US Treasury Yields Near 5%: Why the AI Narrative Hasn't Broken Yet
US money is getting more expensive, yet AI is still inflating while every other sector deflates. The contradiction resolves either by AI genuinely breaking through and pulling rates down, or by AI being falsified — but no one dares slow down first.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Q3 markets: the exhaustion comes from having no direction
In Q3 the S&P rose 2.1%, the Nasdaq was roughly flat, Japan and Korea fell 8-point-something and 18.5 respectively, CSI 300 dropped 9.4, ChiNext fell 23 points, and the Hang Seng rose 8 points while Hang Seng Tech stayed very weak. Commodities were the bright spot: WTI and Brent both rose 40 points over the quarter, and gold gained 8.5. But the host says his own feeling is ‘a strong sense of exhaustion’ — unlike the first half, when hearing a narrative and watching markets move still made you think about what opportunity you had, whether to buy or sell, how to allocate. Now it's ‘tired, forget it.’ Ricky's summary: the first half was making money with your heart in your throat; this quarter was adjusting until you were completely drained.
— David WengThree forces pushing US Treasury yields up
A Clock Tower research note attributes the rise in US long-end yields to three forces. First, energy prices are rising and the Middle East stalemate keeps dragging on. Second, US nominal growth has resilience: Q2 real GDP grew about 1.5 and core PCE inflation was 3.4, which adds up to 5% nominal growth — compare the late 1990s, when the combined level of US Treasury yields was around 5%. Third, corporate bond issuance is expanding rapidly and AI infrastructure buildout keeps advancing, all against a backdrop of already-high US fiscal deficits. The remedies correspond to Middle East de-escalation, tighter macro policy, and a substantive slowdown in the AI investment cycle — but the latter two contradict each other.
— David WengBond yields are credit pricing, not a buy-sell mismatch
Ricky explains it by raising the dimension: a bond yield is fundamentally the cost of funds generated on the basis of credit. After the Soviet failure in the late 1980s and early 1990s, the US entered a long period of low rates, because it was the most powerful country in the world and its borrowing cost should be the lowest in the world. Since 2018 the US has entered the middle-to-late phase of its national fortune cycle, facing more and more doubts, its credit downgraded, and its borrowing cost must rise. In a two-dimensional world you see fewer buyers of Treasuries and more redeemers; in a three-dimensional world you see declining national power driving borrowing costs up. The US decision at the three-dimensional level is to keep pushing on AI, mobilising the whole country for AI buildout while also pushing manufacturing reshoring — so private capital and state capital compete for money, and everyone's borrowing cost rises.
— RickyWill AI infrastructure repeat the ROA trap of big infrastructure
Ricky compares US AI infrastructure to China's big infrastructure buildout starting in 2008: an economic entity dominated by CAPEX will inevitably see ROA decline — more and more assets, no breakthrough in production technology, so the return per unit of asset falls. When ROA drops below the borrowing cost, an asset crisis follows. The biggest worry now is that AI infrastructure will end up with ROA unable to cover borrowing costs, while borrowing costs are rising and ROA improvement is still not obvious. David Weng adds: what's different this round is that deficits weren't this high before — past cycles didn't casually run into trillions the way things do now, and the current magnitude has gone beyond the range of human comprehension.
— RickyThe prisoner's dilemma behind the calls to slow AI down
This week Dario wrote "We must pace the frontier," Musk reposted it saying he agreed, and Sam Altman said in an interview that things should slow down — even saying OpenAI isn't in a hurry to IPO this year. A Reuters piece on September 9 was titled ‘Researchers want AI to slow down, but labs can't afford to.’ Ricky sees three layers. First, right and wrong in the two-dimensional world cannot be overturned in the three-dimensional world — Asimov's Three Laws saw it a hundred years ago, and you can't stop the disaster before it happens. Second, a unified alliance of three or more players can't be achieved; the prisoner's dilemma ends with everyone choosing the worst outcome, and it needs an external force — the government — but the government only abandons its own problems once great-power competition produces enough damage and disaster. Third, the funding side cannot tolerate such a long verification period: if within a year large models haven't proven they have a sufficient application and commercial environment, a lot of money will see runs and even broken funding chains, at a magnitude that would certainly be countless times the 2008 subprime crisis.
— RickyLarge-model revenue $150bn against $200bn of investment
Ricky does the math: across large models, revenue is $150 billion and investment is $200 billion. Split into frontier and open-source models: frontier models bring in about $120 billion in revenue on $100 billion of investment; open-source models take $100 billion of investment and bring in $30 billion. Frontier model revenue alone is only so much, and open-source models are still a long way from actually generating revenue to cover investment — not even profit covering investment. So whether these companies can still post big profit increases next year is under heavy pressure. More critical is the second layer: even if corporate earnings still rise at a low-double-digit pace next year, will the capital market accept it? Because the second derivative is actually heading down, and valuations are set on the second derivative, not on earnings growth. If you can't count on white-collar replacement, the math doesn't work — and then it's definitely expensive.
— RickyMidterms: Trump losing the House is highly likely
From BCA's research: over more than a century of US history, every midterm where performance on livelihood issues and inflation was poor, the probability of losing the House and Senate was almost 100%. The path laid out on the show early this year was that by June, if the US-Iran issue were resolved, energy costs brought under control, and Trump added some fiscal policy support, there was hope of holding the Senate. But neither of those two problems has been resolved, so judging by the polls, Trump losing the House is very likely — many swing states and minorities, such as Hispanics who were originally Trump supporters, are starting to switch to the Democrats. Trump holding the Senate should be the high-probability outcome. If he only keeps one chamber of Congress, he becomes a lame duck.
— RickyA-shares get a Q4 rally, but don't play it too far left
Ricky thinks A-shares can still look forward to a rally in Q4, but it shouldn't be played too far to the left; watch capital flows as the main signal, and if there's a short-term amplification in volume, it's worth participating. After that participation, switch to full defence and wait for the next, larger-magnitude breakthrough — for example, something clear on white-collar replacement, or clearer investment next year in the Huawei supply chain and the DeepSeek supply chain. If this rebound comes, it will certainly be led by tech, because many sectors have been tried without effect, and only tech can drive the market up. In Q3 most fund managers didn't do any work — they were busy writing self-criticisms — so in Q4 they'll have to work no matter what. David Weng is more conservative: cash still comes first, second is dividend low-volatility, third is a narrower slice of tech such as STAR 50, with a position of maybe two or three points.
— RickyIn their own words · checked verbatim
Even though the market really did swing around, I don't know about Ricky, but I myself have a strong sense of exhaustion — it's not like the first half, when you heard some narrative and saw the market move and you'd still think, so what opportunity do I have, do I buy or sell, how do I allocate. Now it just feels like, tired, forget it, that's how it is, that feeling.
虽然一方面市场确实跌宕起子，但是我不知道Ricky怎么样，我自己是有一种很强的疲惫感，就是，就不像上半年，你在听到一些叙事，然后看到一些市场的涨跌的时候，你还会想，那我有个什么机会，我到底是买还是卖，我要怎么去配纸，就现在就感觉累了算了，就这样了，就这种感觉。
David Weng5:03
A bond yield is just the borrowing cost of a bond. All bonds, whether secured or unsecured, are fundamentally a cost of funds generated on the basis of credit. For example, if David Weng's credit is better than mine, then his cost of borrowing must be lower than mine — that's the core.
债券利率就是债券的借贷成本，所有的债券，无论是你有抵押还是无抵押的，它本身其实是一个基于信用产生的资金成本，就是比如说大卫翁的信用比我好，那他借钱的成本一定比我低，这是核心。
Ricky22:17
Figures
| S&P 500 Q3 gain | 2.1% | 2:00 |
| CSI 300 Q3 decline | 9.4% | 2:00 |
| ChiNext Q3 decline | 23 points | 3:00 |
| WTI and Brent Q3 gain | 40 points | 3:00 |
| US 10-year Treasury yield | near 5%, reaching 4.8 the week of recording | 3:00 |
| Japan 10-year Treasury yield | 3% | 4:02 |
| China 10-year Treasury yield | 1.7 points | 4:02 |
| US Q2 real GDP growth | about 1.5 | 18:14 |
| US Q2 core PCE inflation | 3.4 | 18:14 |
Glossary
- CAPEX
- Spending by a company to acquire, maintain or upgrade physical assets.
- ROA
- Net income divided by total assets; measures how much profit each unit of assets generates.
- carry trade
- Borrowing in a low-interest-rate currency to invest in higher-yielding assets and earn the spread.
- hyperscaler
- A tech giant that owns large-scale data centres and cloud infrastructure.
- second derivative
- Here it means the rate of change of earnings growth; valuations are often set on this rather than on absolute growth.
How to listen
Investors and founders watching global macro and asset allocation, and anyone trying to understand the contradiction between rising US Treasury yields and the AI narrative.
The Q4 asset ranking segment after 1:35:32 — the conclusions are fairly casual and can be skipped.