Chamath says much of AI's revenue is still experiments

The numbers were on the table. The host had put Anthropic at a $14 billion run rate and OpenAI at a $20 billion annualized pace. Brad Gerstner, an investor in both Anthropic and OpenAI, had just called Anthropic's $6 billion February a splitting the atom moment, bigger than a year of sales at Databricks or Snowflake, and proof that AI was now competing for labor budgets instead of software budgets.

Then Chamath Palihapitiya, who says he pays millions of dollars a year to use the models in his own work, cut in with a colder reading. "There's not a single good example that we can find of sustained positive margin expansion and impact of AI inside of a true corporate enterprise," he argued.

Chamath asked where that $6 billion comes from, and argued that it is not hard to reach that kind of total when tens of thousands of companies pay $200 or more a month so executives can show their boards an AI checkbox.

Gerstner said roughly half of OpenAI's $20 billion is consumer subscriptions, which he compared to Netflix because customers keep paying without demanding service guarantees.

Chamath's test, he said, is whether AI runs the work that actually creates corporate revenue and profit. Databricks and Snowflake do, he argued, because they sit inside critical production workflows. AI does not yet, he argued, because no hospital will let a model risk a misdiagnosis that brings fines or jail time, and no bank will let a model misallocate a portfolio and then blame the error in a lawsuit.

His prime exhibit was Amazon. After three or four severe outages caused by agent-written code, he said, Amazon required human review and approval for AI code inside its cloud unit. He argued that the unit's extreme reliability comes from humans and deterministic code, not from models that hallucinate.

Gerstner partly agreed, saying he calls it experimental run rate revenue, a monthly pace multiplied out to a year, rather than annual recurring revenue, revenue that repeats every year. But he countered that Palantir, Nvidia, the U.S. military, and other large enterprises would say they are already in full production, and that wartime use makes AI existential rather than experimental.

Chamath said his own bills suggest how early it is. His token spending roughly triples every three months, he said, without tripling his revenue, and he suspects typical big-company teams trail even further behind his own leading-edge group. "There's a lot of test budgets," he said. "It will slowly and methodically emerge into production."