After July, Interrogating the Last AI Bull

“Replacing people is only its starting point. If you have a pile of very cheap geniuses, what that world looks like is something you have no way of imagining from this point in time.” — Tang Yibo, on why he benchmarks this cycle against electrification

Guest: Tang Yibo｜equity analyst at a hedge fund, author of the WeChat account 「奔波儿r」

The US-listed AI commentator Raymond follows regularly hasn’t posted in over forty days.

July was a turbulent month for global markets. Korean equities, US semiconductors and China’s tech board all saw large swings. Once July ended, a lot of people stopped being bullish on AI — not turning bearish, just going quiet. So for this episode, Raymond went and found the one bull left standing in the market.

Tang Yibo is an equity analyst at a hedge fund, in his eighth year in the industry, having covered renewables, then power, then manufacturing, cyclicals last year, and formally picking up AI hardware coverage this year, across global markets. On the side he writes the WeChat account 「奔波儿r」, which started by sharing tutorials on using Claude Code for investment research and later moved to his own judgements on the AI industry — when AI CapEx peaks, what we’re actually discussing when we discuss CapEx, and the rise of domestic Chinese models.

The last time Tang Yibo discussed this publicly was on another podcast in conversation with Xingchen, recorded in exactly the days AI stocks bottomed. Raymond later told Xingchen that the conversation hadn’t been fair to Tang Yibo — a lot of questions hadn’t given him enough time and room to develop. So this time, Raymond plays the AI bear himself for ninety minutes and interrogates him point by point.

Over those ninety minutes the two of them worked through the arithmetic: how much a one-gigawatt data center costs and how much it brings back; what open-source models have actually broken; why, when individual productivity rises tenfold, company output only rises twenty or thirty percent; how much revenue is needed to carry a trillion dollars of annual depreciation; and why he insists on benchmarking this cycle against electrification.

What follows is the full conversation, edited and condensed.

1. “I’m not clear why free cash flow has to be the measure of whether CapEx is reasonable”

Raymond: Straight to it. In the summer of 2026, the big American cloud vendors have spent seven or eight hundred billion dollars on CapEx this year and may spend over a trillion next year, and some companies’ free cash flow is now entirely going into building data centers. Where does this money come from? Is there enough?

Tang Yibo: I think there’s certainly enough. In fact I’m not clear why free cash flow has to be the measure of whether CapEx is reasonable.

Take something recent, like renewables. During the solar buildout a few years ago, why did nobody take a solar company’s CapEx exceeding operating cash flow as evidence that the CapEx was unsustainable? Normally, an emerging industry in its early growth phase must have capital expenditure greater than revenue.

These hyperscalers were already the dominant leaders of the last cloud cycle, the last IT cycle; the money is very deep and the reserves are very deep. If you’re a shareholder in a CSP, a cloud service provider, you’ll think: I bought these CSPs precisely for the stable business model, stable cash flow, annual buybacks plus a bit of dividend, and stable 15 to 20% growth. From that standpoint you feel they must have free cash flow; from outside that standpoint, this is entirely normal.

Go further back, to electrification, to building railways in the first industrial revolution. A technological revolution that significant has to build infrastructure early on. And while building infrastructure, how could you possibly have extremely high returns the moment you invest, or cover CapEx with operating cash flow from the start? Every industry starts with CapEx exceeding operating cash flow and recovers the capital expenditure afterward.

Raymond: Two additions. First, treating hyperscalers as growth stocks is fine, but many people previously treated them as mature companies, so there’s a gap in expectations. Second, the Magnificent Seven previously had such deep reserves that they could carry very large CapEx; when free cash flow runs short and they start asking capital markets for money, people feel the magnitude will be very large. Anxiety about the magnitude, plus the fact of “they’re starting to borrow” itself, is a general fear.

So how many gigawatts can this trillion-plus build? How is the money allocated?

Tang Yibo: Before GB300, building one gigawatt cost roughly $40 billion. The CSPs have put in seven or eight hundred billion, and factoring in sovereign funds, China and other cloud vendors, you multiply by close to two — call it a trillion-plus. Assume $1.2 trillion divided by $40 billion, roughly 30 gigawatts. I’m fairly optimistic on next year, still 80% to 100% growth, so $2.5 trillion, with hyperscalers taking half.

How that $40 billion splits: civil works, power and electrical equipment together are about $10 billion, and the bulk goes to servers. Split the servers further, and the GPU, though its share keeps falling, is still the largest, with storage probably second. So all the AI hardware people are trading sits inside that roughly $30 billion per gigawatt; the power trades — land, HVAC stocks, diesel generators — sit in the remaining $10 billion.

Raymond: So a gigawatt costs $40 billion in. How much comes back?

Tang Yibo: Depends on the basis. $40 billion isn’t the GB300 figure; on the latest numbers GB300 is already at $60 billion per gigawatt, but its revenue is higher too.

Raymond: Sorry, it’s already at $60 billion? I thought it was still in the $40 billion range.

Tang Yibo: Because power consumption fell, so a gigawatt corresponds to more cards, more servers, and CapEx goes up.

Let’s still use the old $40 billion basis. If this gigawatt of servers goes to Anthropic for inference — which it sells as API or subscription, with API at ninety percent of its revenue — the corresponding revenue is roughly $60 billion. Initially people thought it was only twenty or thirty billion, but with new models out and usage up, it’s now $60 billion a year. Its inference gross margin is over 80%, and assume a 60% operating margin, so $60 billion of revenue earns over $30 billion of operating profit. The return is very high.

If this gigawatt runs open-source models, revenue is roughly 40% to 50% lower, running full at around $30 billion — though this is very imprecise, and I discussed it with AI many times without arriving at a precise figure. And open-source models have very low operating margins, so the return is much lower than Anthropic’s.

Raymond: So the basic arithmetic is: a gigawatt costs $40 to $60 billion; run Anthropic’s API on it and revenue might be $60 billion; run open-source models and it might be $30 billion.

Let me add a corroborating data point: a data provider that tracks token prices calculated that since June this year closed-model prices fell from just over $3 per million tokens to close to $2.70 or $2.80. What that feels like in practice is Anthropic letting you use Fable and Codex giving you reset after reset after reset. Over the same period Chinese open-source model prices rose from 0.7 toward 1, so the gap is still more than 2x — which lines up with $60 billion versus $30 billion.

2. “The arrival of open-source models has merely blown up their excess profits”

Raymond: You’ve written that token price is the most important variable in the whole computation — if it falls, everything else is moot. Under pressure from Chinese open-source models, is it in a state of continuous decline?

Tang Yibo: Certainly. Last week OpenAI made the lowest tier free and cut the middle tier substantially; only the top tier’s price is unchanged. There’s also news that Anthropic will shortly launch Sonnet 5, still priced at the Sonnet tier, $15 to $25, but with capability only 10% or 20% below Opus.

But I don’t think this is bad for the industry. Consider what we just said: $60 billion of revenue per gigawatt, 80% gross margin, 60% operating margin. And as Anthropic I don’t have to build a data center at all — I take $20 billion per gigawatt and rent the cloud vendors’ kit; I’ve deployed no capital, I merely rented and resold, and I make an obscene profit. My ROIC feels like it’s in the thousands or tens of thousands.

So I think the arrival of open-source models has merely blown up their excess profits. In any industry, making money like that is in fact unreasonable.

So they have only two paths. One is cutting price, driving down the price of mid- and low-intelligence models. The second is continuing to sprint at the frontier, having the frontier accomplish things you can’t imagine — that’s the ultimate way out. Competing with Chinese models on cost at mid-tier intelligence can’t be won, and it ends up like all manufacturing.

Raymond: So why doesn’t it end up in the kind of ferocious commoditization you used to cover in solar? Second place is a fairly comfortable position, but for first place to keep a lead of six to twelve months or more over everyone else is extremely, extremely hard.

Tang Yibo: Right, and one day it may be that nobody can produce anything new no matter what they do.

Raymond: Then a problem appears: token prices fall continuously, usage explodes, every day is a shopping festival — but the chain leaders’ revenue doesn’t change. P falls, Q rises, and total revenue may be unchanged.

Tang Yibo: That’s an elasticity question: which is faster, the growth rate of demand or the rate of price decline. I don’t remember the solar figure precisely, but roughly, before demand exploded in 2020, the total revenue pool was only a couple of hundred billion yuan — however much volume grew, times that price it was always a couple of hundred billion; it never rose.

But elasticity depends on the value you create downstream, on what you’re replacing. Solar had one very explicit opponent: I absolutely have to get cost below coal power — and because you’re less stable than coal, you have to get below it, running continuously toward that target before you can escape subsidies. You’re really a protected flower.

But AI’s intelligence already exceeds humans’. The value it creates on top I already find somewhat hard to comprehend. Even if that intelligence improves only slightly — it has already raised the proportion on the Riemann hypothesis considerably — if it shifts from coding to creating value, say improving a drug’s development, the marginal value would be enormous. Unlike solar, where I’m only cutting cost for cost’s sake.

So you can’t reach a conclusion at this point in time. One possibility is: intelligence below the level of human genius has been captured by open-source models, and frontier models are only a small notch above human genius — but able to complete some things people can’t.

Raymond: Divinely blessed.

Tang Yibo: Right, as if divinely blessed. Say it suddenly raises a drug development success rate from 5% to 15%; multiply by that drug’s future market and the value is enormous. So it’s possible frontier models have only 5% of usage, even 2% of usage, but ultimately take 50% of the value.

Raymond: So P may be stable or even rising, because it creates value rather than merely substituting. Unlike solar or EVs — however much they advance, a car’s ceiling is where it is and solar’s ceiling is the price of coal power. Intelligence’s TAM ceiling is far higher.

3. “Whoever wants to take one more step forward has to spend another hundred billion”

Raymond: Let me ask a question on Xingchen’s behalf, which is also many people’s question: Anthropic produced SOTA with a fraction of OpenAI’s compute, so is compute really that important? If the cost of obtaining the intelligence tap isn’t actually high, then more and more people will do it — first, the price war gets more dramatic, and second, will training in future no longer need that much CapEx?

Tang Yibo: I read it as something like a latecomer advantage.

Talk to people at model companies and you know that training a model involves large numbers of experiments: ten paths, and you run experiments first to determine which path, and only then train it all the way. That step is critical — you can’t run every path to the end and back out when it fails. So both OpenAI and Anthropic have carried a large share of the frontier-exploration function.

OpenAI was never like Anthropic. Anthropic focused on coding; OpenAI spent two years on multimodality, consumer, all sorts of scattered directions, and cut them one by one afterward, because it found only the coding path is the royal road to the summit of intelligence. OpenAI blazed that trail for everyone. Once it was blazed, Anthropic said fine, I’ll take one path. So of course it’s a fraction of OpenAI.

OpenAI played another role: it said multimodality was the future direction, and then Google went all in on multimodality, which led Google into the ditch. Google was still first tier last October and possibly number one; it’s now dropped out of first place.

As for domestic models, why can costs be that low? Look at the papers Kimi and DeepSeek have published; they’ve made an enormous number of architectural innovations. Architectural innovation isn’t about genuinely being able to train a better model than yours; it’s: my chips really are limited and my memory really is limited, so can I use one tenth of the chips and one tenth of the memory to reach 90% of your effect — you certainly can’t reach 100%, because all compression loses information.

They don’t need to explore the frontier of intelligence, because Anthropic has already explored the frontier for them. At the current stage chips and hardware are both restricted, so all you can do is optimize architecture under limited resources: how to reach 90% of an Anthropic model’s capability with one tenth of the KV cache, the GPU memory used to cache context during inference.

Raymond: In summary: one day China’s and America’s intelligence levels may be entirely parallel. But at that point, whoever wants to take one more step forward has to spend another hundred billion. It’s just that right now America spends first.

By your account, video may be the other path. Sora 2 explored video and Google explored multimodality; now Seedance 2.0 and MiniMax’s H3 are moving forward. Globally the most profitable video model belongs to ByteDance itself, an internal loop. So possibly China then spends enormous amounts on video, and once the trail is blazed the next generation of Gemini copies Seedance and the cost falls instead. But within the whole of humanity, whoever moves forward has to spend the extra hundreds of billions.

Tang Yibo: Yes. So even if Chinese vendors train models at a fraction of the cost, that doesn’t mean you won’t have to spend that much on training to keep raising intelligence. Waste is sometimes necessary; it’s the cost of exploring. Any company’s R&D and any company’s advertising spend could never be a hundred percent accurate. Never mind R&D — R&D is wrong nine times out of ten.

4. “What worries everyone most is that there’s no second scenario”

Raymond: What do you think the market’s main contradiction is after July? I feel whether CSPs have money is only a surface issue — in this last set of results they all said ROI is high and they’ll keep investing, and it still didn’t dispel the market’s doubts.

Tang Yibo: It’s a surface issue. The market’s most deep-rooted worry is the same as in the second half of last year.

In the second half of last year there were no agents, so people said you’re training all day with no use case, this is all burning money. This year the agent coding use case appeared and rapidly generated hundreds of billions of dollars of volume. Now, what people worry about is: coding is such a special scenario — programming, and the users are programmers, so penetration can rise fast — so is there a second scenario?

If it’s only coding, the TAM is only a trillion dollars, whether you benchmark against programmers’ salaries or software’s TAM. Anthropic plus OpenAI may now be $130 or $140 billion, so by year-end that’s 30% penetration. For a growth industry, it ought to become a mature industry.

So what worries everyone most is still demand, or rather worry about there being no new scenario. There will be many scenarios in future, many people believe that — but what if there isn’t one next year?

Raymond: There are two big events in the second half of this year, Anthropic’s and OpenAI’s IPOs. If you were OpenAI’s CFO, investors would ask the same question — how would you answer?

Tang Yibo: They’ve already answered. Anthropic’s Dario keeps saying: our target isn’t coding and isn’t software; our target is every white-collar worker in the world. All the white-collar workers in the world are $40 trillion, and dividing by that number gives you 0.5% penetration. That’s a super, super early state.

So his job is defining that denominator: getting everyone to believe the denominator isn’t a single scenario but $40 trillion, with more value creatable on top. Liang Wenfeng has also told investors AI may account for 20% or 30% of global GDP in future, which is likewise tens of trillions of dollars of room. I’m sympathetic to that view — AI is intelligence, what white-collar workers do is mental labor, and in theory AI can do all of it.

Raymond: Let me rebut as the bear. First, the global SaaS market is somewhere over $300 billion — Adobe, Salesforce, Workday, everything together, just over $300 billion, with US SaaS at $180 billion. The $3 trillion you describe is ten times today’s global SaaS. Second, the Magnificent Seven’s combined revenue, from Amazon to Microsoft to Meta to Nvidia, is under $2 trillion across seven companies. So these two new companies have to create value on the scale of the Magnificent Seven themselves. How many years has Microsoft been around, how many years has Google — their enterprise penetration needs no discussion, and they’ve made a few hundred billion. On what basis can these two new companies make over a trillion?

Tang Yibo: Revenue can’t quite be counted twice, I concede that — I book a hundred at Anthropic and fifty more at Google Cloud, and there genuinely is that in the accounting.

But I feel your benchmark is wrong. SaaS is an auxiliary tool inside enterprise workflow. Within the Magnificent Seven, Microsoft is Office plus cloud, Google is search advertising plus cloud, Meta is social — your benchmark is still the internet, still the information economy. What the internet improved was the speed of information circulation, reducing information asymmetry.

But I don’t know if you’ve ever calculated: how many kilowatt-hours does the world use in a year, times the price of electricity, and how big a market is that? I haven’t calculated it either, but it should be enormous, also several trillion dollars.

So we’re still discussing two different things. If you think a model merely raises the efficiency of processing information — as the internet raised the efficiency of distributing information — then the comparison is reasonable and it may not reach the Magnificent Seven’s magnitude. But what if it isn’t?

5. “Individual productivity rises tenfold and company output only rises twenty or thirty percent”

Tang Yibo: Another thing I’ve been discussing with people lately is the question of product form.

Claude Code, or an agent, is fundamentally still a personal assistant. A personal assistant embedded in enterprise workflow still has problems — you still need people communicating with people, since it isn’t an AI-native company. So although everyone has a personal assistant and everyone’s individual work efficiency has risen tenfold, the enterprise’s output may be only 20% or 30% higher. Because your bottleneck isn’t one person’s output; it may be the friction cost of communication, or some step that can’t be accelerated by an agent.

But if you get to the next stage — say Claude Tag, which Anthropic launched recently — I think that’s a very good product form. In Slack, similar to WeChat or Feishu, anyone in a group can add this Claude Tag, and you assign it work and it does it, and it can read the group’s contents itself. So you don’t need a dedicated person to organize the company’s context for it — it learns the company’s whole workflow in the course of your chatting.

That’s a big shift: it goes from being an individual-level agent to a company-level agent.

Let your imagination run a bit further: much of what a company’s middle management does is relay up and down, reduce friction between departments, and align thinking. If AI becomes the company’s central spindle in future and everyone aligns with the AI — I’m the person executing, and I ask the AI what the CEO actually thinks — then friction cost falls substantially, and you can then eliminate bottlenecks around it. So even if intelligence stops improving, the effect of this product form alone on how work is done could turn a 20-or-30-percent lift in company output into eight to ten times.

Raymond: I’m playing the bear today, but let me play the bull for one minute. Shopify’s CEO has a method for using Claude Tag: across the whole company you may only tag it in public group chats, not in private messages — he wants to make sure every sentence every person in the company says is captured, ensuring the completeness of the context. In theory Shopify should soon find group chat headcount falling.

But there’s a fork here. Over the past stretch a lot of Silicon Valley companies have run similar token-maxing experiments, and productivity gains don’t seem obvious: you can’t see it in results, and layoffs haven’t been that large. It sounds like a fancy experiment — someone invited a Pokémon into the group chat and it’s familiar with your expense policy; so what? That’s where many people can’t connect the dots.

Tang Yibo: That depends on which sample you look at.

The sample I look at is Block, an American financial payments company. It’s aggressive, and even before Claude Tag existed it wanted to build a company-level agent as the company’s whole base layer. In February this year it cut 40% of staff, from a thousand people to six hundred — and after the cuts, operations and maintenance both worked very well. Some startups too, compressing R&D headcount while R&D quality and volume still rise.

This is an organizational-structure question: are you genuinely rebuilding your organizational structure around agents. If you are, you may cut 40% and still improve efficiency; if not, you’re only saying “let’s give it a go, we’re doing AI now,” but I don’t know how to use AI either, and it turns into token maxing — everyone compares who used more tokens and gets ranked — while in reality you’re still stuck at everyone using their own AI. It hasn’t changed the enterprise workflow: it still passes through A, then B, then C, and finally E decides. Everyone’s efficiency rose, and the bottleneck is still in the hands of the person deciding.

So look at many large companies and cutting staff or reorganizing workflow is very hard: office politics, people resisting. Even having done token maxing, that’s only an attempt.

I made a comparison recently: the token-maxing companies are, simply put, factories where I used to drive the line off a steam spindle and I’ve swapped it for an electric one, and that’s it, nothing changed. The genuinely impressive ones are like Ford — I replace the steam spindle with a motor, and then decouple small motors at every station, so I can genuinely optimize at every step and take a car’s production time from 12 hours to 1.5 hours. That requires a whole reconstruction to produce a lift in overall output, not merely everyone giving it a try.

Raymond: So it just becomes buying a more expensive SaaS — a hundred thousand dollars a month with no especially obvious output.

Tang Yibo: For the company as a whole, yes. But each individual’s own experience I think is already very substantial.

6. Sifting three thousand companies to find an MLCC name nobody covers

Raymond: So let’s talk about individual experience. Over the past six months, with 4.6 as the dividing line, what actually made you money?

Tang Yibo: The first is fast screening. Our investment model is like a funnel: look at a lot of things, screen fast, and research deeply if it seems interesting. A fast look at a company used to take at least two or three days, especially in an industry I’d never covered; now it might compress to 10 minutes.

But the thing that genuinely made me think “this is incredible” and completely changed how I work is something else.

As an individual analyst I can only cover three hundred companies — companies I’ve looked at and roughly know what they do; that’s my limit. And then out of those three hundred I find the best ten to fifteen to buy.

I like doing post-mortems and summaries. Last year I looked at overseas power supply companies like Delta Electronics, at optical fiber, and at the overseas copper foil leader Mitsui Kinzoku. None of these companies are in the same industry, and yet the timing and magnitude of their rallies were extremely consistent, up ten to fifteen times after June 2025. Nobody used to cover traditional manufacturing — before it rallied, who knew what these companies even did? Optical fiber, what a lowly industry.

So I summarized it myself: what characterizes traditional manufacturing? AI is under 10% of the business, the industry has been rotten for years, and valuations are low. So I wanted to find companies meeting these criteria that hadn’t yet rallied. Before AI I could only search along the value chain. But now I can say, these are my three criteria, plus one more: hasn’t risen more than 150% in three years.

Raymond: Is this A-shares?

Tang Yibo: Global. Europe, Japan, Korea, Taiwan and so on. I also said no tech companies, because tech companies have all already rallied.

That time it scanned roughly eight hundred companies, took twenty or thirty minutes, and filtered out 20. Ranked first was Taiyo Yuden, sixth Murata, tenth Yageo — MLCC in March, the multi-layer ceramic capacitor, the highest-volume passive component there is.

I had never looked at MLCC. So I said, how does this company score so high, hasn’t risen in three years, and trades at only ten times — let me take a look. I ran my own screen using those criteria plus some finer ones, and I thought, damn, this MLCC really does look like Delta last year.

What did it do for me? It took me from being able to scan only three hundred companies to summarizing the characteristics of winning stocks and having it screen three thousand directly. That MLCC name I researched for only two days, from “what even is MLCC” to “I think we can buy it” and writing the report took two days, and discussing with my boss took a week.

That was March. I asked some friends who cover electronics, and one said: ugh, MLCC, what a garbage industry, never heard a peep about it. I said, can prices rise? He said impossible, what price elasticity is there, those Chinese manufacturers leave their equipment in the warehouse and don’t wheel it out, afraid of capitalizing it into fixed assets. Hearing that made me even more excited.

Raymond: But by August, the degree of diffusion still isn’t high enough. In theory if this skill were used by many people — a million followers, twenty or fifty thousand GitHub stars — it should all have been mined out and priced in.

Tang Yibo: Yes. So the key isn’t the scanning at the back end. With and without AI, the scanning difference is enormous; but if everyone uses AI, the post-mortem and summary at the front are what matter most — how you summarize those winning stocks into three characteristics, or five. Too many and it won’t filter anything. And I don’t know whether that instance was pure luck or a bit of that.

Raymond: I’d say half and half. You happened to catch that move, and you scanned ahead of it. Even following on the right side of the trend I could have caught some, but because you did the research you believe it more readily, so your position is heavier and you hold it longer, because you know what it does. If you’re only chasing after it’s risen, you’re too late.

Tang Yibo: Right, my own experience is excellent. But personally I feel I’m not paying enough right now. It extracts far too little money from me.

Raymond: Because you may have made a hundred thousand out of this and ultimately paid two hundred. That mismatch makes you feel demand is this strong — if it could give me something better, I’d buy more.

7. Building two Three Gorges Dams a year, and a trillion in annual depreciation

Raymond: Push forward a few years. Right now it’s seven or eight hundred billion a year, in 2027 it’s $1.2 or $1.3 trillion, in 2028 maybe $1.3 to $1.5 trillion. Growth is decelerating, but for the foreseeable future it’s over a trillion. The Three Gorges is a 22-gigawatt project, so essentially we now have to build one and a half to two Three Gorges Dams every year.

Tang Yibo: Right, and its depreciation life is very short, so there may be a trillion dollars of depreciation cost annually. I find this world hard to imagine too; a lot of people hear it and think it’s impossible, terrifying.

Raymond: A trillion of depreciation, in a capital-intensive industry, needs revenue of at least $3 to $5 trillion; if it’s less capital-intensive, you might need $10 trillion of revenue to carry it. And that happens somewhere between 2029 and 2030, only four years from today.

Tang Yibo: It’s actually simple. This year, setting open source aside and using only OpenAI’s and Anthropic’s revenue as AI’s bottom-layer outlet, by year-end the two together may be a couple of hundred billion. You only need a 10x to reach $3 trillion.

A 10x is very hard in other industries. Open source will dilute them later; but factoring open source fully in, a 10x by 2029 or 2030, rising a bit more than two or three times a year, is a fairly neutral assumption. I don’t know why you’d think $5 trillion of revenue is hard to achieve — is it simply that the number feels large?

We’ve never seen demand growth like this, and electrification didn’t penetrate this fast either. This year is two to three gigawatts, and extrapolating from contracts already signed, 2028 is roughly 22 gigawatts, which is still conservative. Two hundred billion combined this year, times five or six, and by end-2028 you’re close to a trillion. Then growth keeps decelerating in 2029 and 2030, so reaching $3 trillion in 2030 — that part is imagination.

Raymond: So in the face of a trillion in depreciation, what we’re ultimately betting on is the judgement about TAM — what the sea of stars actually looks like.

8. If Anthropic ends up as a power plant

Raymond: You benchmark this AI revolution against electrification. Let me state my doubt first: if the benchmark is electrification, then P should fall continuously toward zero. Today electricity is something we barely perceive in daily life.

Tang Yibo: You can’t feel it in China. In America residential consumers feel it somewhat, at a bit over ten percent of the household bill. And rising US electricity prices are already causing protests. Trump introduced dual-track electricity pricing in PJM, the northeastern region, which is very like Chinese policy: existing residential customers stay in the old electricity market, and for the incremental part, if you want to build a data center you have to build your own and can’t affect existing residents’ power. New York State has already driven off another data center.

Raymond: So my point is: this TAM is large, but is it ultimately a TAM of people, or a TAM of money?

I can imagine everyone in the world using AI, everyone vibe coding, building eight websites a day — but P may fall continuously and become like the electricity bill. Is it possible that Anthropic becomes a company with three billion daily actives and its market cap and revenue don’t change?

Tang Yibo: I can’t picture it, but I think it’s possible.

At that point Anthropic certainly wouldn’t be a far-and-away leading vendor. We of course can’t see the end of intelligence improvement, but ultimately there’s certainly a wall waiting for us. At that point, like the electricity we use now, you don’t care who generated it — you won’t care whose token this is, since intelligence levels are all about the same. So possibly Anthropic stagnates, and lots of power-plant-like things appear, all identical, commoditized, a bulk commodity.

But even if it commoditizes, I don’t think that matters. Electricity price times volume is certainly a several-trillion-dollar market globally. Electricity, oil, coal are all bulk commodities and all the same magnitude.

Raymond: That TAM is exactly my worry. I certainly can’t say the number of users will fall; that would be wrong. But if I hold that market size fixed, then as the number of people rises the price necessarily falls.

Tang Yibo: Understood, so on that basis it doesn’t reach $5 trillion. But I still think the comparison should be the electricity market — most people in the world can access electricity now and electricity prices are already extremely low, and it’s still a several-trillion-dollar market.

9. “Debt isn’t a bad thing; high leverage is”

Raymond: There’s a post-mortem in your writing — roughly seven big technological leaps, each ending in a large debt event: everyone borrowed a lot of money to build things. So your thesis is that unless you reach a very extreme debt level, the bubble keeps inflating until the debt breaks.

Tang Yibo: Debt, at the outset, isn’t actually a bad thing. If you have something very good and rely only on your own operating accumulation, how many years does it take to scale it and popularize it? So borrowing at the start is very healthy: I have capital from the beginning, scale the good thing, get more people using it, get the flywheel turning. That is the point of capitalism, of stock markets, of financial markets. So you can’t demonize debt.

But why does debt end up the trigger? Because you levered too high — like that genius fund manager who blew up. When debt has accumulated very high and the leverage multiple is very high, the biggest negative effect is that you can’t absorb volatility: a 5% or 10% swing and you’re out. And the external shock could be anything — worse weather, a Fed hike, a public event, some swing you can’t imagine, and you get shaken to death.

Historically I did post-mortems on railways, canals and electricity. Generally their final equity ratio was around 1 to 3, and canals 1 to 1. It always ends the same: after enough of it, people find returns aren’t coming back, and then a run on it bursts the bubble. Debt has to be settled, and equity certainly gets wiped first.

So AI’s current state can bear 30% or 40% of volatility. This time stocks fell 30% or 40% and only levered people blew up; the industry wasn’t really affected: whoever was buying servers is still buying, and no company went bankrupt because the share price fell.

Raymond: You broke out one very specific number in your writing: the peak of CapEx will be $4 trillion — using free cash flow or equity as collateral to borrow more, rolling it into something bigger, bringing more revenue, left foot on right foot, with $4 trillion back-derived from a ratio to GDP. If the peak is $4 trillion, then today’s memory is dirt cheap.

Tang Yibo: Yes, though not absolutely. Memory is a very mixed matter: returns are very high, and yet these people still won’t expand capacity — in a normal cycle high returns bring expansion, which drives it to the social average return. It’s just that AI came too fast this time, and there are only these few players, with some tacit understanding among them.

But from algorithmic optimization you can see DeepSeek’s and Kimi’s attention mechanisms have cut KV cache requirements a great deal. So I think a supply-demand imbalance this large in memory may end in the short term through algorithmic optimization. Even at $4 trillion, memory’s share may be lower than it is now.

Raymond: So you’re expressing extreme optimism about the CapEx supercycle — through 2030, to $4 trillion — while at the same time being less certain about memory. You’re assigning a very high weight to “algorithms can overturn the memory supply-demand gap.”

Tang Yibo: Yes. Because Kimi’s linear attention and DeepSeek’s architecture both exist because they were so short of memory. And foreigners aren’t willing to do this — they’re still exploring the frontier, and they aren’t short of money, so they don’t need to economize. And it’s possible that after optimizing, the next paradigm — say continual learning, where a model adjusts its own weights within the application context — brings a further increase in memory demand. Far too hard to judge.

My worry about memory is simply that its share has become too high. Last year if you’d asked someone “can algorithms optimize memory away,” they’d have thought you were joking — memory was freely available. But now its share of AI hardware may have gone from under 10 points to 20 or 30. Your ROIC is simply too high, and everyone wants to take a slice off you.

Raymond: But at the same time, that sentence applies to Nvidia too, doesn’t it?

Tang Yibo: It does. Look, Nvidia isn’t rising anymore either. Because it was the strongest at the start with a very high share, 50%, so it’s very mixed: profits are rising too, but the GPU’s share is certainly falling — whether squeezed by storage, or optimized by others, or TPUs coming out. That’s because you’re too strong. If you weren’t that strong you might not get suppressed. That’s also a cycle: when you’re very bad you may become very good, and when you become very good you may get hit, and after being hit you may rise again.

10. “Fifteen years of electrification investment, and no productivity gain in the first ten”

Raymond: Back to the core question. Today, from where the CapEx money comes from to changes in token prices, everything reflects one contradiction: whether demand growth is fast enough. And it lands on how big this TAM actually is — if it’s only narrow coding, that market is already fully penetrated and we should just go buy hogs; if this is a vast sea of stars at 0.5% penetration, then hurry up and buy six times more memory. So why electrification, rather than benchmarking against something else?

Tang Yibo: A fairly simple standard is again the magnitude of CapEx.

Electrification’s CapEx was 2% of GDP, sustained for 15 years. And in the first 10 of those 15 years there was no productivity gain, merely factories using more electricity. Because the nighttime use case then was lighting and the daytime use case was streetcars, and nothing else — not that there were no refrigerators, but that there were only electric lights and streetcars; factory motors were swapped in later. That 2% of GDP went into grids, power plants, transmission, that kind of infrastructure. In those ten years, total factor productivity — the part attributed to technology after stripping out capital and labor growth — contributed only 1% to GDP.

For the internet, optical fiber was 1.3% of GDP in CapEx and was only invested in for 4 years, and later people found the fiber they’d built wasn’t being used, so they stopped. AI’s figure this year in America is roughly 2%, which already exceeds the internet.

Another angle: what did the internet actually change? The efficiency of distributing and processing information rose. But with electricity, imagine a world without electricity and a world with it — the difference is enormous. Not just energy and motive power: factory power going from steam to motors produced the assembly line; and after electricity came all kinds of household appliances. Electricity is a very fundamental thing, and a great deal extends off it. Whereas the internet raised efficiency, and what it created wasn’t enough — entertainment, obtaining information, ordering things faster, all based on the efficiency of information circulation.

But what is AI? AI is intelligence — leaving aside that it can also become a robot doing physical labor. If AI is intelligence, I think its magnitude should be the same as electricity’s: we can already see it replacing large amounts of productive behavior; and second, it may change the organizational form of production, as with whether Claude Code will change a whole enterprise’s organizational structure the way the assembly line did.

11. “Replacing people isn’t AI’s limit”

Tang Yibo: And once the price of intelligence becomes lower and lower, entirely new applications I can’t currently imagine may appear.

I especially like this example: the radio appeared after 1920. Electricity appeared, radio appeared, and the company RCA rose more than three hundredfold after 1920. That’s an electricity-native application. But in the decade from 1900 to 1910, investing 2% of GDP in CapEx every year while all you could see was electric lights and streetcars, you couldn’t possibly imagine why a thing like the radio would exist.

The radio is really a packaging of electricity, a wrapper on electricity — based on electric power, plus some radio communication, some technologies combined, and it becomes a very native application whose penetration keeps rising, and the stock went up more than three hundredfold.

I think we will certainly see genuine AI-native things in the next five years — products that couldn’t be built without cheap, extremely strong intelligence.

The idea of replacing people is far too simple. Replacing people isn’t even AI’s limit. Was the electric light only replacing the kerosene lamp? Was the streetcar only replacing the horse carriage? Afterward came cars, trains, high-speed rail, and endlessly varied appliances. Replacing people is only its starting point. Because its intelligence level already clearly exceeds most people’s; it’s only because it’s expensive that you can’t yet replace at scale. If it’s cheap, it won’t only replace, it will certainly create. If you have a pile of very cheap geniuses, what that world looks like is something you have no way of imagining from this point in time.

Of course you could also say I’ve presupposed a conclusion — that I already imagine it exceeding humanity’s limits, and only then went and compared it with electrification. But that’s the picture in my head.

Raymond: Playing your opposite today, my biggest takeaway is this: if I disagree with you, I’m essentially saying I deny the possibility of a big breakthrough in AI in the next six months. That’s an option. If I’m not on your side, I’m really selling an AI call option. I don’t know what it’s worth; it might be worthless — but by what you just said, worthless is also worth something. Thinking about it that way, I find it a bit frightening.

Tang Yibo: All the numerical frameworks we’ve discussed sit in the big CapEx years of 2026, 2027 and 2028, extending on some questions to 2029 and 2030 — so a five-year span.

But we heard a formulation recently from Wang Tianfan. He says human civilization’s first opening and closing took 3,000 years, the second 300 — the industrial revolution — the third 30 — the internet and mobile internet — and the fourth may be three. Because it’s a continuously accelerating process: each civilization’s culmination makes the next civilizational iteration faster and penetrate faster. So it’s possible the cycle we face runs its full course in five to ten years.

Raymond: Then let’s treasure today’s opportunity.

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.