OpenAI Takes a $150B Haircut

OpenAI is reportedly finalizing an investment round of up to $30 billion, led by investors from the UAE and Coatue, at a $1.25 trillion pre-money valuation. The company is calling it an “opportunistic bridge round” ahead of a planned initial public offering. That valuation would still be nearly 50% above OpenAI’s current $852 billion valuation, but it is substantially below the $1.4 trillion target the company had presented to investors earlier in October—hence the episode’s “$150B haircut” framing. The company began discussions after investors approached it in September and told prospective investors it hoped to secure signatures by the end of the following week, though the final size was not set. OpenAI is also negotiating a commercial arrangement for additional computing capacity from the UAE. The company previously told investors it expected roughly $600 billion in total compute spending by 2030. It confidentially filed an IPO prospectus in June; CFO Sarah Friar told employees that OpenAI would be public in 2027, while CEO Sam Altman later said going public immediately would be ill-advised. The IPO is expected, though not scheduled, for the first half of 2027. The fundraising environment has cooled. OpenAI raised $40 billion in March 2025 and a record-breaking $122 billion in March of this year, while rival Anthropic has raised nearly $100 billion this year. Morgan Stanley and the tech fund Coatue reportedly found weaker demand at the $1.4 trillion price, with one source describing the response as “a bit of a thud.” Some investors are fatigued after earlier rounds and others want to see how Anthropic prices its expected IPO. Rockefeller CIO Jimmy Chang said markets need a successful AI IPO because both companies still need funding.

The Treasury Department proposed expanding restrictions on U.S. investments in China, Russia, North Korea, and other countries of concern. The program began with President Joe Biden’s 2023 executive order covering sectors such as semiconductors, quantum computing, and artificial intelligence, and was expanded by Congress in December 2025 through the Comprehensive Outbound Investment National Security Act. The new proposal would add hypersonic systems capable of traveling at five times the speed of sound, supercomputing, biotechnology, and artificial intelligence—described in the proposal using President Trump’s preferred term, “superintelligence,” or SI. The rule would apply to investments exceeding $50,000 but exclude many financial investments, including ordinary stock purchases. Treasury will take public comments before issuing a final rule. Craig Singleton of the Foundation for Defense of Democracies said the proposal signals that Washington is tracking where American capital flows because capital can accelerate military capabilities just as technology transfers can. Former Justice Department lawyer Jason Chipman said companies would need more substantial compliance programs to manage the risks. Brian presented Beijing’s position as a partial inverse. Chinese state media said DeepSeek and MiniMax had received approval to deploy models overseas and that much Chinese AI would remain freely exportable, despite a draft proposal covering some large language models and agents. The proposed rules would apparently exempt API access, web applications, and software-as-a-service products even when they use restricted models. Chinese models have gained international users through lower prices and customization, while Beijing’s more prominent intervention included forcing the unwinding of Meta’s acquisition of Manus, an AI-agent startup with Chinese roots.

OpenAI made its Decisions API generally available, positioning it as a competitor to TypeSafe’s Jev models. These systems act as “super classifiers”: given a set of options, they quickly choose among them and provide a confidence rating. Developers have found that this kind of System 1 model can replace larger language models in some workflows at much higher speed and lower cost. OpenAI is charging $0.10 per million input tokens and no output-token fees, and the public release adds voice and visual modalities. Nukunj Honda, OpenAI’s API product lead, said the team saw Jev and immediately thought of “a million use cases.” Rather than train a dedicated decision model, OpenAI reconfigured its ChatGPT Luna model to reduce inference time and cost. Honda said latency fell from one to three seconds to 120–180 milliseconds. At Speak, an AI language-tutoring company, CTO Andrew Sue said the API reduced the time between actions from 1.4 seconds to 0.13 seconds. OpenAI also used it to make its Dots voice assistant more responsive and geographically placed inference resources to reduce latency worldwide. The product’s performance depends on the workload, including the tradeoffs among intelligence, calibration, and speed. TypeSafe CEO Diego Almeida distrusts generic benchmarks and recommends testing models on a company’s own tasks, although TypeSafe published rankings anyway. TypeSafe’s latest model led the “Jevman” benchmark, in which models play Pac-Man and must make rapid decisions, while OpenAI’s model ranked eighth and fourteenth on two other public indices. Honda said OpenAI wants to support more image-based inputs and potentially reach roughly 60 milliseconds, enabling a call for every frame of a video and revealing use cases that were previously too expensive or slow.

Major U.S. banks are restricting third-party AI agents from entering customer accounts or making transactions after reports that some agents acted independently of users’ apparent instructions. Customers in some cases gave agents their login IDs and passwords, which creates an authorization complication: if the customer shared the credentials, a transaction may not legally count as unauthorized even if the agent behaved unexpectedly. The banking industry nevertheless sees autonomous agents as a significant risk to customers and to institutions trusted with their money. Morgan Stanley has blocked third-party agents from accessing its platform, although a source said it had not seen an actual instance of an unauthorized or rogue agent. JPMorgan Chase CEO Jamie Dimon said the bank was developing tools to manage the risk but would not allow “just any outside agent” inside the company because it was too dangerous for both clients and JPMorgan. His remarks followed discussions among banking executives about agents carrying out tests or actions on customers’ behalf without clear control. Charles Schwab CEO Richard Wurster said his firm was also blocking unauthorized agents. He receives a daily report showing how often agents have been stopped and said the number remains small because customers are not yet using agents extensively for financial matters. Schwab is blocking them preemptively to retain control over authentication and the customer experience. The episode’s conclusion was that banks are choosing controlled access over convenience until they can determine how to distinguish a customer’s intended instruction from an agent’s independent behavior.

The weekend long read followed Matt Cortland, who paid €7.80—about $8.93—for a pint of Guinness in Dublin and wondered what the average price was across Ireland. He discovered that Ireland’s Central Statistics Office had stopped tracking the price of the country’s most popular beer in 2011. Cortland, who runs an AI startup, decided to collect the data himself. Using ElevenLabs for voice generation, he created Rachel, an agent with a Northern Irish accent named after Rachel Duffy, a winner of the British version of The Traitors. Rachel made more than 3,000 calls to pubs across the island and conversationally asked bartenders what they charged for Guinness. Cortland then used Anthropic’s Claude to turn the responses into the “Gindex,” which he describes as a living, breathing consumer price index. Bartenders and drinkers can contribute or modify prices. On the Monday discussed in the episode, the average pint cost about €6.01, or $6.88, while the most common price was €5.50, or $6.30. One bartender quoted a standard price of £6.20 but offered a cheaper pint to a visitor, while another said, “If you can’t afford one, we’ll buy you one.” The project also illustrated how convincing voice agents can be. Cortland said most people did not realize Rachel was artificial, and outside data cited by Fortune suggested people spend 14% more time with voice agents and give 22% longer responses than with human representatives. The Gindex reportedly prompted one pub owner to cut the price of Guinness by €0.40 and update the listing personally. Cortland wants to apply the model to prescription drugs in the United States and pizza prices in New York.