Written by: Techub News
In a recent public dialogue in New York, Dario Amodei, co-founder and CEO of Anthropic, engaged in an in-depth conversation with Andrew Ross Sorkin, a columnist for The New York Times. As an early core member of OpenAI, a key leader in the development of ChatGPT-2/3, and now at the helm of Anthropic, whose valuation has soared, Amodei's perspective combines the insights of a technology pioneer with the practical considerations of an industry leader. In this dialogue, he not only addressed external criticisms of overheated AI investment and the "bubble" theory, but also candidly shared Anthropic's survival strategy within the "cone of uncertainty," and offered his views on pointed issues such as AI's national security, regulatory impact, and job creation.
Technological Optimism and Financial Prudence: Survival Rules in the AI Craze
When asked whether the current massive investments of hundreds of billions of dollars in the AI field constitute a bubble, Dario Amodei gave a dialectical answer. He broke down the question into two aspects: technological prospects and economic realities.
On the technical front, Amodei describes himself as "one of the most optimistic people." His confidence stems from twelve years of observation and practice of the "Scaling Laws." These laws, first systematically articulated by him and his colleagues, demonstrate that with continued investment of computing power and data, the performance of AI models steadily improves across almost all tasks—from programming, science, and biomedicine to law, finance, materials, and manufacturing. "This covers almost every area of value creation in the modern economy," he summarizes. Amodei cites Anthropic's own growth as an example: the company's revenue grew from zero to $100 million in 2023, reached $1 billion in 2024, and is projected to fall between $8 billion and $10 billion in 2025. This tenfold annual growth rate, in his view, is direct evidence that the value of technology is being realized.
However, switching to an economic and financial perspective, Amodei expressed clear concerns. The core of the problem lies in a fundamental "uncertainty": the trajectory of future revenue growth is extremely difficult to predict, but building the computing infrastructure (data centers) needed to support that revenue requires a lead time of one to two years. Enterprises must now decide how much computing power they need to purchase by early 2024 to support the expected business scale by early 2027.
“This forms a cone of uncertainty,” Amodei explained. “I can’t be sure whether revenue a year from now will be $20 billion, $50 billion, or some other number.” He tried to plan conservatively, but risks were everywhere: if the company didn’t purchase enough computing power, it would be unable to meet customer demand and hand over the market to competitors; if it over-purchased, it could fall into financial crisis or even go bankrupt because revenue could not cover the high fixed costs.
Amodei points out that a company's ability to withstand such risk depends on its profit margins. High profit margins provide a buffer. He suggests that Anthropic, with its focus on the enterprise market (rather than the consumer market), has a better business model and higher profit margins, and therefore can manage this risk in a "relatively responsible" manner. But he shifts gears, making it clear that not all players are so cautious: "My concern is that some players in the ecosystem could do terrible things if they just make a timing mistake, or just deviate a little bit." When pressed on which companies are "YOLO" (recklessly taking risks), he avoids the question, but implies the answer is obvious.
In response to OpenAI CEO Sam Altman's goal of turning from huge losses to profitability within two years, Amodei declined to comment on the financial situation of other companies, but reiterated Anthropic's prudent calculation principle based on the "cone of uncertainty": purchasing enough computing power to ensure affordability even in the worst-case scenario (such as the 10th percentile) while remaining competitive.
Amodei defended the legitimacy of the industry's "circular transactions" (i.e., chip suppliers like Nvidia investing in AI companies, which then use that capital to buy their chips). He used a simplified model to illustrate: assuming building a 1-gigawatt data center requires approximately $50 billion in capital expenditure, spread over a 5-year lifespan, the annual cost is about $10 billion. A rapidly growing company with nearly $10 billion in annual revenue might not currently have $50 billion in cash, but partnering with a giant that has capital and is willing to sell chips, obtaining some investment to cover the first year's costs, and then using revenue to pay for subsequent payments, is commercially viable. "In principle, there's nothing wrong with it," he emphasized. The risk only lies in overexpansion when such transactions become extremely large and future revenue expectations are overly aggressive.
Regarding the debate on chip depreciation cycles, Amodei once again demonstrated his conservative stance. He believes the problem lies not in the physical lifespan of chips, but in the rapid iteration and lower cost of new chips reaching the market, causing older chips to depreciate quickly. "This could happen as early as a year after you buy a chip." Therefore, Anthropic's planning assumes a very aggressive chip efficiency improvement curve and makes conservative financial forecasts accordingly. "We believe we can handle almost all scenarios." However, he reiterated that he cannot speak for other companies and hinted that there might be some overly optimistic assumptions.
Enterprise-level approach vs. consumer-level war: Anthropic's differentiated moat
The recent release of Google's new model, which caused a stir in the industry, and the "red alert" issued internally by OpenAI, highlight the fierce competition in the consumer AI market. Dario Amodei expressed his "deep gratitude" for the differentiated path chosen by Anthropic.
Amodei points out that the core battleground for OpenAI and Google is the consumer market. Google needs to defend its search monopoly, while OpenAI's focus is also on the consumer business. Serving enterprise customers is secondary for them. In contrast, Anthropic has focused on the enterprise market from the beginning. Over time, its models have become increasingly optimized for enterprise needs—with the fastest improvement in coding capabilities, and it is expanding into various fields such as finance, biomedicine, retail, energy, and manufacturing.
“These model wars, despite our excellent models…we are developing in a different direction or on a different dimension in some ways.” Amodei believes that this allows Anthropic to be less affected by the fierce battles in the consumer market, placing it in a “privileged position” where it can focus on continued growth and model development without issuing “red alerts.”
So, what exactly is the competitive advantage of AI companies? When model capabilities eventually converge, will users frequently switch simply because someone has the newest and most powerful model? Amodei provides a negative answer from the perspective of the enterprise market.
First, he emphasized the fundamental differences between building models for businesses and for consumers. Business-level models focus more on coding, high-intellectual activity, and scientific capabilities than on user engagement. This different optimization objective leads to significant differences in the "personality" and capability emphasis of the models.
Secondly, even with the advent of Artificial General Intelligence (AGI), Amodei does not believe that all models will converge to the same point. "Specialization exists alongside general intelligence." Just as humans possess general intelligence but excel in different areas, AI models can and will specialize.
Finally, he pointed out the inherent stickiness of the enterprise market: companies build relationships with suppliers and become accustomed to using specific models. Even in seemingly standardized API businesses, enterprise customers find it difficult to switch between different models because downstream customers have become accustomed to the interaction methods and "personalities" of existing models.
When asked whether the existing Transformer architecture and computing power alone could lead to AGI, Amodei gave an affirmative answer. “I believe scaling will get us there,” he reiterated, citing the scaling law and predicting that only occasional minor improvements would be needed in the future. He likened the advancement of AI capabilities to the exponential curve of Moore's Law: models will become increasingly powerful in all aspects, with new models released every few months representing breakthroughs in coding, science, and mathematics. He revealed that researchers within Anthropic have already stated they will no longer write code themselves, but will instead let Claude generate the initial drafts, with themselves only responsible for editing. “This process will continue… what we will see in the future is simply a continuation of the past, only to a greater extent.”
National security, regulation, and employment: The public responsibility of AI leaders
Dario Amodei is known for his outspokenness on AI policy issues, particularly regarding the export of advanced chips to China. Despite Anthropic's partnership with Nvidia and Nvidia CEO Jensen Huang's past dissatisfaction with his remarks, Amodei has made it clear that his views remain unchanged.
He characterized this as a national security issue, not an economic one. Amodei painted a picture: as model capabilities grow exponentially, a "genius nation in the data center" will eventually emerge. The geographical location of this "nation" is crucial. "If it falls into an authoritarian state, I think they can surpass us in every way: intelligence, defense, economic value, R&D… I fear they will be able to oppress their own people and build a perfect surveillance state." Therefore, he believes democratic countries must gain this advantage first; it's an "absolute necessity." Selling the most advanced chips to China "will only increase their chances of arriving at their destination first; that's common sense."
When the topic shifted to the risks of surveillance within democracies, Amodei attempted to elevate the discussion to the level of policy principles, rather than targeting specific individuals or governments. His core principle was: “We should actively use these models in all possible ways, except those that would make us more like our authoritarian adversaries. We need to defeat them, but we cannot take actions that would lead us to become like them.”
Amodei refuted accusations by White House AI chief David Sacks that Anthropic was engaging in "sophisticated regulatory capture" based on "fear marketing" and harming the startup ecosystem. He pointed out that he had written papers on AI risks long before founding his company in 2016, and that major AI bills Anthropic supports (such as SB 53) have exemptions for small companies with annual revenues below $500 million. "These allegations are completely unfounded."
He further explained the differences in stance on regulatory issues. He believes that some people draw parallels between the AI revolution and the internet or telecommunications revolution, trusting that the market will solve everything itself. This view may have been reasonable in the past, but "those closest to AI don't think so." Genuine AI researchers are both excited about the potential and concerned about national security risks, model alignment issues, and economic shocks. He described proposals to suspend all state-level regulations within a decade (in the absence of a federal framework) as as dangerous as "removing the steering wheel from a car because you won't need to steer for ten years."
Regarding the impact of AI on employment, Amodei previously predicted that perhaps half of entry-level jobs could be affected. In this conversation, he elaborated more systematically on three levels of addressing this challenge.
The first tier (private sector-led): Enterprise clients face a trade-off when using AI. They can leverage AI to automate existing processes (such as insurance claims processing and the "Know Your Customer" process), significantly improving efficiency and reducing costs, while drastically reducing the required human resources. Simultaneously, they can also use AI to create substantial new value. Even with AI handling 90% of the work, human employee productivity could be leveraged tenfold, sometimes requiring even more employees to handle a hundredfold increase in workload. Society should encourage businesses to focus more on this second type of value creation.
The second layer (government intervention): Amodei believes that retraining programs are not a panacea, but the government ultimately needs to intervene financially. Rapid AI-driven growth will create a huge economic "pie," and the government needs to use tax policies and other means to ensure that wealth is not excessively concentrated and to support workers in the transition.
The third layer (restructuring of society): In the long run, a society with powerful AI must be different. He cites Keynes' vision of "the economic possibilities for our children and grandchildren," suggesting that in the future, people may only need to work 15 to 20 hours a week. For many, the core meaning of work may shift from economic survival to self-actualization. "Society is flexible... We need to work together to figure out how to function in the age of Artificial General Intelligence (AGI)."
Amodei concluded that his warnings were not intended to spread pessimism, but rather because "early warnings are the first step in solving problems." Only by recognizing the potential pitfalls ahead can society avoid them, rather than blindly stepping into them. This leader in the field of AI, while demonstrating immense optimism about the future of technology, also shoulders the responsibility of urging society to prepare for unforeseen circumstances.
