption, while AI products can still do so. Coding is only the starting point. After large volumes of code are generated, code review, cybersecurity, and other bottlenecks need to be addressed. That does not mean cybersecurity companies can grow at 100% forever; it means those workflows, too, need to be rolled out progressively using AI industrialization methods.
He sees a key difference from 2000: back then, frontend companies were unprofitable and used funding to buy servers; today, AI companies are making money as they expand. Anthropic, Meta, and OpenAI, he says, will use the money they raise to buy GPUs. The most aggressive buyers are willing to pay $50 per watt to lock in compute, above the $35 or $40 per watt typically accepted by CSPs. By his math, 1 GW works out to about $50B. If these buyers secure funding, he sees that as positive for semiconductors.
Jin had previously expected the next three cloud providers to report better results than Meta, because he thought Meta had Muse but lacked a real monetization engine. He expects the next few days and earnings reports to begin providing evidence on cash flow, with subsequent CSP results potentially making the case for cash flow and returns even more clearly. He is also watching whether investors become more willing to put money to work after the midterm elections, while sticking to his caveat: “I’m not sure when the real rebound will come.”