Silicon Valley Bank failed in forty-eight hours this fortnight, and by Sunday night Washington had guaranteed every deposit to stop the failure from spreading, and the striking thing about the death of this bank is that it did not, in the ordinary sense, run out of money. It held assets. They were merely worth less than they had been, and locked up longer than its depositors were suddenly willing to wait, and the moment enough depositors decided, more or less at once, to get their money out, the bank was dead, not because the money was gone but because the belief that let everyone leave their money in had evaporated. I build systems for a living, and this collapse is the purest demonstration I know of the most important piece of infrastructure in any complex system, the piece that never appears on any architecture diagram: trust.
Economists have a precise model of what happened, and it is worth knowing because it is stranger than the morality-tale version. Douglas Diamond and Philip Dybvig won a Nobel Prize for showing, formally, that a bank run is not necessarily a response to a bank being truly insolvent; it can be a self-fulfilling equilibrium, a situation with two possible outcomes, one where everyone believes the bank is sound and therefore leaves their money in and therefore the bank is sound, and another where everyone fears everyone else will withdraw and therefore withdraws first and therefore the bank collapses, and crucially both outcomes are available for the very same bank with the very same assets, the only difference being what everyone believes everyone else believes. The bank's solvency, in other words, is partly a function of the belief in its solvency, which means the belief is not a description of the infrastructure; the belief is the infrastructure. Walter Bagehot understood this in the nineteenth century when he wrote that credit, the whole edifice of modern finance, is a system of confidence, and grows by belief, and collapses by its withdrawal, and that a panic is not a reasoned response to facts but the sudden evaporation of a trust that was holding up more than anyone realised.
Now the counterargument, because there is a real one and I do not want to romanticise the run as pure irrational panic descending on an innocent bank. SVB had made genuine mistakes; it had taken real interest-rate risk, concentrated its depositors in a single jittery industry, and left itself exposed in ways a well-run bank would not have, so the loss of confidence was not baseless mania but a partly rational response to real fragility, and the depositors who ran were not fools destroying a sound institution but people correctly noticing an unsound one. I grant this fully; the trust collapsed faster and further than the underlying facts alone would justify, but it collapsed onto a foundation that was genuinely cracked, and a lesson that treated the run as pure self-fulfilling prophecy would let the bank's real recklessness off the hook. But notice that this actually deepens the point rather than undermining it, because the interaction between the real fragility and the collapse of belief is precisely the danger: a small real weakness, once it becomes visible, triggers a withdrawal of trust that is wildly out of proportion to the weakness itself, and the disproportion is not a malfunction of the system but a property of it, because trust is binary in a way that solvency is not. You either believe or you run, and there is no stable middle, and so a modest crack in the foundation can bring down the whole structure at the speed of belief, which is now the speed of a phone.
This is the thing I try to make people understand about the systems I build, and it applies far beyond banks: that the most load-bearing components are usually the ones that appear nowhere in the specification, the shared assumptions and mutual confidences that everyone relies on and no one maintains because no one quite sees them. A system runs on trust, the users' trust that it will behave, the operators' trust in each other, the quiet confidence that lets a thousand people not all check the same thing at the same moment, and this trust is real infrastructure, as real as any server, except that it is never budgeted for, never monitored, and never maintained, because it is invisible right up until the instant it fails, at which point it fails completely and at the speed of rumour. We pour our attention into the visible machinery and take the trust for granted, and then a bank that had money dies in a day and a half because the one thing it actually ran on, the belief that it would be there tomorrow, was withdrawn faster than any asset could be sold.
So the lesson I take from this collapse is not mainly about banking regulation, though there are lessons there. It is about learning to see the invisible infrastructure before it fails, to notice that the confidence holding up a system is a component like any other, more critical than most, and to understand that it obeys different physics: it accretes slowly and collapses instantly, it cannot be rebuilt at the speed it is destroyed, and it is destroyed not by the exhaustion of resources but by the withdrawal of belief. A bank is a machine for turning trust into money and back again, and this fortnight one of them showed us what happens when the trust runs out before the money does. Everything we build runs on some version of that trust. We would be wise to start treating it as the infrastructure it is, and to notice, before the run, that the most important thing holding up the system is the one thing we never think to maintain.