The Day All Three Went Quiet── Not coincidence, not collusion — one point nobody was watching

When all three went quiet at once, the first thing I suspected was intent. Too neat for coincidence, I thought. What was actually there was not a will. It was one point nobody had been watching.

01Ninety minutes, all three at once

On 3 September 2026, three of the major conversational AI services stopped responding at almost the same moment. ChatGPT, Claude, Grok. It lasted about ninety minutes. More than thirty-seven thousand reports came in from users.

These three are competitors. Different companies build them. The models differ inside. So do the prices, the philosophies, the way they are sold. And the three of them became unusable at the same hour, in the same way.

Reading the news, the first question in my head was whether it was coincidence. The second surprised me a little. Might the three of them be in communication somehow?

02When things happen at once, people read intention

That reflex has a name and an experiment behind it.

In 1944 two psychologists, Fritz Heider and Marianne Simmel, made a short animation. On screen there is a large triangle, a small triangle and a circle, moving around a drawn rectangle. That is all there is.

Asked what happened, nearly everyone told a story. The big triangle is bullying the small one. The circle is escaping. The two are helping each other. Into a film of moving shapes, people read pursuit, flight and protection. Reading intention is closer to a reflex than to a judgment.

David Hume had identified the same working much earlier. The idea of cause is not fitted to the world in advance. A mind that has watched two events sit next to each other repeatedly adds the connection itself. The one making the link is the one who is watching.

Simultaneity is the arrangement that invites that addition most strongly. The moment I heard that three had stopped at the same hour, my head reached for a shared will before it reached for a shared foundation. Which is not different from reading a story into a moving triangle.

03What was actually shared

The cause turned up surprisingly quickly. An outage in one region of one cloud provider. It went down, and what was running on top of it stopped together.

Here is the part that matters. The three companies compete fiercely at the upper layer and share the layer beneath it. Which model is cleverer, which is faster, what it costs — everything we talk about daily belongs to that upper layer. The layer below is not discussed. And what is not discussed was never chosen; it sits there as somebody's default, set long ago.

What was connected was not the products. It was the foundation nobody had been looking at.

04Why "keep a second AI available" did nothing

Plenty of people had prepared for this. Depending entirely on one company is risky, so keep a second service available. That is sound judgment.

For those ninety minutes, the preparation did nothing at all.

The reason is simple. The layer they diversified was the wrong one. They separated the models; what was shared sat underneath. Holding two does not help if both are standing on the same ground. That is holding one.

And one more thing. Summarise the day as "depending on the three major services is dangerous" and the response goes wrong. The danger is not that we depend on three. It is that the three depend on one.

05What collusion would actually require

Take the first suspicion seriously for a moment. If the three really were in communication, what would have to be in place?

None of the four is in place. And that has to be weighed against the fact that one explanation — a shared foundation — already accounts for the whole of what happened. There is no work left for a second hypothesis to do.

Carl Sagan's account of scepticism is useful here: not being suspicious, but keeping a procedure for checking within reach. You are free to hold a suspicion; counting up what that suspicion would require is part of the same piece of work.

06Suspecting a will is simply more interesting

Something worth admitting: collusion is a far more interesting explanation than shared infrastructure.

Interesting meaning easy to remember, worth repeating, complete as a story. "Three AIs went silent in concert" has characters and a motive. "A data centre on the east coast failed" has neither.

Our heads are built to prefer the first. Given that we grant motives to moving triangles, it is hardly surprising. But preference and likelihood live in different places. And awkwardly, the interesting explanation travels faster.

Several posts circulating that day did suggest collusion or censorship. They were not correct, and they were widely read while not being correct. The announcement of a cause arrives after the suspicion. It always does.

07What deserves caution is concentration, not intent

If there is something to take from this, it is neither "AI cannot be trusted" nor "something is going on behind the scenes." It is duller and more structural than either.

Competition happens at the upper layer; dependency is shared at the lower one. The number of models available grows every year, while the number of providers supplying the compute beneath them can be counted on your fingers. The more varied the top looks, the harder the concentration below is to see.

So a stoppage of the same shape will happen again. Whether it stays at ninety minutes, nobody knows. What is worth preparing is not a longer list of AI options but deciding, before it stops, how the time while it is stopped will be spent. That is not a technical question. It is a question of arrangements.

On the day all three went quiet, the first thing I suspected was intent. People appear to be built that way. If we read stories into moving shapes, reading something into three services that stopped at the same hour is, if anything, natural.

But what was there was not a will. It was one point nobody had been watching. Looking for intention is interesting. Counting the foundations you share is dull. And it is the dull one that shortens the next ninety minutes.