Anthropic CEO: “We might be the only private company in the world at some point”
On the August 14 All-In, investor Gavin Baker shared something he says multiple people have told him: Dario Amodei has mused, privately, that Anthropic "might be the only private company in the world at some point. Think about that." Baker summed up the worldview in one line: "there's Anthropic, and then there are governments, and that's it." David Sacks said what most of the internet said. He'd "discourage Dario from saying that ever again to anyone."
Everyone argued about whether the line was arrogant. That's the wrong argument. The useful question is why courts and regulators spent this entire year answering a different one: not how big AI companies will get, but who pays when AI is wrong. Follow that thread and you'll find the real map of where the value goes. It doesn't end at the lab.
Two facts about 2026 look contradictory until you put them together. Anthropic booked roughly $11.5 billion in second-quarter revenue, turned operating-profit positive, and carries Reuters-reported internal projections of $190 to $200 billion by 2028. On the same podcast episode, Jason Calacanis said the founders he backs are moving workloads to open-source models at a tenth of the cost. Frontier intelligence is consolidating and commodity intelligence is flooding in underneath it, at the same time. For anyone building a product, both trends say one thing: intelligence itself is getting cheap and abundant. And cheap, abundant inputs never hold a business up. The value moves to whatever stays scarce.
What stays scarce is accountability. Every consequential workflow reaches a point where someone must answer for the outcome. Someone licensed, insured, and findable when it goes wrong. I call that point the liability seam, and it has a property that should reorganize your roadmap: intelligence doesn't compress it. A model ten times smarter writes a better credit memo. It does not become one ounce more answerable for the loan.
This isn't theory. Look at what actually happened this year.
In May, a German appeals court ruled that companies are on the hook for what their chatbots tell customers, and a Munich court held Google directly liable for false statements in its AI Overviews. Notice what the courts did not ask: which model, whose weights, how good the guardrails were. They asked who deployed it. FINRA's 2026 oversight report told broker-dealers to build procedures for AI agents that act beyond their intended scope. The regulator is not waiting to see how the technology shakes out.
American lending law ran the cleanest version of the experiment. In 2022 the CFPB issued Circular 2022-03, telling lenders that using a complex algorithm doesn't excuse them from giving every rejected borrower the specific reasons for the denial. In 2025 a new administration withdrew that circular in a bulk sweep of sixty-seven guidance documents. What changed for an AI lender? Nothing. The duty was never in the guidance. It's in the Equal Credit Opportunity Act, where it has sat since the 1970s. Political weather changed twice and the requirement didn't move. That's how you recognize a seam: it survives the regulatory mood.
Even Dario's own writing agrees. His essay "Machines of Loving Grace" lists the limits on what more intelligence can do, including, in his words, that "many things cannot be done without breaking laws, harming humans, or messing up society." The considered essay and the podcast anecdote disagree with each other. Bet on the essay.
Own the seam, rent the intelligence
Make it concrete. A model can draft a credit memo, but it can't be the lender of record. It can abstract a lease in forty seconds, but it can't hold the E&O policy when the closing busts. It can price a listing at midnight, but it can't be broker of record, and it can't sit at a kitchen table and be the person a family trusts with the largest transaction of their lives. For one company to absorb all of that, it would need licenses in every jurisdiction and a balance sheet that pays claims for everything that goes wrong everywhere. At that point it isn't a company anymore. It's the economy, and economies get regulated back into pieces. Even the maximalist vision admits this: "Anthropic, and then governments." Governments are the institutions that guarantee seams keep existing.
So here's the playbook for this quarter:
- Map your seams. List every action your product takes. Mark the ones where someone must sign, hold a license, or carry insurance. That marked list is your durable business. If nothing gets marked, you've built a feature, and features are what the labs absorb first.
- Price on risk, not tokens. Charge for the outcome the accountable party needs, at a price tied to the risk you're absorbing. Capability pricing deflates every quarter. Accountability pricing doesn't.
- Keep switching cheap. Run your product against a second model every quarter. The 90 percent open-source discount is real leverage, but only if you can actually reach it.
- Believe the labs about the middle. The copilots, the harnesses, the workflow glue: assume the platform absorbs them. Move your product's weight onto whatever needs a license, a balance sheet, or a customer's trust.
The courts spent 2026 saying you own what your AI does. Regulators spent it reminding everyone that statutes outlive guidance. Put those together and the "only company on earth" line answers itself. The last company on earth would still need someone to sign. Be the one who signs.
If you're working out which of your product's steps are commodity and which are seams, I have that conversation weekly and never tire of it. Write me at hello@talktoone.com.