AI News

Anthropic CEO Amodei on IPO: Compute Is the Deepest Business Challenge

WallstreetcnUpdated Nov 16 20:33

AI summary

Anthropic co-founder and CEO Dario Amodei and JPMorgan Chase CEO Jamie Dimon held a conversation in New York on the topic of IPOs. Amodei said compute is the deepest business challenge the company faces, requiring procurement one to two years in advance, and that it currently adopts a conservative strategy of first locking in part of its compute and then supplementing on demand from the spot market. He mentioned that Anthropic has disclosed roughly $900 billion in forward commitments for compute and infrastructure, with about 80% of compute used to train new models and about 20% for inference, and that revenue grew about 10-fold in each of 2024 and 2025. He also said the choice to go public now is to increase cash reserves, support compute buildout, and seize opportunities in fields such as life sciences and robotics, and that the company will additionally allocate 5% of its equity to a public-interest fund.

Why read this

AI score 58

In an IPO conversation, the Anthropic CEO linked compute procurement, model commoditization, and safe deployment into a single business logic, offering a window into the operating trade-offs of frontier model companies.

Topics

Tags

Body · AI translation

On the midday of November 13, US local time, Anthropic co-founder and CEO Dario Amodei and JPMorgan Chase CEO Jamie Dimon held an IPO-themed conversation in New York.

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

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

English text is being prepared.

Amodei believes that chatbots and coding are merely the markets opened up by current model capabilities, and that fields such as life sciences, robotics, materials, and cybersecurity still hold greater potential. Even if some model applications gradually become commoditized, Anthropic can leverage the technical accumulation of frontier models to train smaller models with lower cost and higher efficiency.

But the stronger the model capabilities, the greater the computing power investment behind them.

Amodei admitted: "Compute is the deepest business challenge Anthropic faces."

Computing resources often need to be purchased one to two years in advance; buying too much increases the financial burden, while buying too little may affect customers' inference needs and the development of next-generation models. Anthropic currently adopts a relatively conservative procurement strategy: first locking in part of the compute, then supplementing it through the spot market according to actual demand.

The IPO provides more funding room for this long-term investment.

Amodei said that increasing cash reserves, supporting compute buildout, 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 while preparing for the IPO, he heard a piece of advice: "Never let the company's valuation be even one dollar higher than what it is truly worth." This is consistent with Anthropic's principle of "under-promise and over-deliver": do not hype model capabilities in advance, but instead show them to the market only after the technology and products are truly mature.

From enterprise customers to compute procurement, from model commoditization to IPO valuation, this conversation actually touched on several questions Anthropic must face after going public.

The following is a condensed transcript of the conversation between Dimon and Amodei:

Talent and culture are Anthropic's moat

Dimon: You once worked at OpenAI and were also 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 several major companies in this industry. Looking back on these experiences and thinking about why some projects succeed while others fail, I have gradually come to realize that the core factor determining a company is talent and culture.

Talent and culture determine technology, technology determines products, and products in turn determine market expansion, sales, and the company's other businesses. A company's public image and external evaluation also come from the people who create the technology and the motivations behind what they do.

When founding Anthropic, I and the other six co-founders decided that talent and culture would be the top priority from day one. Every new employee must undergo a culture interview. Even if their technical ability is outstanding, if they do not share the company's values, are not good at collaboration, or are accustomed to seeking the spotlight, we will not hire them.

This hiring standard makes recruitment difficult in the short term, but in the long run it enables the building of a team that truly works together. Many companies do not lack excellent talent; the problem is that different teams compete with and constrain each other, consuming a great deal of energy on internal strife.

At Anthropic, once an important goal is set—whether it is training a new model, achieving safety goals, developing products, or advancing policy matters—the whole company collaborates around that goal.

Employee retention is a direct reflection of this. All seven co-founders are still at the company, and more than three-quarters of the first 100 employees are still working at Anthropic. We believe the company has one of the better employee retention rates in the industry. Some former employees even rejoin. Two important product leaders once left but came back in less than two weeks.

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

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

Claude enters enterprises, and AI begins to take over more work

Dimon: JPMorgan Chase has 250,000 employees, and every week they use Claude in different ways. How are other Fortune 100 companies mainly applying this technology?

Amodei: Claude is entering every industry. I think one of the most important characteristics of this technology is its generality. As model capabilities continue to improve, more and more work that could not previously be automated is becoming feasible.

Programming was the first field in which enterprises adopted AI at scale. I have written a great deal of code in the past, and many of my co-founders also have programming backgrounds. It is astonishing to see what today's models can accomplish on such complex programming tasks. Nearly all Fortune 100 companies have substantial programming needs, so programming naturally became their initial entry point for adopting AI.

Subsequently, the scope of application gradually expanded to broader knowledge work.

In financial services, Claude can handle both basic paperwork and assist with complex strategic work; in the insurance and legal industries, it can be used for contract drafting, contract review, and identifying potential loopholes.

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

In e-commerce, some customers have deployed Claude as a direct-to-consumer agent. The same is true in manufacturing, where companies have begun using Claude to improve operational efficiency, with applications spanning areas such as semiconductor manufacturing and physical facility construction.

I previously did not know much about manufacturing, but through conversations with manufacturing CEOs and seeing the practical application of Claude in these companies, I have gained a new understanding of the industry.

As 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: First is security. Agent capabilities are improving rapidly and can already handle highly difficult tasks such as life sciences research and complex software architecture. What truly limits technology development, internal deployment, and customer use now is not capability itself, but how to deploy these systems safely.

We do not want agents to lose control. This is also why Anthropic devotes substantial resources to safety research. Safety and alignment must begin at the model training stage, rather than being remedied after the product is released. The company has a large alignment team, and we take safety issues very seriously; almost every few months we reassess our investment and further increase resources for safety, security, and alignment.

Since its founding, Anthropic has intentionally made enterprise customers the focus of its business. This is not only a commercial choice, but also consistent with our safety philosophy.

Enterprises value trust, long-term cooperation, and their own reputations, and they care more than ordinary consumers about whether models may cause serious harm. Therefore, serving enterprise customers allows the company's commercial interests to align with its safety goals. We hope in this way to establish better incentives for developing AI safely.

Dimon: Is it truly possible to limit an agent's permissions? For example, allowing them into sensitive systems but only to perform explicitly authorized operations A, B, C, and D, and nothing else?

Amodei: You can look at the development of autonomous vehicles.

Autonomous driving is a technology of enormous value and is ultimately expected to be far safer than human driving. But to achieve this, the right safeguards must first be established to ensure the system is designed and deployed in a safe manner.

We cannot guarantee that autonomous vehicles will never have an accident under any circumstances. Likewise, we cannot guarantee that agents will never make mistakes in any environment.

The truly feasible goal is to make these systems far safer than human operation—several times, dozens of times, or even hundreds of times safer. This is also the direction we are working toward.

A $900 billion compute bet—Amodei's biggest headache

Dimon: Anthropic has disclosed forward commitments of about $900 billion for compute and infrastructure. The outside world has also noted that currently about 80% of compute is used to train new models and only about 20% for inference, while what customers actually use is mainly inference services. How do you ensure sufficient compute in the future while maintaining cost efficiency?

Amodei: Compute is the most complex and most important business challenge Anthropic faces.

Technological progress in this industry is very fast, but there is great uncertainty about how quickly technological progress can translate into commercial revenue. In 2024 and 2025, Anthropic's revenue grew about 10-fold for two consecutive years. Each year, it is very difficult for us to accurately judge what will happen the next year, and as the scale of the business expands, the impact of this uncertainty grows larger and larger.

Another issue is that compute must be purchased one to two years in advance. How much to purchase depends on our judgment of future demand. Buying too much may put unbearable financial pressure on the company; buying too little means we cannot meet customers' inference needs.

More importantly, compute also affects R&D. The resources used today to train models determine what kind of models can be launched six months or a year from now, and therefore determine the company's ability to generate revenue in the future. Therefore, compute procurement cannot look only at immediate inference demand; it must also reserve R&D resources for the next generation of models.

We mainly rely on two aspects to reduce procurement risk.

First is gross margin. The higher the gross margin, the greater the company's room to adjust when it has purchased too much or too little compute. Simply put, gross margin reflects the difference between the cost we pay for compute and the revenue generated per unit of compute. The gross margin performance disclosed in the prospectus is quite good, especially compared with other players in the industry.

Second is cash reserves. Sufficient cash can help the company withstand large-scale compute investment and respond to changes in future demand. We currently have a large amount of cash and hope to continue increasing our capital reserves.

In terms of specific procurement strategy, we tend to lock in compute relatively conservatively in advance, purchasing a portion of resources with high certainty at a lower cost. If actual demand significantly exceeds expectations, we then make additional purchases through the spot market.

Spot compute is usually more expensive, but we only need to pay this extra cost when business performance is particularly good. This is somewhat like buying insurance: when everything goes smoothly and you don't need extra resources, you instead feel the insurance cost is too high; but when demand exceeds expectations, it lets you seize growth opportunities.

We believe this strategy can both control financial risk when demand falls short of expectations and allow continued investment in R&D when business growth exceeds expectations.

Dimon: As token usage grows rapidly, companies have begun handing simple tasks to cheaper models and reserving expensive models for complex tasks. Some AI services may gradually become standardized commodities. What is wrong with this judgment?

Amodei: First, every capability breakthrough in frontier models may create a new market.

Traditional technology markets often go through a process: early products are scarce, and customers are willing to pay higher prices; as the technology matures and competitors increase, products gradually become commoditized. This trend has indeed already appeared in the chatbot field.

In 2022 and 2023, the industry focused mainly on chatbots. At the time, Anthropic judged that programming would become a larger market, so it put more resources into the models' programming capabilities rather than competing for market share or building competitive barriers around chatbots.

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

Second, even in applications that do not require the most advanced models, we are able to use the technological accumulation of frontier models to train more efficient and cheaper small models. This is related to the compute multiplier technology we use in model training. We can improve the computational efficiency of large models and also apply these technologies to small models.

For example, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. It can provide competitive performance at very low cost on some tasks, and even outperforms other models in some capability dimensions.

Therefore, frontier models and low-cost models are not mutually exclusive. We must continue to improve the capabilities of the strongest models while also making model efficiency high enough that cheap models are equally competitive.

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

In 2023, models even struggled to complete high school math problems. At the time, it was widely believed that although chatbots were good at communication, they lacked mathematical and logical abilities.

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

But as model capabilities continued to improve, the situation began to change. We have seen more and more open problems solved, including even some of the hardest problems in mathematics, with some results approaching the level of research associated with mathematicians competing for the Fields Medal.

I have also seen early signs of similar changes in biology, robotics, and other fields. Experts in the life sciences, especially CEOs of pharmaceutical companies, are often skeptical of AI. They have seen many technologies that tried to solve the difficulties of drug R&D and clearly understand the difficulties this industry faces.

But the development of AI has one characteristic: in the early stages of the capability curve, it may seem unable to do anything well, as if obstacles are everywhere. Some people believe the bottleneck in life sciences is clinical trials, some believe it is manufacturing, and others believe it is regulation. But as technology advances, these obstacles may be overcome one after another.

After Mythos, cybersecurity enters the era of agent offense and defense

Dimon: Cybersecurity was already one of the main risks companies faced before Mythos and agents appeared. Now, AI can discover vulnerabilities at scale, and agents may also continuously attack corporate systems. JPMorgan has thousands of people working in cybersecurity and may need tens of thousands more to patch vulnerabilities discovered by Mythos. How should companies prepare for the next stage?

Amodei: First, we must recognize that discovering vulnerabilities is only the first step.

After Mythos was launched, we did a great deal of vulnerability discovery work. Customers discovered hundreds of thousands of vulnerabilities, and we ourselves also discovered hundreds of thousands. Then the real bottleneck began to shift to vulnerability remediation: the speed of discovering vulnerabilities is already fast, but fixing vulnerabilities is more difficult than discovering them.

Therefore, we are expanding Claude's capabilities in vulnerability remediation so that the model can not only identify problems but also help companies actually fix them.

Second is reducing vulnerabilities at the source. As Claude's programming capabilities strengthen, we are also specifically training the model to write more secure code, reducing the probability of vulnerabilities arising during software development.

Third is building a real-time defense system. This has not yet been widely adopted, but I think the potential is great.

Cyberattacks usually occur on time scales of hours, minutes, or even seconds. In the future, companies can deploy clusters of agents to continuously patrol servers and network systems, identifying and resisting attacks in real time.

If attackers begin to use agent clusters to launch attacks, defenders will also need to use agent clusters for defense. I expect this approach to defense to see more adoption in the next 6 to 12 months.

Why go public now? Amodei lays out his IPO principles

Dimon: Anthropic is about to go public. Why choose now? How do you personally view going public? Will it make the company better, or make management more difficult?

Amodei: The most important reason for going public now is to provide more funding for the company's future development and compute investment.

There is enormous economic value in biology, robotics, cybersecurity, and the broader field of knowledge work. We need to invest enough resources to seize these opportunities.

Returning to the question of compute procurement, there is a buffer space between buying more compute in advance and maintaining a relatively conservative procurement strategy. The more cash the company has, the larger this buffer space is, and the better it can cope with uncertainty in future demand.

As for the specific month and day of the listing, and the market environment at the time, we did not agonize over it.

Before starting the listing process, I had never experienced anything similar, so I sought advice from many people. The piece of advice that impressed me most was: never let the company's valuation be even one dollar higher than its true value.

This is similar to the principle Anthropic has always adhered to: under-promise, over-deliver.

We do not talk up in advance what capabilities a model will have, nor do we promote a model before it is finished. We build the model first, confirm it is good enough, and then officially release it. The same goes for safety systems.

I hope the IPO follows the same principle and does not over-promise in pursuit of a higher valuation.

Dimon: Anthropic announced it will establish a public benefit fund and transfer 5% of the company's equity into it. Why did you decide to do this? What areas will the fund focus on in the future?

Amodei: The public benefit fund is part of Anthropic's long-term social responsibility arrangements.

In addition to the company contributing 5% of its equity, several of our co-founders have also pledged to donate 80% of their personal wealth. Since the company's founding, we have also had an employee stock donation matching program: if employees donate a certain proportion of their stock to charity, or set it aside specifically for public benefit purposes, the company will provide additional stock to match.

We do this because AI is a technology with enormous impact. It has already created tremendous economic value for Anthropic and the entire industry, and may bring even more change in the future.

But at the same time, there is also an issue worth paying attention to: if a large amount of newly created wealth all comes from the same technology and is concentrated in the hands of a few companies, society may face a situation in which economic growth coexists with highly concentrated wealth.

We hope to mitigate this risk as much as possible, and the public benefit fund is one such arrangement.

How the fund will specifically operate and which issues it will focus on solving are still being studied. It may involve the labor transition brought about by AI development, helping society smoothly adapt to changes in the employment structure; it may also involve biosecurity, ensuring that while AI drives biological development, related technologies are not misused.

We hope to first identify the areas truly worth investing in and then proceed step by step, rather than promising a series of goals that have not yet been clearly defined.

Dimon: If you look back on today 15 years from now, what do you most hope AI will achieve, and what do you most hope to avoid? Is there anything else we have not discussed that you would like to emphasize to the audience here?

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

AI carries very high risks, including being used for biological attacks and cyberattacks, causing economic shocks, and humanity losing control over agents. These are not talking points; they are issues Anthropic has taken seriously since its founding.

We hope not only to develop AI responsibly ourselves, but also to set an example for the entire industry. AI companies, customers, and suppliers all need to take these risks seriously and jointly establish corresponding safety standards to ensure that risks during technological development are kept under control.

I hope that one day, heart disease, Alzheimer's disease, and cancer will all become diseases of humanity's history, like the Black Death and smallpox are today.

This may sound unrealistic, but I truly believe it is possible. I cannot be sure whether it will take 5 years, 10 years, or 15 years, but we will do our utmost to push this goal forward as early as possible.

Source: Tencent Technology Risk Disclosure and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this article is at the user's own risk.

Source

Wallstreetcn