BIS Warns the AI Bubble Will Take the Economy With It

The Central Bankers Said It Out Loud

The Bank for International Settlements, an institution that normally communicates in sedatives, used the words bubble and global financial stability in the same warning. Returns on AI capex are too weak to justify the spend, and a funding pullback would not stay contained to Nvidia's order book. It would propagate through hyperscaler balance sheets, private credit, sovereign wealth allocations, and the retirement accounts of people who have never heard of a transformer.

This is the most serious thing a central banking body has said about AI to date. It is also late. The capex has already been committed, the data centers are already breaking ground, and the equity concentration is already a structural feature of every major index.

Which brings us to Australia. Reporting on superannuation funds revealed that balanced portfolios, the default option for millions of retirees, now carry up to 12 percent exposure to tech and AI names. Members did not opt in to a thematic bet. They opted in to balanced. The word has quietly come to mean overweight US megacaps with a side of Nvidia. If the BIS is right, a lot of Australians are about to find out what their fund manager meant by diversified.

Infrastructure Keeps Pouring While the Warning Light Blinks

Firmus, an Australian AI infrastructure firm, announced a Nvidia partnership and 30 billion dollars in committed agreements for a first data center project in Indonesia. The timing is almost funny. Capital is still being deployed at a pace that assumes the curve only goes up, even as the body that monitors systemic risk publishes the opposite thesis. This is how every late-cycle infrastructure boom has worked. The last shovel goes in roughly six months after the smart money has already left.

The Financial Times also ran a useful corrective arguing that physical robots, not chatbots, will deliver the actual economic transformation. Manufacturing automation has measurable throughput. Conversational AI has measurable churn. If the productivity case for generative models continues to underwhelm, the BIS warning gets sharper teeth.

Models, Music, and the Slow Normalization of Surveillance Evidence

Zhipu released GLM-5.2, an open-weight model that matches Western competitors on cybersecurity benchmarks and lags on general tasks. This is a narrow win with a wide implication. Chinese labs do not need to match GPT-class models everywhere. They need to match them in the domains that matter to states. Cybersecurity is one of those domains.

Suno launched an incubator for unsigned artists, offering grants and marketing support in exchange for, presumably, training data and goodwill. The company is being sued by every major label for ingesting copyrighted music. Standing up a program that pays human artists is both a legitimate business move and a legal hedge. Both things can be true.

In California, prosecutors won an arson conviction using the defendant's ChatGPT search history as evidence in a case tied to a deadly wildfire. The conversational interface has been quietly absorbed into the discovery process. People treat chatbots like a diary or a therapist, then learn in court that they were typing into a logged database owned by a third party with no privilege protections. Expect this to become routine within eighteen months. Defense attorneys will start telling clients the same thing they say about text messages, which is do not put anything in there you would not want read aloud.

The Cognitive Cost Nobody Is Pricing

The Guardian published a quiet piece arguing that AI answers arrive faster than humans can generate doubt about them. The contemplative gap, the half-second where a person used to wonder if a claim was true, is being compressed to zero. This is not a regulation problem or a capex problem. It is a cognition problem, and it is the one that will be hardest to reverse.

The BIS is worried about your portfolio. It should also be worried about the thing reading the warning.