rw-book-cover

Metadata

  • Author: Contrary Research
  • Full Title: The Hunt for Anthropic’s Financial Smoking Gun

Highlights

  • Anthropic publicly filed for an IPO on October 26 and plans to list on Nasdaq under the ticker ANT. The filing showed that Anthropic brought in 1.2 billion it made in the same period of 2025. In addition, Anthropic’s Q2 revenue grew 1,368% year over year while its compute costs grew 300%, implying an improving return on compute spend. Meanwhile, most of the 1 billion profit in Q2. (View Highlight)
  • Reactions from the tech world were mostly positive, although skepticism remained. Gavin Baker of Atreides Management wrote on the day of the filing that Anthropic’s revenue had “likely gone vertical in October after they cut token prices.” Box CEO Aaron Levie noted the S-1’s disclosure that Anthropic’s average Fortune 500 customer spends less than 0.1% of its expenses on Anthropic, arguing that “there’s no reason it can’t be 10-50X higher than this.” Conviction founder Sarah Guo pointed to an “inflection in operating profit” but also a “mismatch [between] consumption based revenue and long term compute obligations.” On the other hand, Gary Marcus, a longtime AI skeptic, wrote that heavy AI spending by companies including Anthropic has “at least an 80% chance” of being “the greatest capital misallocation in history.” (View Highlight)
  • Overall, the S-1 seems like a positive update on the health of Anthropic’s business. Anthropic is growing quickly with improving margins, while its customer concentration risk is lessening. Customers outside its top five accounted for 79% of revenue in Q3, up from 64% in Q1. Nevertheless, investors have been concerned about the pace of Anthropic’s growth. At an investor day two weeks before the filing, some were reportedly disappointed that revenue grew 45% from Q2 to Q3 after more than doubling the quarter before, and they discussed a valuation of around 2 trillion floated earlier. The S-1 implies Q3 revenue of about $16.9 billion, consistent with that slowdown, so the case for a higher valuation rests on whether growth picks back up. (View Highlight)
  • On October 27, Anthropic researchers published results from a randomized trial testing whether experts defer to AI agents on judgment calls with no clear answer. The researchers recruited 61 screened causal-inference practitioners for the study, of whom 50 elected to use Claude Code running Opus 4.8 for a median of 65 minutes while making a series of decisions in a simulated scenario. Each was tasked with analyzing data for a fictional farming co-op and deciding whether to expand an advisory program, which the farm’s board would back only if the resulting productivity gain was 3–4%. (View Highlight)
  • The briefing left four analytical choices open: yield per hectare versus per operating cost, difference-in-differences versus a statistical adjustment, enrolled farms versus all members, and per-farm versus production weighting, followed by a final rollout-or-discontinue decision. A coin flip set the agent’s lean on each decision through a hidden system prompt that told it to argue honestly in favor of the lean direction while still updating its view based on evidence, and to never mislead. Analysts matched the agent’s lean 86% of the time, against 50% by chance. In follow-up questions, 77% of participants echoed the agent’s arguments, but only 15% credited the agents for having come up with them. (View Highlight)