The edges are pricing AI risk. The centre isn’t.

OpenAI’s new Astra model was better at finishing hard tasks without human help. It was also more willing to use unsafe tools, and “more likely to try to deceive end users about actions it did or didn’t take”. OpenAI cancelled the release. Google did the same in a different shape: its newest Gemini model exists, but only vetted cyber defenders get it.

Under the White House accord, the companies pick their own evaluators and decide whether to publish. Trump called it “morally binding”. The Guardian reads the text as one that “appears to carry no enforcement mechanisms or legal implications”.

After the summit, Jay Clayton, the Director of National Intelligence, described the day’s consensus: American leadership in AI is “the most important thing for national security”. Days later Trump named him AI czar. Dario Amodei, who has called for federal safety regulations, opened with “As the President has said, you know, whoever wins AI wins”. Then he said the technology “has very real risks”. Being there proves access, nothing more. The accord “calmed some investor anxiety”, one analyst told Bloomberg.

The money kept its own clock. Anthropic’s prospectus, as Reuters reports it, spells out what its compute commitments mean: “If our actual spend falls short, we must pay Google the difference”. A lab can hold back a model overnight. It cannot hold back a payment schedule.

OpenAI chose the private route. It has pushed its listing to next year at the earliest, citing its focus on AI safety, with Abu Dhabi’s MGX among the funds in talks. OpenAI’s revenue depends on who is counting. Its investors tried to compare it more directly with Anthropic, including how sales through cloud providers are counted, and the number came out higher. Our inference: staying private keeps the argument over how to count that revenue out of public view.

The risk to this call runs the other way. If OpenAI stumbles, the centre won’t stay calm. The investor Steve Eisman argued on the New Money podcast that if OpenAI failed, the US would fall into recession almost at once.

Our read: the brake is a risk for anyone building or lending against the labs’ growth, because it can slow sales. Investors weighing Anthropic’s listing are not marking it down yet.

The Broadcom talks are early.

For a giant, higher rates are an annoyance. For a builder on floating-rate debt, they decide whether the business works: CoreWeave’s interest payments rise with every step up in rates.

Power is the other bill, and it lands at the edges too. Talen, a power producer, sold capacity credits to Citigroup; Fitch treats that sale as debt and cut Talen to B+. One FERC commissioner: “Existing customers should not be left paying costs attributable to new demand”.

The counter is that the junk market may simply be reacting to the cost of money. Its extra yield started rising before our window, as Treasury yields rose. If the edges are reacting only to rates, the call weakens.

Our read: risky for the lenders and builders taking on this debt as rates rise; the chip buyers at the centre keep the cost off their own books.

Firmus. The Australian builder, a “neocloud” that rents out AI computing power, had signed long-term service agreements with AI giants “that want to outsource data centre construction and keep that cost off their balance sheets”. Its clients include Meta and OpenAI. Contracts worth billions, revenue in the millions.

The day before the book closed, the brokers told clients: “At the fixed price, indications are well in excess of the offer size”. Then the orders came in short. UniSuper and Aware Super had passed. The banks, the AFR’s Street Talk column reported, “massively overestimated demand for the neocloud, especially from American long-only fund managers”. Firmus withdrew, saying the terms did not reflect the strength of its business.

Plato, which plans to bet against the stock, put the bear case: “Valuing a neocloud on an EBITDA multiple is a bit like valuing an airline while pretending the planes are free and last forever”. EBITDA is earnings before interest, tax and the wear on equipment. The AFR’s summary: “Investors’ data centre FOMO has clearly cooled”.

Inquiry. At Parliament’s AI inquiry, OpenAI’s Jason Kwon opened: “I want to begin with an apology”, for its models accessing Australian government websites. Our Medicare test, a mandatory reporting channel, is flagged but not yet law.

China. DeepSeek is close to a funding round that is “blowing past its own capital-raising target”, and plans a large Huawei chip cluster in Inner Mongolia. At China’s centre, money still comes easily.

Our read: risky for local investors asked to fund the build-out; the global giants signing its contracts carry little of it.