Anthropic this week closed a $30 billion Series G at a $380 billion valuation and revealed annualised revenue of $14 billion. That makes it the fastest-growing software company in history. The trajectory is staggering: $0 to $100 million in 2023, $100 million to $1 billion in 2024, $1 billion to $14 billion through 2025 and into early 2026. Three consecutive years of roughly 10x growth. No software company has ever done that. Not Salesforce, not Slack, not OpenAI. Claude Code alone, launched less than a year ago, is now at a $2.5 billion run rate. Customers spending over $1 million a year went from 12 to over 500 in two years. Eight of the Fortune 10 are now Anthropic customers.
In a wide-ranging interview with Dwarkesh Patel this week, Anthropic CEO Dario Amodei was candid about where all this is heading. He expects growth to “slow” to 3-4x in 2026, targeting around $30 billion in annual revenue. The reason is not that demand is softening. It’s that you can’t physically build data centres and procure chips fast enough to serve continued demand. Demand that is increasingly creating a new paradigm. As Anthropic co-founder Jack Clark put it back in December, by summer 2026 he predicted “the AI economy may move so fast that people using frontier systems feel like they live in a parallel world to everyone else.” The revenue numbers suggest that parallel world is already forming.
Amodei remains resolutely supportive of the LLM’s transformer architecture, and offered a perspective on how he sees AI models actually learn. They are not as “sample efficient” as humans, he concedes, requiring far more data to pick up a new concept. But within their context window, they can learn and adapt remarkably well. And the comparison to human learning, he argues, is less clear-cut than people assume. The human brain is not a blank slate at birth. It arrives pre-loaded with millions of years of evolutionary architecture, instincts and inherited structures that took billions of parameter-equivalents of natural selection to build. When you account for all that embedded prior knowledge, models start to look less like slow learners and more like a different kind of fast learner, one that trades our biological head start for raw scale and speed.
Amodei argues that even without solving continuous on-the-job learning, in-context learning within a million-token window gets you most of the way there, and that’s enough to generate “trillions of dollars of revenue.”
Takeaways: There is a persistent narrative that AI infrastructure is being overbuilt. Anthropic’s numbers suggest the opposite. A company that tripled revenue to $14 billion and is still constrained by how fast it can pour concrete and rack GPUs is not a company outrunning demand. Rather than a supply glut, the reality, for now, is a fundamental supply constraint.