Anthropic filed its confidential S-1 with the SEC on Monday. By Tuesday, three separate data points landed that complicate the revenue trajectory that filing has to survive.

Axios reported on June 2 that corporate buyers are hitting what it describes as an “AI sticker shock” phase precisely as Anthropic moves toward a public listing. The timing is not incidental. It is the central credibility problem the IPO will have to answer.

The sharpest data point comes from Bain & Company, whose survey of nearly 1,000 companies found that 40% reported AI cost savings below 10% after making substantial investments. That is not a finding about AI failing to work technically. It is a finding about AI failing to produce the financial return that justifies the spend. For a company whose primary customer base is enterprises paying premium rates for Claude access, a survey showing that four in ten of those enterprises feel underdelivered is a direct challenge to the pricing floor.

The anecdotes inside the Axios piece are equally pointed. An early Anthropic investor described companies “waking up to how much they’re spending on Claude” as a risk worth tracking. An AI consultant told Axios that a CFO client had “accidentally spent half a billion dollars on Claude in a single month.” That figure is striking even discounting for the looseness of a secondhand anecdote. Accidental spend at that scale signals a procurement failure, and procurement failures produce audit responses and contract renegotiations, not repeat purchases.

The competitive context sharpens the exposure. Anthropic surpassed OpenAI in business customers for the first time in April, according to Axios. Business clients produce higher revenue per seat than retail users, so this metric is load-bearing for the valuation case. But business customers are also the ones with CFOs who read Bain surveys. The same buyer profile that drove Anthropic’s revenue advantage is the one most likely to respond to cost scrutiny by switching toward cheaper closed models or open-weight alternatives.

This is where the story connects directly to the open-versus-closed debate Nathan Lambert documented at Interconnects this week. Lambert’s argument is about the supply side: open models are catching up to frontier capability faster than the labs assumed. The Bain survey and the Anthropic CFO anecdotes are the demand-side confirmation of his thesis. Enterprises do not need open models to be equal to Claude for the closed-lab pricing premium to compress. They only need open models to be good enough, at a moment when the bill for “better than good enough” has become visible on the income statement.

Sam Altman acknowledged the tension directly. In a CNBC interview this week, he called corporate concerns about AI cost “the most fair criticism of AI so far.” That is a notable concession from OpenAI’s CEO at the moment a direct competitor is pricing an IPO on enterprise AI growth. It also signals that the concern is not limited to Anthropic’s customer base.

None of this means the IPO fails or that Anthropic’s fundamentals are misrepresented. The company’s Q2 revenue may be strong. But an S-1 is a document that asks investors to underwrite future growth, and the argument for future growth depends on enterprise buyers deepening their AI commitments rather than pulling them back. Three signals in 48 hours suggesting those buyers are under budget pressure is not a refutation of the thesis. It is a stress test that the roadshow will need to address explicitly.

Enterprise AI buyers who have not yet run a line-item audit on their Claude spend should do so before Q3 budget cycles close. The companies in the Bain survey that found returns below 10% did not all have weak use cases. Some of them had weak unit economics on their API consumption.

Reported by Axios (axios.com), published June 2, 2026. Cost-savings data from a Bain & Company survey of nearly 1,000 companies. Sam Altman’s quote sourced from a CNBC interview the same week.