Do AI labs spend far more than they earn?

Epoch AI reports that at each of the three AI companies it examined, "spending at each of these labs is 2-3x higher than revenue." The three diverge sharply on whether or when that gap closes: "Minimax and Z.ai both indicate that they will remain unprofitable for the foreseeable future, [while] Anthropic expects to become profitable by 2028."

The figures come from the same breakdown described in Do compute costs really dominate AI company budgets?. Anthropic's numbers come from The Information's reporting of the company's "optimistic" projections as of December 2025 — $5.2B in EBITDA losses against $4.5B of revenue, implying total expenses of $9.7B — which Epoch AI expects to be close to final since they were reported near year-end. Minimax's and Z.ai's numbers come from the detailed financial documents each released for its January 2026 IPO. Epoch AI notes that "second-hand reports of financial statements carry a degree of uncertainty over the exact financial accounting being used," and says its Anthropic estimates in particular "are somewhat speculative."

The gap Epoch AI describes — AI companies collectively spending multiples of what they earn — sits upstream of the adoption-side numbers in Why do individual AI gains not translate to business results?: even if more enterprises eventually report EBIT gains from using AI, that says nothing about whether the labs supplying the compute reach profitability themselves. It also echoes Do AI customer service agents actually reduce total support costs?, where cost reductions from deploying AI didn't translate into falling costs at the user company either — the gap between spending and return shows up on both sides of the transaction, not only at the labs. And it complicates the picture in Are enterprises actually spending more on AI that works? of steadily growing enterprise AI budgets: that growth is the revenue line the three companies in this note are still running multiples behind.

Epoch AI's profitability figures rest on the same thin, partly secondhand disclosure base as its cost breakdown: a single press-reported projection for Anthropic and two IPO filings for Minimax and Z.ai. "Expects to become profitable by 2028" is Anthropic's own forward projection as relayed by The Information, not an audited outcome, and Minimax's and Z.ai's "foreseeable future" unprofitability is likewise self-reported rather than independently verified. Whether the 2-3x revenue gap narrows depends on inference prices falling faster than usage grows and revenue scaling faster than compute spend — a dynamic this excerpt documents the cost side of but does not itself project forward.