Two Anthropics

×

Anthropic’s safety mission began with a lab that would pull rivals upward. Five years later, it was one of the most aggressive frontier scalers, with a $380 billion valuation. Its founding thesis treats capability at that scale as the reason safeguards are necessary.

A personal note first. This post is an outtake from a 14,000-word profile of Dario Amodei that I published on May 9, 2026. I don’t like Anthropic noticeably more than I dislike the other giant cloud platforms and AI model providers. Amodei is probably the AI chief executive whose language and thinking land closest to mine. During April and early May 2026, I worked through a dozen of his interviews and a stack of his essays. I also watched many hours of him on YouTube. The longform is the portrait. This post is the structural argument that fell out of it. If you want the character work, family backstory, and scenes that a paradox piece can’t carry, read the full thing.

Dario Amodei founded Anthropic in 2021 with six other former OpenAI researchers. Their artificial-intelligence (AI) thesis had two halves.

First, powerful AI will arrive whether safety-aligned labs reach the frontier or not. The frontier is the leading edge of general-purpose model capability. Safety-aligned labs therefore have to remain there.

Second came the load-bearing claim: competition on safety would pull rivals upward in a “race to the top.” Both halves remain in Anthropic’s public materials. The first half is doing fine. The second is colliding with the company Anthropic became.

A Series A is an early institutional funding round. Anthropic’s Series A raised $124 million in May 2021. The Lightspeed round added $3.5 billion in early 2025. By September 2025, investors valued Anthropic at $183 billion. The valuation reached $380 billion by February 2026.

Revenue rose from zero to roughly $10 billion annualized within three years, with tenfold year-on-year growth each year. Annualized revenue extends the current pace across a full year. The character work is in Inside the Mind of Dario Amodei. This post makes the strategic argument that emerged from it.

Anthropic’s race to the top, on paper

Anthropic registered as a Public Benefit Corporation. This structure requires directors to consider a stated public benefit alongside shareholder returns. Anthropic describes itself as “an AI safety lab that is also an AI lab.”

A lab that genuinely cares about safety must compete commercially at the frontier. Otherwise, labs that care less about safety set the frontier. Frontier status lets Anthropic publish practices and hire the best alignment researchers. Alignment research studies how to make AI systems follow intended goals. Frontier status also gives Anthropic policy credibility. Rivals see its practices work and copy them. The industry shifts. Race to the top.

In January 2026, Amodei published the roughly 22,000-word The Adolescence of Technology. It defines five risk categories: misalignment, individual misuse, state misuse, economic disruption, and indirect or unknown effects. Its anchor claim on geopolitical risk includes the following fragment.

…autocracy is simply not a form of government that people can accept in the post-powerful AI age.

This is the language of someone who sees the company’s mission as civilizational and says so under his own name. It is also the language of a company positioned as the democratic-world frontier lab. That position shapes both its customers and its adversaries.

The strongest external signal against pure positioning came in November 2023. OpenAI briefly fired and reinstated Sam Altman. Its board then offered Amodei the chief executive role or a merger with Anthropic. He declined both.

I read that refusal as the most expensive credibility signal Amodei could send. He walked away from the top job at the world’s most valuable AI company. He had left it less than three years earlier. The refusal suggested that safety was the actual thesis, not a brand exercise. Roughly fourteen OpenAI researchers had followed him out two years earlier.

There is a softer counterweight worth keeping in mind. Nicolai Tangen asked Amodei in 2024 about scaling timelines. Amodei said:

frankly, although I invented AI scaling, I don’t know that much about that either. I can’t predict it.

What scaling produced

By May 9, 2026, five years after its Series A, Anthropic carried a $380 billion valuation. Annualized revenue stood at roughly $10 billion. Revenue reached $100 million in 2023, $1 billion in 2024, and a $10 billion run-rate by late 2025. That is three consecutive years of 10x growth.

Investor decks project $26 billion for 2026 and $70 billion for 2028. These figures are forecasts, not reported revenue. Even so, the projected path is unusually steep for enterprise software.

The capital behind that revenue is even more telling. Amazon has committed roughly $8 billion in total. Lightspeed wired the first $1 billion of its $3.5 billion early-2025 round on the Monday Nvidia fell 17%. That was the DeepSeek shock. Deliberately or not, the timing signaled conviction when everyone else flinched.

A year later, Singapore’s sovereign wealth fund GIC and Coatue led the $30 billion Series G in February 2026. A Series G is a late-stage funding round. It gave Anthropic a $380 billion post-money valuation. Post-money valuation includes the new investment. This was the second-largest technology funding round on record.

The Fluidstack data-center deal was $50 billion. Project Rainier, the Amazon-anchored compute build, was $11 billion. Compute is the processing capacity used to train and operate models. Anthropic’s compute commitments through 2028 total roughly $78 billion. Headcount reached roughly 2,500 employees by late 2025, up from a few hundred two years earlier.

The customer list reads like that of a sales-led enterprise-software company, because that is what Anthropic became. Pfizer. United Airlines. Novo Nordisk. AIG reports an eightfold to tenfold speed-up in insurance-underwriting workflows during an 18-month pilot.

These are not safety partnerships. They are revenue contracts in regulated industries that came down to a procurement bake-off. Claude Code shipped as a developer-tier product in February 2025. It gave Anthropic a per-seat footprint in enterprises that had bought only through an application programming interface (API).

Anthropic claims internally that it earns 2.1 times as much revenue per compute dollar as its largest rival. This remains a claim rather than a verifiable number. Still, it fits a company that optimizes seriously for unit economics, or revenue and cost per unit.

As of May 9, 2026, Anthropic was a frontier lab competing for every contract that crossed procurement. It was no longer a research lab with a corporate appendage. It was a frontier company with a research function. That function inherited the company’s constraints, not the other way around. None of this is bad on its own. The question is whether the founding thesis still describes what the organism does.

Why both cannot fully be true

To pull rivals upward, Anthropic must remain competitive at the frontier of the thing it calls dangerous. Faster scaling makes the safety thesis depend on rivals copying Anthropic rather than Anthropic slowing down. Rivals’ incentive to copy weakens precisely as Anthropic becomes a credible competitor for revenue and contracts.

One empirical hedge is worth flagging. OpenAI’s Preparedness Framework and DeepMind’s Frontier Safety Framework share structural features with Anthropic’s Responsible Scaling Policy. This convergence may reflect parallel development under similar regulatory pressure. Anthropic may not have caused the diffusion. The race-to-the-top claim is harder to falsify than it looks.

Amodei has publicly acknowledged the tension. He told Fortune that Anthropic struggles to balance its safety mission with commercial pressure. Three forcing functions show the operational shape of that struggle.

1. The Department of Defense

On March 26, 2026, a federal judge granted Anthropic a preliminary injunction in its lawsuit against the Department of Defense (DoD). A preliminary injunction restricts conduct while litigation continues.

The dispute began when Pete Hegseth’s department sought to renegotiate Anthropic’s existing acceptable-use restrictions. These restrictions banned Claude from mass domestic surveillance and fully autonomous weapons. Anthropic refused the change. DoD then designated the company a “supply-chain risk.”

Judge Lin held: “Punishing Anthropic for bringing public scrutiny to the government’s contracting position is classic illegal First Amendment retaliation.”

The court also found violations of due process, which requires fair government procedures. My point is not the merits of either side. It is the structural change. At frontier scale, a safety constraint becomes a federal court fight rather than a research-policy choice.

2. The Pottinger op-ed

In January 2025, Amodei and former deputy national security advisor Matt Pottinger co-authored a Wall Street Journal opinion piece. Trump Can Keep America’s AI Advantage argued for tighter chip-export controls against China.

This is policy advocacy from a national-security actor, not a research lab. A coherent line connects the safety thesis to chip controls. Frontier capability in the wrong hands is a geopolitical risk in Amodei’s taxonomy. Still, that posture differs from “we publish safety research and hope rivals copy us.” Anthropic is now a constituency in great-power competition.

3. The Huang feud

In August 2025, Amodei and Jensen Huang traded public criticism over export controls. Amodei accused Huang of an “outrageous lie.” He became visibly emotional while discussing his father’s preventable death. That thread appears in nearly every long-form interview and forms the emotional spine of my profile.

The Nvidia chief executive is one of two or three people who can directly affect Anthropic’s compute supply. Picking that fight publicly is a political actor’s choice, not a research lab’s. The founding thesis would not have predicted it in 2021.

The point is not that these choices are bad. Each is defensible on its merits. The point is that “race to the top” framed a 2021 company. The 2026 company is a different organism.

It has federal-court fights and named geopolitical adversaries. The Senate voted 99-1 against a proposed ten-year moratorium on state AI regulation. Amodei opposed the proposal in a June 2025 New York Times op-ed. Investor decks imply roughly 60% annual revenue growth from 2026 through 2028.

None of this resembles a safety lab. All of it resembles a frontier company that takes safety seriously.

Three scenarios

A. The thesis holds

Frontier labs converge on Anthropic-style safety practices, and race-to-the-top works as advertised. Anthropic then earns a durable safety-narrative premium. If its safety position remains credible, it can carry extra strategic or financial value.

This case needs a European Union AI Act enforced with teeth. It also needs a US transparency framework that gives Anthropic’s published practices federal legal backing. The Senate’s 99-1 rejection of the proposed ten-year moratorium on state AI regulation offers some support. It suggests that the political ground is not hostile to oversight.

There are reasons to discount this scenario. The Senate vote was defensive, not a positive endorsement of federal standards. As of May 9, 2026, no such US framework existed. Amodei’s June 2025 New York Times op-ed, Don’t Let A.I. Companies off the Hook, did not change that. The required regulatory tailwind has not appeared at the speed the thesis needs.

B. The thesis becomes a constraint, not a durable advantage

I consider this the most likely scenario. Anthropic loses raw-capability ground to less-constrained competitors. These include xAI, a more permissive next-generation OpenAI, and leading Chinese labs. Safety becomes a self-imposed handicap rather than a market-shaping lever.

Anthropic’s seven cofounders made an 80% wealth pledge, disclosed in the January 2026 essay. I treat this personal pledge as a partial offset to wealth-concentration concerns. It is not a formal corporate-governance constraint, and it does not reverse the competitive dynamic.

The DoD fight, Pottinger op-ed, and Huang feud offer early evidence. At frontier scale, a safety stance creates adversaries that Anthropic cannot route around. The pattern accelerated in early 2026. Time reported that Anthropic dropped its flagship safety pledge for a non-binding framework. The company said this framework “can and will change,” the kind of operational adjustment that I treat as consistent with scenario B.

The 2021-2023 safety premium should therefore compress. Investors paid for a safety lab that would pull rivals upward. It is becoming a frontier company that constrains itself relative to less-constrained rivals.

C. The paradox dissolves because the scale itself ends

AI capital expenditure (capex) hits a Jevons-paradox-for-labor wall. Jevons paradox describes cases in which efficiency gains increase total use because rebound demand exceeds the savings. Model commoditization then compresses margins as competing models become more interchangeable. Frontier scale becomes uneconomic. Anthropic looks like a research lab again because every lab does.

The paradox would then describe a phase, not a company. Two considerations weigh against this scenario. Anthropic projects $26 billion of revenue in 2026 and $70 billion in 2028. Those forecasts make retreat harder for Anthropic than for more capital-intensive large cloud providers. Gross margin is the percentage of revenue remaining after direct service costs. Commoditization also reaches Anthropic’s gross-margin profile later than labs that run on rented compute.

What Anthropic’s safety mission means for pricing AI labs

A capital allocator decides where to deploy funds. Three observations follow.

1. Safety narrative is not a moat at this scale; it is a constraint. A moat is a durable competitive advantage. A counterfactual is an alternative world used for comparison. That constraint belongs in the valuation. Investors paid the 2021-2023 premium against a counterfactual with no safety-aligned frontier lab. That world no longer exists. Anthropic is the frontier lab.

Safety now binds the company through legal exposure, slower product cadence in some categories, and fewer customers it can serve. No single constraint breaks the company. Together, they change what an investor buys.

2. Compare how quickly frontier capabilities spread with how quickly rivals adopt safety practices. This comparison shows whether race-to-the-top is happening. If safety practices spread faster, the thesis works. If capability spreads faster, the thesis becomes a story that Anthropic tells about itself.

Evidence through May 9, 2026, leans toward the second case. One European Union enforcement action or one US framework could still change this comparison.

3. Anthropic-the-company and Anthropic-the-thesis are different things to evaluate. They have different ranges of possible outcomes. The company has revenue, named customers, and a personal wealth pledge from its cofounders. I treat that pledge as a partial offset to wealth-concentration risk. Its numerical trajectory is hard to bet against.

The thesis is a 2021 framing under growing structural strain. My longform argues that Dario sees the tension. This piece argues that the tension exists. What anyone does with that observation is their own call.

Based on twenty-five hours of Dario Amodei’s on-record interviews and six long-form essays. The full reported profile, Inside the Mind of Dario Amodei, runs 14,000 words at the author’s newsletter.