On the eve of an exponential explosion in AI technology, Anthropic's leader Dario Amodei has presented a startling prediction for the industry: we are in the "twilight of exponential growth," and as early as 2026, humanity will witness the emergence of a "country of geniuses in a data center," composed of tens of thousands of top-tier minds. In a recent in-depth conversation with Dwarkesh Patel, the CEO of the large model unicorn Anthropic, Dario Amodei unusually disclosed the company's astonishing revenue growth expectations and provided detailed commentary on the timeline for AGI (Artificial General Intelligence), the financial logic behind compute investment, and geopolitical risks. Amodei believes AI technology is at a critical turning point from quantitative to qualitative change, and the next 2-3 years will determine the course of the next two hundred years for humanity.
AI is Nearing the End of its Exponential Growth Phase At the outset of the interview, Dario Amodei pointed out that we are approaching the end of the AI exponential growth curve, a qualitative shift the world has yet to fully grasp. From GPT-1 to today's professional-grade models, AI has leapt from the level of a "bright high school student" to that of a "PhD," even surpassing human capabilities in areas like programming and mathematics. The underlying scaling laws have held firm, with continued investment in compute and data delivering clear returns. The magic of exponential growth lies in its final explosive phase. Dario stated that Anthropic's tenfold annual revenue growth, Claude Code doubling engineer productivity, and rapid breakthroughs in model context length and generalization ability all signal that the "endpoint is near." This growth is not merely about stacking parameters but represents an upgrade in the essence of intelligence—shifting from data fitting to autonomous generalization, with AI assembling the final crucial pieces of the puzzle.
"Country of Geniuses in a Data Center": Redefining 2026 Amodei introduced a highly impactful concept during the interview: "A Country of Geniuses in a Data Center." Reflecting on the technological evolution of the past three years, he believes AI models have progressed from "smart high school students" to "professionals." He boldly predicts that by 2026 or 2027, the intellectual level, depth of knowledge, and logical reasoning capability exhibited by a single model will not only be equivalent to one Nobel laureate but to the collective output of tens of thousands of top geniuses working in concert. Regarding the certainty of this timeline, Amodei expressed high confidence: "I am about 90% confident this will happen within 10 years; for it happening within the next 1-2 years, I see it as a 50/50 possibility." He noted that the only variables might stem from geopolitical disasters, such as disruptions in the chip supply chain, or severe social upheaval.
Revenue Surge: The "Terrifying" Curve from $100 Million to $10 Billion Regarding the financial metrics most watched by the market, Amodei disclosed Anthropic's staggering growth curve. He revealed the company is experiencing "bizarre 10x per year growth." Amodei stated plainly in the interview: "In 2023, we grew from $0 to $100 million in revenue; in 2024, from $100 million to $1 billion; and for 2025, we anticipate reaching $9 to $10 billion. This exponential growth roughly aligns with my expectations; even in the first month of this year, we added several billion dollars in revenue." Amodei emphasized that despite the lagging effects of economic diffusion—where enterprise adoption of AI involves lengthy processes like legal review and compliance checks—the improving capability of the technology itself is driving this frenetic growth curve.
The Compute Gamble and Bankruptcy Risk: A CEO's Financial Balancing Act Faced with such a certain technological outlook, why not borrow trillions now to hoard chips? Amodei provided a highly practical financial explanation: compute expansion must be tied to revenue growth and forecast accuracy, otherwise it risks catastrophic failure. "If I predict trillion-dollar demand in 2027 and thus pre-purchase a trillion dollars worth of compute, but the demand surge is just one year late, or the growth rate dips slightly from 10x to 5x, there is no hedge that can prevent the company from going bankrupt," Amodei explained. He noted that the return on investment, based on these "logarithmic return laws," requires precise calculation. He indicated that Anthropic's current strategy is "responsibly aggressive," meaning the scale of compute investment is sufficient to capture massive upside potential, but if the market explosion is delayed, the company can still survive thanks to the high margins and cash flow from its enterprise-grade business. He anticipates Anthropic could achieve profitability around 2028, at which point AI will become one of the most profitable industries in history.
The Endgame of Software Engineering: From Writing Code to Replacing Engineers Regarding specific application scenarios, Amodei identified programming as the first fortress AI will conquer. He outlined three stages of evolution for AI in software engineering: Stage 1: Models write 90% of code lines (already achieved). Stage 2: Models handle 90% of end-to-end tasks, such as fixing bugs, configuring clusters, and writing documentation. Stage 3: Models possess "workplace experience," enabling them to understand the context of complex codebases and set technical direction. Amodei predicts that within 1-3 years, AI will be capable of performing all the duties of a senior software engineer. "This doesn't mean engineers will be unemployed, but rather there will be a massive explosion in productivity. What models can do now is not just complete code, but directly take over high-difficulty tasks like writing GPU kernels."