Market Overview — September 1, 2026

Summary

• Jin is skeptical of mapping the current market directly onto the 2000 bubble: historically, “no bubble has ever happened when everyone thought it was a bubble.” He considers himself bullish, observing that “the bears around me think it’s 2000 and about to blow up, while the bulls think it’s 1998,” yet “nobody thinks this AI is an industrial revolution—as if we’re in 1950.” In 2000, by contrast, nobody thought the internet was a bubble; everyone saw it as an industrial revolution.
• Kevin Warsh’s hawkish remarks at Jackson Hole have put the market at roughly a 60% probability of a rate hike, but Jin sees a September hike as unlikely. He notes that the probability data may not have been updated or may have been updated incorrectly. Since 1982, the Fed has never hiked after 2 consecutive negative nonfarm payrolls reports—it “hasn’t happened in 40 years.” A more likely explanation is that Warsh expects the data to be very weak and therefore felt comfortable talking tough. Forward indicators—job openings, wage growth and current retail sales—are all weakening, while the $40T-$41T debt ceiling approaches at year-end and the downward leg of the K is proving persistent.
• He explicitly rejects the popular claim that Warsh’s hawkishness is meant to help the White House cap the 10-year yield: the Fed has controlled the long end for 15 years, and “there are 100 ways” to push it lower. It could adjust the repo facility, alter its Treasury holdings and extend maturities at the long end. The real issue is that, once the force of QE is removed, the long end should converge with nominal GDP at 6%-6.5%; with the 10-year currently around 4.74%-4.78%, inflation and growth would need to fall by roughly 200bp in aggregate before the long end could move below current levels. “I think that’s very difficult”—this inflation cycle is driven by supply-side inputs including memory, chips, tariffs and oil, none of which rate hikes can solve.
• In a K-shaped economy, rate hikes are chemotherapy: “AI is a cancer; going in with chemotherapy kills the other cells and makes the cancer cells grow faster.” Data centers have a 2.2-year payback period, according to an Alibaba report, and do not care whether interest rates are 6.5% or 4.5%. The damage falls on brick-and-mortar businesses such as new Starbucks and bubble-tea shops, whose payback periods exceed 10 years. When the economy deteriorates, companies cut people before tokens, “because tokens are cheaper than people.” If Warsh insists on hiking, it will be a policy error destined to fail; a year-end hike would bring bear steepening, and the 10-year still could not be controlled next year.
• On valuation, “the market is extremely surreal and crazy”: the K-shaped split has made new productivity the downward leg of valuation. Jin starts with Starbucks at 36x P/E versus Nvidia at 12x P/E, then compares a company growing 75% with a 74% gross margin at 12x P/E against one with a 15% gross margin, 15% growth and 30x P/E. The S&P 2026 forward P/E has already fallen from 22-24x to below 20x, around 20x, “not very expensive.” The conclusion: stay away from slow-growth companies—“the biggest valuation premium is slow growth.”
• Positioning and credit suggest that the market’s short-term structure is relatively healthy, with liquidation shocks far smaller than in June and July. Semiconductors have shifted from the momentum long leg to the short-term short leg, while fundamental-book positioning has fallen to Liberation Day levels. Jin questions BofA’s report that the bull-bear index and positioning concentration are both high; at least Goldman’s data show positioning is nowhere near high. IG absorbed large debt sales from Microsoft, including $20B, and Google, including $5.5B, while HY remains at low levels. He has bought CDS and made money, but stresses that “the real problem isn’t today—it may be next year.” AI-basket volatility has fallen from roughly 120 to 66; excluding cross-asset correlation, vol-control strategies are actually adding to positions.
• Jin believes AI’s real constraint may not be rates but power and permitting: based on figures cited from a Fangda report whose key term is unclear in the original transcription, truly available compute is about 50GW, versus 35GW the US expects to build, while only about 11.4GW was added over the past 12 months and just 11-14GW has permits and power. “It doesn’t matter how much money you spend.” Of the 43GW of abandoned projects, 78% were killed by policy and only 1% were abandoned by developers themselves. GE’s large generators are booked through 2031, and 60%-70% of data-center delays are power-related. 黄仁勋 therefore declared that “China wins” the AI race: China can deploy power in a year, with fewer permits, community objections and approvals. If China is not power-constrained while the US becomes power-constrained, compute equipment could flow to wherever the electricity is, potentially reshaping the competitive landscape. Jin sees the economics of AI as difficult to derail, but further policy restrictions remain uncertain.

Deep dive

1. The Market Is Mapping to 2000—but That Does Not Make It 2000

• Jin set the tone at the outset: among the funds around him, “the bears think it’s 2000 and about to blow up, while the bulls think it’s 1998.” He considers himself bullish, but “nobody thinks this AI is an industrial revolution—as if we’re in 1950; nobody feels that way.”
• The reflexivity argument is the episode’s anchor: “No bubble in history has ever happened when everyone thought it was a bubble.” In 2000, nobody thought the internet was a bubble; everyone thought it was “especially powerful, especially impressive.” The fact that so many investors now map the market to 2000 from a position of fear is itself worth watching.

2. Jackson Hole: Warsh Is Hawkish, but the “Cap the Long End” Thesis Is Wrong

• The backdrop: the latest payrolls report was “indeed very bad,” yet Warsh continued to focus on inflation and said the data were not sufficient to establish that inflation would certainly decline. The market read him as highly hawkish and priced roughly a 60% probability of a hike; Jin also warned that the probability data may not have been updated or may be wrong. The immediate reaction was a curve flattening: the 10-year barely moved, while the 2-year and 5-year yields rose.
• Jin differs sharply from many macro investors. The popular explanation is that Warsh is talking tough to help Bessent, who cares more about the 10-year, push down the long end. “I think that is definitely wrong.” The Fed has controlled the long end for 15 years; change the repo settings, adjust the Treasuries on its balance sheet and increase long-end rollovers, and “the long end comes down immediately.” There are 100 ways to do it.
• The real anchor is that, once QE is removed, the long end ultimately has to align with nominal GDP—more than 6% today, or roughly 6.5%. The 10-year is currently around 4.74%-4.78%; even if oil, core inflation and tariffs—all supply-side factors—were brought down, inflation and growth would still need to fall by about 200bp in aggregate before the yield could move below current long-end levels. Jin elsewhere puts the gap, absent long-end control, at 180-190bp: “That’s difficult.” The pressure on oil demand is “all coming from China,” where the economy is in severe deflation, and “has nothing whatsoever to do with you hiking rates.”

3. K-Shaped Economy: Rate Hikes Are Chemotherapy That Kills Healthy Cells

• The K has 2 dimensions: private versus public sectors, and AI versus non-AI. The private sector—mortgages, auto loans and opening a new Starbucks with a payback period of more than 10 years—is highly rate-sensitive; the government is not. During the Biden-era hiking cycle, fiscal policy actually became looser, partly offsetting monetary policy. This structural conflict is extremely difficult for US policymakers to resolve.
• Alibaba’s report provides the AI-side comparison: a compute center pays back in 2.2 years. “If something pays back in a little over 2 years, do you care whether the interest rate is 6.5% or 4.5%? You don’t.” But a new Starbucks cares whether the rate is 0.5% or 4.5%. Hence the “not entirely appropriate” but precise metaphor: “AI is a cancer. It absorbs more resources; if you go in with chemotherapy, you kill the other cells and make the cancer cells grow faster.”
• The transmission sequence makes rate hikes ineffective. Raising rates to 10% has zero impact on continued data-center construction. When the economy weakens, brick-and-mortar businesses first cut people rather than tokens; the hit to token demand comes later, “because tokens are cheaper than people.”

4. Warsh’s Fork: Fixating on Inflation Means Policy Error; Tough Talk and Soft Action Are More Likely

• Path A—actually hiking—is bound to fail. This inflation cycle is supply-driven and cannot be solved by rate hikes, while the K-shaped economy makes the damage highly uneven: “The government doesn’t get hurt, Bessent will borrow just the same, AI doesn’t care about borrowing costs, and the real economy gets hit 3 times as hard or more.” The long end may look temporarily contained—the 30-year has already fallen—but a genuine year-end hike would produce bear steepening, and the 10-year still could not be controlled next year. Jin also dismisses the “US debt is blowing up” argument: the dollar’s essence is having a lot of debt, and Britain won the Napoleonic Wars by building a government-bond market.
• Path B is more likely in Jin’s view: Warsh may have seen that the next payrolls report is likely to be negative again and that the economic data will deteriorate, giving him the confidence to talk tough while leaving rates unchanged in September. Since 1982, the Fed has never hiked after 2 consecutive negative payrolls reports. Even including AI, employment is already negative; excluding AI, it would be worse. Job openings, wage growth and retail sales are all weakening. Tax refunds provided an earlier lift, but once the $40T-$41T debt ceiling arrives at year-end, some fiscal funds will no longer be available.

5. The K-Shaped Valuation: New Productivity Is the Down Leg, Old Productivity the Up Leg

• Jin starts with Starbucks at 36x P/E and Nvidia at 12x P/E. He then contrasts a company with 74% gross margins, 75% growth and a 12x P/E against another with 15% gross margins, 15% growth and a 30x P/E, asking: “Is this a bubble? What kind of bubble is it?” New productivity is the downward side of the valuation K, while Coca-Cola, Walmart and Starbucks are on the upward side. “The market is extremely surreal and crazy.”
• At the index level, valuations are not expensive. The S&P’s 2026 forward P/E has been trending down from 22-24x to below 20x, around 20x, and could still reset to 2018 levels. If monetary policy becomes excessively tight, valuations could return to the levels seen during the era of outright money printing. The market is getting cheaper because earnings at the K’s upward-sloping companies are rising. The trade: “Stay away from slow-growth companies—the biggest valuation premium is slow growth.”

6. Positioning and Credit: Much Healthier Than June and July; the Real Pitfall Is Next Year

• Momentum positioning has improved. In June and July, the long leg was entirely semiconductors and highly concentrated; one round of liquidation could hit the entire AI complex. Today, investors remain structurally long AI but have turned tactically short AI, so even another liquidation event “couldn’t possibly have the same impact as in June and July.” On a 1-2 month horizon, system leverage is not high; on a 6-month horizon, it is low. Gross leverage is around its December level relative to the peak, while fundamental-book positioning—mutual funds, long-short hedge funds and others—has fallen to Liberation Day levels. Jin directly questions BofA’s report that the bull-bear index and positioning concentration are high: “I don’t know where that data came from.” At least Goldman’s data show positioning is nowhere near high.
• Credit markets are absorbing new issuance well. 博康—name transliterated from the original—says it needs to raise $60B, but its data center has not actually been built. Microsoft raised $20B at the beginning of the year, while Google issued $5.5B of debt in Australia; all of it came through investment grade, and the market continues to absorb the supply. HY remains at low levels. IG borrowing costs are generally rising, though they have been flat recently; as financing volumes increase, Jin expects them to rise again. The credit spreads of the 4 hyperscalers and their market impact need to be assessed by aggregate issuance, which may not truly frighten the market until next year. For now, investors can trade the expectation and position in CDS: “I bought CDS too, and I made money.” But the real problem may only arrive next year.
• Volatility confirms that there has been no forced deleveraging. Volatility in the AI basket has fallen from roughly 120 to 66, and the metric does not account for cross-asset correlation. After volatility is cut in half, vol-control momentum strategies mechanically add risk. Even after the Anduril news-driven trade, there was no deleveraging and no visible impact on equities, IG or HY.

7. Power and Permitting May Be AI’s Real Bottleneck: Capital Will Flow to Where the Power Is

• The core numbers from a Fangda report, attributed in the original transcription to “COVUS [?],” are stark: truly available compute is about 50GW, versus 35GW the US expects to build; only 11.4GW was added in the past 12 months, and just 11-14GW has both permits and power. “That number cannot be increased; no amount of money will help.” Of the 43GW of abandoned projects, 78% were killed by policy and only 1% was abandoned by the developers themselves—“because it’s very profitable.”
• Building private power grids has become the alternative. Developers are repurposing turbine generators from aircraft or using Bloom Energy systems that do not require permits. Jin digresses to say, “You still have to admire Nancy Pelosi; she understands exactly where the political pressure in this industry lies.” Large GE generators may require a 4-5-year wait, and GE’s production schedule is already booked through 2031. The 11-14GW that could come online next year all comes from applications filed in 2025; applications filed today may not deliver until after 2031. Power accounts for 60%-70% of current delays.
• The source of the public backlash is straightforward: compute centers and AI itself do not create many jobs, while voters get noise, higher electricity bills and disappearing white-collar jobs. Jin also says China is creating large numbers of accounts on TikTok, Xiaohongshu and X to encourage Americans to oppose compute centers. Ahead of the midterm elections, pressure from state governments could intensify. His metaphor: “The world will become 0.1% gods and 99.9% bugs. Are you the pig being raised, or are you going to use more compute to do smarter things?”
• 黄仁勋 declared that “China wins” the AI race because China can deploy power in a year, with fewer permits, community objections and approval hurdles. Previously, both the US and China were short of chips rather than electricity, and China could not outbid the US for chips. If China remains power-abundant while the US becomes power-constrained, 50GW of compute equipment could gradually be deployed overseas: “Wherever there is power, that’s where I am. Making tokens means making money.” That could reshape many aspects of the competitive landscape. Jin believes the economics of AI are difficult to derail, but whether policy opts for further “chemotherapy,” decoupling or development restrictions remains to be seen.