BTC $64,281.77 -1.78%
ETH $1,796.29 -1.64%
BNB $588.95 -1.20%
XRP $0.9432 -2.23%
SOL $70.88 -1.47%
TRX $0.3024 -1.70%
DOGE $0.0582 -1.00%
ADA $0.1517 -1.18%
BCH $205.24 -1.65%
LINK $9.09 -1.64%
HYPE $66.69 -1.56%
AAVE $108.71 +1.92%
SUI $0.7179 -1.24%
XLM $0.1370 -1.48%
ZEC $886.81 -2.71%
AAPL $358.20 +0.36%
AMZN $264.75 +0.44%
GOOGL $294.56 +0.63%
MSFT $527.26 +1.20%
META $635.91 -0.19%
NVDA $159.47 -0.88%
TSLA $384.55 -0.72%
SNDK $831.30 -2.72%
INTC $56.79 -2.00%
SPCX $139.46 +1.37%
MU $737.79 -2.64%
AMD $508.38 -2.82%
BTC $64,281.77 -1.78%
ETH $1,796.29 -1.64%
BNB $588.95 -1.20%
XRP $0.9432 -2.23%
SOL $70.88 -1.47%
TRX $0.3024 -1.70%
DOGE $0.0582 -1.00%
ADA $0.1517 -1.18%
BCH $205.24 -1.65%
LINK $9.09 -1.64%
HYPE $66.69 -1.56%
AAVE $108.71 +1.92%
SUI $0.7179 -1.24%
XLM $0.1370 -1.48%
ZEC $886.81 -2.71%
AAPL $358.20 +0.36%
AMZN $264.75 +0.44%
GOOGL $294.56 +0.63%
MSFT $527.26 +1.20%
META $635.91 -0.19%
NVDA $159.47 -0.88%
TSLA $384.55 -0.72%
SNDK $831.30 -2.72%
INTC $56.79 -2.00%
SPCX $139.46 +1.37%
MU $737.79 -2.64%
AMD $508.38 -2.82%

Dario Amodei in conversation with JPMorgan CEO: Anthropic's competitive advantages, computing power challenges, and future bets

Core Viewpoint
Summary: Computing power needs to be purchased in advance, but the future is difficult to predict. On the eve of the IPO, Dario Amodei discussed how Anthropic seeks to find a balance between financial risks and technological ambitions.
ChainCatcher Selected
2026-11-16 17:44:44
Computing power needs to be purchased in advance, but the future is difficult to predict. On the eve of the IPO, Dario Amodei discussed how Anthropic seeks to find a balance between financial risks and technological ambitions.

Guest: Dario Amodei, Co-founder and CEO of Anthropic

Host: Jamie Dimon, Chairman and CEO of JPMorgan Chase

Compiled by: Jiahua, ChainCatcher

As Anthropic prepares for its IPO, CEO Dario Amodei engages in a public dialogue with JPMorgan Chase CEO Jamie Dimon.

Dimon mentioned that Anthropic has disclosed approximately $900 billion in forward commitments related to computing power, leasing, and data centers, and questioned how the company ensures it will have sufficient computing power in the long term while controlling costs.

Amodei referred to computing power as the "deepest commercial challenge" facing Anthropic: the company needs to procure computing power one to two years in advance, but it is difficult to determine when technological advancements will translate into revenue. Over-purchasing could put financial pressure on the company; under-purchasing could not only fail to meet customer demand but also impact the development of the next generation of models.

In addition to computing power and the IPO, the two also discussed Anthropic's corporate culture, the commercial applications of Claude, agent safety, price competition among AI models, and whether AI can drive advancements in medicine and scientific research.

Key points:

  • Computing power is the biggest commercial challenge. Amodei stated that Anthropic's revenue is expected to grow about tenfold in both 2024 and 2025, but the company still struggles to predict demand for the following year, while computing power needs to be procured one to two years in advance.

  • Anthropic tends to conservatively procure computing power. The company locks in a portion of computing power at a lower cost in advance, and if business growth exceeds expectations, it supplements from the higher-priced spot market. He likened this strategy to purchasing insurance.

  • A significant reason for the IPO is to increase cash reserves. Amodei believes that the more cash the company has, the greater its buffer in various computing power demand scenarios, and the more capable it is of seizing opportunities in fields like biology, robotics, and cybersecurity.

  • The competition in AI goes beyond chatbots. Amodei believes that every breakthrough in the capabilities of cutting-edge models could create new markets. Following chatbots and programming, scientific discovery, life sciences, robotics, and cybersecurity may become larger markets.

  • The key limitation for agents to be deployed is safety. He stated that Anthropic chooses to serve enterprise clients partly because businesses value trust, long-term relationships, and reputation, which makes it easier for the company's commercial incentives to align with safety goals.

  • The bottleneck in cybersecurity is shifting from discovering vulnerabilities to fixing them. Amodei noted that both clients and Anthropic have identified hundreds of thousands of vulnerabilities, but the pace of fixing them lags behind the pace of discovery. In the next 6 to 12 months, the application of defensive agent clusters may increase.

  • Anthropic plans to allocate 5% of the company's equity to a public welfare fund. Amodei is concerned that the new wealth generated by AI is concentrated in a few companies, and the fund may focus on issues like workforce transformation and biosecurity, with specific directions still under research.

Talent and Culture are Anthropic's Competitive Advantages

Jamie Dimon: You previously worked at OpenAI and are an early participant in this field. With many companies globally working on AI, how does Anthropic differentiate itself from others?

Dario Amodei: I have worked at several major companies in this field and have been thinking about why some projects succeed while others do not. After many years, I concluded that talent and culture are the most important factors.

Talent and culture create technology, technology creates products, and products drive market expansion, sales, and other business operations. How the public perceives a company is also related to the people who create the technology and their intentions.

When founding Anthropic, my six co-founders and I decided to prioritize talent and culture from day one. Every candidate must undergo a cultural interview. We also insist on rejecting technically strong individuals who do not fit the company culture.

We do not want individuals in the team who disrespect others, seek the limelight, or cannot collaborate with others; we also hope employees recognize the public interest. In the short term, adhering to these standards can sometimes make hiring difficult, but in the long term, it allows everyone to be on the same side.

A company's difficulties often do not stem from a lack of excellent people, but rather from excellent people not working together. Different teams fighting against each other ultimately drain energy internally.

At Anthropic, when an important goal arises—whether it is training models, achieving safety objectives, developing products, or advancing policy work—the entire company can collaborate around that goal.

We also see results from employee retention. More employees from other companies have joined Anthropic than those who have left; all seven co-founders are still with the company; more than three-quarters of the first 100 employees are still here. Even if someone leaves, they often come back. Once, two key product personnel returned less than two weeks after leaving, saying they had not realized what they had before.

I spend about 40% to 50% of my time on talent and culture. Every two weeks, I participate in an internal event company-wide called Dario Vision Quest. I candidly inform employees about what is happening in the company, do not hide issues, and try not to use official language, encouraging other managers to do the same.

Management can trust employees, and employees believe that management will communicate honestly. This trust has significant value.

In Which Industries is Claude Entering?

Jamie Dimon: JPMorgan Chase is one of Anthropic's large clients. Claude has brought us significant changes, with 250,000 employees using it in various ways multiple times a week. How are other Fortune 100 companies using Claude?

Dario Amodei: Claude is being applied across various industries. One of the most surprising features of this technology is its versatility. You can think of it as a "Moore's Law" for cognition or intelligence.

Through communication with clients, I have also encountered many industries I was previously unfamiliar with and seen what role Claude can play in those sectors.

The earliest and fastest-adopting field is programming. Most of my co-founders have written a lot of code in the past, and it is still astonishing to see the tasks the model can accomplish. Almost every Fortune 100 company has significant programming needs, so programming is usually where they first use Claude.

Subsequently, the application scope has expanded to broader knowledge work. In the financial services sector, Claude can handle simple clerical tasks and assist with more complex strategic work. In insurance and legal fields, it can draft contracts and identify issues within them.

Life sciences is a field I pay particular attention to, with applications including writing clinical research reports, assisting with basic research, predicting protein binding, improving manufacturing efficiency, and aiding in fundamental biological research.

In e-commerce, some companies deploy Claude as a direct-to-consumer agent. The manufacturing sector is similar, with more companies beginning to use Claude to enhance efficiency, with applications including semiconductor manufacturing and physical construction.

Agents Entering Enterprises Must First Address Safety Issues

Jamie Dimon: You mentioned agents. For regulated enterprises, what preparations need to be made before allowing agents to access sensitive information?

Dario Amodei: The most important factor is safety. The capabilities of agents are rapidly improving. Whether they can conduct life sciences research or write complex software architectures is not the only challenge; the real limitation on technological development, internal deployment, and client usage is how to use these capabilities safely.

We do not want agents to lose control. Therefore, Anthropic has invested significant resources to ensure that appropriate safety and alignment mechanisms are established from the model training phase.

We have a large alignment team that remains vigilant about these issues. Every few months, we reassess our safety, security, and alignment efforts and increase our investments.

Anthropic has focused on enterprise clients from the beginning. We know that businesses often prioritize safety issues more than consumers do, and enterprise operations align with our safety goals.

We value safety because we believe that this technology should be developed in the right way. At the same time, we are also considering what business models can provide us with stronger incentives to pursue safety. The answer is enterprise business: enterprises value trust, long-term cooperation, and their own reputation, and do not want models to cause serious harm.

This aligns our commercial incentives more closely with the societal goals we wish to achieve.

Jamie Dimon: Can we truly control the permissions of agents? For example, allowing them to access sensitive systems while ensuring they can only perform operations A, B, C, and D, and nothing else?

Dario Amodei: This is somewhat akin to the development of autonomous driving technology.

Autonomous driving has enormous potential value and could ultimately be much safer than human driving. But first, the technology must reach a sufficiently reliable level, establish the correct protective mechanisms, and design the system in the right way.

We can never provide a 100% guarantee. You cannot say that an autonomous vehicle will absolutely never have an accident, no matter how many people use it. But we can strive to make the system significantly safer than human operation, achieving safety levels several times, hundreds of times, or even higher.

Behind the $900 Billion Commitment: The Challenges of Over and Under Purchasing Computing Power

Jamie Dimon: Anthropic has disclosed approximately $900 billion in forward commitments related to computing power, leasing, and data centers. I also saw a statement: about 80% of computing power is used for training new models, while only about 20% is used for inference, and we, as clients, mainly use inference services.

How do you ensure you will obtain the necessary computing power in the future while maintaining cost efficiency?

Dario Amodei: Computing power is the deepest commercial challenge facing Anthropic.

The technological development in this field is very rapid, but we do not know how quickly technological advancements will translate into commercial value, leading to significant uncertainty.

In 2024 and 2025, Anthropic's revenue is expected to grow about tenfold each year. But every year we ask ourselves: what will happen next year? We do not know. As the company grows, the amounts involved also increase.

At the same time, computing power must be procured one to two years in advance. How much to procure depends on how much cost the company can bear.

Purchasing too much could place a heavy financial burden on the company; purchasing too little would not provide sufficient inference services for clients. More importantly, insufficient computing power could also impact research and development, which determines what models we can launch in six months or a year and affects the company's future revenue-generating capabilities.

We primarily expand our buffer space through two aspects.

The first is gross margin. The higher the gross margin, the greater the adjustment space when purchasing too much or too little computing power. Simply put, it reflects the difference between the cost of computing power and the revenue the company generates from utilizing that power. The prospectus shows that our gross margin is quite good, especially compared to other companies in the industry.

The second is cash on the balance sheet. We currently have a significant amount of cash and hope to continue increasing our cash reserves.

However, even with these buffers, the company's forward commitments remain substantial. Although these commitments span multiple years and revenue is growing rapidly, the amounts are still considerable.

Therefore, the strategy we adopted is to conservatively purchase computing power: buying a portion of it in advance at a lower cost. If business performance falls short of expectations, we have controlled costs; if performance exceeds expectations, we can purchase additional computing power from the spot market.

Spot computing power is more expensive, but we will only buy it when business performance is particularly good. This is somewhat like buying insurance: to ensure coverage in different situations, one must pay for additional flexibility.

We believe this strategy allows the company to continue investing in research and development and seize growth opportunities when needed, while also providing protection when business performance does not meet expectations.

Are cheaper models becoming more common, will AI become a commodity?

Jamie Dimon: More and more companies are purchasing Tokens. People quickly realize that many tasks do not require expensive models and can be accomplished with cheaper models and Tokens. This part of the service may gradually become a standardized commodity. What is wrong with this judgment?

Dario Amodei: First, every time the capability boundaries of cutting-edge models are expanded, it can create entirely new markets.

The familiar market development path is: initially, there are many startups, and customers are willing to pay high prices; as the market matures, products gradually become commoditized, and discussions about moats and cheaper alternatives begin.

In a specific application of AI, this situation may indeed occur. Chatbots are an example. But AI is not just one application or one market. As intelligent capabilities continue to improve, new markets will constantly emerge.

In 2023, everyone is focused on chatbots. Anthropic, after seeing the capabilities of the models, believes that programming will become a larger market, as the models have started to become smart enough to handle programming tasks.

Therefore, we focus on the cutting edge of capabilities rather than competing for the chatbot market or building moats around chatbots.

Now, AI programming is gradually maturing, but there is still a long way to go. Looking at the penetration rate of Fortune 500 companies, there is still a lot of work to be done.

Meanwhile, I believe the market's focus will shift to new areas. These areas may be larger than programming, just as the programming market is larger than the chatbot market. Scientific discovery, life sciences, physics, materials, robotics, and cybersecurity all belong to this category. The challenges they face are more complex than ordinary programming, so we will continue to focus on the cutting edge of capabilities.

Second, even in tasks that do not require cutting-edge models, we can leverage large models to train smaller, more efficient models.

By improving the efficiency of computing power usage, we make large models more efficient and can also apply these technologies to smaller models. For example, Haiku 5.5 is priced at $0.10 per million input Tokens and $0.50 per million output Tokens. We are also able to create low-cost models and compete with other vendors in terms of efficiency.

Can AI drive breakthroughs in medicine and science?

Jamie Dimon: Can AI help humanity cure cancer and other diseases, or drive major breakthroughs in nuclear fusion, quantum technology, etc.?

Dario Amodei: We repeatedly see similar technological advancement processes in different fields. For example, in mathematics, in 2023, models struggled to solve even high school math problems. People generally believe that while chatbots excel in communication, they do not excel in math and logic.

By 2024, models are able to perform high school math problems well. By 2025, they begin to solve some previously unsolved mathematical problems. Initially, there are only one or two cases, and people might say these problems could be in the training data, or the models just got lucky.

But then, the number of problems solved by the models increases, and they begin to involve more important and difficult issues.

Now, we see early signs of a similar process in new fields such as biology and robotics. Experts are often very skeptical, especially in the pharmaceutical industry. They have seen many technologies attempting to solve industry problems in the past and know how difficult these problems are.

In the life sciences, people say the bottlenecks are clinical trials, manufacturing, or FDA approvals. But we see that AI is gradually helping to solve problems in these areas. I believe that in the coming years, this change may become more apparent.

We cannot determine the outcome in advance. In a specific field, skeptics may also be right. But I believe the potential for AI to drive significant progress in biology is very high.

After discovering vulnerabilities, the harder part is fixing them

Jamie Dimon: Before the emergence of Mythos and agents, cybersecurity was one of the biggest risks facing JPMorgan Chase. We have thousands of people working in cybersecurity, and we may need more people to fix the vulnerabilities discovered.

Now, continuously running agents may attack various security boundaries of enterprises. How should companies, vendors, and the entire industry prepare for this?

There are also some key industries that lack sufficient protective capabilities, and we need to help them catch up.

Dario Amodei: First, discovering vulnerabilities is just the first step.

After Mythos was launched, we invested a lot of work into finding vulnerabilities. Customers discovered hundreds of thousands of vulnerabilities, and we found hundreds of thousands ourselves. The next bottleneck became how to fix these vulnerabilities.

Fixing is harder than discovering. We found many vulnerabilities, but neither customers nor we could fix them quickly enough. Therefore, Claude is gradually improving its ability to fix vulnerabilities, and we are also starting to offer more related services.

The third step is to write more secure code from the beginning. As Claude's programming capabilities improve, we are also specifically training it to be better at writing secure code.

The fourth step is real-time defense. Cyberattacks can happen in hours, minutes, or even seconds. If a group of defensive agents can be deployed to continuously patrol enterprise servers and systems and defend against attacks in real-time, it would be very valuable.

This approach has not yet been widely adopted, but I believe that in the next 6 to 12 months, there will be more and more applications. Because attackers will also use clusters of agents, defenders will also need clusters of agents.

Why choose to go public now?

Jamie Dimon: Anthropic is about to go public. Why now? Going public has its pros and cons; do you think it will make the company better or make management more difficult?

Dario Amodei: There are several reasons for going public, the most important of which is the investment in computing power.

There are huge economic opportunities in biology, robotics, cybersecurity, and broader knowledge work fields. We want to invest enough resources to ensure we can seize these opportunities.

The more cash-rich the company is, the greater the buffer between low and high computing power demands. This is also an important reason for going public.

As for the specific month, day, or market environment in which to go public, we are not particularly concerned.

Before starting the IPO process, I had never experienced anything like this. Anthropic has taken about six years to grow from a startup to where it is today. I sought advice from many people, and one of the most resonant pieces of advice was: do not let the company's valuation exceed its true value by even a dollar.

This is very close to one of Anthropic's principles: under-promise and over-deliver.

We typically do not talk extensively about what capabilities a model will have before it is completed. We first create the model, confirm it is good enough, and then release it. The same goes for safety systems. I hope the IPO will follow the same principle.

Allocate 5% equity to establish a public welfare fund, concerns about AI wealth concentration

Jamie Dimon: Anthropic also announced the establishment of a public welfare fund. Many people may not understand why this is being done. What are your thoughts on the fund?

Dario Amodei: We decided to allocate 5% of the company's equity to the public welfare fund.

In addition, the co-founders have committed to donating 80% of their personal wealth. Since the company's inception, we have also established an employee donation matching program: if employees donate a certain percentage of their shares to charities or reserve them for public welfare purposes, the company will provide additional shares for matching.

These arrangements share a common idea: AI is a very powerful technology. It has already created tremendous economic value for Anthropic and the entire industry, and we hope future opportunities will be even greater, but we cannot be certain.

I hope AI can help humanity cure cancer and Alzheimer's disease, provide cheaper and more abundant energy, better materials, and allow more people to enjoy a richer material life.

But we are also concerned that if a large amount of new wealth comes from the same technology and is concentrated in a few companies, society may experience rapid growth alongside a high concentration of wealth.

We hope to alleviate the problem of wealth concentration as much as possible, and the public welfare fund is part of that.

The specific direction of the fund is still under research. We will not announce a long list of problems to solve right now; instead, we hope to first clarify which issues are most worth investing in.

Possible directions include helping society smoothly cope with the labor transformation brought about by AI and biosecurity. AI may bring significant advances in biology, but like network technology and other technologies, it can also be misused. We need to prevent misuse and establish effective defense mechanisms.

15 years from now, I hope diseases become history and AI does not go out of control

Jamie Dimon: Last question. If you look back at today in 15 years, what do you hope will have happened, and what do you hope will not have happened? Is there anything we haven't talked about today that you would like to say to everyone?

Dario Amodei: The most important thing is to ensure that this technology is developed in the right way.

I have talked about many technological risks, including being used for biological or cyber attacks, causing economic shocks, and humanity losing control over agents. These risks are very serious and not just marketing talk. Anthropic has taken these issues seriously since its founding and has invested resources to address them.

We hope to set an example for the entire industry on how to address risks. But it is not enough to do well only within Anthropic; we also hope that other companies, customers, and suppliers in the AI industry act responsibly.

Second, we need to truly realize the positive value of AI.

In biology and medicine, we have just begun to see the potential of this technology. There are still many skeptical voices from pharmaceutical companies, early biotech firms, and academia. We cannot know the outcome in advance; skeptics may be right, but I hope to ultimately prove that their concerns do not come true.

I hope that one day, heart disease, Alzheimer's disease, and cancer can become diseases of the past, like today's Black Death and smallpox.

This may sound idealistic, even crazy, but I truly believe it is possible. I do not know if it will take five years, ten years, or fifteen years, but we will do our utmost to push for this goal to be realized as soon as possible.

warnning Risk warning
app_icon
ChainCatcher Building the Web3 world with innovations.