This week, Google Cloud went down, and the trouble did not stop at search results or video queues. It reached into Britain’s banking system, where several banks and payment services built on that cloud went dark with it. Customers could not move their own money until the outage ran its course, according to an ABC News report. A bank did not build the data center that failed. It did not choose the hardware inside it, and it still owns every customer complaint that followed. So whose failure was this, and why does a bank get to call it someone else’s problem?
The Google Cloud failure was mechanical: a piece of infrastructure inside Google’s own network broke, and every service wired straight into it broke with it. Alphabet ($GOOGL) rents that infrastructure out through Google Cloud as a utility, and banks lease it the way a factory buys electricity, trusting that the current will be there without asking too many questions about the wiring.
Compare that with a warning from Microsoft ($MSFT): its security chief has said that frontier AI models cannot yet be trusted to behave as instructed, as a widely shared Bluesky post relayed. Startups are now pitching banks core-banking integrations built on those same models. One failure is a machine that went still. The other is a machine that keeps running while no one can prove, ahead of time, that it will do only what it was asked. Banking has already handed its uptime to the first kind of dependency, and those startups want it to hand over the second kind too.
These look like two separate problems, one mechanical and one about trust, each fixable with the usual tools: better redundancy for the outage and better oversight for the models. In fact, banking made the same decision twice. It outsourced the thing it cannot afford to get wrong, first its uptime and now its judgment, to a vendor it cannot fully inspect.
Josh Kaufman’s The Personal MBA has a name for the failure the Google Cloud outage exposed: a tightly coupled system. As he puts it, “When one domino fails to knock over the next, the entire System grinds to a halt.” A bank’s app is one domino in a chain that ends inside a cloud provider’s own network closet. An AI model wired into core banking would add a second kind of domino, one whose behavior cannot be fully predicted before it tips. Banking is not buying two different kinds of risk. It is buying the same risk twice, from two vendors, in two currencies: downtime and unverifiable judgment.
If that holds, the exposure will outlast one bad week. Years out, UK banks would probably keep their own settlement records on infrastructure they fully control, independent of whichever cloud hosts the rest of the stack, and regulators may make that a requirement rather than a choice.
Or the dependency could run the other way. Google Cloud, having already proven itself the floor under British payments, might end up deciding which AI systems banks are allowed to connect to it, and billing for every incident report along the way. Either path hands the decision about what a bank can trust to someone other than the bank, the same arrangement that already put a country’s banking system at the mercy of one provider’s bad day.
The outage and the warning are two installments on the same debt. Banking traded control for convenience once, when it moved its uptime onto someone else’s cloud, and now the startups are asking it to make the same trade again, this time with judgment instead of hardware. A system does not need to be malicious to fail you. It needs only to be the one piece you never checked yourself.
Leave a Reply