Market Overview — November 3, 2026
Topics
Macro Asset Allocation in High-Interest Regimes
Trillion-Dollar AI Capex and Bubble Risks
AI Chip Memory Wall and Bandwidth Bottlenecks
Compute Financialization and Futures Hedging
Physical Bottlenecks in AI Clusters
Pioneers
Shanghao Jin
Summary
- Shanghao Jin admits he underestimated in October how liquidity would overwhelm strong AI demand and earnings. Rising rates and rate volatility, along with deteriorating European credit, had already sent warning signals, yet he still expected earnings season to deliver: “We were a little complacent.” A pullback in 10-year yields does not prove liquidity has recovered; when risk appetite contracts, good results get little reward, while even a minor miss sent Intel down roughly 35%–40%.
- Extremely light positioning and lower valuations leave room for a rebound, but the trigger is easing funding stress. He says semiconductor hedge fund positioning has turned negative and market shorting is near the most extreme levels of the past decade. At the same time, earnings have grown since the start of 2026 while share prices have come under pressure, pulling the S&P multiple down from 23x to 18x. The key watch points are Anthropic completing its financing and whether the US continues fighting: improved liquidity could spark a “violent rally,” while further deterioration could spread to Apple, Walmart, Starbucks and Goldman Sachs.
- In his view, the AI funding debate is about the credibility of out-year demand, not whether the financial system as a whole has enough capital. He points to the $18T private credit market as ample capacity for AI investment, but says investors cannot see demand clearly in 2028–2029 and are applying a steep discount. Capital markets price future compute as depreciating, while industry players are willing to lock in more expensive compute ahead of time—a stark divergence. Nvidia’s expected 70% growth next year against a PE in the teens is the clearest example of that disconnect, he says.
- Slower CapEx growth does not mean a broad semiconductor glut, but DRAM may be first to face budget-driven pricing pressure. He expects cautious upstream capacity expansion and says GPU, CPU and connectivity products will remain in short supply for the next 2–3 years; the market could still favor sellers in 2028–2030. DRAM, by contrast, is a commodity that has already seen a major price surge. As power, permitting and financing all compete for budgets, “the best case for memory is that prices stay where they are.”
- Constraints on compute buildout could push rents higher, benefiting owners of existing capacity. His premise is that token demand will keep growing and enterprise process changes are hard to reverse: “Once the road is built, there’s no going back.” He cites growth of 103% at GCP, 39% at AWS and 48% at Azure. If cloud providers are right about demand, financing and construction bottlenecks will limit new supply and make existing compute even scarcer.
- His long-term view on Intel is unchanged: external customers and cash have resolved the two existential problems facing its foundry business. Yields on 18A and 14A are improving; Google and MTK have confirmed EMIB packaging orders, while 14A could attract Apple and Chinese customers. After 2 rounds of financing, he believes Intel has the cash to bring its processes to volume. “Intel will survive,” he says—a clear judgment on the company’s prospects, not a guarantee about its near-term share price.
- He attributes most of the drag in Intel’s guidance to a sharp drop in Chinese PC demand, while server remains supply constrained. He admits he had expected results to be good and did not anticipate such a severe deterioration in consumer demand. He rejects the idea that domestic CPUs or AMD took share. Higher yields are helping high-margin server chips, but shifting production from PC to server takes about 6 months; he says the lines could start changing over in the first half and be fully converted by Q4, with yields reaching 90% by year-end.
Deep dive
1. October’s mistake was letting AI fundamentals drown out liquidity warnings
- Shanghao Jin’s postmortem starts with an admission: in early October, he could already see rates and rate volatility rising and European credit deteriorating, yet he still expected robust AI demand and month-end earnings to support the market. “We were a little complacent.” He now stresses that when liquidity interacts with other factors, “liquidity is definitely the bigger short-term driver.” Short-term investors should have been more alert, or even stepped aside.
- He distinguishes a recovery in bond prices from a recovery in the overall funding environment. Moves in 10-year yields reflect both convexity hedging and buying by investors deleveraging or moving into bonds in a risk-off market. US high-yield and investment-grade credit affect equities, while European debt and FX carry feed into broader macro liquidity. For now, he says, “the catalysts for further liquidity deterioration have temporarily stopped appearing.”
- The environment has amplified the asymmetry in earnings reactions: AEHR reported very strong results but its stock barely rose, while a minor blemish sent Intel down roughly 35%–40%. He says funding stress cannot be blamed entirely on the Fed: volatility in long-dated US rates, the ongoing war and a lack of market certainty have all weighed on risk appetite. Intel’s results were not bad enough to warrant that kind of selloff.
2. AI financing is draining liquidity as shorts reach extremes and wait for a turn
- He attributes some of the recent selling to financing needs: OpenAI raised money last Friday, followed by Anthropic. The latter is expected to finish in mid-to-late November, possibly around the 20th, though he says the timing is uncertain. His explanation is that financing draws cash out of equities as investors sell stocks to meet funding needs; more of that money is coming from within AI, leaving AI and 3B-related sectors particularly short of capital. He points to CoreWeave’s CDS borrowing rate, which may have reached roughly 15%, as another sign of how cold the capital markets have become.
- Positioning is extreme: he says hedge fund shorting is near its most concentrated level in the past decade. A similar setup may have appeared in early 2026—he recalls it may have been around the start of the Iran conflict—and during Trump’s tariff push. Even as the market kept falling on Thursday and Friday last week, hedge funds overall were still covering shorts. The semiconductor positioning he may have shown earlier had already turned negative, meaning the market was net short.
- The valuation compression has not come from flat earnings. In his account, earnings have risen since the start of 2026 while share prices have fallen, taking the S&P multiple from 23x to 18x—a contraction of more than 20%. But he warns that when liquidity is scarce, “they can knock it down no matter what multiple it’s on.” He sees the liquidity shock as a short-term phenomenon because a prolonged squeeze would do far more damage to the fundamentals of the US and European economies.
- He says the shock has not yet spread to the rest of the S&P. Tight macro liquidity hits credit, then neocloud, then AI financing; bond markets are also under pressure in France and elsewhere in Europe, but financing stress is worse in semiconductors and AI. If it persists, Apple, Walmart, Starbucks and Goldman Sachs could also fall. If liquidity improves, semiconductors’ light positioning could fuel a “violent rally.”
- He is watching 2 events for signs of a recovery: Anthropic completing its funding round and the US deciding whether to go to war. A war-driven rise in oil prices could trigger a second round of liquidity stress. Positioning, valuations and the prevailing narrative are all in what he calls a “downside bubble” phase; the narrative is being driven by falling prices, not by fundamentals.
3. Industry and capital markets are putting opposite prices on the same future compute
- On funding capacity, he says the $18T private credit market can absorb AI investment, and the corporate bond market may also be able to handle $400B in hyperscaler financing. But another $400B of issuance next year could push large companies up against their financing limits, forcing them to turn to private credit or structured products. Ample financing channels do not mean the money will automatically be available.
- The real disagreement is that funders cannot see demand clearly. Even if CapEx reaches $1.1T next year, he says, it is unclear how much will be needed the year after or the year after that. Financial markets know almost nothing about 2029 and may know nothing about 2028 either. The more uncertain the forward curve, the larger the discount—one reason the market started talking about bubbles and a CapEx inflection so early. He points to the railroad cycle as a reference: concerns, mergers and restructuring only emerged once railroad investment exceeded 5% of US GDP, equivalent to perhaps $1.2T today. Before that, the market was relatively optimistic.
- Capital markets are pricing compute in backwardation: future capacity is expected to depreciate, demand is uncertain and supply could eventually exceed demand, so semiconductors get a cyclical valuation. His starkest example is Nvidia: the market believes growth will be 70% next year but gives it a PE in the teens. Even if investors accept that supply will be tight in 2027, they apply a steep discount to the years after.
- Industry, by contrast, is trading in contango: “Compute will cost more in the future.” Anthropic and cloud providers are booking capacity; they are also securing power and chips, while fabs are ordering optical modules, memory and upstream equipment. He sees the entire supply chain locking in capacity ahead of time. The industry’s implicit bet is that the gap between token demand and available compute will widen, though that does not guarantee semiconductor stocks will take off immediately.
4. Slower second-derivative CapEx growth changes bottlenecks and pricing first
- The CapEx growth path he cites is 100% this year, 50% next year and about 17% the year after. That can still be above prior consensus while representing a marked slowdown. But the further upstream you go—TSMC, ASML, Lam Research, AMAT—the more cautious capacity expansion becomes. Even with more investment, companies may not be able to buy enough chips. The market could therefore still favor sellers in 2028 and 2029, and even in 2030.
- Memory is no longer the only short board in the barrel. Storage could still be a bottleneck, but the list of constraints now also includes regulatory approvals and land, power equipment and grid connections, grid supply, local opposition, financing, and GPUs, CPUs and memory themselves. He understands the argument that compute is scarce but semiconductors are not in a major cycle; he thinks it may overstate how much semiconductor supply can expand.
- He names DRAM as the only area that could be directly hurt by slower CapEx. As a commodity, it can rise 10x when it is the sole bottleneck; if bottlenecks spread across the system, it gets a smaller share of the budget. His view is that the “best case for memory is prices holding at current levels.” The market will not assign it a high multiple, and the stock could trade around current levels. Demand for GPUs, CPUs, optical and copper connectivity could still go unmet for 2–3 years.
- The budget trade-off is concrete: if a company could previously spend 50 per watt of compute but financing constraints bring that down to 45, it has to ask suppliers to give up the 5. In his account, CPU, GPU and optical communications have not raised prices; memory, which saw the biggest price increases earlier, is therefore first in line for pressure. Memory prices can weaken even as compute remains broadly scarce.
5. Enterprise transformation is hard to reverse; constrained new supply raises the value of existing compute
- He uses a highway analogy for AI adoption: once cars are on the road, and supply stops, gas stations and toll booths are built around them, you cannot go back to horse-drawn carriages. “Once the road is built, there’s no going back”(路一旦修出去,就没有回头路了). As companies rework processes around tokens and configure their IT and cloud services, post-train models and deploy them, AI also replaces labor. He saw coding tools that could do real work emerge in January this year, including Claude Opus. He had not yet seen large-scale job losses then; over the past few months, he has started to see about 4,000 people a month lose jobs because of AI. Blue-collar jobs created by data center construction have temporarily offset some of the employment impact. He also cites an article about Hengdian background actors, saying material about them was post-trained into DeepSeek; he thinks actors and other industries will also be affected. Just as hand-sewn bags are displaced by machine-sewn ones, a few people may still prefer handmade goods, but most will switch to machine-made products.
- He cites growth of 103% at GCP, 39% at AWS and 48% at Azure. Google’s unease about AI’s impact on its traditional search and advertising businesses, he says, will strengthen its resolve to invest. The debate ultimately comes down to whether these companies are right about demand. If not, the impact could reach the credit and CDS curves of Amazon, Microsoft and Google; if they are right, owners of existing compute will continue to benefit.
- The mechanism is straightforward: the more that financing, permitting or grid connections constrain new supply, the wider the supply-demand gap and the higher rents can go. He says GPU prices have continued rising in recent months, even as capital markets worry that GPUs will last only 2–3 years, not enough to support 3–5 years of depreciation, and that compute will eventually be in oversupply. The tradable medium- to long-term debate is about the demand curve for compute power driven by tokens, not CapEx growth alone.
6. Intel’s foundry now has the customers and cash it once lacked
- Looking back at the pessimism when the stock was at $18–$20, he says Intel genuinely could have been acquired, gone bankrupt or been restructured. The company had invested in its fabs for 5 years, and internal yields appeared to have reached 70%, but no serious external customer was willing to use its PDK. More than $103B had gone into the foundry business, with another $10B due this year. A lack of customers to tune the process and a lack of money to keep investing were two problems that locked each other in place.
- Now, he says, yields on 18A and 14A are improving materially, and the external customer picture has changed. EMIB packaging involves TPU v9, and Google and MTK have confirmed orders. 14A could attract Chinese customers and Apple; Google may even put in funding. Apple and Google pushing for a second source to TSMC is, in his view, an important condition for Intel to keep moving forward. With customers involved, getting the process to work looks more like a matter of time.
- The other change is that Intel has completed 2 rounds of financing. His view is unequivocal: regardless of the share price, the company now has the money to bring 14A and 18A through. Using the Go term for making a group live, he says both conditions—customers and funding—have been met, so Intel will survive. That is the basis for his long-term optimism.
7. Intel’s near-term disappointment came from PCs; the server recovery still depends on yields and a production shift
- He sticks to his pre-earnings surprise: “I even thought this quarter would be good.” His main concern had been rumors that MTK’s yields were terrible, perhaps even zero. He found zero yields “hard to fathom,” thinking it could point to a design problem; he did not believe the rumor but was worried about the market reaction, so he had told friends asking whether to buy to wait. What he did not expect was the scale of the deterioration in PC demand.
- On servers, customers want to buy but cannot get the product; he cites Tan Lip-Bu’s regret that Intel cannot meet demand. Using 128-core server CPUs as an example, he explains the yield leverage on high-margin chips with output rates of about 10% at 83% yield and potentially 70% at 98%. At Intel, yields are still improving and could reach 90% by year-end, lifting data center gross margin.
- Switching production also takes time: AMD outsources production to TSMC, while Intel has to shift its own lines from PCs to servers, a process that takes about 6 months. He says Intel could start switching lines in the first half and complete the change by Q4, with financial performance affected in the interim. He also cites the bombing of Intel’s plant in Israel as a factor affecting future expectations.
- He traces the core PC problem to weak Chinese consumer demand: the channel guidance given in Q2 did not materialize. Over the 3 months from August onward, sales data for PCs, phones, consumer electronics and new-energy vehicles were all weak, and may have worsened recently. He rejects the claim that domestic CPUs or AMD are taking share, and does not expect AMD’s PC business to be particularly strong either. In a tight-liquidity environment, the market is focused on the PC downdraft and overlooking data center growth and the gross-margin improvement from better yields.