The same May weekend Anthropic was about to close $50 billion at a $900 billion valuation on a run rate built almost entirely from coding agents displacing entry-level engineers, Greg Abel — Warren Buffett’s chosen successor and Berkshire Hathaway’s first new CEO in 60 years — held his first annual meeting in Omaha and laid out a doctrine that takes the other side of that trade.
Abel called it “narrow AI.” He said it twice on stage Saturday: “Berkshire will never pursue AI for AI’s sake.” For a trillion-dollar conglomerate that owns BNSF Railway, GEICO, the second-largest power utility in the United States, and roughly 10 million shares of Amazon, that sentence is a position. A $900 billion AI lab, a $135 billion Meta capex line, and a $25 billion Tesla AI/robotics budget make “AI for AI’s sake” the consensus trade. Abel just printed the dissenting vote.
Four exhibits
The Buffett deepfake. Abel opened the Q&A by playing what looked like a video message from Buffett — sitting in a chair, talking to the camera, asking shareholders to hold Berkshire shares for the long term — and then revealed the clip was AI-generated. The choice of demo is itself an argument. At a moment when the rest of corporate America is showing investors revenue charts that go up because AI replaces workers, Abel showed his investors a chart that goes up when AI deceives them. The most measurable near-term ROI of AI for Berkshire’s businesses is defense against fraud, not displacement of labor. Buffett, sitting in the front row, said only one thing about AI on the record: “It’s scary.”
Ajit Jain’s “years” timeline. Ajit Jain, Berkshire’s insurance vice chairman and the most respected underwriter in American finance, told the room that AI doing complex pricing or claims decisions is “years away.” That is a direct counter to Customers Bank’s Sam Sidhu cloning his cognition into a custom OpenAI model, to the 27.2% of US workers who told the Oracle/TIME survey their employer had explicitly trained AI to replace them, and to the entire Anthropic enterprise pitch. Insurance is the test case where AI gets tens of billions of dollars in unit-economics leverage if the labor-substitution thesis is true. Jain just said the leverage isn’t there on Berkshire’s underwriting book — and if it isn’t there for the underwriter most carefully optimized for it, the case is shakier than it sounds.
BNSF as the use case. Abel pointed to BNSF Railway, where Berkshire is building targeted AI tooling internally for operations and safety. Two things to notice. First, “internal” — Berkshire isn’t writing a check to OpenAI or Anthropic for seat licenses. Second, “operations” — the tools sharpen logistics, not replace conductors. The labor-substitution coefficient on Berkshire’s largest US workforce is being held deliberately low. The contrast with Tesla’s Fremont line being torn down to build 1 million Optimus units a year is this week’s contrast with the same labor-substitution thesis priced very differently.
The energy unit is the only unambiguous AI win. Abel’s Berkshire Hathaway Energy unit projected “a 50% expansion of data center load over the next five years” in its Iowa service territory, where data centers already make up roughly 8% of peak load. This is the single most enthusiastic sentence Abel said about AI all weekend. Note what it is. BHE makes money when the Mag-4 build data centers, not when the data centers’ models replace BHE’s customers. Berkshire monetizes the AI capex line through its energy infrastructure, then sits out the labor-substitution leg of the same trade — long the picks and shovels, neutral on whether the gold-rush metaphor pans out.
$397.4 billion in cash is its own commentary
Berkshire’s cash and Treasury pile hit a record $397.4 billion at quarter-end — up from $373 billion at year-end 2025 — because Berkshire was a net seller of stocks for the largest two-year stretch since the 2008 crisis. Abel inherited the cash mountain, did not deploy it, and at his first annual meeting did not signal he was about to. A $397.4 billion cash pile in May 2026, with the S&P 500 at one of its highest forward multiples in history and the Mag-4 collectively spending $650 billion on AI capex this year, is the loudest quiet position in American finance. The mountain is the position.
What this means for the labor-market story
LostJobs’s working thesis through this AI capex cycle is that labor compensation is being structurally compressed and Anthropic’s revenue curve is the measurement. The Berkshire counter-thesis, distilled from Saturday’s stage, is sharper than anything we’ve seen from a major capital allocator to date: AI’s near-term ROI is ambiguous, the customer-side bet (energy infrastructure for data center demand) is more defensible than the substitute-side bet (replacing white-collar headcount), and at current multiples the labor-substitution leg is overpriced.
Anthropic at $900 billion and Berkshire at $397.4 billion in cash are the same weekend’s two simultaneously-printed prices. One is the price of believing the labor-substitution bet generalizes. The other is the price of not believing it. Both are unusually large numbers; only one of them can be right at this magnitude.
The dry coda
Greg Abel’s first annual meeting contained no acquisition announcement, no buyback boost, no new strategic disclosure. It contained one Buffett deepfake, three “narrow AI” sentences, one Ajit Jain timeline correction, one BHE data-center projection, and a $397.4 billion cash pile.
The same May calendar that holds Meta’s 8,000-person layoff effective date, Microsoft’s 8,750-employee buyout cohort, and the Tesla Fremont retool now also holds the first concrete vote against the labor-substitution multiple from a capital allocator with $397 billion to back it up. The Mag-4 are going to keep printing the layoff calendar. Anthropic is going to keep printing the run-rate curve. But the most patient capital in American finance just announced, very quietly, that it disagrees with the multiple.
That dissent is itself news.