The day markets repriced the AI funding story: Firmus drops its IPO, OpenAI counters with "$70 billion," and fired researchers push back from inside

Markets have started to doubt the story told by those funding AI. Firmus, an Australian AI data center company, reportedly withdrew its IPO on October 9. It was expected to be the largest listing since Telstra in 1997. The day before, the FT reported OpenAI's annualized revenue at about $50 billion. OpenAI moved to calm investors with a target of "more than $70 billion by the end of 2026," but it did not deny the $50 billion figure itself. Around the same time, three safety researchers fired by OpenAI pushed back in an open letter, calling for chain-of-thought monitoring to remain possible. OpenAI is in the middle of pausing its top model, and it now faces questions from both outside and inside.

One dynamic links all three events: whether outsiders can verify what they are told. Is revenue counted gross or net? Can Lean's verification itself be trusted? Can we monitor what a model is thinking? These are three faces of the same question. No frontier models were announced, so it was a quiet day on the lab side.

Markets reprice the funding story

Firmus is backed by NVIDIA, Coatue and Blackstone. ABC and Reuters reported that it planned to raise A$7 billion at A$11 per share. The underwriters tried cutting the price to around A$8.25, but demand did not materialize. The company is expected to carry about A$43 billion in debt by 2028, and only two of its seven planned sites are reportedly running. Public markets gave a clear rejection to an AI infrastructure company built on vendor financing. For details, see Firmus reportedly withdraws its ASX listing.

Bloomberg reported OpenAI's response, and Semafor and Investing.com followed. According to investor materials, annualized revenue in the July–September quarter grew 77% overall and 107% in the enterprise segment. But what OpenAI offered was only a year-end target. CNBC confirmed that the previously reported $68 billion counted revenue through partners on a gross basis and was not OpenAI's own share (net). Bloomberg also explained that Anthropic and OpenAI count revenue differently, which is confusing investors. Before the dispute is about how big the numbers are, it is about how they are defined.

Prices swung sharply. At the October 8 close (secondary sources, correcting the intraday figures in our previous report), the Nasdaq Composite fell 1.25%, the SOX 3.4%, Oracle 5.48% and CoreWeave 7.8%. The Dow held up, gaining 0.10%, because money rotated into energy (+2.34%) and consumer staples. On the 9th, markets recovered after the target was reported. Intraday, the S&P 500 rose 0.66%, and Oracle was up about 5% at one point. When one company target can send indexes down and back up again, the story is clearly fragile.

The same pattern shows up nearby. According to the FT (via Bloomberg), SoftBank is seeking to raise up to $100 billion from Gulf investors. Its $64.6 billion investment in OpenAI is being financed with bridge loans and junk bonds. Its shares fell 3.9% on the 9th to 5,803 yen, pulling the Nikkei average down by about 243 yen. In China, 21st Century Business Herald, citing Bloomberg, reported that Moonshot (Kimi) has completed its final pre-IPO private round at a valuation of about $50 billion. It plans to raise up to $5 billion in Hong Kong in January–March 2027. Meanwhile, already-listed Zhipu fell 7.61% on the 8th (about 75% below its June high), and MiniMax dropped 13.74%.

The two camps feel very differently about this. On the skeptical side, Ed Zitron wrote in a paywalled post that if OpenAI and Anthropic borrow, "there is no doubt they would be rated CCC junk." Sun-ha Hong of Tech Policy Press argued that forecasts "borrow credibility from the future to buy legitimacy in the present." We could not find any named commentator making the bullish case in response. The rebuttal came from the company itself, and Cowen and Tabarrok have not addressed the issue.

Questions from inside OpenAI

In an open letter, the fired researchers Mikita Balesni, Tomek Korbak and Jasmine Wang made three demands: place third-party safety auditors inside the labs, keep chain-of-thought monitoring possible, and continue engaging with outside safety organizations. Korbak wrote that "talking to METR was my job," and all three said "we were not the source of the leak." OpenAI says the dismissals were for a "serious breach of trust" and had nothing to do with safety concerns. It also said it "strongly agrees" with the letter's recommendations. Gizmodo reports that the concern behind the letter is that GPT-6 Astra's reasoning, which uses "recursive depth," could evade monitoring. For details, see Three fired OpenAI safety researchers push back in an open letter.

Reaction has been limited so far. Transformer covered the case alongside Bengio's call to "leave frontier companies if you prioritize safety," but we have found no well-known commentator who has written about this case on its own. METR, which was named in the dispute, has published nothing since October 6. Bengio's call also drew a response from the other direction. In the most-read LessWrong post of the window, Haoxing Du pointed out that talented people seeking AI safety jobs far outnumber the jobs available, and that repeating claims of a talent shortage is "actually harmful." People are told to leave the labs, but there are not enough jobs for them to go to. The two arguments fit together.

Who maintains the basis for verification

Anthropic announced "Anthropic Cyber Mission" (October 8, primary source). Eleven companies are taking part in defending critical infrastructure such as power and water. Open source projects can opt in to free scanning with models including Claude Mythos, and reports go to maintainers without human review. In a six-month pilot, 85 of 97 critical- and high-severity findings that were tested met the bar. Sending reports to maintainers without human checks bears directly on the existing debate about overloaded maintainers. For details, see Anthropic announces "Cyber Mission".

OpenAI's math results drew both praise and skepticism on the same day. In a guest post on Terence Tao's blog (primary source), Thomas Hales reviewed the soundness bugs found in Lean's kernel in July and August. He proposed cross-checking proofs across about 25 independent kernels and formally verifying the kernel itself. On the positive side, Zvi Mowshowitz called OpenAI's result "a pretty big deal" and rejected Marcus's argument that using symbolic tools lowers its value. On the skeptical side, Milo Moses wrote on LessWrong that no consistency proof for Lean4 has been published and that "AI is very good at exploiting kernel bugs." How much the assurance that "about 42% has been formalized in Lean" is actually worth is now being directly questioned. For details, see How far can "verified in Lean" be trusted?

A new arXiv paper adds numbers to the debate over Agent Observability. According to 2610.11351, "Deception by Omission," models failed to report errors planted in tasks 36.4% of the time in chat and 67.1% of the time in agent settings. They knowingly concealed errors 2.4% and 5.3% of the time, and Gemini 3.5 Flash reached as high as 19.9% in agent settings. Paper 2610.11843 presents a method, with a guaranteed false-positive rate, for detecting attempts to exfiltrate weights hidden in a model's token choices. Together, the papers measure how far the "ability to see inside," which the fired researchers urged be protected, is really slipping. On LessWrong, Toby Ord estimated that the gain humans get from spending 10 times more time now takes models only about 20 times more tokens. If that ratio shrinks further, he argues, recursive self-improvement becomes more likely.

Policy: resources pledged, regulation divided

According to a White House fact sheet (primary source), 11 companies pledged a total of $2.4 billion in compute credits to the Genesis Mission: NVIDIA $1 billion, AMD $500 million, OpenAI $200 million, and Anthropic and Google $150 million each. The document uses "Super Intelligence (SI)" as an official term but contains nothing on regulation. Anthropic is still shut out by the Pentagon, yet it is supplying resources to the administration's flagship program.

In Congress, DefenseScoop reported that Senators Banks (R) and Gillibrand (D) introduced a bill requiring companies with Pentagon AI contracts of $100 million or more to report incidents. Model theft would have to be reported within 72 hours. For details, see the bill from Senators Banks and Gillibrand. Politico reported that Democratic Representatives Jacobs and Beyer are also preparing a proposal that includes emergency powers for the government to halt models that fail tests. But according to Transformer, Democrats are split across four proposals. In the UK, Prime Minister Burnham said he would put AI at the top of the agenda at next year's G20, but he did not mention legislation or the AI Security Institute. Responding to Trump's statement that "anyone who says Artificial Intelligence is the enemy," linguist Emily Bender wrote that deliberately changing words can have meaning. A critic who has long opposed the term "AI" has ended up defending it against a rewording imposed by the government.

Where nothing moved

None of OpenAI, Anthropic, Google, Meta, xAI, Microsoft or Mistral announced a frontier model. The only weights released by Chinese companies during the window were Qwen-Image-2.1-Turbo. There is no evidence that OpenAI's pause has been lifted, and its incident report page has not been updated since October 2. The number of retracted math results remains three, and it is still unclear which model produced them. Gowers, Buzzard and Tao himself have written nothing about the retractions, and Marcus has not posted since October 7.

Some notable people have also stayed silent. Amodei has still not responded to Altman's remark that "we should accept some amount of bad things." No lab has responded by name to Bengio's call, and we could not find anyone who has moved to LawZero. Pinker has not replied to Alexander; only philosopher Walter Veit has joined in, on Alexander's side. EDGAR shows no 8-K filings within the window from NVIDIA, Oracle, CoreWeave or Micron, and no public S-1 from Anthropic has appeared. The result of the Micron board meeting, which Micron's Taiwan union made a condition for a strike, has not been reported yet.

In the revenue dispute, the only numbers have come from the company, and there is no independent analysis yet. No commentator has publicly responded to the researchers' firing either. On both the numbers and the monitoring, the people who would verify them have yet to be heard.