Anthropic Technical Leads Peg AGI at "Next Couple of Years," Defend $80 Billion Run-Rate Against Distillation and Regulatory-Capture Critics
American Optimist podcast, Napa, August 2026 — Nick Marwell and Schult Douglas, Anthropic's reinforcement-learning tech leads, lay out the company's AGI timeline, its economic defense of frontier pricing, and why it walked away from a multi-month model lead
Two of Anthropic's most senior technical staffers used a long-form conversation on American Optimist to make a set of claims that will matter to anyone pricing the AI infrastructure trade: models "as or more capable than all humans" are, in their words, "very likely" within the next couple of years, and the economics of frontier intelligence are exponential rather than linear. Nick Marwell, who helps lead RL science, and Schult Douglas, a reinforcement-learning tech lead, also addressed Anthropic's roughly $80 billion run-rate — a figure that has leaked but which the company does not officially confirm — along with the distillation fight with Chinese labs, the regulatory-capture criticism increasingly aimed at the company, and where it believes cyber and biological risk are headed over the next two years.
The Exponential Case for Frontier Pricing
The most consequential argument in the conversation is economic rather than technical: Marwell and Douglas contend that each marginal unit of model intelligence is worth exponentially more than the one before it, which is why they believe frontier labs retain pricing power even as lagging-edge capability gets commoditized within months. Marwell pointed to coding as the proof case. "The first unit of intelligence in coding that was useful at all to people was tab auto-completion," he said. "You moved from this into the next marginal unit of intelligence where you made agentic coding a reality in your terminal with Claude Code. And this was a 100,000x more valuable capability." His conclusion is blunt: "You can't even really have a company based on tab auto-completion at this point."
That framing matters for how investors should think about margin durability across the sector. Total AI revenue across the industry is, in Douglas's estimate, "somewhere north of $100 billion into a tens and tens of trillions of dollars economy" — implying the category is still a rounding error relative to its addressable base, and that the leading edge, not the lagging edge, is where value concentrates. Marwell's historical reference point is drug development IP protection, but with a twist: AI's "shelf life of the frontier is measured in months," making the deflation curve far steeper than pharma, with capability-per-dollar falling roughly 10x per year according to Artificial Analysis benchmarks cited in the discussion.
Inside the $80 Billion Run-Rate and the Shift From Tool to "Entity"
Pressed on the leaked run-rate figures that reportedly place Anthropic ahead of OpenAI in growth, Douglas and Marwell declined to confirm numbers directly but offered a granular account of internal capability progress that underpins the growth story. Douglas described the shift in how engineers inside Anthropic now work with their own models: eighteen months ago he was "typing all of my lines of code by hand," then "intervening every few minutes" to correct a model that would "go off course." Today, he said, "I can ask models to do a day or two days of work independently and drive progress basically like a junior team member." That compounding autonomy — from tool to semi-independent contributor — is the operational basis for Anthropic's view that it should be thought of as deploying "entities" rather than software tools, a framing Douglas said is "a useful frame" because these models increasingly "take actions on your behalf" rather than requiring constant supervision.
Why Distillation Is an Existential Threat to the Business Model
The pair gave one of the more detailed public explanations yet of why Anthropic treats distillation — the practice of training new models by harvesting outputs from frontier systems — as a direct threat to the capital structure of the industry, not merely an IP nuisance. Marwell argued that distillation "is not a method that's capable of bringing you past the frontier of current capability" and that allowing it unchecked breaks the economics that justify the enormous capex now flowing into training runs. "We're going to be in regimes... where there are $10 billion, $100 billion, $1 trillion training runs," he said. "You can only support that kind of R&D investment if you're able to then make money selling the thing that you invested in." If a rival can copy a frontier model "for a fraction of the cost" and release a near-equivalent "in a week or two," that rival becomes "an economically advantaged competitor" without having borne the R&D load.
Crucially, the two acknowledged Anthropic cannot fully stop distillation through technical means alone. Douglas described the dynamic as "a cat-and-mouse game": giving developers access to model outputs, reasoning traces, and local inference — all features customers demand — simultaneously makes distillation easier. There is no clean technical fix, only a tradeoff between product openness and defensibility. That admission is notable for investors trying to underwrite long-term moats in foundation models: the barrier is less a wall than friction, reinforced largely by the pace of frontier advancement itself rather than by enforcement.
Offense-Dominant Today, Defense-Dominant in Cyber Within Two Years
On safety, the two offered a specific and somewhat bullish prediction: cyber risk, currently offense-dominant because attackers can exploit AI faster than defenders can patch, should flip to defense-dominant within roughly two years as organizations use the same models to proactively find and fix their own vulnerabilities. "Right now both cyber and bio are offense-dominant," Douglas said, "but I think that over the next two years cyber becomes defense-dominant." Biology is a harder problem. Douglas was candid that reversing offense-dominance in bio "is going to require hundreds of billions of dollars of infrastructure" and realistically depends on "dramatically advanced capabilities in robotics in the 2030s" to build out that defensive infrastructure at scale — a much longer and more capital-intensive runway than cyber.
Marwell separately flagged a less-discussed bottleneck for AI's application in life sciences: physical lab capacity, not model capability, is now the constraint. "I am almost more concerned if we have the physical and lab infrastructure required to take advantage of this than I am about whether AI will be able to be a useful intellectual tool," he said, adding that U.S. lab infrastructure is "clearly lagging behind places like China" on quality metrics — an implicit call for biotech-adjacent capex that current public-market biology plays are not yet pricing.
Answering the Regulatory-Capture Critique
Addressing the accusation — increasingly voiced in venture and open-source circles — that Anthropic's safety advocacy is a disguised strategy to entrench its lead through regulation, Marwell offered a specific counterexample: the company's handling of what he referred to as the Fable (formerly Mythos) model rollout. At a moment when Anthropic reportedly held a model lead with "nobody" else close, the company chose to "go work with the government to roll it out in a way the government was comfortable with," rather than exploit the gap. "This was just a decidedly bad economic decision for the company," Marwell said, "but one that was aligned with our values about how we think models should be deployed and regulated." He also pushed back on the idea that open-weight regulation disproportionately targets smaller players, noting that the biggest open-weights developers today are Nvidia, Amazon, Microsoft, and well-funded Chinese state-linked labs — not scrappy open-source communities — undercutting the claim that regulation is designed to disadvantage underfunded rivals.
On the broader skepticism that CEO Dario Amodei is "trying to scare everyone to create more rules to slow things down," Douglas's response was direct: "So far we've slowed ourselves down much more than anyone else." He cited DeepMind CEO Demis Hassabis's proposal for closed-model labs to jointly adopt pre-release safety thresholds as evidence the restraint is voluntary and self-imposed rather than a lobbying tactic.
No Unilateral Slowdown Against China
Despite the safety rhetoric, both were unambiguous that Anthropic would not support the U.S. unilaterally slowing its AI development while China continues. "I don't think any scenario that we're comfortable with would involve slowing America and letting China go ahead," Douglas said, adding that any coordinated pause "would have to be fully coordinated in a way that we fully trust every counterparty." On open source specifically, Anthropic's stated position is not opposition but parity: the company wants capability-based release thresholds — for example, around cyberattack capability — applied equally to open and closed models, rather than restrictions targeted at open-weights developers specifically.
Compute Trajectory and the 2028 Checkpoints
On infrastructure, Douglas sketched a capex trajectory worth tracking against hyperscaler guidance: industry-wide AI capex is roughly $1 trillion this year, and if the historical pattern of 2x to 3x annual compute growth holds, that could reach $2 trillion next year and $4 trillion by 2028. Extrapolated further with robotics, Douglas suggested the early 2030s could see global GDP effectively doubling — a claim he flagged as requiring "a lot of things to go right" rather than a base case. For 2028 specifically, he expects a terawatt of AI compute capacity online, "tens of thousands" of humanoid robots performing basic household tasks, and said he would not be surprised if an AI model wins a Fields Medal before the end of the decade, though he was skeptical of AI reaching a Pulitzer or Nobel Prize-caliber achievement on a similar timeline.
On near-term labor data, Marwell pointed to a counterintuitive signal: Philippine business-process-outsourcing employment and U.S. software engineering employment have both continued to rise despite AI automation, which he attributes to a transitional phase where AI makes existing workers more valuable rather than redundant. He cautioned against reading too much into that trend, however: "At some point... we're going to go from a world where you need a human paired with an AI model to get all the value out of it, to a world where you don't. And I think that's the world where we start to get very nervous about the consequences" — drawing a direct parallel to chess, where human-plus-machine superiority over pure machines lasted roughly three decades before collapsing.