AI and Robots Will Change the World. Who Pays for It?

“Whichever ending you get, long-end rates fall.” — Mark Tang, on the chart where both roads lead to the same place

Guest: Mark Tang｜co-founder of Caidazi AI, host of the podcast Fācái Dāzi

Mark Tang had just come down from the Alps. He spent a week in Chamonix at the end of August running the UTMB trail-running world final. The infrastructure there is absurdly good — there is signal on the summit — but he still went feral for a week and then spent the week after that catching up.

What stopped him was the landing. Flying into Geneva, the airport was wall-to-wall Nike, Brooks, adidas and Hoka. Flying home, the jet bridge — an absurdly long one — was papered end to end with ads for AI office assistants. Same person, one week apart, two different worlds.

This is the monthly review. Markets crawled back in August from July’s selloff and stopped falling, but the mood never came back with them — what Raymond calls “a meat-grinder tape that goes back and forth.” Nvidia’s earnings landed like a sedative, and only a sedative. Meanwhile the bubble argument reignited in the US: Nvidia pulled Wall Street’s biggest institutions into a syndicate to fund data centers, Michael Burry came out saying tech-company depreciation is fake, Chinese open-source models pushed prices down to a line nobody can survive under, and the long end of the Treasury curve is at its highest since 2007.

This episode takes those apart: where Jensen Huang’s residual-value guarantee holds and where it breaks; why an eight-year-old A100 is still going up in price; how many of a thousand robotics startups actually live; when exactly “distillation” turned into a dirty word; and the chart nobody enjoys — that whether AI succeeds or fails, long-end rates end up falling either way.

What follows is the full conversation, edited and condensed.

1. Nvidia’s earnings were a sedative, not a stimulant

Raymond: August is done. Markets recovered from July’s big drawdown and stopped going down, but plenty of investors are still uncomfortable — or maybe they aren’t, maybe they just took August off. You spent a week in the mountains. What did catching up feel like?

Mark Tang: I did a review back in July, when the whole global AI hardware chain sold off hard. In August sentiment recovered, but people are still carrying that fear around — bubble, hesitation, worry. That kind of feeling doesn’t disappear quickly.

The tape came back a little; the mood didn’t. Then Nvidia’s earnings set the rhythm, and it was like injecting a fresh sedative into the market.

Raymond: But it’s a sedative, not a stimulant. Definitely not a stimulant, and the effect wasn’t even obvious — it popped after the print and then fell straight back. It wasn’t a sustained move.

Mark Tang: You can clearly see that the CEOs of these companies, the ones whose share prices are still elevated, are anxious. Everyone is trying to manufacture a new narrative, or hand out higher guidance, to keep sentiment going.

Raymond: Nvidia’s numbers were genuinely good, and next year’s guidance was good too — that isn’t in dispute. It still looks like a large, fairly stable company. That’s as far as I’ll go; this is definitely not investment advice.

But some of its recent moves are unusual, and two things are being conflated into one. One is that it pulled a lot of people in to lend money. The other is its own circular financing.

The lending part is Nvidia going to the very large Wall Street institutions — BlackRock, Brookfield, Goldman Sachs, KKR — and putting a financing together with them. In plain terms: I bring a crowd of rich backers along to invest in data centers with me. It turns the data center into something very close to a building — bundle them all together, then go to the backers for money. From the sound of it plenty of people are willing.

Mark Tang: The demand really is there. We watch it every month, sometimes every week, including token-usage data on OpenRouter — all of it is still surging.

Raymond: My first reaction was surprise that so many institutions were willing to back it, though of course the ones that declined don’t come out and say so. And then there’s Nvidia providing the guarantee.

2. “When one company defaults, it’s usually the moment the whole industry is in trouble”

Mark Tang: Jensen has discussed this publicly before. His argument is residual value: if a company goes bankrupt or defaults, the chips can be resold to another data center as residual. His example is the eight-year-old A100, which just signed another five-year lease — the fixed asset should have been fully depreciated, and instead its life keeps extending.

Raymond: Two things to add.

First, I saw a statistic somewhere saying that oversupply of just three to five percent is enough to break the entire market’s pricing. I don’t know what it was actually based on, but the direction is right, because pricing is set by the marginal seller.

Mark Tang: It runs the same way in reverse. Undersupply works identically — one extra bidder and the price doubles. Memory is exactly that; the gap isn’t large and the price is up.

Raymond: Right, a small gap and they can ask for double, triple or more, because people need it badly.

The second point is that you have to separate the single case from the system. A single bubble bursting and a systemic problem are different things. Jensen’s residual argument is entirely coherent at the single-case level: one company defaults, you take the chips back, re-lease them, sell them secondhand, and you’re covered. But when one company defaults, it is very likely the moment the whole industry is in trouble. At that point chip residual values are falling hard too, and equity and credit sell off together.

Mark Tang: Like Chinese property over the past few years. That gets very ugly, and you can’t hold it back — it isn’t one building marked down, it’s the dam giving way and the flood running straight through.

3. The short-seller earning a hundred million a year writing emails

Raymond: The loudest voice behind all this is Michael Burry. Let me introduce him briefly.

He’s the original in The Big Short — the fund manager who worked out subprime would break before 2008 and shorted it successfully. But across the last few decades he has spent far more time bearish than bullish, and apart from that one big call in 2008, how many of the others were actually right is an open question. He’s been saying the AI bubble is about to pop for two years now, and started shorting Nvidia and Palantir something over a year ago — I’ve forgotten the exact timeline.

He has actually deregistered his fund; AUM is around 150 million dollars. What actually makes money is his Substack, Cassandra Unchained — over 300,000 paying subscribers at a bit over three hundred, close to four hundred dollars a year. He makes a hundred million dollars a year writing that newsletter.

I’m genuinely envious of someone who can write a hundred million dollars’ worth of email.

His indictment has two parts. First, circular financing, which he thinks eventually breaks. Second, and harsher: he broadly believes every tech company’s depreciation is fake.

His logic is that a chip is a consumer electronic, not a means of production. Consumer electronics should be written off fast, like an iPhone, in two or three years. These companies are amortizing over five or six, which artificially pushes cost into the future and inflates profit. So he says the profits are all fake.

4. Why an eight-year-old A100 is still going up in price

Raymond: There’s a fact that contradicts him.

A100 demand is expanding — and it’s the expanding demand that lets the price rise and hold, which I find genuinely remarkable. An eight-year-old card, with rental and purchase prices both firm and liquidity that’s good. A consumer electronic that genuinely deserved to be written off would not look like this.

Mark Tang: Because it isn’t one batch of cards being used for one thing.

Compute demand used to be driven mostly by training, pre-training especially. But an increasing share of token consumption has shifted from training to inference at the point of use, and inference doesn’t need top-end cards. The top cards eat pre-training; the older cards drop down into the lower tiers of inference. When you use Codex or Claude and there’s a fast tier and an ultra-fast tier, those tiers are different silicon underneath.

Raymond: It’s easy to understand. If I want to ask whether you put salt or sugar in tomato and eggs, I don’t need the most expensive model — anyone can answer that. The customer base is tiered too: AI for science, materials science, that whole set doesn’t need heavy compute, and neither do local deployments. So the A100 isn’t obsolete and unwanted; it descended into a different customer group.

Mark Tang: So there are now probably two layers of routing.

The first is at the model layer: different tasks, or even different sub-tasks inside one agent run, get routed by difficulty to different models. The second is at the silicon layer: the same model may run on completely different cards in China versus the US, because electricity prices and total cost of ownership are completely different in the two places.

Electricity is the raw material for producing tokens.

Raymond: A more concrete version: Amazon serves its customers with one set of models, and Volcano Engine or Alibaba Cloud serves Chinese agent customers with another — same capability, entirely different arithmetic on chips and power.

Mark Tang: And a lot of this never shows up fully in the data. The single most important thing about researching AI right now is that you have to use AI yourself, get your hands on the products, and feel what’s actually changed versus a year ago.

Raymond: So a trip around Silicon Valley should tell me.

5. If the bubble is going to break, which indicator moves first?

Raymond: Suppose they’re right and this is a bubble. Which indicator shows it earliest?

Mark Tang: What worries me most is round-tripping inside the chain.

The circular financing we discussed, and AI applications generally, come with a lot of unwritten practices — wash trading, people buying from each other. I buy your service, you buy mine, both receivables go up, both sets of numbers look great, and neither of us made money. We may even owe extra tax. That loop can sustain itself inside the chain for quite a while.

So what you have to watch is demand from outside the chain. If demand outside the chain doesn’t genuinely rise, nothing inside it matters — it’s just an internal loop.

Raymond: But those indicators lag badly. With a startup like OpenAI you simply can’t see it; with a hyperscaler, high CapEx is normal, reporting is quarterly, it arrives another two months after that, and they have every incentive to sugarcoat.

Mark Tang: On the financial side you can also watch operating cash flow, plus inventory and receivables — those matter a lot.

Raymond: But all of it is hard to read and currently very murky — the numbers can be made to say whatever. By the time you can see it in the filings, it’s very late.

So I think of two indicators that might be earlier.

The first is the marginal move in credit default swaps on CoreWeave and Oracle. A CDS price is the market’s real-time pricing of whether they can pay, and it’s far faster than a filing.

The second is A100 volume and price — I’d guess that could be a decent leading indicator. Not the GB300: that’s hard currency, Chinese buyers will say send it over and I’ll take all of it, so there’s no information in the price. But if A100s start piling up in the US, do Chinese buyers want them? I don’t think they do. That makes the A100 a more marginal, more sensitive asset.

As for circular financing itself, I’m less worried. Somebody asked Jensen Huang on CNBC: why is this circular — meaning, does it still work if you don’t hand him the money? The reality is that if you don’t, he’ll borrow from somebody else and buy the cards anyway. Buyers in China can’t get cards and want them badly. The real demand is there.

6. “Don’t assume that because you use Doubao, we’ve stopped fighting”

Raymond: One more thing worth mentioning. David Sacks said something on the All-In podcast that I strongly agreed with: that AI is fundamentally an arms race between the US and China, at the level of nuclear weapons.

If it’s at that level — you’re researching the nuclear weapons of a new era, so shouldn’t the Defense Department be paying? What’s happening instead is that private companies cover it out of their own free cash flow, and when they can’t, they borrow. That’s an awkward way to run it.

Don’t assume that because you use Doubao, we’ve stopped fighting.

So I keep thinking that the biggest defense item in America right now should be getting interest rates down. Get rates down and the hyperscalers don’t have to borrow so painfully. You can’t say this is a national-security-grade race and simultaneously leave the long end at its highest since 2007. Those two things contradict each other.

Mark Tang: The All-In crowd has been consistently vocal about wanting government to push this along.

7. Robots: the endgame, and fake it until you make it

Raymond: Have you been watching Unitree?

Mark Tang: Hasn’t it already halved?

On robots, I approach this from the endgame.

If robots genuinely work, production efficiency approaches unlimited, and the consumption side changes — you get whatever you want. So it may not be UBI, universal basic income, but UHI: universal high income. Which admittedly sounds like science fiction.

Raymond: What I care about is how much of that GDP increase comes from printing money and how much from real productivity. It only counts if the GDP lift is real.

Mark Tang: What are robots actually doing right now? Holding sports meets — performing martial arts, slicing fruit, throwing their backs out running. The things that actually go into factories are robot arms, and an arm is more like an ASIC than a GPU: it does one thing, but it does that thing well.

So this stage does need fake it until you make it. You have to build something and show it, even manufacture some of the traction — genuinely fake it until you make it.

Raymond: I think I see it differently from you.

The sports-meet stuff isn’t fake it, it’s marketing — it’s part of marketing.

Mark Tang: I remember All-In making this point too: Americans will play up how terrifying robots are, and China does the opposite — back the pillar industry first, then get ordinary people gradually comfortable with it.

Raymond: So you mean the Chinese government is building public affinity for robots. Which means that at the level of top-down design, China is friendlier to robots and the US and Europe are less friendly, and public attitudes end up reflecting that.

8. The dishwasher paradox: how many of a thousand robotics startups survive?

Raymond: On acceptance — I posted one line on Jike, the social app, a couple of days ago. I said: an enormous number of Chinese people don’t use a dishwasher, so how are they ever going to accept a domestic robot walking into their home and doing chores?

That one line got several hundred comments and several hundred likes, which on Jike is trending-level. The comments were uniformly negative.

The first group said dishwashers don’t clean well or fast enough, that you have to pre-rinse, that you have to load it — might as well wash by hand. A lot of people take pride in washing dishes better than a machine does.

The second group said who says Chinese people don’t use dishwashers, we love them. One example was wonderful: he’d use a dishwasher if it could wash vegetables and fruit. That was the first I’d heard that a Chinese manufacturer had added a produce-washing mode to get people to buy one. Isn’t a dishwasher supposed to run scalding hot and sanitize? If you wash produce in it, is that a hot pot — do you just eat it afterwards? I looked it up and found I was simply behind the times.

The third group said there’s nowhere to plumb it and nowhere to put it, this enormous thing. Plus the older generation won’t get used to a machine, and young people order delivery anyway.

But not one person could rebut the actual line: don’t all of those objections apply, one for one, to a domestic robot?

Mark Tang: A human helper also makes mistakes, and it’s hard to find one you actually get along with. Expectations for a humanoid are very high — it should guess what I want to eat today, and put a blanket over me when I doze off on the sofa. That kind of thoughtfulness is hard to get from a person too.

Raymond: Because expectations for robots may be too high. If a robot collapses down to the single scenario a robot arm handles — sorting parcels, say — people actually like it. The moment you make it humanoid, expectations about generalization go through the roof, which makes the thing harder to push forward.

Mark Tang: I don’t think our views conflict. My point is it has to land in a concrete scenario people can see. Not “I know it can do everything, so I’ll buy it,” but “it’s very good at the thing I do all the time” — then I’ll buy. If you can’t do one scenario well and you’re claiming to do everything, nobody buys.

Raymond: Punch through one scenario, have an obvious use case, generalize later.

I’ll add one more thing: there is clearly a bubble in Chinese primary-market robotics. That’s a fair statement.

Mark Tang: In Chinese robotics there are now something like a thousand-plus startups, with arrangements with local industrial parks and governments, buying from each other to prop things up — and your delivery numbers go up. Each of them shares a similar upstream supply chain, with some like Unitree doing their own motors and vision.

But the volumes are simply not large. Most companies including Unitree produce a thousand-odd, maybe two thousand units a year. That’s nothing; it’s easy to absorb. So the question is how many can survive to the technical inflection three or five years out, or have enough capital to keep accumulating through it.

Of a thousand startups now, half or more are going to die. They have no choice — they get eliminated, or they vertically integrate.

Raymond: Here’s what I think happens: people have goodwill and expectations for robots, the sector fails to meet them, funding gets harder, and quite soon you get a wave of deaths.

But I am highly optimistic about the sector long term. It’s a bit like how I’m very optimistic about Chinese electric vehicles without necessarily being optimistic about any single company — those are two different things. The state certainly wants the EV industry as a whole to do well and beat everybody worldwide, but individual companies can be in serious trouble.

Mark Tang: The sector is fine; the question is who can raise enough capital to stay alive. I heard a well-known investor make this point — Kathy Xu, I think — about investing in NetEase in the early 2000s and then hitting the dot-com bust. Most companies died, but NetEase had the money to keep going. In that environment most competitors died, so surviving was normal, because competition shrank — and then it took off. Frontier tech sectors mostly go through one of these shakeouts, and the winner-take-most effect is very pronounced.

9. The kill line: when switching costs are zero

Raymond: Let’s do another brutally competitive piece of frontier tech: large models. First — can you keep all the model version numbers straight, which one is which, which point-release?

Mark Tang: More or less, because we use them every day.

Raymond: You are an AI founder, after all.

Mark Tang: It’s how I eat. A thing that happens constantly in our work is that a new model ships today. We look at what dimensions the evals are OK on, then guess whether it can replace whichever cheap-and-good model we’re using in some scenario.

Raymond: For context: Mark Tang co-founded Caidazi, which is an AI application. You can think of an AI application as a refrigerator — the fridge uses a lot of electricity, and the cheaper the electricity the better. So the person in the world who best understands which model is good, cheap and stable is the founder of an AI application company. He lives on it daily.

Mark Tang: It’s my cost line. I have to pick the model that produces the best output for the least money, which is what’s best for my margin.

Raymond: The chart that went viral on Twitter and Chinese social media in August is the kill-line chart — plot every model on capability and price, draw a line, and models below it basically have no reason to exist. I won’t put it in the show notes, because it changes daily.

The situation now is that DeepSeek V4 is an important benchmark line: you either beat V4 on capability or you beat it on price. Which basically isn’t possible. So most models genuinely have no reason to exist.

This is unlike Chinese EVs. With a car, if you buy wrong and end up below the kill line, you bought it, you drive it for ten years, it’s a one-time purchase. But software is infinitely easy to switch — ten minutes and you’re gone, and your memory and skills come with you, your whole personal context comes with you, with no friction at all. That makes it brutally competitive.

Mark Tang: That line is very real for us. We would never use something more expensive than DeepSeek V4 Flash that’s also weaker. There’s no point.

So my range of choice is now very narrow. Either I want extremely high task-completion, accuracy, hallucination-first in some particular scenario, and I take the very top model, which is probably closed-source; or I go to the cheaper side but pick something better than most.

My recent definition is “good enough.” Most requests you send a large model just need to be good enough — don’t make dumb mistakes, get the answer right. It may never be shown to a user at all; it may sit in a back-end data pipeline. Do that step solidly and it’s fine. What gets tested after that is the harness of our whole application — the shell that schedules models and organizes tools and context.

Chinese open-source models have developed tiers too. Kimi’s K2 is a 2.8T-parameter model, so deployment cost is high — it needs sixty-four cards for a decent deployment. DeepSeek is six cards. I think. Six cards will do it; four cards will run it.

Raymond: Hold on — is it K2 that needs sixty-four cards, or V4 that needs four?

Mark Tang: Four cards. Four or six, I don’t actually know, it’s from the official deployment guidance. Anyway, the gap is that big.

So K2’s cost is higher. Its launch pricing was basically aligned with the GPT-5.6 tier in the US, and more expensive than GPT-5.6’s cheapest model, Luna. It doesn’t qualify as a value open-source model at all.

Our old definition of open-source was cheap — cheap, good, plentiful, eco-friendly. That’s no longer the definition; there are tiers now. Conversely, Zhipu’s GLM-5.3 Flash is a small model with low deployment cost, so its price still sits on the kill line, and it’s smart enough.

American anxiety about this is considerable. Closed-source capability is strong, everyone applauds, no problem — but most enterprises can no longer afford the best models, because the volumes are too large. Recently, friends of mine at gaming companies, at big tech, at all sorts of companies, have all started saying their employer is capping token usage.

Raymond: You mean Silicon Valley firms capping Claude usage?

Mark Tang: I mean Chinese big tech mostly.

Raymond: Chinese firms cap usage of Chinese models too?

Mark Tang: They do. Once a company reaches a certain size you start thinking about data security, and then you won’t let employees use a Codex coding plan for real work. Internally they may have their own harness shell, roughly the same underneath, and let employees pick different models with different credit multipliers. You combine them yourself, but when the month’s quota is gone it’s gone.

Raymond: Hold on — they cap quota on Chinese models too?

Mark Tang: I was talking about closed-source models. These big Chinese companies all plug into the top closed-source APIs.

Raymond: Right, now I understand, I misheard you. You mean employees at a large Chinese company can no longer burn Opus without limit. I thought you were saying they couldn’t use DeepSeek.

Mark Tang: Correct. And Silicon Valley has started capping too.

Raymond: What does that force people to do?

Mark Tang: You go ask your manager for more quota, and you have to say: my output is twice what it was, so I want twice the quota, is that unreasonable? My ROI per head went up.

It used to be the token-maxing era — I used twice as many tokens as last month, aren’t I impressive. Not anymore; now you have to actually produce results. My wife runs this policy at her company, which is funny, because as a manager it’s genuinely hard to measure what people underneath actually produce.

Raymond: Here’s something I’ve never worked out. If Chinese open-source models are this cheap, why hasn’t the global price come down visibly?

There’s a framing I read: Kuwait has the lowest oil extraction cost in the world, but it isn’t the most influential producer, because its Q isn’t big enough — very low P, but not enough volume. In theory, if Americans could also use open-source DeepSeek, Anthropic’s price should come down to match it. Clearly it hasn’t. Is that a Q problem?

Mark Tang: Domestic Chinese inference cards are being adopted faster and faster. I don’t know the real situation, but from media and financial reporting it has clearly grown, including training moving from mostly Nvidia to hundred-thousand-card domestic clusters. Inference is theoretically simpler. So I think it’s fine for now, nowhere near a supply wall.

10. Is “distillation” a dirty word?

Raymond: Distillation got discussed on a lot of podcasts in August, so we don’t need to explain what it is. Give me your reaction.

Mark Tang: Everyone is distilling everyone else, across the US and China. Nobody gets to say the other side distills and nobody gets to say they don’t.

Chinese open-source models don’t just release open weights and stop — they publish technical reports, and everything’s in there. Those reports genuinely spell out new architectural improvements: DeepSeek’s KV cache work, Kimi’s attention-mechanism work. That’s architecture-level.

Raymond: From DeepSeek’s recent interview transcript, it looks like they published some of their results and not all of them. So in practice they’re still much better than others, and their cost is still much lower.

Mark Tang: But overseas models also reference Chinese models’ technical approaches and architectural contributions — so I’ll distill a bit, isn’t that normal?

Raymond: What I actually want to get to is Zhang Yiming’s remarks — ByteDance’s founder.

Let me state my own position first. I did an episode about relay stations — the intermediaries who resell the big labs’ APIs — and the question I most wanted to ask that operator was: have you sold your data to anyone, or has anyone come to buy it?

Because to me that’s a positive in an investment case. A model company that’s willing to distill others, and does it well, or aggressively goes after that data by every available means — I think that’s good for the company. I’d be willing to buy it.

Zhang Yiming disagrees. His argument has two parts. First: if you’re a student copying the teacher’s homework every day, you will never out-copy the teacher — you can only chase, you can approach infinitely closely, but you can never get past. Second, on organization: what you end up with is a company full of people who are good at distilling, a lot of hackers; while the real work — data cleaning, pre-training preparation, post-training, that whole apparatus — is where your organizational capability fails to develop.

That may be an entrepreneur holding himself to a higher standard. But that remark also turned distillation into a negative word.

In tech, distillation wasn’t negative before. The air conditioner in my living room is a huge model; if I want it on a phone or a laptop, all those Lite and Flash variants are a form of distillation, going from large to small — collecting the intelligence of the big thing into a small footprint. Those companies do it to themselves constantly, distilling their own big models into small ones and saying mine is cheaper and better. That was a value-for-money move.

I personally don’t think it’s a negative word. I posted another example on Jike a couple of days ago: do you think Pinduoduo is a distillation company?

What I mean is, take China’s e-commerce environment from 2015 to 2016 and 2017, and assume it was the same as the US–China environment today. The US–China environment is: I have the chips and I won’t let you use them, I hold you there permanently, you can’t develop. And the Chinese e-commerce environment then had “pick one” exclusivity — every platform was simply locked out, genuinely locked out, so JD.com went first-party, and first-party could survive while other platforms couldn’t.

Under that exclusivity, Pinduoduo rose using its “cut a slice” referral mechanic. I don’t know whether the word is right, but you could say it distilled Tencent’s social graph — or that it used a workaround to get itself a seat at the table and a way forward, which is what eventually produced the moment its market cap passed Taobao’s.

Mark Tang: When I first heard that I thought the distillation in my head was a different thing from yours. Your first one is still more of a social-viral play; it did distill WeChat’s social graph.

But Pinduoduo’s later move into Temu is the second kind of distillation: a brand used to build its own site and sell its own goods, and Temu abstracts that capability out, takes the essence, turns it into a white-label product, fully managed. It distilled your supply chain. Distillation is extraction of essence.

Raymond: I should clarify that what I described about Pinduoduo is definitely not the model-technical sense of distillation. What I’m describing is: when you meet someone in business who is bigger than you and who imposes constraints beyond your competitive ability, what routes let you break through and work around it, stay alive and grow — because it’s only once you reach a certain size that you have any possibility of surpassing him. Imitate first, surpass later.

So my subjective expectation is that startups facing both big incumbents and a chip shortage have to find more ways to overtake on the curve. Because you may have to stay on the curve to survive at all.

Mark Tang: And our companies here — whether in AI and robotics or in cars — are to a large extent leading the world now. Frontier tech goes around the curve, around the curve, and then it’s simply out in front.

Raymond: But you have to stay at the table.

Mark Tang: From a model-iteration standpoint you can split this into three pieces: architectural optimization, data optimization including synthetic data, and optimization from distillation. Distillation and the data piece have a lot in common, because a lot of synthetic data can equally be called distillation.

Those three run in parallel, and what they test is the organization’s management capability — how much of your headcount is weighted toward architectural improvement. That doesn’t stop me from separately doing some distillation to close the post-training gap fast. I don’t see a problem with that. It’s also why we do see continuous architectural improvement out of the frontier Chinese labs, and that improvement then gets actively adopted by the two American closed-source companies.

Raymond: If what you distill out is closed-source, that’s a big problem — a bigger problem, I’d say. But what they distill out is open-source, which is why half of America supports it and half doesn’t.

I hear Fable 5.1 added an anti-distillation mechanism and made it harder. Is that real?

Mark Tang: They may have added some mechanisms, but you can absolutely keep doing it — the difficulty rose, it isn’t impossible. It’s often the same as with relay stations: it can happen inside a coding plan rather than by calling the paid API directly, and there’s no way to block that.

Raymond: So with GLM-5.3 now this close to Fable, is the relay-station business still fine?

Mark Tang: It’s fine. Because employees at big companies still habitually use the best thing. It’s like telling me to stop using Codex and switch to WorkBuddy, the domestic office AI assistant —

Raymond: I’d have to think about that too.

Mark Tang: It’s a habit problem — the switching cost we discussed last time.

11. “Chinese people are imprisoned by work”

Raymond: Let’s do China’s AI office software in August. The ninety-minute conversation we did before got editors’ pick and an enormous number of people heard it, and a lot of big-tech employees gave me very different reactions. Some said it isn’t nearly as bad as you claim, you’re spreading rumors, my company does great work. Others — former employees at big companies — said privately that the reality is worse than what we described on the show.

First the data. Cross Mobile put out an interesting set of numbers on the latest monthly actives at several big-tech apps as of July, all of them domestic AI office or coding assistants: WorkBuddy at six-million-plus, Qclaw at two million, TRAE at two million, Qoder at just over two hundred thousand.

If they all doubled in August, the leader might reach ten million. But other numbers I’ve heard suggest its daily actives may only be in the low millions. Given the scale of the ad spend, I don’t know whether that meets their expectations.

Mark Tang: Speaking of ads — on this trip, and this has nothing to do with investing: the airports in China and in Europe run completely different advertising.

Before boarding in China I saw AI office assistants, then Doubao, Alibaba Cloud, Volcano Engine, that whole database-and-cloud category. Then I landed in Geneva, and because the week I went was the race, it was all sports brands — Nike, Brooks, adidas, Hoka, trail-running ads everywhere. Two completely different worlds.

Coming home I got another dose, because that jet bridge is genuinely too long, and it was papered end to end with AI office ads. I think Chinese people are imprisoned by work.

Raymond: In America it’s probably all the companies that want to raise money, buying up buses and billboards.

The fact that Chinese airports are full of office software, databases and cloud ads reflects that they’ve measured it and it works. Airport advertising isn’t cheap; they only keep paying because downloads and usage genuinely go up.

Chinese people really do love work. Very upward, very driven.

Which reminds me of something recent. At Stanford’s commencement, Google’s CEO Sundar Pichai attended, and the students weren’t especially fond of him. I find that interesting, because Stanford is already the most AI-native university in America.

The reported picture is that many students have strong negative feelings about AI — AI replaces people, AI is dangerous, AI takes my job. That’s widespread in America; there are even protests against data centers. None of that exists in China. I very rarely see anyone on Chinese social media saying we should resist AI. Chinese people are the opposite, embracing it wholeheartedly, and the biggest influencers now are AI-tutorial influencers: how to use AI better, how to make money on the side, how to work a second job.

Mark Tang: You asked why the two places differ.

Europeans are “this has nothing to do with me.” AI has nothing to do with us, our Mistral could collapse any minute, I just want to float back down the Geneva river after work and hike up the mountain. Don’t talk to me about any of this — I didn’t hear that vocabulary once over there.

Americans can access this stuff, but their way of life puts a lot of weight on balance, so many of them feel like they can’t afford not to grind.

Chinese people are used to grinding and are extremely driven — he wants to move up, he wants to succeed. So the frame is: how do I make myself more competitive by any means available.

Raymond: I was watching a cooking show the other day and one line was very funny. A chef said of another chef, who was seventy: he’s at exactly the right age to go out and make his name. The old master really is very good — seventy, going on variety television, cooking those excellent dishes, probably also promoting the restaurant. “Exactly the right age to go out and make his name” — if a European heard that, it’d be a court case.

Mark Tang: Elder abuse.

Setting aside investing and technology for a second, I do think people could look at more varied ways of living. Grinding away in an office using an AI assistant is not the only way to have a good life. There are a lot of ways to live and people could explore more of them, not be boxed in. I think everyone in our country tries too hard and is too driven, and may lose a lot of chances to go find something good because of it.

12. Give the person wings, or make the person unnecessary

Raymond: You’ve just said a lot of very un-grindy things, so now I’m going to make you grind again.

I imagine Pony Ma and Jack Ma might both be listening. How would you assess where the office-software war has got to, and what the endgame looks like? Some wars are good when they end — the payments war ended with clear moats and a clear result. Some end badly; the food-delivery war has visibly ended in a mess.

Mark Tang: I genuinely haven’t thought about it. But I think this war has only just started.

We’ve raised this before: office AI probably goes from assistance, to replacement, to genuinely becoming a digital employee that expands headcount — three stages. And we’re nowhere near completing stage one. It’s still assisting people, pushing individual efficiency to its limit.

There are newer product forms abroad — the Grok bot category is a new way of working; or Claude Tag; or what Slack is attempting — and all of them are already trying to move toward digital employees. Very good Chinese AI application startups are doing this category too, and started very early. I can’t recall the name right now; I’ll put it in the show notes, out of respect for them.

But the mainstream in China still defines office AI around the general-agent co-work scenario. That’s a generation behind.

Raymond: But Chinese application-layer teams catch up fast.

No — it isn’t a question of catching up. They’re two different ways of thinking. One is: I give this person wings, that’s one way of thinking. The other is: this person simply doesn’t exist — the digital employee, the Claude Tag category. It carries a sense of replacing a person with a general knowledge base; it puts the thing and the person on the same tier. It isn’t a tool, it is the person.

Mark Tang: Some people in China may go down that road, but every office war being fought today is fought around tools, not around digital employees. That’s a generational gap in product design and product definition.

Claude Cowork came out at the end of last year; big Chinese firms held meetings at year-end, ran into Chinese New Year, and three or four months later the corresponding products shipped. Equally, if Grok bot’s numbers over the next few months validate, big Chinese firms will start holding meetings about whether that’s the new good form.

Raymond: So from vibe coding, from Cursor, to Claude, to Cowork, to Claude Tag — three generations, three forms — China looks, learns, and fights the next war.

Let me narrow it. What’s being fought now is office software, the AI-empowered era of WPS. My subjective view is that this war isn’t very worth fighting. The Claude Tag category, the digital-employee category, is worth fighting.

I did an episode with Michael, formerly of Final Round, precisely about digital employees. The listen numbers were terrible. But I think that episode was excellent and important — what he described is how the next generation of products should be built, and his company itself walked through all three generations. I’ve never understood why nobody listened.

I pay a great deal of attention to whether any Chinese startup goes straight at the third generation. My subjective view is that China’s second generation is hard to make work — products where the human is the protagonist and the agent is the assistant will have low daily actives, low monthly actives, no paying users. Going straight at digital employees may have more room.

Mark Tang: And that may not be big tech’s opportunity — it may be a startup’s.

Because once it becomes a digital employee, you’re dealing with the whole relationship network of human-to-human, agent-to-agent, human-to-agent, and infrastructure requirements go up enormously. Today I use my assistant and you use yours, and the two don’t know each other at all, and don’t need to.

What annoys me most now is someone telling me in a meeting that AI wrote this. You don’t have to tell me what you wrote it with; the work you hand in is your work.

But in the future, if everyone has a Wukong — the Monkey King — then it’s Wukong talking to Wukong, Bajie talking to Bajie, and eventually maybe my Bajie talking to your Wukong. Inside an organization it’s that; between a supplier and a customer it’s that. Human-to-human communication gets very complicated.

Raymond: The category I mean is every form where the human is the protagonist and the agent is the assistant. Its ultimate problem is that it’s a technical solution fitted onto the existing world, and the existing world may itself be broken. Force the fit, and users don’t like using it, and after using it they feel it’ll replace them — a long, awkward state.

Mark Tang: If you overreach and go straight at third-generation digital employees, you may pull something. And it’s hard for big tech to justify — to prove something works you point to revenue, actives, retention, conversion. But with a digital employee the person you have to persuade is the one paying, and that’s a harder decision.

Buying tools today is obviously about efficiency and headcount reduction. To do digital employees you have to design group strategy and organizational planning, which is a far more complex chain. So at the start it takes a few marquee companies doing it as a case, and those companies may not even be paying — they’re doing it for the sake of doing it.

Raymond: Part of why I’m going to Silicon Valley is to look at this, because the larger firms there, thousand-plus-person companies, are certainly ahead of many Chinese companies on how to actually use AI. You want to see how other people do it, because organizational relationships and organizational method matter enormously.

Very few people discuss this, because it usually requires the person at the top to explain why their own company isn’t doing it well, and most won’t. But it’s genuinely valuable to society — when a tool this unusual appears, what should organizations change, what’s short-term pain, what’s social cost. That deserves a full discussion.

To echo what you said: Michael, out of a whole range of digital-employee possibilities, chose marketing, because marketing’s entire workflow is dealing with numbers — placements, YouTube ads — with no need to touch real people or the physical world. The other categories all had barriers he couldn’t clear.

13. Whether AI succeeds or fails, long-end rates fall

Raymond: Last question, on macro.

Labor Day has passed. The yen has run hard, Treasuries are at very high levels — right now, whether or not you cut, the Treasury market is hiking for you. Oil is back at relative highs, in the nineties. Wars continue daily even if they’re off the front page. Then there’s Bessent’s treasury twist, issuing short and buying long. Plus the recent hawkish FOMC language.

All of it is volatile and none of it is clear. Which macro factor in August made you think this one feels different?

Mark Tang: I saw a chart the day before yesterday that I found interesting, and it’s AI-related.

It says AI has two possible endings: the bubble pops, or it’s extraordinarily successful. Whichever ending you get, long-end rates fall.

Why? If AI is extraordinarily successful, penetration goes very high, productivity rises sharply, and you get some degree of deflation — and after deflation the long end falls. The second case is the bubble popping, in which case every stock across the AI chain and the Nasdaq blows up, and after that investors buy Treasuries, so rates come down too.

So either way, long-end rates fall necessarily by around the end of 2027, once AI’s ending becomes somewhat clear. It’s just that along the way, current fiscal pressure and inflation pressure keep short rates elevated.

The Treasury has been trying to press the long end down with fiscal tools, and all of it feels like drinking poison to quench thirst. It doesn’t solve the actual problem; the market is using trades to force the Fed to hike.

Raymond: I think the probability of a hike is still fairly low, for many reasons — unpacking it is another hour. But I can feel investors pricing “US debt crisis” more and more heavily.

What’s interesting is that we’re in a boy-who-cried-wolf state. Each time it’s Trump, or the Fed, or Scott Bessent taking a turn to say something. The first time, it moves the ten-year or thirty-year a few basis points and everyone thinks the market is obedient; the second and third time less; by the fifteenth, the decay is complete and nobody listens.

This treasury twist — issue short, buy long — got a reaction on day one, then day two, day three, and a week later it had all come back. Which is why the US long end is at its highest level since 2007.

Mark Tang: This about-to-blow-but-not-blowing state has gone on a long time. Bridgewater’s Ray Dalio has been talking about US debt for well over a decade. If they can keep pushing the impact out with timing arbitrage, the patient stays alive and you can hold on a while longer.

But if rates keep rising, the effect on the hard-tech growth names we follow is very large, because it directly raises everyone’s cost of capital. Look at Google — with these big firms’ free cash flow already turning negative, everyone needs financing, equity or debt, and either has a cost. As overall US rates get pushed up, everyone’s financing cost rises, potential ROI falls, ROIC falls, and every story turns negative. People watch macro news because they’re really watching whether it drags down big tech.

Raymond: You can think of it as: today’s share prices are prices under a thirty-year at 5.3. If the thirty-year fell to 4.7 or 4.8, what should those prices be? Possibly something entirely different.

Mark Tang: You also can’t look only at the long end — an inversion would be bad too. But an inversion feels difficult right now. The current trend is that Bessent’s moves are meant to push it toward inversion, and it’s clearly —

Raymond: Not listening.

This month we may see Anthropic prepare to list. Under the chase from the Chinese model line — and it isn’t a kill line, it’s a chase-down line, it chases you down — the thing it’s chasing is Anthropic. I’m extremely curious how Anthropic explains, on an IPO roadshow, that this chase-down line has caught up.

Mark Tang: He’ll definitely tell you he’s been sitting on something big and hasn’t shipped it. And then ship a Fable 6.

Raymond: That’s how it always goes — because if it isn’t explosive, you don’t ship it.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.