HomeArticle

Anthropic CEO "recommends stocks via live stream", the entire screen is filled with disclaimers.

36氪的朋友们2026-11-16 08:12
Dario himself answers what the exact valuation of Anthropic is

At midday local time in the United States on November 13, Dario Amodei, co-founder and CEO of Anthropic, and Jamie Dimon, CEO of JPMorgan Chase, held an IPO-themed dialogue in New York.

For Anthropic, which is preparing to land on the public market, this dialogue is not only a listing roadshow, but also a window for the outside world to observe its business model, competitive strategy and long-term development plans.

Amodei attributes Anthropic's competitive advantage to "talent and culture".

In his view, talent and culture determine technology, technology determines products, and products then determine market expansion and sales. The emphasis of enterprise customers on security and trust also forms positive incentives that align with Anthropic's business objectives. He believes that security is not an additional condition for the commercialization of enterprise AI, but the foundation for customers to choose long-term cooperation.

The boundary of the AI market is also constantly expanding. "Every time we expand the frontier, we will create a brand new market," Amodei said.

Amodei believes that chatbots and programming are only the markets opened up by the current model capabilities, and there is greater potential space in fields such as life sciences, robotics, materials and cybersecurity. Even if some model applications are gradually commercialized, Anthropic can use the technical accumulation of cutting-edge models to train small models with lower training costs and higher efficiency.

However, the stronger the model capability, the greater the computing power investment behind it.

Amodei admitted frankly: "Computing power is the deepest business problem Anthropic is facing."

Computing power resources usually need to be procured one to two years in advance. Buying too much will increase the financial burden, while buying too little may affect customers' reasoning needs and the R&D of next-generation models. Anthropic currently adopts a relatively conservative procurement strategy, first locking in part of the computing power, and then supplementing it through the spot market according to actual needs.

The IPO provides more capital space for this long-term investment.

Amodei said that increasing cash reserves to support computing power construction, and seizing growth opportunities in fields such as life sciences and robotics are important reasons why the company chose to go public at this time.

He also mentioned that he heard a piece of advice when preparing for the IPO: "Never let the company's valuation be even one dollar higher than what it is really worth." This is in the same line with Anthropic's principle of "underpromise and overdeliver": do not hype the model capability in advance, but show it to the market after the technology and products are truly mature.

From enterprise customers to computing power procurement, from model commoditization to IPO valuation, this dialogue actually touches on several key issues that Anthropic must face after listing.

What Amodei hopes investors can see is not only the current growth rate of a model company, but also whether it can control risks while expanding its business scale, and convert AI capabilities into long-term economic value.

The following is the refined version of the dialogue transcript between Dimon and Amodei:

01

Talent and culture are Anthropic's moat

Dimon: You used to work at OpenAI and are an early participant in the AI field. Now more and more companies are entering this field. How does Anthropic differentiate itself from other competitors?

Amodei: I have worked at many major companies in this industry. Looking back on these experiences and thinking about why some projects succeeded while others ended in failure, I gradually realized that the core factor that determines a company is talent and culture.

Talent and culture determine technology, technology determines products, and products then determine market expansion, sales and other businesses of the company. The public image and external reputation of the company also come from the people who create the technology and the starting point of their work.

When founding Anthropic, the other six co-founders and I determined that we should put talent and culture first from day one. Every new employee has to go through a culture interview. Even if someone has excellent technical capabilities, we will not hire them if they do not agree with the company's values, are not good at cooperation, or are used to stealing the show.

This recruitment standard will make recruitment difficult in the short term, but in the long run, it can build a team that truly works in collaboration. Many companies are not short of outstanding talents. The problem is that different teams compete with each other, restrict each other, and consume a lot of energy in internal struggles.

At Anthropic, once an important goal is determined, no matter it is training a new model, achieving security goals, developing products, or promoting policy affairs, the whole company will collaborate around the goal.

Personnel retention is a direct manifestation. All seven co-founders are still in the company to this day, and more than three-quarters of the first 100 employees who joined the earliest still work at Anthropic. We believe that the company has a relatively good employee retention rate in the industry. Some former employees will even rejoin. There were two important product leaders who left the company, but they came back in less than two weeks.

I spend about 40% to 50% of my time on talent and culture building. Every two weeks, I will participate in an internal activity for all employees called "Dario Vision Quest" (DVQ). I will directly introduce the actual situation of the company to the employees, without hiding problems or using official jargon, and encourage other managers to do the same.

I want employees to believe that the management is candid, and the management can also trust the employees. This two-way trust is very important for the company.

02

Claude enters enterprises, AI begins to take over more work

Dimon: JPMorgan Chase has 250,000 employees, who use Claude in different ways every week. How do other Fortune 100 companies mainly apply this technology?

Amodei: Claude is entering all walks of life. I think one of the most important characteristics of this technology is its general applicability. As model capabilities continue to improve, more and more work that could not be automated in the past is becoming feasible.

Programming is the first field where enterprises adopted AI on a large scale. I wrote a lot of code in the past, and many co-founders also have programming backgrounds. It is amazing to see that current models can complete such complex programming tasks. Almost all Fortune 100 companies have a large number of programming needs, so programming naturally becomes the entry point for them to initially adopt AI.

Subsequently, the scope of application gradually expands to a wider range of knowledge work.

In the field of financial services, Claude can not only handle basic paperwork, but also assist in complex strategic work; in the insurance and legal industries, it can be used for contract drafting, contract review and identifying potential vulnerabilities.

Life sciences is a field I am particularly optimistic about. Claude has been used to write clinical research reports, assist basic research, predict protein binding, improve manufacturing efficiency, and support basic biological discoveries.

In the e-commerce sector, some customers deploy Claude as a consumer-facing agent. The same is true for the manufacturing industry, where enterprises begin to use Claude to improve operational efficiency, with applications covering semiconductor manufacturing, physical facility construction and other fields.

I didn't know much about the manufacturing industry before, but by communicating with CEOs of manufacturing enterprises and seeing the practical application of Claude in these enterprises, I have gained a new understanding of this industry.

03

Agents enter enterprises, security becomes the first threshold

Dimon: For regulated enterprises, what preparations should be made before allowing agents to access sensitive data and systems?

Amodei: The first priority is security. The capabilities of agents are improving rapidly, and they are already able to handle highly difficult tasks such as life science research and complex software architecture. What really restricts technology development, internal deployment and customer use now is not only the capability itself, but how to deploy these systems safely.

We don't want agents to lose control. That's why Anthropic devotes a lot of resources to carry out security research. Security and alignment must start from the model training stage, rather than remedying after the product is released. The company has a huge alignment team, and we attach great importance to security issues. We reassess our investment almost every few months to further increase resources in security, safety and alignment.

From the very beginning of its establishment, Anthropic intentionally took enterprise customers as the focus of its business. This is not only a business choice, but also consistent with our security philosophy.

Enterprises value trust, long-term cooperation and their own reputation, and pay more attention to whether the model will cause serious damage than ordinary consumers. Therefore, serving enterprise customers can keep the company's business interests consistent with security goals. We hope to establish better incentives for the safe development of AI in this way.

Dimon: Can we really restrict the permissions of agents? For example, allow them to enter sensitive systems, but only perform clearly authorized operations A, B, C and D, and do nothing else?

Amodei: We can refer to the development of self-driving cars.

Self-driving is a technology of great value, and it is expected to be much safer than human driving in the end. But to achieve this, we must first establish correct protection mechanisms to ensure that the system is designed and deployed in a safe manner.

We cannot guarantee that self-driving cars will never have accidents under any circumstances. Similarly, we cannot guarantee that agents will not make mistakes in any environment.

The truly feasible goal is to make these systems much safer than human operations, reaching a safety level several times, dozens of times, or even hundreds of times higher. This is also the direction we are striving for.

04

900-billion-dollar computing power bet, Amodei's most troublesome problem

Dimon: Anthropic has disclosed long-term commitments for computing power and infrastructure of about 900 billion US dollars. The outside world has also noticed that currently about 80% of the computing power is used to train new models, and only about 20% is used for reasoning, while what customers actually use is mainly reasoning services. How do you ensure that you get enough computing power in the future while maintaining cost efficiency?

Amodei: Computing power is the most complex and important business problem Anthropic is facing.

The technological progress in this industry is very fast, but there is great uncertainty about how fast technological progress can be converted into business revenue. In 2024 and 2025, Anthropic's revenue grew by about 10 times for two consecutive years. Every year, it is very difficult for us to accurately judge what will happen in the next year, and as the business scale expands, the impact of this uncertainty is also growing.

Another problem is that computing power must be procured one to two years in advance. The quantity of procurement depends on our judgment of future demand. Buying too much may cause unbearable financial pressure to the company; buying too little will not be able to meet the reasoning needs of customers.

More importantly, computing power is also related to R&D. The resources used to train models today determine what kind of models can be launched six months or a year later, which in turn determines the company's ability to generate revenue in the future. Therefore, computing power procurement cannot only focus on the immediate reasoning needs, but also reserve R&D resources for the next-generation models.

We mainly rely on two aspects to reduce procurement risks.

The first is gross profit margin. The higher the gross profit margin, the more room the company has for adjustment when it procures too much or too little computing power. Simply put, the gross profit margin reflects the difference between the cost we pay for computing power and the revenue generated by each unit of computing power. The gross profit margin performance disclosed in the prospectus is quite good, especially compared with other players in the industry.

The second is cash reserve. Sufficient cash can help the company withstand large-scale computing power investment and cope with future demand changes. We currently have a large amount of cash and hope to continue to increase our capital reserves.

In terms of specific procurement strategies, we tend to lock in computing power in advance in a relatively conservative way, and purchase a part of resources with high certainty at a lower cost. If the actual demand significantly exceeds expectations, we will make additional purchases through the spot market.

Spot computing power is usually more expensive, but we only need to pay this part of the extra cost when the business performance is particularly good. It is a bit like buying insurance: when everything goes well and no additional resources are needed, you will even feel that the insurance cost is on the high side; but when the demand exceeds expectations, it allows you to seize growth opportunities.

We believe that this strategy can not only control financial risks when demand is lower than expected, but also continue to invest in R&D when business growth exceeds expectations.

Dimon: With the rapid growth of Token usage, enterprises have begun to assign simple tasks to lower-priced models, and reserve expensive models for complex tasks. Some AI services may gradually become standardized commodities. What's wrong with this judgment?

Amodei: First of all, every capability breakthrough of cutting-edge models may create new markets.

The traditional technology market often goes through a process: in the early stage, products are scarce, and customers are willing to pay higher prices; as technology matures and competitors increase, products are gradually commoditized. This trend has indeed emerged in the chatbot field.

But AI is different from other technologies. It is not just an application or a market, but on a continuously expanding intelligent capability curve. As model capabilities improve, new applications that could not be realized in the past will continue to emerge.

In 2022 and 2023, the industry mainly focused on chatbots. At that time, Anthropic judged that programming would become a larger market, so we invested more resources in the programming capability of the model, rather than competing for market share or building competitive barriers around chatbots.

Now programming applications are still in the development process, especially among Fortune 500 companies, and there is still much room for penetration. But the bigger opportunities in the future may come from scientific discovery, life sciences, physics, materials, robotics and cybersecurity. The tasks in these fields are more complex than ordinary programming, and the potential market may also be larger.

Secondly, even in applications that do not require the most advanced models, we have the ability to use the technical accumulation of cutting-edge models to train smaller models that are more efficient and cheaper. This is related to the computing power doubling technology we use in model training. We can improve the computational efficiency of large models, and we can also apply these technologies to small models.

Take Haiku 5.5 as an example. Its price is 0.10 US dollars per million input Tokens and 0.50 US dollars per million output Tokens. It can provide competitive performance at a very low cost in some tasks, and even outperform other models in some capability dimensions.

Therefore, cutting-edge models and low-cost models are not mutually exclusive. We must not only continue to improve the capabilities of our most powerful models, but also make model efficiency high enough to make cheaper models equally competitive.

05

From mathematics to biology, AI is approaching scientific breakthroughs

Dimon: Can AI help humans cure cancer and other diseases in the future? Can it promote major scientific breakthroughs such as nuclear fusion and quantum technology?

Amodei: We have already seen the process of continuous improvement of model capabilities in the field of mathematics.

In 2023, models could hardly even complete high school math problems. At that time, it was generally believed that although chatbots were good at communication, they lacked mathematical and logical capabilities.

In 2024, models have been able to complete high school math tasks well. By 2025, they began to solve mathematical problems that had no answers before, initially only in sporadic individual cases. Some people thought that these problems might appear in the training data, or it was just accidental success.

But as model capabilities continue to improve, the situation begins to change. We are seeing more and more open problems being solved, even including some of the most difficult problems in the field of mathematics, and some of the achievements are close to the research level involved in the Fields Medal that mathematicians strive for.

I have also seen early signs of similar changes in fields such as biology and robotics. Experts in the life sciences field, especially the CEOs of pharmaceutical companies, tend to be skeptical about AI. They have seen many technologies that tried to solve the difficult problems of drug R&D in the past, and are well aware of the difficulties faced by this industry.

But the development of AI has a feature: in the early stage of the capability curve, it may seem that it can't do anything well, and there seem to be obstacles everywhere. Some people think that the bottleneck of life sciences is clinical trials, some think it is manufacturing, and others think it is regulation. But as technology advances, these obstacles may be broken one by one.

I believe that in the next few years, more obvious changes may occur in the life sciences field. Of course, we cannot guarantee that every field will make the same progress, but I believe that AI has the potential to promote biological research by a very large margin.

06