Anthropic's Jack Clark Says the "Singularity Is Here," Citing an Internal Economic Model That Implies $44 Trillion in US GDP by 2030
Anthropic co-founder's keynote at AF14, October 18, 2026, pairs a bold growth scenario with a push to build lab and robotics capacity
Anthropic co-founder Jack Clark told the AF14 audience that the company's in-house economists have modeled a scenario in which US GDP reaches $44 trillion by 2030. He conceded the figure "sounds very outlandish," but argued it is "implicated by the technology continuing to get better." The model assumes, as "people at the labs do," that AI systems will autonomously complete a broad range of tasks and that diffusion into the economy accelerates. Clark gave no detail on baseline assumptions, adoption curves or the other trajectories Anthropic simulated. Investors should treat the number as a scenario output from an interested party, not a forecast. Still, it shows how aggressively the company is framing the economic stakes.
Clark added a logical corollary. If GDP expanded that much, "economic doctrine suggests that new industries and inventions would need to have been created to power such a major economic expansion." In his telling, that is exactly what autonomous AI scientific inquiry would supply, which makes science rather than chatbots or coding the main driver of the thesis.
The Core Argument: Cheap, Copyable Researchers Break the "Ideas Are Getting Harder to Find" Problem
Clark's economic case rests on the long-running decline in researcher productivity, with headcount rising rapidly while breakthroughs get costlier. He argued AI changes both variables in that equation: the number of researchers and the time needed to train them. He framed it personally: "There is one of me and it took me about 40 years to get here. What would it mean to be able to run hundreds or thousands of copies of me simultaneously?" Training a new cohort of human specialists takes "10 or 15 years of targeted education and millions of dollars," he noted, while AI systems "can be created immediately in arbitrary numbers as long as there is enough computing capacity."
His read is that human scientists become more valuable, not less, as each supervises "hundreds or thousands of like-minded experts." He did not address how this squares with labor displacement outside frontier science. The speech was silent on the distributional costs of the growth it describes.
Evidence: Frontier Science Benchmarks and a Telling Physics Anecdote
Clark said 2026 has brought the first signs that AI can do science at the frontier. He cited AI contributions to proposed solutions on problems including the Navier-Stokes Millennium Prize, the Jacobian conjecture and several open Erdos problems, and said AI supplied "core insights" in many cases. In biology he pointed to AI-designed viruses and drug candidates, with some drugs in trials that were "in large part designed by AI systems." He also noted AI increasingly operates laboratory equipment autonomously.
The benchmark data he offered was stark. On Humanity's Last Exam, top systems scored 8% at the January 2025 launch, and the leader this week is Claude Opus 5.5 at 67.7%. On FrontierMath, best scores rose from 2% at the November 2024 launch to 94% now. Clark said he has assembled roughly 100 graphs in the last week showing the same pattern across scientific fields, with acceleration beginning "in the last two years, sometimes just one."
His most memorable evidence was a paper released last week, TheoPhys-Eval, built from 50 unpublished theoretical physics problems that the authors' own research had produced. Clark quoted one author, Indian physicist Suat Raju: "I don't know of any individual who can solve any significant fraction of our final collection of problems without AI assistance." Initial scoring put GPT-6 Astra in the late 80s, followed closely by Opus 5.5. But the models, especially Astra, insisted the expert-supplied answers were wrong. On review, the experts found "subtle errors buried deep inside their calculations," and after they revised their answers Astra reached 98%, or 49 of 50. Clark's takeaway is that frontier science implicitly tests creativity, one of the last unproven AI capabilities, and that success there signals generalization, "including certainly AI research itself, which could have an explosive effect on innovation." Notably, he cited a rival lab's model as a top performer, which lends the evidence some credibility.
Coding Productivity and the Cyber Double Edge
Clark offered one internal operating metric. Anthropic's researchers now produce almost 10 times the code they did in the company's first four years, and the increase "basically directly tracks advancing AI capabilities." He said AI systems are now "more capable and efficient than the vast majority of the world's software engineers."
He was candid about the downside. The same coding gains have made models much better at both cyber defense and cyber attack, and he stressed that the two are "bound up inseparably." He called this "a taste of things to come."
Infrastructure: The Next Build-Out Is Labs, Equipment and Robots
Clark said his decade of lobbying governments to build electricity capacity and data center space was largely ignored: "Not many policy makers listened. I think now we deal with the consequences of that." Capacity is concentrated in the US, China and increasingly the Middle East, and supply chains are so contested that "prices are going up."
He argued the compute build-out was only phase one. The next phase is expansion into the physical world, with AI running experiments and operating robots. With roughly 10 million scientists worldwide per UNESCO, he said there are already too few humans, laboratories, biological labs and large scientific instruments to meet existing demand. The prospect of a billion or even 10 billion virtual scientists means "we will run very quickly out of capacity and we'll see history repeat itself." The read-through for investors is a new demand vector in laboratory automation, scientific instruments and robotics, plus raw materials supply. Clark suggested developing countries supplying basic inputs could "leapfrog."
He also flagged a downstream bottleneck. "An avalanche of proposed new drugs is of little use if only a handful can ever be tested," he said, calling on policymakers to speed the path from lab to clinic, factory and field. That implicitly concedes that regulatory and clinical throughput, not discovery, may cap the realized economic gains.
Governance: Restraint Over AI-Driven AI Research
The most consequential admission may be about AI research itself. Clark said labs like Anthropic "will need to make decisions about how much oversight we apply to AI-driven research," which could accelerate progress but means "we will need to move extremely carefully here because otherwise we could lose control of the technology." He called for the strictest human oversight in biological research and said AI's creativity is already forcing humans to spend time decoding its mathematical advances. He also raised the question of whether humans must understand scientific equipment that AI designs for itself within "a few short years."
Clark closed by calling for "unusual and serious rules for this technology and the companies building it," arguing the world will need to "go slowly and deliberately" on development even as it builds infrastructure to use it. He also said the window to act is "a handful of years, less than a decade easily." That position sits uneasily with a $44 trillion growth scenario that depends on rapid capability gains and diffusion. Clark acknowledged the tension without resolving it.