SAN FRANCISCO — Anthropic told investors Wednesday that third-quarter revenue hit $16.8 billion, a 45% jump that puts the company on a run rate few enterprise software firms have matched at any stage of their history. The number, disclosed at an investor meeting ahead of a planned initial public offering as soon as next month, is the clearest evidence yet that enterprise AI spending has moved past pilot budgets into recurring, contracted revenue.
For context: Salesforce took 11 years to cross $1 billion in annual revenue. Anthropic is reporting quarterly numbers nearly 17 times that figure, five years after its founding. The investor meeting is widely read as the final rehearsal before a listing that bankers have been pricing informally in the hundreds of billions.
The backdrop is less tidy. OpenAI, Google and Anthropic this week announced a joint effort to combat what the industry euphemistically calls "model theft" — the extraction of proprietary weights and training techniques through API probing and fine-tuning attacks. The timing is not coincidental. As frontier model revenue scales into the tens of billions, so does the incentive to steal it. Model weights that cost hundreds of millions of dollars to train are, structurally, no harder to copy than a PDF once they leave a secured environment.
The coalition announcement arrives alongside a less comfortable disclosure: an OpenAI engineer charged with planting a hidden device inside company infrastructure had also advised Saudi government agencies and sat as a listed partner at a venture firm connected to a Saudi prince. OpenAI has not detailed what the device did or how long it operated undetected.
None of this changes Anthropic's top line. But it does sharpen the question investors will ask before any IPO prospectus is finalized: revenue growth of 45% is real money, but the defensibility of the underlying model — against insiders, against extraction, against geopolitical entanglement — remains a number nobody has figured out how to put in a quarterly filing.
