Market Overview — November 18, 2025
Summary
The speaker rejects the view that AI is already a 2000-style bubble, because this parallel-computing revolution has been driven by profitable megacap tech companies from day one. Google and Meta’s ad recommendations are already lifting both revenue and margin, while hyperscalers are continuing to spend because they are making money. A Goldman Sachs chart roughly estimates that AI adds 1.5% to efficiency each year, or 15% cumulatively over a decade, implying around $8T in present value of future profits. The real cost is a K-shaped divergence: US college-graduate unemployment has reached 9%, and “AI’s industrial revolution is also a story of machines eating people.”（AI工业革命也是机器吃人的故事）The speaker does not see a massive bubble today, but thinks one is very likely next year.
The immediate risk is not an AI demand collapse, but a vacuum in monetary policy and data, alongside a fiscal-policy hiatus. Tariffs are estimated to have lifted inflation by roughly 0.5 percentage points, removing about two Fed cuts’ worth of room, while the government shutdown has degraded the quality of surveys such as payrolls and unemployment. Even if the Fed cuts in December, it will have fewer “bullets” in the first and second quarters of next year. The trading signal will shift from “bad news is good news” to “bad news is bad news,” but fiscal stimulus has only been delayed, not eliminated.
The speaker expects this S&P correction to look more like last July than this April, because there is no broad crowding across equities, volatility, rates and credit. Investment-grade CDX is roughly flat year to date rather than showing a major tightening. NVIDIA’s results on the 19th will “definitely beat, definitely double beat,” but the outcome is already priced in and the market will lack a catalyst once earnings season ends. His positioning framework is “reasonably bearish over the next few months, but fundamentally bullish.”
Next year’s bullish case still rests on three pillars: fiscal stimulus, token demand and financial deregulation. Fiscal measures in the US, China and Japan may take time to work through, but hyperscaler capex has not been cut. The global AI arms race is pushing shortages beyond GPUs into memory, ASICs, optical modules, networking and server CPUs. The speaker estimates that eSLR changes could ultimately free up $4T-$5T of balance-sheet capacity; banks will need 6-12 months to adapt, with the effects potentially emerging after roughly 6 months, or more likely around 8 months.
In the speaker’s view, Oracle CDS rising above 100 basis points is a warning about financing structures, not evidence that the company or the AI cycle is about to break. His hotel analogy is that high returns lead hyperscalers to fund their own data centers first with free cash and then with debt. The real bubble signal would come when financing moves off balance sheet and is sliced into senior and subordinated tranches, until “your pension fund is buying their bonds.” He also calls Oracle’s roughly $12.5B acquisition of TikTok’s US business “daylight robbery”; the value of the data, video-training material and user gateway is enough to change the conclusion one would draw from leverage alone.
Bitcoin has fallen to the source’s “9T” level, with the unit ambiguous, largely wiping out its gains for the year, but the speaker neither treats it as a core asset of this major cycle nor argues for an extremely bearish view at current levels. Liquidity deteriorated sharply after October 11, while the 10-year Treasury yield is around 4%, far above the 0.7%-0.8% range during the last crypto bull market, showing that Bitcoin remains “a liquidity asset.” Financial deregulation in February could bring money directly back into markets, so the selloff looks more like a liquidity and niche-asset problem than a break in the AI thesis.
With the US 10-year yield approaching 4.2%, the speaker believes most of the upside in the corresponding rates trade has already been realized. He expressed the view that yields would struggle to break materially below 4% through a swaption when they were around 4.0%, then closed the position near 4.14%, while still viewing 4.2% as fair value. Bid-to-cover was 2.4 at a 4.074% yield, and yields rose after last week’s Treasury issuance, underscoring how persistent supply weighs on bonds. The most likely bubble over the next year remains semiconductors at the intersection of token demand and supply, not a pure liquidity asset.
Deep dive
1. AI’s K-shaped divergence is a real cost, but not a 2000-style bubble
The speaker defines AI as a “K-shaped economy”: strong companies get stronger while weak links deteriorate. US unemployment among college graduates has reached 9%, reflecting both a weakening economy and technological displacement. His sharpest formulation is that Britain’s Industrial Revolution was a story of sheep eating people, and AI’s industrial revolution is a story of machines eating people.（英国工业革命是羊吃人的故事，AI工业革命也是机器吃人的故事）
Goldman Sachs offers a rough framework in which AI generates roughly 1.5% efficiency gains per year, or 15% cumulatively over a decade, implying around $8T in present value of future profits for the companies at the front end. The speaker stresses that this is a “very broad” macro estimate based on a discount rate. The tangible impact is developer productivity: work that once required multiple people across the front and back ends can now be handled by one person with AI.
The difference from 2000 is that the internet ecosystem then had to be built almost from scratch. Traditional companies were not even guaranteed to make the transition to the web, utilization was around 20%, and many winners were newly created internet-native companies. This cycle is being driven by the megacap tech companies born out of the previous internet revolution: “it was making money from day one.”（从第一天开始它就在赚钱）The speaker does not see a massive bubble today, but thinks one is very likely next year.
The application layer remains highly uneven. OpenAI, GPT and FSD use cases have yet to generate real revenue and are still in training, while Google and Meta’s ad recommendations are already monetized with high returns on investment. AI investment is increasing revenue while also lifting margins. Greater efficiency brings layoffs, but existing profits fund the next round of capex, creating a textbook K-shaped divergence.
2. Tariffs and the government shutdown push monetary, data and fiscal policy into a vacuum or hiatus
The speaker expects economic data to weaken decisively in the fourth quarter and next year. After Trump imposed tariffs arbitrarily, inflation is roughly 0.5 percentage points higher, equivalent to removing about two cuts from the Fed’s room to maneuver. The market will eventually realize that rate cuts cannot immediately offset a recession, and that “bad news is good news” will become “bad news is bad news.”
A month-long government shutdown has degraded survey-based data such as payrolls and unemployment. The Fed is like “driving through heavy rain”: it cannot see the road clearly and does not have many bullets left. The speaker still leans toward a December cut, but asks what policymakers can do when the economy weakens further in the first and second quarters. Whether the Fed skips December or cuts and then faces a weaker economy in Q1 and Q2, he sees a possible monetary-policy vacuum.
Fiscal policy is also in a hiatus. After the BBB bill passed, tax refunds that were supposed to go out in November were delayed, potentially pushing the fiscal impulse that was expected to turn neutral or positive in February or March back by another month. The temporary funding deal runs through January 30, and the two parties may fight again. Japan’s GDP is around 1.8%, while tariffs are hitting exports and housing is weakening. The speaker maps Japan’s property market onto China’s: population decline means that even massive money printing is unlikely to produce a sustained property bull market after a recovery. Further Japanese stimulus could weaken the yen, expand the carry trade and channel capital into the US.
3. Earnings have not deteriorated; what is missing is a catalyst after earnings season
On NVIDIA’s results on the 19th, the speaker is not debating whether they will beat: “the beat is certain, the double beat is certain, but there is no catalyst after the beat.” The upper end of the K-shaped economy continues to deliver strong earnings, but the outcome is already priced in, and the market is approaching a narrative vacuum as earnings season ends.
The vacuum will amplify concerns about Oracle CDS and the AI bubble, but it also creates the buying opportunity the speaker has been waiting for. He expects the worst economic data to arrive as late as February or March, but the market will bottom in advance. That makes it impossible to pinpoint whether the low comes in December or January.
He expects the S&P correction to look at worst like last July, not this April. The April episode required simultaneous crowding among CTAs, vol-control funds, vol sellers, and rates and credit investors. Investment-grade CDX is roughly flat year to date rather than showing a major tightening, and there is no broad imbalance like last year-end, when “every position was long and extremely crowded.”
The resulting positioning is short-term bearish and long-term bullish: “reasonably bearish over the next few months, but fundamentally bullish.” Economic deterioration is not an end-state signal. With the AI earnings thesis intact and policy stimulus merely delayed, “every decline is an opportunity.”
4. The token arms race and eSLR form the next liquidity cycle
Hyperscalers did not cut capex in their latest annual reports, and data-center investment is spreading from the US to markets including India and China. The speaker sees this as an industrial-revolution-scale, even US-China-competition-scale AI arms race, making a sudden collapse in overall training demand next year difficult to imagine.
The shortages are propagating layer by layer: from GPUs to memory, then ASICs, optical modules, networking and servers. Even Intel’s CPUs showed signs of shortage in this year’s annual reports and could remain scarce early next year. The speaker calls it “an across-the-board shortage,” rooted not in speculation around any single component but in the aggregate supply-demand gap created by data-center construction.
After the eSLR rule hearing concluded in late August, the speaker estimated that banks could receive more than $200B in capital relief, corresponding to $4T-$5T of balance-sheet expansion capacity. After accounting for bank dividends and other factors, the amount actually released could be $2T-$3T. The effects could emerge after roughly 6 months, more likely around 8 months, or by the end of Q1 or early Q2 next year; banks typically need 6-12 months to adapt. Alongside energy and crypto deregulation, this forms the third pillar of his bullish framework.
5. Oracle’s risk lies in leverage evolution, not its current CDS quote
The speaker compares hyperscalers to hotel groups that build and operate their own properties. When returns are high, they spend free cash first and borrow once that is insufficient; CDS moving above 100 basis points simply means the market has started to scrutinize financing costs. The truly dangerous next step would be moving projects off balance sheet and securitizing them. For now, “they are still on the balance sheet.”
His bubble test is highly specific: independent data-center companies issue senior and subordinated debt, hedge funds buy the subordinated tranches and pension funds buy the senior tranches. The cycle would be approaching its investment-bubble endpoint when “your pension fund is buying their bonds.”
Oracle cannot be assessed through liabilities alone. The speaker calls Oracle and Silver Lake’s roughly $12.5B acquisition of TikTok’s US business “daylight robbery” and values the asset at around $200B. More important, Oracle gains video-training data and a major user gateway, with video becoming the next training material after text and images. TikTok’s gateway may be larger than YouTube’s; the main remaining owners of internet gateways are Google, TikTok and Meta. He also emphasizes that Oracle is a major GOP donor and a major financial backer behind Trump.
6. Bitcoin, gold and Treasuries each have their own drivers; none is the core AI-cycle asset
Bitcoin fell to the source’s “9T” level, with the unit ambiguous, nearly erasing its gains for the year. The speaker says liquidity deteriorated sharply after October 11. During the previous bull market, the 10-year yield was only 0.7%-0.8%, with the low below 1%; today it is around 4%. Bitcoin therefore remains a liquidity asset rather than a “big-cycle asset.”
He also avoids an extreme bearish view at current levels. Bitcoin previously benefited from crypto deregulation, and financial deregulation in February could bring money directly back into the market, allowing it to benefit again. But he is candid that its rise also reflected Binance-related manipulation and arbitrage mechanisms; price performance should not be equated with the macro thesis.
On the US 10-year, the speaker used a swaption to bet that yields would struggle to break materially below 4% when they were around 4.0%. He had largely closed the position by 4.14% and views 4.2% as fair value. Bid-to-cover was 2.4 at a 4.074% yield. Yields rose after last week’s bond issuance, showing that persistent supply continues to weigh on Treasuries under the fiscal supercycle.
Gold’s distinctive support comes from the PBOC’s continued purchases as a “forced buyer,” while Bitcoin has deregulation and arbitrage mechanisms behind it. Neither is driven by the core industrial cycle. The speaker ultimately places the center of next year’s potential bubble at the intersection of token demand and supply: the semiconductor supply chain.