Excess Returns Sep 20, 2026 55m 37m saved
With Jason Hsu, Founder and Chief Investment Officer at Rayliant Global Advisors
Rayliant's portfolios are built from more than 200 factors. The index most Americans hold, on Jason Hsu's reading, is built from about seven stocks.
The standard defense of passive investing is that the index is the diversified option. Hsu's position is the reverse: an S&P 500 holder is making one concentrated bet on the AI build-out, and is making it without having decided to.
"The S&P has 500 stocks, but really it's got seven stocks."
Jason Hsu, Founder and Chief Investment Officer at Rayliant Global Advisors, on Excess Returns, co-founded Research Affiliates with Rob Arnott, where factor research became investable products, and has spent the last 15 years building strategies for Chinese markets. It was his third appearance on the show, and the first since he put China's AI models at 20% behind the US.
The full interview is covered here so you can skip it. 55 minutes of audio, 18 minutes of reading.
Here are the 12 arguments that matter.
Key Takeaways
Chinese models went from 20% behind to level with the US in a year, and he now assumes parity on model development
His argument against US-only AI regulation: it stunts US firms and changes nothing about what China ships
China's advantage is electricity, not code — multi-source supply, the world's largest solar build and a national grid against 50 state ones
Most of the value will not sit in the models at all, because open weights are close enough that buyers stop paying for closed ones
The rent goes to hardware for about 10 years, then to whoever integrates the cheap models into products
Export controls created the competitor they were meant to prevent: "We've given them the necessity to invent"
The AI capex arms race is spending that does not earn, because each company's spending cancels the next one's, though none of them is close to a solvency problem
The S&P 500 is "predicated on the bubble not bursting", which makes a passive holder an unwitting active bet
Momentum crashes are the price of the factor working, and the crash is what keeps the trade from getting too crowded
Inefficient markets are not easy money — the mispricing gets worse before it corrects, and staying solvent through it is the hard part
1. The Gap Is a Hair
The episode opened on the Anthropic chief executive's weekend letter and the argument that AI development should slow down. Hsu's answer was that the risks are real and the regulatory remedy does not exist.
Safety is a real problem; national regulation is not the answer to it
So I think the genie's out of the bottle. Yes, we need to think about safety, but that then becomes a global coordination. And if the G2 countries, meaning US and China, don't play ball here, then it doesn't work.
Jason Hsu
Regulating US companies is possible, he said; regulating Chinese ones is not, and Chinese firms are no longer far enough behind for a US slowdown to be safe. A US pause would simply let them pass.
The host put the measured gap to him: Hsu had said on an earlier appearance that DeepSeek was about 80% of the way there at a twentieth of the cost, and Stanford has since put the gap at under 5%.
A year took China from 20% behind to a hair's length
They've gone from maybe 20% behind last year to now kind of just a, hair length away. I think, yeah, we should just assume today that they are on par with the US when it comes to model development, and perhaps, have an edge when it comes to kind of real world model.
Jason Hsu
His explanation for why this was predictable is the recruiting pool. The top research teams at US labs are already full of Chinese and Indian scientists, so the assumption that the same talent could not do the same work at home was never a safe one.
2. Who Builds China's AI
Asked to name the Chinese companies doing the work, Hsu started with the one exception.
DeepSeek came out of a hedge fund with spare compute
So I would say other than DeepSeek, which, kind of the funny story is that DeepSeek actually was, first of all, a high-frequency hedge fund, right? And they just had so much money, so much compute power. This was like a side project
Jason Hsu
Everything else, he said, sits inside the largest Chinese technology companies, with the cash flow and the scientists to match, which makes the structure a mirror of the US rather than a departure from it.
The follow-up question was whether China would sign up to a mutual model-inspection regime of the kind Elon Musk had floated in an interview the host watched that morning. Hsu's answer was that China has less to lose from it than the US does.
Open weights already put Chinese models under world inspection
Well, right now if you look at the China models, right? They are open source, meaning any and everyone can look at it and check it. So they already open themselves to world's inspection. Now, that's because they're not trying to sell software, they're trying to sell hardware, right?
Jason Hsu
An inspection agreement would give China a look at closed US models it cannot see today, so he expects it would be welcomed. He also pushed back on the idea that Chinese labs are indifferent to safety: Beijing's fear is that a more capable US system could reach sensitive Chinese data centers, which he thinks puts AI security higher on its agenda than on that of US companies.
3. Energy Is the Real Edge
The host raised the power constraint on US data centers, which has become a political problem as well as an engineering one.
Hsu's answer was that this is the one area where central planning has paid off. Decades of war-gaming an embargo left China with supply from Russia, the Middle East, Oman outside the Strait of Hormuz, and land pipelines into Central Asia, on top of the world's largest solar build and its nuclear fleet.
Diversified supply means a price shock changes the cost, not the capacity
So they have gone multi-source, well-diversified, and they can turn on a slightly more expensive resource if oil isn't flowing.
Jason Hsu
The second advantage is delivery. Cheap solar generated in Xinjiang moves across the country on a high-efficiency grid, while the US runs decentralized, state-specific and old networks that nobody is incentivized to fix. The host's own example was power lines strung under oak trees in Connecticut.
The constraint on AI turned out to be power, not software
And who thought, like, in the world of AI, it's not ultimately technology that's the difference maker, it's actually the power that drives the technology.
Jason Hsu
That, on his reading, is why the Chinese strategy is to give the technology away and charge for the hardware and the electricity underneath it.
4. The Models Converge
Asked where value accrues between frontier and open-source models, Hsu gave the answer that runs through the rest of the interview.
In equilibrium the model layer is not where the money is
I would say in equilibrium, most of the value is not gonna be in the models.
Jason Hsu
The differences that remain, he said, will be matters of taste rather than capability, and buyers will notice: a developer who can inspect and deploy an open model will not pay a license fee for a closed one that does the same job. What is left is an application business that picks whichever model is cheapest and wires it into a product.
The end user is buying energy, not intelligence
You just know they're using the cheapest model, and you're really just paying for the energy token, right?
Jason Hsu
The host had the same observation from his own work. Excess Returns uses AI to write episode titles, which is a harder language problem than it sounds, and for a long stretch ChatGPT and Claude were clearly better at it than anything else. Kimi K3 now matches them.
Past a certain quality, the last 5% does not sell
Like, a difference in 5% just doesn't matter, right? Like, we all hire people, right? Like, when you have top graduates who are smart with experience, like, a 5 percentage of difference in IQ is, like, indiscernible, and you're not gonna fuss over it, especially when the price difference is large.
Jason Hsu
5. Hardware Gets 10 Years
The host's objection was historical: in the internet build-out the value did not stay with the people who built the infrastructure. Hsu agreed about the destination and disagreed about the timing.
The sequence he described starts with DRAM, GPUs, server builders and data centers, and they capture the rent because supply cannot answer demand quickly. A data center is billions of dollars and a five-year project, down to waiting on generators.
Hardware holds pricing power until everyone's capacity arrives at once
But there's gonna be probably a good 10-year period where hardware is going to capture most of the rent.
Jason Hsu
After the overcapacity arrives and infrastructure gets cheap, the application companies build on top of it and take the next phase. The host added the point that this cycle may run longer than earlier ones because the industry is not even meeting current demand, let alone building ahead of it. Hsu's reply was three words: "No, not even close."
6. Export Controls Backfired
Asked for an update on chip restrictions, Hsu treated the policy question as settled and the consequence as the story.
China is not close to Nvidia's GPUs or to TSMC's ability to fabricate at the top node. What changed is the rate of closure, which he expects to follow the model timeline: a 20% gap becoming a 5% gap, then parity, because the engineering talent is globally available to anyone who pays for it.
The restrictions produced the domestic capability they were meant to delay
So, yeah, I think we have basically encouraged China to become self-sufficient in areas where the US and the US allies dominate and China couldn't be bothered because they could simply pay and buy it, right?
Jason Hsu
The policy supplied the motive
We've given them the necessity to invent.
Jason Hsu
7. Retail Alpha in China
Asked to explain the Chinese market's last two years, Hsu described a 2025 in which every catalyst fired at once: the real-estate fear lifting, the DeepSeek moment forcing a revaluation of China's position in AI, and earnings growing off a low base. This year is a wait-and-see market, held up by the same questions US investors are asking about the Middle East, the Fed and whether AI capital spending is a bubble, with a more volatile reaction to each of them.
What makes it a working ground for factor strategies is who owns it.
Ownership is 80 to 90% retail, and trading is about half
But because you have so many retails, right, short-term overreaction, underreaction, long-term mispricings are abundant.
Jason Hsu
Alpha has to come from somebody, he said, and in China it is coming from uninformed trading rather than from other professionals. Other inefficient markets exist, India and Latin America among them, but they are too small to deploy into. China is one of the few that is both large and retail-dominated, alongside Korea and Taiwan at an order of magnitude less size.
Most of the institutional money that has gone in is high-frequency, hunting intraday momentum and reversal with machine-learning models on price and volume. Long-horizon mispricing, he said, is still largely untouched.
Korea came up as the extreme case, and the difference is regulatory.
Korean retail runs on margin at a level he has not seen elsewhere
But in terms of utilization of margin, South Korea is an order of magnitude more aggressive than any other market I've studied.
Jason Hsu
When Chinese regulators see speculative excess they tell brokers to cut margin. The Korean president, by contrast, publicly encouraged investors to go all in on Samsung and SK Hynix. The host noted the two were about 50% of that market at one point.
A two-stock index is not an efficient index
Well, like, if you look at the Korean market index, it's got two stocks in it, right? And they do exactly the same thing.
Jason Hsu
Hard to beat, in other words, is not the same as efficient. It is hard to beat because it is two correlated stocks in a technology boom.
8. A Barbell Market
Asked what the Chinese market is actually made of, Hsu split it in two, with little in the middle that matters. At one end are the technology firms built by founders he compared to Musk, Gates and Jobs. At the other are state-owned enterprises: banks, utilities, power, alcohol and tobacco, operating in regulated markets at enormous scale.
Nobody buys China for dividends, which is why the dividends are there
So you really have some of incredibly high dividend paying value stocks in the state-owned.
Jason Hsu
Foreign money goes to China for growth, so the state-owned half trades on low price-to-earnings multiples and high yields, which he described as close to a fixed-income substitute sitting inside an equity allocation. The two ends together give an investor a barbell that the US market, in his view, no longer offers.
He also answered a question about status: engineers and founders are admired in China without needing rehabilitation. The US, he pointed out, needed a television show, "The Big Bang Theory," to make being a nerd respectable, while in China the student with the best exam results was always the one with standing.
9. An Unhappy Marriage
On whether US-China cooperation improves from here, Hsu gave a macro answer rather than a diplomatic one. US government spending runs above 30% of GDP, more than 80% of it welfare, financed by printing, which is inflationary. China is the world's manufacturing base and gets more productive every year, which is deflationary. The trade exchanges one for the other.
Both sides got rich on the arrangement, and the US got the better of it
Our national wealth, right, has gone from $17 trillion from when we started trading with China to now $170 trillion, right?
Jason Hsu
The US bought cheap, high-quality goods and sold them on to the world with its own brands, marketing and distribution, he said, which is why the idea that China captured all the gains does not survive the arithmetic. The structural need, on both sides, is still there.
Neither can leave, and neither wants to stay
It's almost like a, an unhappy marriage where you can't divorce each other, but you don't like each other. That's what we're gonna have.
Jason Hsu
10. The S&P Is 7 Stocks
The host brought the AI story back to the factor evidence: companies that spend heavily on capital investment historically underperform, though a recent guest had argued the spending should be judged against the size of the firms doing it.
Hsu split the spending into a healthy half and an unhealthy one. The unhealthy half is an arms race, where each company spends because the others are spending, and every additional dollar earns less.
Spending that only exists because a competitor is spending
And so there is a lot of capex expenditure today that isn't bringing more money. It's because someone else is overspending, right?
Jason Hsu
His software analogy was the office-suite era, when every product got more feature-rich and none of them could charge more for it. What he does not expect is financial distress. These companies generate enormous cash flow and are borrowing against future cash flow at attractive rates rather than out of need.
Nvidia's $20B of debt is a choice, not a constraint
They're going out and floating $20 billion debt because people know they're so profitable, they're good for it, right?
Jason Hsu
The risk he does flag is who is holding the position without knowing it. An active investor making an AI bet has chosen the bet; a passive one has been told the index is diversification.
The index is a position on the bubble continuing
Then you look at your S&P 500. It's only that, right? It is predicated on the bubble not bursting.
Jason Hsu
His instruction to passive investors
So I would say passive investors, this is time to be un-passive and start to look at your portfolio
Jason Hsu
The underlying point is that market beta is not a fixed thing to own.
Market beta has been a different bet in every decade
Like, 20 years ago, yeah, 2000-ish, right? The market beta was just internet. Today, the market beta is just AI.
Jason Hsu
Before the financial crisis it was real estate and financials. The remedy he offers is a deliberate multi-factor portfolio holding low volatility, value and quality, so that more than one idea is represented, in the way the Chinese barbell does.
11. Momentum Always Crashes
Asked about the momentum crash that came with the semiconductor reversal, Hsu described the shape of the factor rather than the event. Momentum produces steadily and then breaks, so its returns are skewed rather than normally distributed, and the attempts to time the break do not work.
The crash is the reason the trade keeps working
So the crash just, it's, part of what keeps the momentum trade from getting too crowded, right? It crashes, reset, and then that trade's available again.
Jason Hsu
The pairing he recommends is value, which he says does not cancel momentum out but runs on a different cycle.
Value tends to pay out exactly when momentum breaks
'Cause usually when you have a momentum crash, you'll tend to see a really, really strong pop up in the value factor.
Jason Hsu
Value, in his description, has a fat right tail: long stretches of nothing and then a year or two of a lot, which is the mirror image of momentum's profile. The host added the measured version — momentum beats value on consistency over one, three and five years, and value's weak periods run longer.
On whether factors still work in the US at all, Hsu's answer was that the question is usually asked about too few factors. Standard packages report value, size, momentum and some quality.
Value and small cap were taught to everyone, and the premium went
And so I would say, yeah, so value and small cap probably doesn't work in the US, right? It's just sort of cyclical.
Jason Hsu
What survives is what is not in the textbook
But stuff that aren't taught in the US, I would say, those are still, I would say, reliable and still things I would count on to work.
Jason Hsu
The reason Rayliant can use a much wider set is computational. The statistical methods are 30 years old, but the machines to validate them are not: hyperparameter tuning, which tests whether in-sample results survive out of sample, used to take a month per run and now takes hours.
200-plus factors, with the overfitting check as the real work
We use 200-plus factors. We have machine learning to figure out different ways to piece them together to look at nonlinear behaviors.
Jason Hsu
The portfolios that come out hold fewer names than the benchmark, scored stock by stock and combined so the resulting factor exposure is diversified, with little weight on value in the US because it has not paid there.
The index's long tail is friction, not diversification
The S&P has 500 stocks, but really it's got seven stocks. So it's got, like, 400-plus stocks just creating transactions costs and round lot issues because, they don't have any weight.
Jason Hsu
Asked whether the outperformance comes from avoiding bad stocks or owning better ones, he said both, and in different market conditions: avoiding overpriced story stocks is what protects the portfolio when the market turns rational, while enough quality-screened growth is what keeps it in the race while the market is running.
12. Stop Picking Stocks
The last stretch was about what financial advisers should do, and Hsu's answer was to stop competing where they cannot win. His list of counterparties was Citadel, Jane Street, Two Sigma and fund managers with 30 or 40 years of cycles behind them.
The competition is not the client
And if you're not there, you don't have an edge, right? This is not the poker table you should be at.
Jason Hsu
The edge advisers do have is the relationship: trust, and the ability to keep a client from reacting to noise. He was explicit that treating investing as entertainment with clients creates anxiety rather than returns.
His illustration was a dinner with Jack Bogle, who had been at an event full of billionaires and hedge fund managers, none of whom ran anything the size of Vanguard and all of whom earned many times what he did. What Bogle said he had that they did not was enough.
The job is to make a number feel like enough
if they can get a client to feel that's enough, that is more value you can create for a client, right?
Jason Hsu
Asked the show's closing question, which is the belief his peers would disagree with, he named one he used to hold himself.
Inefficient markets have more alpha and are harder to harvest
So don't think of inefficient markets as sort of easy picking when it comes to alpha. It's hard to fight against a very irrational market.
Jason Hsu
Fifteen years of running money in China taught him that a mispricing usually gets worse before it corrects, and that the dry powder set aside to double down is rarely enough. The opportunity is real; surviving the wait is the part nobody prices.
Bonus Insights
China's AI security worry is the mirror of America's
Every reported incident of a US model misbehaving is read in Beijing as evidence of capability, he said, and the fear is an American system reaching sensitive Chinese data centers.
The host's own AI usage is the evidence for the convergence argument
Jack Forehand writes the show's episode titles with whichever model does it best, and said the ranking has flattened out in the past few months.
Jensen Huang made the same split at the All-In Summit
The host referenced Huang's argument the night before that closed models and open models will both matter, with developers largely on the open side.
Nobody is building ahead of demand this time
On the host's observation that earlier technology cycles built in advance of demand while this one is not even meeting it, Hsu's whole answer was that the industry is not close.
Hsu's bottom line is that the AI trade and the China trade are the same trade viewed from two ends: the models are converging on free, the rent moves to hardware and then to energy and integration, and an investor who owns the S&P 500 and calls it diversified is holding the most concentrated version of that bet available.
Products, Companies & Tools Mentioned
Rayliant Global Advisors (Hsu's firm. Runs multi-factor ETFs built from more than 200 factors, with machine learning used to test for overfitting)
Research Affiliates (The firm he co-founded with Rob Arnott, where factor-based approaches were turned into investable products)
DeepSeek (Started as a side project inside a high-frequency hedge fund with spare compute; the model he previously put at 80% of US capability at a twentieth of the cost)
Moonshot AI's Kimi (The open model the host says now matches the closed ones at writing the show's episode titles)
ChatGPT and Claude (The two closed models that had been clearly ahead on that task until recently)
Nvidia (The GPU supplier China cannot yet match, and the example of a cash-rich company floating $20B of debt to pull future cash flow forward)
TSMC (The foundry capability China is still far from replicating, and the second target of export restrictions)
Samsung and SK Hynix (The two stocks the Korean president urged investors into, and roughly half of that market's index)
Vanguard (Jack Bogle's firm, the setting for the "enough" story he tells about him)
Citadel, Jane Street and Two Sigma (The counterparties he tells advisers they would be trading against if they pick stocks)
Caterpillar (His shorthand for the generator bottleneck that makes data-center capacity a five-year project)
Books & Resources Mentioned
The Big Bang Theory (His example of what the US needed to make technical work socially respectable, against a Chinese culture where it already was)
Excess Returns' full transcript of the episode (The show publishes the whole conversation in text)
Listen to the full episode
🔴 YouTube
Watch the full episode:
More Jason Hsu interviews and podcast appearances
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: