After two weeks spent on who holds the AI labs to account, this week turned to who pays for them. On Thursday the Financial Times reported that OpenAI's annualized revenue is about $20 billion lower than the figure that had been circulating, and chip and cloud stocks fell until a second report on Friday steadied them. Public investors walked away from a $5 billion data centre listing in Australia. Broadcom, Oracle and SpaceX turned to private lenders for tens of billions of dollars to buy chips, and private funding rounds carried on at full size. Prices for buyers kept falling, with small-model rates down as much as 90%, and Google gave its new work agent an email address of its own. Elsewhere, OpenAI published 722 mathematics papers and withdrew three within a day, a senior Senate Democrat proposed independent audits before advanced models ship, a poll found two in three Canadians want limits on data centres, and HubSpot cut 7% of its staff in a year when revenue grew 20%.
1. OpenAI's revenue came in $20 billion below the reported figure
The Financial Times reported on October 8 that OpenAI has told investors its annualized revenue was close to $50 billion at the end of September, not the roughly $70 billion widely reported a week earlier; CNBC's source says the larger figure included gross revenue from OpenAI's partners, a method intended to allow a direct comparison with Anthropic's reported $65 billion run rate. Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8% that day before recovering on Friday when Bloomberg reported OpenAI expects to reach the $70 billion pace by year-end, a reminder that "annualized revenue" is an unaudited figure vendors do not all calculate the same way.
2. An Nvidia-backed data centre firm scrapped its $5 billion IPO
Firmus, an Australian AI data centre operator backed by Nvidia, shelved on October 9 what would have been the second-largest new share sale in Australia's history after lukewarm demand, with Reuters reporting that investors were concerned about roughly $30 billion of debt and about terms allowing existing holders to sell more than half their stock at a $30.6 billion valuation. The company made $51 million in revenue last financial year and has completed 46 megawatts of a 912-megawatt pipeline, and it will now seek private capital instead, which Reuters reported as a sign that investors are looking harder at the economics of AI investments.
3. Broadcom, Oracle and SpaceX went to private lenders to pay for chips
The Wall Street Journal reported on October 7 that Broadcom is arranging more than $50 billion in financing for the custom AI chip it is developing with OpenAI, with Apollo and Blackstone in early talks, while Oracle explores having private investors fund an off-balance-sheet entity that buys chips and leases them back and SpaceX seeks to borrow directly, each in deals worth tens of billions of dollars. The Journal says the hundreds of billions of dollars in bonds already issued for the AI build-out have pushed public debt markets to their limits, so a week after lenders questioned what Nvidia chips are worth as collateral, the borrowing is moving to private credit where it is harder to see.
4. Private investors kept writing billion-dollar cheques for AI
Cloud provider Lambda is raising up to $4 billion at a $14.5 billion valuation led by Coatue and Blackstone and has pushed its IPO from this year to 2027, TypeSafe AI raised $870 million at a $7.5 billion valuation, and Manus closed more than $500 million in its first round since Meta was forced to unwind its acquisition. Bloomberg reports DeepSeek is close to securing about $12 billion ahead of a listing early next year and the Journal reports Amazon is in talks to buy Decart for around $7 billion, so the money has not left AI, but more of it now sits in private markets where prices are not tested every day.
5. The price of routine AI work fell by as much as 90%
Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens, matching OpenAI's GPT-6 Luna and coming in 90% below the previous Haiku for requests under 100,000 tokens, though the price is five times higher above that threshold and VentureBeat notes it is not the cheapest model on the market. Small models handle the high-volume work of summarizing, classifying and answering routine queries, so the cost of that work keeps falling for buyers, which is also one more reason vendor revenue is hard to forecast.
6. Google's new work agent gets its own email address and account
Google Cloud introduced the Gemini agent on October 8, a single AI agent for work that plans and completes tasks across business systems, has its own Google Workspace account and email address, and writes an audit trail under its own name rather than a person's. It can run on Anthropic's Claude models as well as Gemini and works inside Microsoft 365, Slack and ServiceNow, with businesses getting it first, which makes agent activity easier to audit and also adds one more account whose permissions someone has to set.
7. OpenAI published 722 maths papers and withdrew three a day later
On October 6 OpenAI released 722 manuscripts claiming progress on 372 mathematical problems, produced by an unreleased internal model, then withdrew three the next day because of a sign error and revised 14 others; a spokesperson told Retraction Watch that roughly 50% of the results were released unconfirmed. MIT mathematician Andrew Sutherland said many in the field are "unhappy with how OpenAI has behaved up to this point" while expecting the corrections to be viewed positively, and the episode shows that when AI produces work faster than experts can check it, verification becomes the slow part.
8. A top Senate Democrat proposed audits before any advanced AI ships
A week after six AI companies promised to audit themselves, Senator Maria Cantwell, the senior Democrat on the Commerce Committee, set out six principles on October 7 that include enforceable safety standards for frontier models and an independent audit before advanced models are released, "similar to outside auditors examining a company's financial statements," Politico reports. Cantwell would chair the committee if Democrats win the Senate in November, current chair Ted Cruz has repeatedly postponed a markup of AI bills, and Representatives Brett Guthrie and Ted Lieu said on October 8 that a kill-switch law could pass in the lame-duck session, so the shape of US rules now depends heavily on the election.
9. Two in three Canadians want provinces to limit AI data centres
An Angus Reid Institute poll of about 4,400 adults found 64% think their provincial government should limit large data centres against 18% who said governments should encourage them, with opposition highest in Manitoba and Atlantic Canada at 72% and lowest in Quebec at 50%, and about half of opponents naming water use as their main concern. The numbers arrive as Meta builds a $13 billion site north of Edmonton, two First Nations challenge Bell's AI data centre south of Regina over consultation, and Emera agrees to buy Alberta's Canadian Utilities for C$14.3 billion, which means the capital for Canada's build-out is arriving faster than public consent for it.
10. HubSpot cut 7% of its staff while growing 20% a year
HubSpot said on October 6 it will cut nearly 660 jobs, about 7% of its workforce, a quarter after reporting revenue of $911.7 million, up 20% on the year, with chief executive Yamini Rangan saying the company is moving from "building software that helps customers grow to delivering outcomes for them with AI" and that the cuts were not driven by AI efficiencies. Amazon confirmed fewer than 1,000 cuts the next day, mainly in its Stores division, and the pattern of a growing company reorganizing around a different product is harder for employers and staff to plan for than automation replacing individual tasks.
