The AI Reality Check: Four Scandals, 60,000 Exit Offers, and a $2 Trillion IPO

What you are looking at

Data cut-off is 4 November 2026. Everything below links to primary documents or major reporting. Where a number comes from a leaked document or a single source, we say so. The sarcasm is ours. The facts are theirs.

The pattern across this month is easy to state. Money goes in. Headcount goes out. The adjectives go up. Below: four scandals, one layoff wave, one IPO, and a short list of what to do about it.

Scandal 1: OpenAI’s $20 billion revenue gap

On 8 October, OpenAI told investors it had reached roughly $50 billion in annualised revenue at the end of September, as first reported by the Financial Times. The figure everyone had been repeating a week earlier was $68 billion. A person familiar with the matter told CNBC the larger number included gross revenue from OpenAI’s cloud partners.

The market took it personally. That Thursday Nvidia fell 3%, Oracle nearly 6%, CoreWeave nearly 8%, and Broadcom and AMD about 4% each. Investing.com’s summary of the FT report adds the key detail: Anthropic counts sales made through AWS and Google Cloud in its run-rate, while OpenAI leaves that partner revenue out.

Same word, different ruler.

Two things to notice. First, “annualised run-rate” is the only metric where you take your best recent period, multiply by twelve and call it a year. It is a slide, not a statement. Second, OpenAI still told investors it grew run-rate 77% in the quarter, and CEO Sam Altman said in September that right now would be an “ill-advised moment to go public”. The company is reportedly eyeing 2027.

Your takeaway is cheap: whenever someone quotes an AI revenue number, ask “gross or net of partners, and is it audited?” We covered the underlying economics in how OpenAI makes money and the financing loop in AI circular financing.

Scandal 2: Grok and the deepfake lawsuits

Since January 2026, xAI’s chatbot Grok has been the defendant in a growing pile of legal actions over sexualised deepfakes. Women and girls are suing in US courts, including a class action in the Northern District of California alleging Grok created and shared sexualised images of women, and a later class action by three girls alleging the tool was used to generate sexual abuse material from their photos. The City of Baltimore sued X Corp, xAI and SpaceX, alleging Grok was deceptively marketed as a safe, general-purpose assistant. UK MP Jess Asato filed a High Court claim in June, and a running list of the suits and bans also notes a California Attorney General cease-and-desist letter and an EU privacy investigation of X.

Important caveat: these are allegations. No verdicts, no public settlements, no finding of liability. The ABA Journal frames the real question well: should AI companies be held liable for what their models generate?

The irony writes itself. A model pitched as an unfiltered assistant turned out to be an unfiltered assistant. Guardrails are a product feature, not a press release.

For you, the lesson is boring and expensive. If you wrap someone else’s model in your own product, you inherit the output. Logging, filtering and abuse handling belong in your launch checklist, not your post-mortem. Start with AI agent security and prompt injection.

Scandal 3: The consultancies that hallucinated

Management consultants sell AI strategy for a living. Three of them have now been caught shipping AI-generated errors.

• KPMG, June 2026. The firm pulled its report Redefining Excellence in the Age of Agentic AI after research group GPTZero flagged problems and the Financial Times verified them. GPTZero found only 5 of 45 citations were accurate, and organisations named in the case studies, including UBS, the NHS and Swiss Federal Railways, disputed the findings. The report had been public for about eight months, per the Irish Times.
• EY, May 2026. According to the same coverage, EY withdrew a loyalty-programme report with apparently fake footnotes the month before.
• Deloitte Australia, October 2025. Deloitte agreed to partially refund an A$440,000 government report that contained fabricated references and a made-up quote attributed to a judge. It had disclosed use of Azure OpenAI.

A report about agentic AI, hallucinated by AI, billed by the hour. If you wanted a case study of “why humans stay in the loop”, you could not commission a better one. You could only ask a model to invent it, and then you would have to check the citations.

Which is the whole point. Fluent text is cheap. Verified text is not. If your content, contracts or reports touch an LLM, add a citation-check step before publication. This site’s own rule is that every statistic links to a primary source, and we applied it here. For a pipeline that catches regressions like these, see agent evaluation and observability.

Scandal 4: The government renamed AI “Super Intelligence”

On 22 September, President Trump said at the UN General Assembly that the US would call AI “super intelligence”. On 29 September he signed an order titled “Inaugurating the Era of Super Intelligence”. Per the White House fact sheet, executive-branch agencies must use “Super Intelligence” and “SI” in official communications and “no longer acknowledge the terms ‘Artificial Intelligence’ or ‘AI’”. The science adviser has to propose a federal definition of SI. ABC News reported a “Super Intelligence Force” task force on 4 October, and in late October Trump said he wants to ban the term “AI”.

His reasoning, per PolitiFact: “It’s not artificial. We all agree on that.”

PolitiFact’s fact check is the useful part. In the AI industry, superintelligence means AI that exceeds human cognition in nearly all fields, and the industry generally agrees nothing has reached that level. In the order, it is a rename of existing technology. Same word, two meanings, one of which does not exist yet.

Our reading notes, offered in a spirit of constructive nomenclature:

• Rebranding is the only known technique that improves a model at zero GPU cost.
• Same weights, same hallucination rate, one extra adjective.
• A statistical LLM predicts the next token. The federal government has now predicted the next noun.
• A “Super Intelligence Force” is presumably the first service branch to launch with a naming committee.
• If your agent still cannot count the r’s in “strawberry”, it now fails with federal approval.

We will keep writing “AI” on this site, because that is what you search for. Sorry, Washington.

The serious point: a label is not a capability. When you sign an AI vendor contract, define what the system must do, at what error rate, with what recourse. Do not define what it is called.

Microsoft’s 60,000 exit offers: replaced by autocomplete?

On 26 October, Microsoft offered a voluntary exit package to about 60,000 employees in the US, Canada, India and the UK at experience level 70 or below, about 27% of its 223,000 workforce. Employees decide in early December and those who accept leave on 31 December. It follows a first-ever voluntary retirement programme in April, which more than 30% of eligible staff took, and a July plan to lay off 4,800. This is an offer, not a layoff, and Microsoft’s own framing is “a costly artificial intelligence infrastructure build-out with capital efficiency”.

The company says it does not expect 60,000 people to leave, and expects under 5% uptake. It also says AI is not the driver of its buyouts, as Fortune reports. Fortune also has the numbers that explain why everyone is asking:

• Spending on property and equipment was $115.9 billion in the fiscal year to 30 June, up from $64.6 billion.
• Capital expenditure in the September quarter rose about 30% to $45.2 billion, roughly three-quarters of it on short-lived assets such as chips and networking gear. The CFO expects more than $50 billion this quarter.
• Over the same fiscal year, headcount shrank by about 5,000.

Economist Jay Zagorsky told Fortune: “Microsoft’s announcement is a clear indication they want fewer workers.” Joseph Phillips of Seattle University says Microsoft is “becoming a more capital-intensive employer”. Both can be right.

Now zoom out. Challenger, Gray & Christmas counted 97,006 US job cuts in May 2026, with 38,579 attributed to AI, about 40% and a record for the category. AI was the top-cited reason for the third month running, after 26% in April. The same report says market and economic conditions (69,645 year to date) and closings (66,733) are also large, and Challenger’s own line is that “AI isn’t yet the jobpocalypse some predicted.”

So who is replacing whom? Be precise. A statistical language model predicts text. It does not do a job. What changed is where the budget goes: toward GPUs and away from salaries. The replacement happens in the spreadsheet, not in the model. “We are investing in AI” is a more flattering sentence for investors than “we hired too many people in 2021”, and that is our opinion rather than a documented fact. The honest test: does the cut name a task that software now does at equal quality, or a budget that moved?

If you are a founder, the practical read is about tasks, not jobs. LLMs absorb drafting, triage, summarising and first-pass code. They do not absorb accountability, judgment or the cleanup afterwards, which is why we wrote about AI cleanup after vibe coding.

Anthropic’s IPO: the safety company meets the roadshow

The IPO stopped being a rumour on 26 October, when Anthropic publicly filed its S-1. It plans to list on Nasdaq as “ANT”, with Goldman Sachs and Morgan Stanley leading, as Bloomberg had reported in June. We have not read the filing itself. These numbers come from Fortune’s summary and Yahoo Finance’s live coverage, which agree with each other.

• The asterisk: Anthropic has eight compute contracts worth $657 billion, of which $498 billion is noncancelable, per Fortune. Spread over roughly seven years that is about $70 billion a year. Q2 revenue annualised is about $46 billion, which is our arithmetic, not the filing’s.
• Net loss: $80.7 billion in the first half, which Fortune attributes mostly to non-cash fair-value changes on convertible notes. The 2025 net loss was $42 billion, the same figure Reuters had reported from the leaked prospectus.
• Valuation: Wall Street expects $1.8 to $2.0 trillion, with trading possible as early as mid-November.
• Revenue: $16.3 billion in the first half of 2026, against $4.6 billion for all of 2025. Q2 revenue was $11.6 billion, up from $4.7 billion in Q1.
• Profit: Q2 operating income of $974 million, the first operating profit reported. Q1 was a $1.7 billion operating loss.

Scandal 1 is still in the room. Anthropic counts sales made through AWS and Google Cloud in its run-rate; OpenAI left that partner revenue out when it told investors about $50 billion, not $68 billion. A prospectus reports audited revenue, not slides, but it does not settle which ruler is fairest.

Skeptics who said the revenue was imaginary were wrong on existence, not on structure. Independent research firm New Constructs still called it the “most ridiculous IPO of 2026” and valued Anthropic at $150 billion before the filing, against a roadshow range of $1.8 to $2.0 trillion. The filing shows real scale and real losses in the same document. The fight that remains is the bill: the S-1 itself warns that new compute can compress margins before revenue arrives, and the commitments are owed whether or not the capacity is used.

The macro backdrop is not friendly. CNBC notes rising rates are weighing on AI stocks, chip shares fell on concerns about AI financing durability, and Broadcom’s CEO said access to capital could be the biggest constraint on AI companies.

Three things we will be watching, with scepticism and popcorn:

1. Commitments versus revenue. Does revenue keep outrunning a roughly $70 billion-a-year bill? That is the whole thesis, bull and bear.
2. Who pays for the compute. The loop of chip makers investing in labs that buy chips is the one we mapped in AI circular financing.
3. What a public Anthropic does to its customers. Quarterly targets change pricing, rate limits and product roadmaps. If you build on Claude, or any single model, read how Anthropic prices and bills once it has shareholders to please, and keep a second provider warm.

Nothing here is investment advice. It is a reminder that an IPO turns a vendor into a company with obligations to shareholders, which is a different thing from a lab with obligations to its mission statement. For the sharpest critic of this whole financing structure, see our profile of Ed Zitron and the AI bubble case.

What to do with all this

• Define the ruler. In any contract or board deck, state whether revenue is gross or net of partners, run-rate or recognised, audited or not.
• Own your outputs. Whatever model you wrap, you inherit its mistakes. Log, filter, review, and set an abuse path before launch.
• Cite-check automatically. Add a pipeline step that verifies every URL and quote. The consultancies above did not.
• Separate tasks from jobs. Automate tasks with measurable quality. Do not announce “AI replaced X” until you can show the quality metric.
• Plan vendor risk. A public vendor raises prices when the quarter requires it. Keep your prompts, evals and data portable.

If you want help turning that into a plan for your product, talk to us, or start with our durable AI value versus speculation framework. And see our 2027 predictions for where this goes next.

Bottom line

The technology is real and so is the hype, and the two are billed on the same invoice. This month’s stories are the same story: output shipped faster than verification. Revenue without a ruler, deepfakes without guardrails, citations without checks, and a rename without a definition.

Meanwhile 60,000 people were offered an exit while a company spent $45 billion in a quarter on hardware, and a lab listing at a valuation its critics call absurd while its filing shows real revenue. None of it proves AI does not work. It proves that someone should check the numbers before you copy the strategy.