2026-11-16 08:26
Introduction: Enterprise clients and safety alignment are driving Claude’s commercialization. On the afternoon of November 13, U.S. time, Dario Amodei, co-founder and CEO of Anthropic, held an IPO-focused dialogue with Jamie Dimon, CEO of JPMorgan Chase, in New York.
At noon on November 13, U.S. time, Dario Amodei, co-founder and CEO of Anthropic, engaged in a high-profile dialogue with Jamie Dimon, CEO of JPMorgan Chase, in New York.
For Anthropic, preparing for public market entry, this conversation was not merely a roadshow—it served as a window into its business model, competitive strategy, and long-term vision.
Amodei attributes Anthropic’s competitive edge to “talent and culture”.
In his view, talent and culture shape technology, which in turn determines product quality, and products then drive market expansion and sales. Enterprises’ emphasis on security and trust aligns directly with Anthropic’s core mission. He believes security is not a mere add-on condition for enterprise AI commercialization—it is the foundation upon which lasting client partnerships are built.
The boundaries of the AI market continue to expand. “Every frontier expansion creates entirely new markets,” Amodei stated.
He argues that chatbots and coding represent only the initial applications unlocked by current models. Far greater potential lies ahead in life sciences, robotics, materials science, and cybersecurity. Even as certain model applications become commoditized, Anthropic can leverage its cutting-edge model R&D to train smaller, more cost-efficient, and higher-performing models.
Yet, stronger model capabilities demand exponentially greater compute resources.
Amodei candidly admitted: “Compute is the deepest commercial challenge facing Anthropic.”
Compute capacity often requires procurement one to two years in advance. Overbuying burdens the balance sheet; underbuying risks failing to meet customer inference demands or hinder next-generation model development. Currently, Anthropic adopts a relatively conservative procurement strategy—securing baseline compute upfront, then supplementing via spot markets based on actual demand.
The upcoming IPO provides crucial additional capital to sustain this long-term investment.
Amodei emphasized that increasing cash reserves, supporting compute infrastructure, and capturing growth opportunities in fields like life sciences and robotics were key reasons for choosing this moment to go public.
He also recalled advice received during IPO preparation: “Never let your company’s valuation exceed its true worth by even a single dollar.” This echoes Anthropic’s enduring principle: “Less promise, more delivery”—avoiding premature hype around model capabilities and instead showcasing mature technology and products only when ready.
This dialogue, touching on enterprise clients, compute procurement, model commoditization, and IPO valuation, actually confronts several critical challenges Anthropic must address post-IPO.
Amodei hopes investors see beyond just current growth velocity. They should assess whether Anthropic can scale commercially while effectively managing risk and transforming AI capabilities into sustainable economic value.
Below is a condensed transcript of Dimon’s conversation with Amodei:
Dimon: You worked at OpenAI and were an early participant in the AI field. With so many companies now entering, how does Anthropic differentiate itself from competitors?
Amodei: I’ve worked at several major companies in this industry. Reflecting on those experiences—and why some projects succeeded while others failed—I came to realize that the core determinant of a company’s success is talent and culture.
Talent and culture shape technology, which defines product quality, which in turn drives market expansion, sales, and other business functions. A company’s public image and external reputation stem directly from the people who build the technology and their foundational motivations.
When founding Anthropic, my six co-founders and I made it clear from Day One that talent and culture would be paramount. Every new hire undergoes a cultural interview. Even if someone has exceptional technical skills, we won’t hire them if they don’t align with our values, lack collaboration skills, or thrive on ego-driven competition.
This hiring standard makes recruitment harder short-term, but ensures long-term team cohesion. Many companies lack no talent—they simply suffer from internal fragmentation, where teams compete rather than collaborate, wasting energy on internal politics.
At Anthropic, once a strategic goal is set—whether training a new model, achieving safety milestones, building a product, or advancing policy—all employees rally around it.
Employee retention speaks volumes. All seven co-founders remain with the company. Of the first 100 hires, over 75% are still employed at Anthropic. We believe we have one of the best retention rates in the industry. Some former employees have even returned. Two key product leads left but rejoined within two weeks.
I personally dedicate about 40% to 50% of my time to talent and culture. Every two weeks, I host a company-wide internal event called “Dario Vision Quest” (DVQ). I speak directly to employees—no spin, no platitudes—about the company’s real state, encouraging other leaders to do the same.
I want employees to trust leadership, and leadership to trust employees. This mutual trust is fundamental.
Dimon: JPMorgan Chase has 250,000 employees, all using Claude in various ways each week. How are other Fortune 100 enterprises applying this technology?
Amodei: Claude is entering every sector. One of its most significant traits is generality. As model capabilities grow, increasingly complex tasks previously deemed non-automatable are now feasible.
Programming was the first domain where enterprises adopted AI at scale. I wrote extensive code myself, and many co-founders share that background. Seeing today’s models handle highly complex programming tasks is astonishing. Fortune 100 firms have massive coding needs, making programming a natural starting point for AI adoption.
From there, applications expanded into broader knowledge work.
In financial services, Claude handles both basic documentation and complex strategy tasks. In insurance and legal sectors, it aids contract drafting, review, and identifying potential vulnerabilities.
Life sciences is particularly promising. Claude is already used to draft clinical reports, assist in foundational research, predict protein binding, enhance manufacturing efficiency, and support fundamental biological discovery.
In e-commerce, some clients deploy Claude as consumer-facing agents. Similarly in manufacturing, companies use Claude to boost operational efficiency—from semiconductor fabrication to physical infrastructure construction.
I had limited prior exposure to manufacturing, but after conversations with manufacturing CEOs, seeing Claude’s real-world impact has reshaped my understanding of the sector.
Dimon: For regulated enterprises, what preparations are essential before allowing agents access to sensitive data and systems?
Amodei: First and foremost, security. Agent capabilities are advancing rapidly—now capable of handling high-stakes tasks like life sciences research and complex software architecture. The real bottleneck isn’t capability itself, but how safely these systems can be deployed.
We absolutely do not want agents to go rogue. That’s why Anthropic invests heavily in safety research. Alignment must begin at the training stage—not retrofitted after product release. We maintain a large-scale alignment team and continuously reassess our security investments, frequently increasing resources dedicated to safety, security, and alignment.
From day one, Anthropic deliberately positioned enterprise clients as its primary focus—not just a business decision, but aligned with our safety ethos.
Enterprises prioritize trust, long-term partnership, and reputation far more than average consumers. They care deeply about whether models could cause serious harm. Serving enterprise clients aligns our commercial interests with safety objectives. We aim to create better incentives for developing safe AI.
Dimon: Can agent permissions truly be constrained? For example, allow them to enter sensitive systems but only perform clearly authorized actions A, B, C, D—nothing else?
Amodei: Consider the evolution of autonomous vehicles.
Autonomous driving holds immense value and could eventually be far safer than human driving. But achieving this requires robust safeguards ensuring systems are designed and deployed safely.
We can’t guarantee autonomous vehicles will never crash under any circumstance. Similarly, we can’t ensure agents will never fail in any environment.
The realistic goal is to make these systems dramatically safer than human operation—by factors of tens or even hundreds. That’s exactly our target.
Dimon: Anthropic disclosed approximately $90 billion in forward commitments for compute and infrastructure. Observers note that currently ~80% of compute is used for training new models, while only ~20% supports inference—yet customer usage is primarily inference-based. How do you ensure sufficient future compute access while maintaining cost efficiency?
Amodei: Compute is the most complex and critical commercial challenge for Anthropic.
Technological progress is rapid, but translating it into commercial revenue remains highly uncertain. In 2024 and 2025, Anthropic’s revenue grew roughly tenfold each year. Each year, we struggle to predict what the next will bring—and as our scale expands, this uncertainty amplifies.
Another issue: compute must be procured one to two years in advance. The amount depends on our forecast of future demand. Over-procurement strains the balance sheet; under-procurement risks failing to meet customer inference needs.
Even more critical: compute impacts R&D. Today’s training resources determine what models we can launch in six to twelve months—directly shaping our future revenue-generating capacity. Thus, procurement can’t focus solely on immediate inference demand—it must reserve resources for next-gen models.
We mitigate procurement risk through two main strategies.
First, gross margin. Higher margins provide greater flexibility when overbuying or underbuying compute. Simply put, margin reflects the gap between cost per unit of compute and the revenue generated per unit. Our disclosed gross margins are strong, especially relative to peers.
Second, cash reserves. Abundant cash enables absorbing large-scale compute investments and adapting to future demand shifts. We hold substantial cash and plan to increase reserves further.
On procurement strategy, we favor a relatively conservative approach—locking in baseline compute at lower costs upfront. If actual demand exceeds expectations, we supplement via spot markets.
Spot compute is pricier, but we only pay this premium when business performance is exceptionally strong. It’s akin to buying insurance: when everything runs smoothly and extra resources aren’t needed, the cost feels high. But when demand surges, it enables capturing growth opportunities.
We believe this strategy controls financial risk when demand falls short, while enabling continued R&D investment when growth exceeds expectations.
Dimon: With token usage rising rapidly, enterprises are already delegating simple tasks to cheaper models and reserving expensive ones for complex work. Could some AI services eventually become standardized commodities? What’s flawed about that assumption?
Amodei: First, every breakthrough in frontier models opens entirely new markets.
Traditional tech markets typically evolve: early scarcity leads to high prices; maturity brings competition and commoditization. Chatbot markets indeed show this trend.
But AI differs fundamentally. It’s not just one application or one market—it sits on a continuously expanding curve of intelligence. As model capabilities grow, entirely new applications emerge that were previously impossible.
In 2022 and 2023, the industry focused on chatbots. Anthropic predicted coding would be a larger market, so we directed more resources toward enhancing model coding ability—not competing over chatbot market share or building moats.
Programming applications are still evolving, especially in Fortune 500 firms, where penetration remains low. But future opportunities may lie in scientific discovery, life sciences, physics, materials, robotics, and cybersecurity—tasks far more complex than general coding, with potentially larger markets.
Second, even in applications not requiring the most advanced models, we can leverage frontier model R&D to train smaller, more efficient, and lower-cost models. This ties to our compute-doubling techniques in model training. We improve efficiency in large models and apply these gains to small models.
For example, Haiku 5.5 charges $0.10 per million input tokens and $0.50 per million output tokens. It delivers competitive performance at very low cost—outperforming others in specific dimensions.
Thus, frontier models and low-cost models aren’t mutually exclusive. We must simultaneously push the limits of top-tier models and optimize efficiency so even inexpensive models remain competitive.
Dimon: Can AI help cure cancer and other diseases in the future? Can it accelerate major scientific breakthroughs like fusion or quantum computing?
Amodei: We’ve already seen steady improvements in model capabilities in mathematics.
In 2023, models struggled with high school math problems. Many believed chatbots, though fluent in dialogue, lacked mathematical and logical reasoning.
In 2024, models began solving high school math tasks effectively. By 2025, they started resolving previously unsolved mathematical problems—initially isolated cases. Some doubted these results appeared in training data or were mere lucky outcomes.
But as capabilities advanced, patterns changed. More open problems are being solved—including some of the most difficult in mathematics, with partial results approaching the level of research recognized by Fields Medalists.
I’m seeing similar early signs in biology and robotics. Life science experts, especially pharmaceutical CEOs, often remain skeptical. They’ve seen countless technologies aimed at drug discovery and understand the field’s deep challenges.
But AI has a unique trait: in early stages, it may appear incapable of anything—every barrier seems insurmountable. Critics point to clinical trials, manufacturing, or regulation as bottlenecks. Yet as technology advances, these barriers fall one by one.
I expect clearer changes in life sciences over the next few years. Not every field will progress equally—but I believe AI has the potential to drive biology forward at an unprecedented scale.
Dimon: Cybersecurity was already a major enterprise risk before Mythos and agents emerged. Now, AI can mass-detect vulnerabilities, and agents may perpetually attack corporate systems. JPMorgan employs thousands in cybersecurity and may need tens of thousands to patch vulnerabilities discovered by Mythos. How should enterprises prepare for the next phase?
Amodei: First, recognize that vulnerability detection is only the first step.
After Mythos launched, we conducted extensive vulnerability discovery. Clients identified hundreds of thousands; we did too. Then, the real bottleneck emerged: remediation. Detection speed is fast, but fixing vulnerabilities is far harder.
So we’re expanding Claude’s capabilities in vulnerability remediation—enabling models not just to identify issues, but to help enterprises fix them.
Second, reduce vulnerabilities at the source. As Claude’s coding abilities improve, we’re specifically training models to write more secure code—lowering the probability of defects during software development.
Third, establish real-time defense systems. Though not yet widespread, the potential is enormous.
Network attacks occur on timescales of hours, minutes, or even seconds. In the future, enterprises can deploy agent clusters to continuously monitor servers and networks, detecting and repelling attacks in real time.
If attackers use agent clusters, defenders must do the same. I anticipate increased adoption of such defensive agent clusters within the next 6 to 12 months.
Dimon: Anthropic is preparing for an IPO. Why now? How do you personally view going public? Will it make the company better—or more complex to manage?
Amodei: The primary reason is securing more capital for future growth and compute investment.
Fields like biology, robotics, cybersecurity, and broader knowledge work hold massive economic value. We need sufficient resources to capture these opportunities.
Returning to compute procurement: there’s a buffer between buying more capacity upfront and maintaining a conservative strategy. The more cash we hold, the wider this buffer becomes—better enabling us to navigate future demand uncertainty.
As for exact timing—what month, day, or market conditions—we haven’t over-optimized.
Before initiating the IPO process, I’d never experienced this. So I sought advice from many. The most memorable was: “Never let your company’s valuation exceed its true worth by even one dollar.”
This aligns with Anthropic’s core principle: “Less promise, more delivery.”
We won’t hype model capabilities prematurely. We won’t market models before they’re complete. Instead, we’ll build them first, confirm they’re strong enough, then officially launch. Safety systems follow the same rule.
I hope the IPO adheres to the same standard—no excessive promises driven by valuation goals.
Dimon: Anthropic announced establishing a philanthropy fund, allocating 5% of company equity to it. Why make this decision? What areas will the fund focus on?
Amodei: The philanthropy fund is part of Anthropic’s long-term social responsibility framework.
Beyond allocating 5% of equity, our co-founders have pledged to donate 80% of their personal wealth. From inception, we’ve maintained an employee stock donation matching program: if employees donate a portion of their stock to charity or designated causes, the company matches with additional shares.
We do this because AI is a transformative technology that has already created immense economic value for Anthropic and the entire industry—and will likely generate even more.
I envision AI helping cure cancer and Alzheimer’s, providing cheaper, abundant energy, creating superior materials, and raising overall material well-being.
But a critical concern remains: if vast new wealth stems from a single technology concentrated in a few companies, society risks extreme wealth inequality alongside economic growth.
We aim to mitigate this risk. The philanthropy fund is one mechanism.
How it operates and which issues it prioritizes are still under study. It may address workforce transitions due to AI, helping society adapt to changing employment structures. Or it could focus on biosecurity—ensuring AI accelerates biological advancement without enabling misuse.
We want to first identify truly impactful areas, then proceed gradually—not pre-committing to vague goals.
Dimon: If we look back 15 years from now, what do you most hope AI achieves—and what do you most wish to avoid? Is there anything we haven’t discussed but you’d like to emphasize to the audience?
Amodei: Above all, ensure this technology develops in the right way.
AI carries extremely high risks: misuse in biological or cyberattacks, economic disruption, loss of control over intelligent agents. These aren’t rhetoric—they’ve been central concerns since Anthropic’s founding.
We want not only to develop AI responsibly ourselves, but to set an industry benchmark. AI companies, clients, and suppliers must take these risks seriously, jointly establishing safety standards to keep technological progress under control.
At the same time, we must fully realize AI’s positive potential—especially breakthroughs in biology and medicine.
Currently, we’ve only begun to glimpse AI’s potential in biology and medicine. Many in pharma, biotech, and academia remain skeptical. But I hope we can prove these doubts unfounded in the future.
I dream of a day when heart disease, Alzheimer’s, and cancer are remembered like plague and smallpox—historical diseases conquered by humanity.
Sounds unrealistic? Perhaps. But I genuinely believe it’s achievable. Whether it takes 5, 10, or 15 years, we’ll do everything possible to accelerate that future.
Source: AI Observer
Disclaimer: Contains third-party opinions, does not constitute financial advice
The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups
10 days ago
Yushu Surges 629% on First Day of Trading, Humanoid Robots Sell 5,500 Units
08-19
Yushu's 61 Billion IPO: Who Will Make Money?
08-07
AI stocks down 40%-60%, yet NVIDIA's valuation is at a 10-year low
08-06
HBM Bottleneck Migration, NVIDIA Also Aims to Downgrade
08-06
AI Servers Fuel Demand for Minor Metals: Prices Surge Over 70% in Half a Year
08-06
AI is not a search bar—it's a supercomputer costing just a few dozen dollars
08-05






