Anthropic’s IPO Push: Is a $2 Trillion Valuation Expensive?

Models that depreciate the moment they are landed · Existing customers spending nine times as much · A one-person drug company · The next coding opportunity is still coding

“The large-model business is very, very brutal. I think of it more as the seafood business: the moment the catch comes out of the water, it starts losing value.” — Raymond, on the shelf life of frontier models

Guests: Aaron (Zhou Jiuzhou) | Host of 涉市未深 and public-market investor; Raymond Huang | Host of 苔藓之火 (Mossfire) and technology investor

Host: Diane (Ding Jiao) | Host of What’s Next｜科技早知道 and co-founder of Shengdong Huopo

A Brazilian company came to Shanghai to look for Chinese AI models. It mainly uses Anthropic and OpenAI today, but plans to move half its token usage to open models by the end of 2027. Customer service, sales, image editing: save money wherever a task already works, and reserve the expensive models for harder development work. Raymond listened and asked: if Anthropic releases a smarter model tomorrow, won’t your engineers still want the best one? The company did not answer.

That calculation is much the same one running through this conversation. Is a cheaper model good enough? How much smarter is the expensive one, and how much more will customers pay? Only then do you get to the question of why anyone is discussing a $2 trillion valuation for a five-year-old company.

Anthropic publicly filed its S-1 on October 26, 2026. First-half revenue was nearly $16.3 billion, about 14 times the year-earlier level. Raymond, a former investment banker, sees the pricing advantage of listing first—but also a “seafood business,” where a model’s shelf-life clock starts ticking as soon as it is landed. Aaron, a longtime public-market investor, does not entirely agree. He sees the rapid replacement cycle as a phase, and is more interested in model companies developing drugs themselves, solving problems alongside customers, and sharing in the resulting returns.

Diane brings both investors back to practical questions: if there is plenty of cash, why list? She also wants to understand how quarterly operating profitability and a first-half net loss exceeding $80 billion can appear in the same financial statements. When the conversation turns to Chinese models, the two people who had just been agreeing part ways again. This is a crossover between What’s Next｜科技早知道, 苔藓之火, and 涉市未深. The $2 trillion figure is a valuation scenario under discussion, and each investor has his own full-year revenue forecast. Whether those expectations materialize is what the argument is about.

The following conversation has been translated from Chinese, edited, and condensed for readability.

1. If Cash Is Plentiful, Why Race to Be the First to List?

Diane: We used to think of Anthropic as the industry’s number two. Now it seems to be moving toward the lead. Why choose this moment to go public?

Aaron: People assume OpenAI is number one largely because it appeared earlier. But look at a specific business, such as AI coding, and Anthropic is the leader. I see this IPO as the coding leader going public—and it will not be confined to coding.

My understanding of the company has changed. It has an opportunity to become the underlying infrastructure for how industries operate. As for timing, I think there is an element of racing to get there first: being the first listed company in a category carries reputational value in itself.

Zhipu and MiniMax were also racing to list. Beyond raising money, companies consider their place in the capital markets. Getting the highest possible valuation right now may not be Anthropic’s only consideration.

Raymond: Let me add something. I spent a long time in investment banking, mainly helping these companies go public. One example that stuck with me was in 2014: JD.com and Alibaba were both preparing to list. Both were good companies. Why were they still discussing who should go first?

Because the first company has more pricing power. To public-market investors, you seem to have sprung out of a rock: I could not buy this company before, and now I can. You start the roadshow. You tell the story, define the category, and first explain which metrics investors should look at. Once investors believe it and buy, you have established the reference point for that market.

The second company has an awkward position. Investors ask: I already bought the first one. You came later, so shouldn’t you offer a discount? Otherwise, why should I buy you? Within the same category, listing first often helps the valuation a little.

The pressure on the other side is OpenAI. It is still raising private capital and lifting revenue expectations. Anthropic can say it is in no hurry to list. But if OpenAI raises the money and pushes forward, the pressure comes back. The floor underneath keeps rising. Your decision not to list does not mean your competitor stops moving.

Diane: But it only completed a financing round in May, and still has plenty of cash. So this is more about capturing the valuation high ground?

Raymond: I think that is part of it. But there is something else about listing: people tend to focus only on how much money is raised in the IPO. More important, becoming a public company opens up more financing channels.

Some companies do not even sell new shares when they list; they only sell existing shares and raise no new money at all. But once listed, you may obtain a better credit rating and find it easier to borrow. With good terms, debt can be more direct and more economical than issuing equity. When you need money later, you have more routes available.

Aaron: Let me pose a question: is the large-model industry asset-light or asset-heavy?

Raymond: I think it is asset-heavy. Across society as a whole, it is extremely heavy—a large manufacturing business. The question is simply: who pays? Some companies let someone else buy the hardware and hold it on their balance sheet; others are willing to swallow the whole thing themselves.

Aaron: I also see it as a very asset-heavy business. My judgment is that it needs to accumulate sufficient funding, and may eventually make heavier investments itself.

Its partners are simultaneously distribution channels, capital providers, shareholders, and infrastructure suppliers, with competing products of their own. A model company can hardly hand its entire fate to any one of them. Over these years of AI development, upstream chips and data centers have captured a lot of the profit. Nobody can do this without money. A little more cash today may mean a little more room in tomorrow’s competition.

2. How Do You Decide Whether a Company Is Worth $2 Trillion?

Diane: Investors do not have to buy Anthropic. There are other technology companies to choose from. Can it support a valuation of $1.5 trillion to $2 trillion?

Aaron: I disagree with the premise: I do not think there are that many alternatives. It is not directly comparable to cloud-service providers. Other model companies are the better comparison. OpenAI is comparable, but most public-market investors cannot buy it yet.

Since MiniMax and Zhipu are listed in Hong Kong, we can compare them. I will put an estimate on Anthropic’s full-year 2026 revenue: roughly $60 billion. The prospectus shows more than $16 billion in the first half, and I estimate more than $40 billion in the second. That is my own revenue forecast.

MiniMax and Zhipu are growing too. Taking a somewhat optimistic view, assume MiniMax does $500 million to $600 million for the year, and Zhipu more than $1 billion. Rather than comparing profits, let us simply calculate price-to-sales, or P/S: equity valuation divided by annual revenue.

If Anthropic is valued at $1.5 trillion, dividing that by $60 billion gives 25 times sales. Based on my estimates of the other two companies’ revenue and valuations at the time, Zhipu is near 30 times and MiniMax is lower, at around 13 times.

Revenue scale and model capability have to be considered together. Assume $600 million for MiniMax against $60 billion for Anthropic: that is a hundredfold gap, so pricing will differ. I can understand MiniMax trading at a lower multiple. But if Anthropic is at 25 times sales and Zhipu close to 30, what would I choose with dollars to invest? I would choose a global number one or number two at the lower valuation multiple. On that calculation, $1.5 trillion is supportable.

Raymond: I very much agree with Aaron. A normal person hears $1.5 trillion or $2 trillion and thinks: I have never heard a number that big. Your company has only existed a few years—what makes you worth that much? Why hasn’t my salary gone up?

But think about it seriously: revenue is very high too. By September or October, after the summer, US equity investors are usually already looking at next year’s revenue forecasts. Calculate P/S on this year’s revenue, then calculate it on next year’s, and you may get two completely different impressions. We cannot just sit there stunned by the $2 trillion number.

My guess is that annualized revenue could reach roughly $100 billion by the end of this year. I would use that figure to project into next year, then look at the corresponding P/S.

Also, reaching this scale is difficult in itself. Why do public-market investors often give big companies different valuations? Because by reaching that size, they have demonstrated things smaller companies have not yet demonstrated. The certainty around the business is different.

Then think about the ultimate buyers. Besides individual investors, there are big long-only funds and pension funds. In Boston, you might find many such funds on a single street, each wanting meaningful exposure to AI. Previously, there was no independent model company sitting there for them to buy. Now there is. That is a very different choice for them.

3. “The Moment the Catch Comes Out of the Water, It Starts Losing Value”

Diane: Looking at technology and competition, can it maintain its current advantage?

Raymond: We were just doing something public-market investors know very well: linear extrapolation. A hundred billion this year, $150 billion or $200 billion next year, then keep calculating forward. But public-market investors can certainly be wrong. In a technological wave this large, a lot of the time we have no idea what is going on.

The large-model business is very, very brutal. It is a large manufacturing business, but I think of it more as the seafood business: the moment the catch comes out of the water, it starts losing value.

The prospectus disclosed something I had never seen disclosed this way: more than 80% of third-quarter revenue came from model versions released in the preceding six months. Think about that—six months. If the next round goes badly, revenue could be significantly affected.

It combines the disadvantages of manufacturing with the disadvantages of seafood. You have to spend a lot up front, and once the product is made, it keeps becoming obsolete. So if you ask me whether Anthropic can stay ahead for years, I cannot be certain.

We have only a few underlying assumptions: it has done very well, accumulated compute, and can run more experiments; its people are among the world’s best, and we still see talent from other labs moving in. It is already involved in many customer workflows, and is accumulating data. In theory, it has a chance of remaining on the throne. But whether it can sit there securely, I do not know. That may be the biggest investment risk.

Aaron: The fact that 80% of revenue comes from models less than six months old is striking. But I think it is a phase, not necessarily a permanent condition.

Frontier models are changing too fast right now. Never mind six months: one, two, or three months can bring a transformation. Naturally, the leading provider earns most of its revenue from its most advanced models. The space for older models can be left to Chinese models, open models, or cheaper products from other providers.

One day, when iteration slows, things change. We used to talk about Moore’s law in chips: doubling every 18 months, again and again. Eventually that doubling slows. Then you find that 7-nanometer, 6-nanometer, and 5-nanometer chips can all be used, rather than becoming obsolete within a year or two. I think the present situation in models is a phase.

I have always had a question mark over the idea that “new kings” will constantly emerge. Whether a company keeps its position is not directly determined by when it lists. But its position when it lists matters a great deal.

Kuaishou is listed and ByteDance is not. That does not mean listing lets Kuaishou change its competitive position in short video. Conversely, if you already lead, capital-market funding and greater influence can help you strengthen that lead.

In this wave, we still keep counting Microsoft, Google, and Amazon. Look back twenty years: weren’t their names there during the internet bubble too? They lived through that wave. Unless they make strategic-level mistakes, leaders may experience periods of decline and resurgence. The new king does not appear; the old king does a few push-ups and, when things are going well, bounces around a little. That is the reality of US equities.

Raymond: Yes, that is important. Looking back at mobile internet, how many startups actually became new kings? Forget a new king—very few even became a “little prince.” Uber, Instacart, DoorDash: those count as major successes, but most of the new market value was still captured by the incumbents. Facebook was founded in the PC-internet era; it was not a new company born out of mobile internet.

I think this may only intensify over the next ten years. Large models require greater capital investment. Without tens or hundreds of billions of dollars on your balance sheet, you cannot even get a seat at the table. Technology moving fast is one thing. Who has enough money to keep running is another.

4. Customers Are Spending Nine Times as Much—What Are They Buying?

Diane: Another figure in the prospectus is net revenue retention of about 900%. It measures revenue changes within the same customer cohort. Those customers are spending more.

Raymond: SaaS—subscription software companies—often looks at this metric. Suppose you bought Zoom last year for $100. This year it sells you another feature and you spend $120. Revenue from that same customer is now 1.2 times as much.

The software company would already be delighted. Add a feature this year, another next year, and after a few years $100 becomes $200. Great.

But Anthropic’s ninefold figure is more about usage changing. Previously, you asked AI to write a little code. Now you hand it very long, complicated tasks. Make a website, make a video, then keep doing other things. Tokens—the units used to measure content processed and generated by models—get consumed in large quantities as those tasks run. The more complete the work you ask it to do, the more different usage becomes.

So I think token consumption will keep going up, possibly very quickly.

Aaron: The number is striking, but its business model is not exactly traditional SaaS. Fixed-price subscriptions account for only part of revenue. A larger portion comes from API calls—other products and corporate systems directly calling the model.

Last year, a customer may have just been experimenting and used very little. This year, having found that it solves a problem, the customer starts using it in production, and usage jumps.

Diane: So the low base last year is also part of the explanation?

Aaron: Yes. Moving from experimentation to real use can produce a huge increase. The number is impressive, but what matters more to me is that customers have really started using it heavily.

People keep talking about an AI bubble, but that is too vague. Could there be pockets of excess in data-center construction? In some chip stockpiling? Possibly. But Anthropic’s customers have gone from trying it out to using it intensely over the past year. That is actual demand.

It is hard to put a bubble label on the entire AI industry on that basis. Read this prospectus and you see that the explosion in real demand is extraordinary.

5. After Selling Models, Why Start Developing Drugs Yourself?

Diane: We used to discuss what business models could really make money beyond coding. Aaron, did reading this S-1 change your thinking?

Aaron: I think my earlier understanding of model companies was mistaken. There were too few companies we could research deeply using public information. Looking closely at how Anthropic and OpenAI do business changed my thinking.

I do not see it as a simple model provider. Take one step further and it is a foundation for how the world operates. Take another, and I am not exaggerating when I say it could become a business partner and co-creator across industries.

It moves from selling you a model to take home and use yourself, to working with you to solve something using the model, and then sharing the business proceeds.

There are already signs of that. Anthropic is researching new drugs with pharmaceutical companies, and has acquired a one-person company called Mithril Medicine, beginning to participate in drug development itself. It has moved from providing tools to doing the work itself.

The company I imagine in the AI era might have one, three, or five people, yet build a very large business operation.

There is also Ode with Anthropic, an AI services company formed with other institutions to help industries apply models to actual problems. This goes a step beyond an FDE, or forward-deployed engineer: not merely deploying something for you, but solving the problem with you.

Imagine that one day it develops a treatment for a rare disease, and external partners then handle clinical trials, approvals, and commercialization. What share of the returns should it receive?

Today, some rare-disease treatments attract insufficient development investment because the economics do not work. If AI makes it possible to advance many more development programs, they may become economically viable. What share should Anthropic receive? That depends on the practical value it actually creates.

Investment banking is similar. Many teams are still making PowerPoint presentations and preparing roadshow materials. The work that truly requires people dealing with other people may account for a smaller share than assumed. Other industries face similar changes. So I see its potential as more than a company selling models: it could be a foundational company involved in many businesses.

Raymond: I agree with Aaron, but with a slight difference. I strongly believe large models will underpin many businesses, like water, electricity, and gas. But those utilities are not necessarily all monopolized by one company. China has a highly developed power system, yet different grids; different countries have their own utilities. Anthropic will be an important provider, but may not capture everything.

Drug development and FDEs moving further into the business also show model companies looking for new ways to monetize. Today, APIs account for the bulk of revenue, but opening an API exposes a model to imitation and distillation. The product turns into seafood as soon as it comes out.

Suppose you keep your most advanced model in-house rather than offering an API, and use it to develop drugs, protein structures, or new materials, charging for the results. Could that extend its shelf life? What you sell changes, and the business model may change with it.

6. “The Next Coding Opportunity May Still Be Coding”

Raymond: Diane’s earlier question is really, “What comes after coding?” That is a trillion-dollar question. I often say: the next coding opportunity may still be coding.

Many people understand coding as a programmer writing software with AI helping to autocomplete. I think that is too narrow. Much of what you do in the digital world can ultimately be accomplished through programming.

For example, I build a chatbot to keep my parents company. The thing they experience is the result of coding. But do my parents feel, “I am using a programming tool”? Making videos, building tools, and automating workflows can be similar.

People therefore often compare the ceiling for the coding business with programmers’ wages. I think that is the wrong comparison. Look across white-collar work and corporate R&D spending. The pool you are comparing it with is no longer the same pool.

Coding itself still has a very high ceiling, nowhere near reached. Beyond it, there are many things we have not yet imagined.

Diane: So the addressable market includes programmers using AI, ordinary people accomplishing tasks through coding, and model companies entering fields such as drug development. What other figures make you more positive about this business?

Raymond: One number made a deep impression on me. In April 2026, the value of Anthropic’s internal Claude usage at official list prices exceeded its monthly personnel expenses excluding stock-based compensation.

Many companies may eventually reach a similar point. How? One route is fewer people and lower payroll; another is spending more on AI. Buying closed models or using open ones are both ways of embracing a new form of productive capability. As AI takes a growing share of spending, the share spent on people may fall.

My prediction is that technology companies may begin reaching this point in 2027, with other industries perhaps doing so in 2028. We can discuss the social consequences separately. But I think there will be major changes in how companies divide work between people and AI, and how they spend money.

Aaron: I also look at revenue per employee. Using a rough calculation on the basis available at the time, Anthropic’s revenue per employee was much higher than that of the other two model companies—a very large gap.

That raises a question. If all these companies competed in one market without barriers, that scale advantage would be very powerful. But the real world has different markets and customer groups, so individual companies may still find room for themselves. We cannot forget that premise when comparing China and the US.

Diane: There is customer diversification too. Revenue outside the top five customers rose from 64% in the first quarter of 2026 to 79% in the third. The company is becoming less dependent on a handful of big customers.

7. Quarterly Operating Profit, but an $80 Billion-Plus Half-Year Loss?

Diane: Second-quarter operating profit was already close to $1 billion, yet first-half net losses exceeded $80 billion. How should we understand those two numbers?

Aaron: A large portion is an accounting, non-cash loss. Of the more than $80 billion net loss, over $70 billion came from fair-value adjustments—losses at the accounting level.

Raymond: Some accounts use “Anthropic lost another $80 billion-plus” as their big headline. Whenever I see that, I unfollow. They have not explained the accounting relationships at all; they just use a giant number to scare people.

Aaron: Well, they want clicks. Without that headline, people might never open the article.

Raymond: Our podcast could say, “Why is Anthropic losing $80 billion, yet going public at $2 trillion?”

Aaron: “Shocking: Anthropic’s massive $80 billion loss.”

Diane: Two big numbers, and you have a headline.

Raymond: Without reading the prospectus, you really do not know how it got here. In 2023, revenue was only a little over $40 million, and R&D spending more than $400 million—ten times revenue.

More than $400 million hardly sounds like money today. It was different back then, with revenue only a little over $40 million. Investors would ask: you spent all that money and can predict the next token, so why can’t you predict where the revenue is? Who is the next customer? You can’t predict the next customer?

Now look at the first half of 2026: R&D spending was 59.8% of revenue for the same period. Is that still high? Of course. But going from ten times revenue to 59.8% is already an enormous change.

Aaron: Another number I find important is how much revenue a dollar of compute spending brings in.

If you invest one dollar and get only fifty cents back, you are continually bleeding cash. By the second quarter of 2026, roughly one dollar of compute spending was producing two dollars of revenue. That is an important turning point.

If that quarter’s situation and trend can continue, further compute purchases and long-term contracts have support. But we cannot draw the conclusion from one quarter alone. We have to see whether it is a flash in the pan.

8. If a Shareholder Is Also a Supplier, Is That Automatically Circular Financing?

Diane: Roughly 48% of revenue is generated through cloud-platform partners. Could such a high channel share be a risk?

Aaron: There is certainly potential risk, but I still see an environment where cooperation and competition coexist. Anthropic is not entirely tied to one provider, and no single provider is entirely tied to Anthropic. Both sides are developing their own models and channels.

Think of it as direct sales and distribution. Early on, using major cloud platforms to reach more customers seems perfectly reasonable. Cooperation among these companies alone cannot prove that the business is circular financing, much less establish financial fraud.

If Anthropic’s own channels become strong enough, it can reduce its reliance on others. The question to watch is whether it is expanding its own customer relationships.

Raymond: People read the news and immediately wonder: why is the same company an investor in Anthropic, an upstream supplier, and downstream as well? It makes chips on one side and helps sell models on another—is something wrong? But inside a big company, those decisions are often made by different departments.

The investment department handles strategic investments and asks which model companies are worth investing in. The cloud business needs to sell things and generate revenue. It may invest in several providers and sell several providers’ models. Make more money where you can; there is no need to pick a camp every time and exclude everyone else.

If everything had to be exclusive, you would end up with separate islands. Every company would have to do everything itself. What would be the point of cooperation then?

Diane: What do you see as the biggest risk? The prospectus’s risk factors even discuss extreme consequences AI could bring.

Raymond: Let me first explain why risk factors are there. Prospectuses include that section, and lawyers advise what needs to be disclosed clearly. If investors lose money later, they may say: had you told me, I would not have bought; you did not disclose it, so now I am suing. Companies therefore list all kinds of possibilities, and lawyers also reuse templates in many places.

What I really care about is whether new use cases and large revenue opportunities will appear. Who exactly will pay you serious money?

After all, it has signed very large compute contracts with suppliers, and has a lot of money to spend in the future.

Diane: Customers have not signed contracts of the same length with it, while its compute commitments are long-term.

Raymond: Right. In Chinese financial terminology, it is a maturity mismatch. You have made long commitments upstream, without the same assurance about customers’ future usage.

But building a data center really does cost a lot. You cannot wait until someone has built it and only then say you want to buy capacity. They might not sell it to you. Some of those commitments have to be made in advance.

If revenue keeps rising, the pressure of absorbing those commitments will be different.

Aaron: I think the more likely risk is not that all AI demand suddenly disappears. Customers try it for six months, find it completely useless, and collectively stop using it next year—that seems relatively unlikely to me.

What seems more likely is that demand is still growing quickly and adoption is real, just not as fast as people imagined. For a company carrying very high growth expectations, even 30% revenue growth next year could fall well short of expectations.

Revenue can keep growing while the valuation falls. I lean toward that scenario: real demand exists, but it develops more slowly than imagined and ends up seriously disappointing expectations.

9. Beyond Multiples of Annual Growth, What Should Chinese and US Model Companies Be Compared On?

Diane: MiniMax and Zhipu are also growing revenue quickly, but their scale is clearly different from Anthropic’s. Aaron, you have followed both for a long time. How do you compare them?

Aaron: Year-on-year growth alone is not that meaningful. When a new industry moves from almost zero penetration to 10%, the growth numbers are usually extremely high. That was true of new-energy vehicles early on.

Between 10% and 30%, growth remains fast and people begin discussing risk. Beyond that, more competitors enter, competition intensifies, growth slows, and profit margins may narrow. We are still early today. Growth of 1,000% versus 400% does not directly tell you who will survive.

I care more about several questions. First, in which quarter can cash flow turn positive, and can it stay positive? That determines whether a company can survive when financing conditions deteriorate.

You cannot assume you can raise money every year. When financing is easy, companies may rush to secure it because they know the market may not be as favorable after the next turn. Will the same volume of capital support data centers next year? Will public markets offer equally attractive financing terms? Neither is certain.

Second, can the money invested actually bring revenue back? Is the relationship between compute spending and revenue improving? If it cannot cover the spending, the business model needs more iteration.

Diane: Does that mean some already-listed companies are still at a very early stage?

Aaron: Different stages of a company’s development call for different standards. Angel investors and early-stage venture capital can accept very little current revenue alongside large R&D spending. In public markets, investors need another set of standards.

Hong Kong’s stock market is looking more and more like venture capital. If things go further, it might start looking like an angel-investment market.

Raymond: I think it depends on where the money is. The US private market has enough capital to support a company until it becomes very large—even reaching a trillion-dollar valuation—before it lists. Some Chinese companies go through angel, Series A, and Series B rounds, but private funding cannot support that kind of scale, so they have to keep raising money in public markets.

Both are called listings, but the US company may already be much further along, while some Chinese companies are effectively still raising a growth round. They have reached different stages. Listing alone does not make them identical businesses for comparison.

Another difference is whether users will pay and how much companies will pay. There is a substantial difference between China and the US, inherited from the SaaS era. That reality is part of why revenue looks as it does today.

What the market liked in the first half of this year was that Chinese companies could rapidly approach the frontier with fewer chips and lower costs. That is impressive. But looking forward, can they keep catching up this way? Model capability, data, and access restrictions are changing. Whether yesterday’s methods will keep working tomorrow deserves a question mark.

Aaron: Another analogy needs unpacking. We often call AI a utility, but I do not think every tier of model will become a public utility everyone can afford.

Ordinary chat can be broadly accessible. More capable models that generate more commercial value may come with higher prices and barriers to use. Think of refined fuel at the 97-octane grade. If I cannot afford it, I can use 93. If I do not want 93 either, can I use something cheaper, like diesel? There are options.

So what concerns me more is how capability becomes something customers will pay for, and whether that money can support continued investment.

10. Is the Opportunity for Chinese Open Models Just That They Are Cheaper?

Raymond: Ever since DeepSeek burst onto the scene, some people have said Chinese models are so cheap and US frontier models so expensive that nobody should use the American ones anymore. That is not what happened. There is still strong demand for smarter models.

Think about it: if a company has the budget, wouldn’t it want to hire a better student? If you can get a Peking University PhD who really produces different results, you will pay more. Today we may still be hiring elementary-school students. With each model update, we move to middle school, then high school, then undergraduate level, only later distinguishing among China’s Project 985 and Project 211 universities, then moving on to master’s degrees and PhDs. Capabilities will keep advancing, and companies will pay for those differences.

Diane: But some companies have built good products with Chinese open models. Not every task needs the strongest closed model.

Raymond: I happened to meet a Brazilian company last week. Most of its model spending is on Anthropic and OpenAI. By the end of 2027, it plans to put 50% of its—not spending, token usage—on open models.

That is why I met a Brazilian company in Shanghai. They were there to evaluate Chinese models and make purchases.

A lot of their workflows already work: customer service, sales, image editing, video. Today’s models can handle those things. Switching to open models is about lowering costs and improving efficiency. Harder internal development would still use frontier models. That arrangement is easy to understand.

But I asked: if Anthropic’s next model is clearly more capable, won’t your employees and engineers still ask for the best one? They did not answer that question.

If you assume Chinese and American models are only two or three months apart, then ordinary tasks on cheaper models and difficult tasks on expensive ones make a neat calculation. But what if the gap widens? What if the frontier model suddenly gains a capability we never imagined—as if it had not only earned a Peking University PhD but instantly completed postdoctoral work at Tsinghua, Peking, Fudan, and Shanghai Jiao Tong as well? Then the economics change again. Even someone coming to Shanghai today to buy at scale has to consider whether Chinese models can stay close enough.

Diane: Many startups already work that way: the best models for R&D, cheaper ones for everyday operation. You do not need a PhD for work a middle-school student can do. I think Chinese open models will have room of their own for some time.

Aaron: My hesitation is about where their valuations come from.

When the market calls a company “China’s Anthropic” and gives it a high valuation, it is because investors believe it may approach the world’s first tier. The high valuation is not a reward for being cheap; it reflects the belief that the company can catch up.

If the gap subsequently stops narrowing, you have to ask again: is the valuation based on price-performance, wider usage, or the ability to solve concrete business problems?

I originally had greater hopes for the third route. A company adopts a model to solve real business problems, not merely to run a program. In the end, it has to calculate costs and returns. If that works, value can be created.

But Anthropic is also moving down that path today—working with businesses, solving operational problems, and even developing drugs itself. Open-model providers will also face competition from frontier closed models in practical deployment. If the latter solves the same problem better and generates a larger return, customers may still be willing to pay the higher price.

That makes me re-examine my judgment. I was very enthusiastic about Chinese models narrowing the gap. But I also have to ask whether the signals we now see are moving in the direction I originally imagined.

Raymond: I may not be that pessimistic. I do a quarterly review of Chinese models and go through the major models each time. The impression really changes from quarter to quarter.

Earlier this year, the gap felt enormous, with no end in sight: compute was constrained, data was constrained, and the more you failed to catch up, the further behind you seemed. Then new versions came out and got very close, before the gap widened again. Every quarter, we discuss how many months apart they are now. It moves in waves.

I still believe in China’s density of talent and technical ability. Teams such as DeepSeek are also willing to open-source many inference and engineering methods. I think there is much more ingenuity there. Stocks are another question, but technically I am still inclined to be optimistic. There is room to keep moving forward.

Aaron: Everyone likes asking that question.

Diane: Think about phones too. Android and Apple have open and closed ecosystems, and the needs they serve are not exactly the same.

At a Qualcomm event recently, I saw open models already running locally in glasses and small AI devices, solving basic everyday problems. You do not need to call Anthropic or ChatGPT for everything. These two kinds of ecosystem will reappear in different fields.

Raymond: Yes. But I think this industry is still early. I often compare today with 2014. The iPhone had been around for a few years, and everyone was still upgrading furiously. Until about 2018, I was buying a new phone every year. Then I stopped changing it so often, because upgrading seemed little different from not upgrading. Everyone may have a moment like that. I do not think models have entirely reached that point yet.

Diane: If I am only using it for chat, the improvement feels less dramatic than it did a few years ago. But agents—AI systems that execute tasks—still differ greatly on long tasks, including how well their harness works: the system organizing tools, context, and execution around the model. Different use cases really do produce different perceptions of improvement.

11. Does Being Positive on Anthropic Mean Buying at the IPO?

Diane: If you managed a fund, would you allocate to Anthropic after it lists?

Aaron: It depends on the valuation. A decision to buy is made at a particular moment. You have to consider the market that day, how excited people are, and what price they assign to the company.

Even something as scarce as SpaceX does not automatically mean “you must buy at the IPO.” Good companies also experience excessive expectations and price swings. You have to see how excited people are before listing and what price they are offering.

Some pre-listing trading can serve as a thermometer. If prices are already very high and sentiment very hot, it is worth taking a closer look.

Over ten years, I lean toward holding companies like this. Over the first month or two after listing, it depends on whether enthusiasm has gone too far. If sentiment is excessively excited, I would wait a little. There are always opportunities.

Raymond: My view is similar to Aaron’s, though perhaps more optimistic. I would lean toward holding it; the difference is whether to be overweight or underweight.

In a technological wave this large, arguing over whether it is worth $1 trillion, $2 trillion, or $3 trillion today matters less to me than participating. Of course, the price will be volatile. It could move sharply up and down over the next several weeks.

This is an entirely new company in an entirely new category, so judgment is difficult. The market may suddenly drive it down and find a narrative explaining why Anthropic is finished. But that could also be a buying opportunity. Or the market may suddenly push it much higher, which is a different kind of opportunity.

Today we have discussed both the reasons to be positive and the things that worry us. It is inherently complicated, so I believe the market will keep offering opportunities.

Precisely because it is complicated, I think it is worth spending more time understanding it properly.

Original podcast: Anthropic’s IPO Push: Is a $2 Trillion Valuation Expensive? Claude’s Growth, Costs, and Competition | S10E34

Sources: Anthropic S-1 prospectus; Anthropic’s announcement on drug development and Mithril Medicine.

This article preserves the views and judgments expressed at the time of recording. It is for informational discussion only and does not constitute investment advice.

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.