Anthropic’s Revenue Growth Slowed as IPO Expectations Rose

Anthropic reported $16.8 billion in revenue for the third quarter, up about 45% from the second quarter. That single quarter amounted to more than 3.6 times the company’s revenue for all of 2025. But the growth rate was sharply lower than the previous quarter’s: revenue had risen more than 140% from the first quarter to the second, during the surge in demand for Claude Code.

For Alex Kantrowitz, the distinction is central. A company can post extraordinary growth and still disappoint investors if they have come to expect the earlier pace to continue. The question is whether the Claude Code boom marked a durable point on an accelerating curve or a burst of demand that has since cooled.

Ranjan Roy called the quarter-on-quarter growth “incredible,” but said it was hard to judge without the company’s full financial disclosures. If the reported pace continues, he estimated, Anthropic could reach roughly $75 billion in annual revenue; reaching $80 billion or $100 billion would require growth to accelerate. Those estimates matter because investors have been discussing valuations as high as $2 trillion. At that level, Roy said, even remarkable numbers may not be enough.

The comparison is complicated by a revenue source Anthropic reportedly told investors accounted for 5% to 7% of revenue: distillation. In this context, distillation means other companies querying Anthropic’s models at scale and using the answers to train their own models. Roy said the practice may have contributed disproportionately to the first-to-second-quarter surge, when demand for Claude Code was also exceptionally strong. Anthropic is trying to rein it in, Kantrowitz noted, which could remove some revenue. Roy said that would be a meaningful headwind, though he did not think it automatically changed the company’s broader prospects.

Both speakers treated other indicators as suggestive, not conclusive. Kantrowitz cited Apptopia estimates that Claude’s monthly U.S. downloads climbed from nearly 670,000 in January to 7.42 million in March, then fell to 3.26 million in September. App downloads are not a measure of Anthropic’s whole business: Roy noted that Claude use is concentrated among expensive enterprise and professional customers, rather than being primarily an everyday consumer app. Still, he agreed the download pattern could be one signal that the spring’s excitement had moderated.

The same figures produce opposing readings. Skeptics see decelerating growth and a harder path to the valuation that has been discussed. Optimists see 45% growth on a much larger base. “Everything we see” has become a Rorschach test, Roy said: people read the same numbers according to their prior view of the company.

His frustration was less with the disagreement than with the way investors were being asked to interpret incomplete information. Anthropic had discussed financial figures with potential investors while in confidential IPO filing, and reports and clarifications about the figures added to the confusion. A Reuters report, as Roy described it, attributed a $110 billion annualized revenue figure to investors who attended an event; the reporter later clarified that it was an estimate by sell-side analysts based on guidance, not a figure disclosed directly by Anthropic. For Roy, that episode showed how easily an estimate could be mistaken for company guidance and then feed a market narrative.

He argued that Anthropic should release its S-1—the filing that lays out a company’s financial and other information ahead of going public—rather than leave investors to infer the business from leaks and partial reports. Roy cited investor and former banker Imran Khan’s criticism of holding a large investor meeting while the company was in confidential filing, particularly after its S-1 had reportedly leaked. The process, Roy said, was unusual and made it harder to separate company figures from interpretations of them.

Even a full filing may not settle the debate. The speakers said the economics will depend on details such as how Anthropic classifies training costs and what it includes in adjusted profitability. Roy also cautioned that the business itself is still taking shape: it is not yet clear whether expensive frontier models, cheaper models, or some other set of products will account for most revenue, or how the company will compete with open-weight models.

Kantrowitz framed that uncertainty as a feature of the AI trade. The products have shifted quickly, from chatbots to coding tools, workplace assistants and agents; the eventual form of the technology, and of the businesses built around it, remains unsettled. Roy agreed that uncertainty would persist, while hoping that fuller disclosure could at least make the underlying economics easier to see.

The speakers differed on whether the current anxiety around Anthropic was an overreaction. Roy thought there was enough interest in a successful offering that it would take a lot to derail the IPO. Kantrowitz said the concern was justified. He pointed to possible commoditization of frontier models, competition from OpenAI in enterprise, reports of large customers pulling back on token spending, and the cooling of enthusiasm around Claude Code and Claude Cowork. He still considered Anthropic a strong company, but rejected the idea that its position was inevitable in an industry where fortunes can change quickly. He predicted the company would go public by year’s end; Roy was less certain about the timing.

The uncertainty around AI equities was also weighing on companies that build and lease computing infrastructure. Alex Kantrowitz pointed to a CNBC report on falling neocloud stocks, rising trading in bearish options, and CoreWeave bonds reaching yields around 15%—with reports of yields briefly reaching 16%. CoreWeave shares fell 15% in the period described, while Nebius and Iren also fell. Options volume surged, and traders bought nearly 200,000 CoreWeave puts against 106,000 calls.

Ranjan Roy said the outlook for these infrastructure providers looked more fragile than Anthropic’s. An AI company reporting billions in quarterly revenue offers evidence of substantial customer demand, he said. By contrast, some neocloud valuations rest on plans to build infrastructure with large amounts of borrowed money. He cited the proposed Australian IPO of Firmus, a company reportedly valued at $44 billion and backed by Nvidia and others. Roy said Firmus began as a Bitcoin-mining company in 2019 and described concerns about its investors’ favorable terms, which could allow them to block an offering or withdraw if the valuation fell short of a threshold. He also questioned the company’s leadership credentials and its progress on infrastructure. The details, in his view, illustrated why investors might scrutinize neoclouds more closely than the headline demand for AI computing alone would suggest.

Borrowing costs add a more direct pressure. Roy’s point was that the expanding demand for capital to finance data-center construction can itself raise the cost of that capital. That cost is not abstract for companies that depend on debt to build their businesses.

Kantrowitz asked whether rising interest rates and more expensive borrowing could squeeze the economics of supplying computing capacity. Roy agreed that this mattered, while distinguishing the yield on a bond in the market from the interest rate the borrower originally agreed to pay. CoreWeave’s bonds had been issued at 9.625%. A reported 15% yield did not mean CoreWeave was now paying 15 cents on every dollar of its existing debt. Rather, it meant investors were valuing the bonds at a lower price in the secondary market, demanding a higher yield to hold them. For existing bondholders, that means selling at a loss; for the market, it signals greater concern about the risk of default.

Roy called the bond market’s response one of the few signs of financial discipline amid the AI investment boom. Capital is finite, he said, and investors were beginning to price that constraint. The contrast with the broader enthusiasm around AI was important: a company may have a compelling growth story, but its ability to finance construction also depends on the cost and availability of capital.

A separate report concerned an OpenAI employee accused of hiding a recording device in an executive conference room. According to the Wall Street Journal account read on air, Abdulrahman Alfozan had worked on OpenAI’s product team for more than a year. Prosecutors alleged that he spent weeks examining a room reserved for executives and their assistants, then placed a small circuit board and microphone in a USB port behind a wall-mounted television. The device was wrapped in toilet paper.

Alex Kantrowitz distinguished the reported allegations from the charge. Alfozan was arrested and charged with using or attempting to use a device to intercept oral communications. The reporting described the room as a place for leadership meetings, discussions with government officials and business partners, and conversations about research and products. Ranjan Roy said the speakers did not know whether the alleged conduct was connected to a foreign government or intelligence service.

The reported details made the case striking partly because the equipment was so accessible. Roy said reporting had identified the device as a small, inexpensive development board with an attached microphone. Prosecutors also alleged that Alfozan had used an AI tool to ask about leadership meeting rooms, while a chatbot account associated with his spouse had asked questions about recording, remote access and data encryption. OpenAI said it used its models to understand the device and trace it to an employee. Roy wondered whether that meant something as simple as asking a chatbot to identify an unfamiliar piece of electronics.

The same accessibility cuts both ways. A device that might once have required specialist knowledge or spy-film gadgetry can now be assembled with inexpensive components and guidance from a chatbot. Roy said he had looked at the components himself and could imagine trying a recording project as a weekend hack. The ease of making such a device makes it harder for companies to assume that surveillance attempts require elaborate tools.

The Journal’s account said OpenAI reviewed security-camera footage and saw Alfozan near the room several times before the device was found, despite having no apparent reason to access it. Prosecutors also said that a bag left unattended elsewhere in the office contained Alfozan’s wallet and a similar recording device. Those were among the details reported about the investigation.

Kantrowitz noted that a U.S. attorney called the case a matter of national security, while Alfozan’s lawyer said his client was charged with violating a federal recording statute, “not stealing trade secrets or working as a foreign agent,” and urged people to withhold judgment until the facts were known. The distinction matters: the charge described by the lawyer was about recording, while the broader questions about motive and potential intelligence links remained open.

Kantrowitz recalled reporting years earlier on two Saudi nationals at Twitter who were accused of helping someone associated with the Saudi royal family access information about Saudi dissidents. He also recalled a source at a technology company saying that people connected to nation-state intelligence agencies were present inside Silicon Valley companies. The speakers wondered how much sensitive information might be sought inside technology companies, and by whom. The alleged device appeared intended to capture talk about strategy, research and partnerships, rather than to extract model weights directly.

The recording-device story sits alongside a more familiar kind of corporate-security concern: attempts to get information through relationships. Alex Kantrowitz cited a Wall Street Journal report that OpenAI had warned employees about outsiders who might approach them in public, listen in cafés, offer money, or try to draw them into romantic or sexual relationships. The report said employees at Anthropic and Google had received similar warnings.

The Journal described a scenario in which a stranger befriended an OpenAI employee at a playground while both were with their children, then asked questions about the employee’s work. It also described “honeypots,” agents who use the offer of sex or romance to get close to a target. Ranjan Roy said the basic idea was familiar from spy stories, but the business context makes even small disclosures potentially valuable. In a fast-moving field, information about a company’s direction or concerns might be useful sooner than a conventional product leak would be.

The speakers joked about how an employee might be approached, and about the implausibility of an overly direct interrogation on a date. But Roy’s point was that a skilled approach would not begin with a demand for secrets; it could start as ordinary conversation and build gradually. That is why, in their view, the warnings are less about one dramatic incident than about recognizing a range of ways information can be elicited.

Kantrowitz opened the food-image debate with a New York Times report about restaurants and fast-food businesses using generated images that look uncanny or simply wrong: pasta with unexplained black dots, sandwiches whose shapes do not resolve into recognizable ingredients, and a gray burrito marked by repeated holes. He argued that this kind of “food slop” was unpleasant enough to merit regulation.

Ranjan Roy disagreed. He recalled seeing a halal cart near his home in New York with obviously AI-generated images and exuberant names for its burgers and chicken and lamb dishes. To him, the appeal was not that the pictures were flawless. It was that a small business owner could make promotional material without finding, briefing and paying a graphic designer—work that might previously have been out of reach. A rough image, he said, could still represent a new opportunity to make something.

That argument shifted Alex Kantrowitz’s view. His remaining complaint was that many generated food signs look alike. He suggested image tools could offer a randomizer or otherwise prompt users to make the result more distinctive. Roy agreed that this would be a straightforward way for AI companies to help small businesses without stripping the process of its creative possibility. Both treated better models and better prompts as part of the answer, rather than assuming that every awkward image was a reason to ban the practice.

Their example was a photo, found on X, of a spa sign that appeared to show a sign outside a spa—with another sign depicted inside it. The likely instruction had asked for a sign advertising services such as foot, neck, shoulder and back massage. The generated image took the instruction literally enough to place a sign in the scene; the result was then printed and put inside an actual sidewalk sign.

Kantrowitz saw a model misunderstanding its brief. Roy saw a piece of accidental visual comedy that did exactly what sidewalk signage is supposed to do: attract attention. A polished campaign might have reached the same recursive idea through a creative agency, he joked, but the low-cost version gave someone without that route a chance to produce something memorable. His defense was not that every generated menu or poster looks good. It was that imperfect output can still be useful, expressive, and available to people who previously could not easily produce it.