Dario Amodei thinks there is a 50 percent chance that within one to two years, the world will have something he calls "a country of geniuses in a data center" — an AI system smart enough to do any intellectual task a human can do, at superhuman scale. He puts the odds at 90 percent within a decade. He is not writing this in a blog comment section. He is the CEO of Anthropic — valued at approximately $380 billion with a net worth to Amodei of roughly $7 billion, one of the most consequential companies in the world, and he has spent the last year trying to get Washington, policymakers, and the public to take this seriously before it arrives.
The primary evidence for why Amodei might be right is the growth curve. Anthropic's annualized revenue went from essentially zero in 2023 to $100 million, then to $1 billion in 2024, then to $9–10 billion in 2025. In January 2026 alone, it added several billion more. By March 2026, annualized revenue had reached $19 billion — a 1,167 percent year-over-year increase, according to Sacra, with Yahoo Finance also confirming the surge to $19B. These are not projections. They are the actual numbers that came in.
"Every year we predict exactly what the demand is going to be, we'll be profitable every year," Amodei said in the interview. "Because spending 50 percent of your compute on research, roughly, plus a gross margin that's higher than 50 percent and correct demand prediction leads to profit." The implication — that Anthropic is deliberately underbuilding compute relative to demand — is itself a bet on how fast the world will absorb AI capability.
The revenue numbers matter because they fund the infrastructure that Amodei believes will produce the "country of geniuses." The compute math is stark. Amodei estimates the AI industry is currently running 10–15 gigawatts of compute capacity. At a rough cost of $10–15 billion per gigawatt per year, that is $100–225 billion in annual spending today. He expects the industry to triple compute each year — a projection Zvi notes means roughly 30–40 gigawatts in 2027, around 100 gigawatts in 2028, and approximately 300 gigawatts by 2029. If the math holds, the industry will be spending roughly $3–4.5 trillion annually on compute within three years — a figure that approaches the entire US defense budget. That figure represents total industry-wide compute spending, not any single company's budget. Amodei then illustrated what being wrong about demand by even one year would mean for a company making that level of commitment: "If my revenue is not $1 trillion, if it's even $800 billion, there's no force on earth, there's no hedge on earth that could stop me from going bankrupt if I buy that much compute." The personal bankruptcy scenario he describes is a hypothetical for a company that over-committed at industry scale — not a claim that he or Anthropic is personally spending trillions. "In 2028, we get the real country of geniuses," Amodei said. "It is hard for me to see that there will not be trillions of dollars in revenue before 2030."
The technical case rests on what Amodei calls the Big Blob of Compute Hypothesis, a framework he wrote in 2017 — before GPT-1 — arguing that only a handful of variables determine AI progress: raw compute, quantity and quality of training data, training duration, and a scalable objective function. Everything else, he argues, is noise. What has changed in the current moment is that the same log-linear scaling gains Amodei observed in pre-training are now appearing in reinforcement learning. "We are seeing the same scaling in RL that we saw for pre-training," he said. The OSWorld benchmark for computer use climbed from roughly 15 percent to 65–70 percent over approximately one year, a trajectory that mirrors earlier pre-training improvements that eventually cleared that threshold and kept going.
Amodei has elaborated on what "country of geniuses" means in practical terms: cluster sizes by approximately 2027 will allow millions of AI instances operating at superhuman speed, equivalent in knowledge output to 50 million Nobel Prize winners — a concept Amodei introduced publicly in a January 2026 essay that drew wide attention for its framing of AI's eventual scale. He has also said AI is already writing much of the code at Anthropic to build the next generation of systems, substantially accelerating progress.
Anthropic is building Claude Code — which by February 2026 had reached $2.5 billion in annualized revenue, according to industry analysis — partly to validate its own models internally before shipping them externally. Some Anthropic engineers now write no code at all, relying entirely on Claude to produce it. Amodei estimates that coding models currently provide a 15–20 percent total factor productivity improvement, up from roughly 5 percent six months ago. The improvement curve, he argues, is the kind of thing that starts slow, barely registers, and then crosses a threshold where it becomes impossible to ignore.
The economic risk is asymmetric in a way that Amodei has thought through carefully. Amazon plans to spend $200 billion this year alone, while Alphabet projected up to $185 billion and Meta up to $135 billion in capital expenditures — Fortune reported that Anthropic itself committed $50 billion to US data center infrastructure in Texas and New York. "We bought an amount that allows us to capture pretty strong upside worlds," Amodei said. "It won't capture the full 10x a year. Things would have to go pretty badly for us to be in financial trouble." The responsible compute stance — the thing that separates Anthropic from the "YOLO" approach he attributes to some competitors — is in this framing not caution but math.
There is a second risk that does not fit neatly into the spreadsheet. In late 2025, the US Department of Defense demanded that Anthropic support "all lawful uses" of Claude, including for military operations. Anthropic refused. In February 2026, Amodei wrote an internal memo calling OpenAI's reported Pentagon deal "safety theater," The Atlantic reported. In March 2026, the Trump administration moved to blacklist Anthropic from federal contracting, making it the first American company to be publicly designated a supply chain risk under a statute historically reserved for foreign adversaries. Anthropic executives have said the blacklisting could cost billions, CNBC reported. Judge Rita Lin blocked the blacklisting on March 26, ruling the action likely unlawful First Amendment retaliation. The DoD has reportedly been extensively using Claude for military operations, including in target selection and analysis of missile strikes in its war against Iran, The Guardian reported. But the confrontation revealed something the revenue curve cannot capture: that the "country of geniuses" is not just an engineering problem. It is a national security problem, and the two may be on a collision course.
Anthropic's Chief Financial Officer has said the company surpassed $5 billion in cumulative GAAP revenue through December 2025, Reuters Breakingviews noted, with annualized revenue at approximately $14 billion as of mid-February rising to $19 billion by month-end. SaaS-focused analysis put Claude Code's run-rate revenue above $2.5 billion, having more than doubled since the start of 2026.
"The governments of the world may have to work together to make it happen," Amodei said when asked about governance. "We may have to talk to AIs about building societal structures in such a way that these defenses are possible." He was not being whimsical. He was describing the actual gap between where the technology is going and where the institutional infrastructure exists to govern it. "My worry," he said, "is that this is happening all so fast. So maybe we need to do our thinking faster about how to make these governance mechanisms work."
The most honest thing in the interview may be the prediction Amodei made about his own uncertainty. "If we had the country of geniuses in a data center, we would know it," he told Dwarkesh. "Everyone in this room would know it. Everyone in Washington would know it." The interview was recorded in February 2026. The 50 percent chance window — one to two years — closes sometime in 2027 or 2028. Whether that moment arrives on schedule, overshoots, or comes early because someone found a way to accelerate RL scaling on a cluster that should not have been ready yet, the revenue numbers and the judge blocking the Pentagon blacklisting are both telling you the same thing: the world is not prepared, and the clock is running.
