The AI bubble is breaking the wrong things

Alphabet reported on 28 October. Revenue $127.6 billion, up 25%. Operating income $40.6 billion, up 30%. Google Cloud $30.7 billion, up 103%, carrying $11.1 billion of operating profit. Capex for the quarter $51.4 billion, against $42.3 billion of cash from operations. Free cash flow minus $9.1 billion, against a contracted backlog of $567.7 billion, most of it cloud.

The stock fell 12.6% the next day.

And that pair explains the whole argument. The business got better and the bill arrived in the same quarter, and the distance between those two things is where the story sits. Not on the exchange.

Two waves, one mechanism

China's wave ran first. The STAR 50, Shanghai's index of chip and AI names, is down 42% from its June peak, and about $480bn of market value left with it over four months. October brought the American version: the Nasdaq lost 2.9% on the last trading day of the month, its worst month since March 2025.

So, two markets, one shape. Expectations got priced first, and the results arrived later at a different size.

The industry was partly its own customer

The Bank for International Settlements looked at 1,246 AI firms between 2021 and 2025. Of the investment those firms received, 55.2% came from other AI firms, and in 46.4% of the deals between AI firms the two sides were also supplier and customer (BIS Bulletin 137, 1 October).

And somebody outside pays

AI data centres took the memory, so every other product bid for what was left. Memory and storage went from 16% to 42% of a laptop's component cost (Omdia). A machine that retailed at $699 in early 2025 now carries $699 of parts alone. Also, global PC shipments fell 21% in the third quarter (Counterpoint).

But read that list once more. The first thing this build visibly broke was a personal computer.

The same mismatch, one floor down

But now walk inside a company. The same argument runs there, in a smaller room, with fewer numbers attached.

And a project got funded a year ago because someone senior read the same headlines you did. It has a demo, a vendor, and a quarterly deck with usage numbers in it. What it lacks is a sentence saying what one decision costs to run, and a name attached to the outcome.

I have watched this pattern in companies and agencies for years. When the people running a business cannot tell what to do next, they spend. A budget, a platform, a licence, another headcount. It is the fastest move available, it reads as decisive, and it works often enough to stay the default.

But the slower work loses that contest every time, because none of it has an owner in the org chart or a deadline in the quarter. What the system should do, who is allowed to do what, where the record lives, how anyone would know it went wrong. So it gets deferred, and the deferral stays invisible until someone asks a question in a review.

So ask what a decision costs to run, and the room goes quiet. Ask what the demo can do, and everyone has a slide.

Money arrives on a purchase order. Judgement does not.

Money has a process. A purchase order clears in days, a budget moves in a quarter, a licence renews on autopilot. Considered decisions have no equivalent, which is why they cannot be added at anything like the same rate. You can fund the tenth AI project this year without a new thought entering the building. And funding it does not make thinking about it any faster.

Inside an organisation of a hundred thousand people, that difference decides the outcome. The money moves, and the thinking queues. By the time the architecture gets reviewed, the platform is bought and the team has built against it. Nobody in that sequence is careless. So the sequence itself produces the wrong result.

The money moves first. The decision arrives to find the platform already bought.

So cost is worth reading properly, because money is where this becomes visible. Rarely as one dramatic write-off. Usually as the platform bought for a decision nobody defined, the model bill nobody can attribute to a product, the licence renewed because cancelling it needs a meeting.

The number I would put in front of any review is a ratio: what a decision costs, against how long it takes to arrive. Both halves count. And a cheap decision nobody reaches in time costs as much as a slow one.

Speed comes from what you built before

And if the constraint is throughput, more analysis changes nothing. What changes it is a decision that arrives already checked.

And that is not a theory I picked up somewhere. It is why I can move quickly on my own systems. Under them sit months of accumulated structure: a standing set of rules that travel with every session, a memory file per project so a decision made in August is still visible in October, checks that run themselves and fail a build instead of waiting for someone to remember, and a small number of points where a person actually signs.

Of course none of that was glamorous, and it took months, most of it tuning. The return shows up as speed. I can settle in an afternoon what would otherwise need a dozen meetings, three sign-offs, and a hundred people waiting to find out.

Three conditions hold it up, and all three are built things.

1. 1The risk of a decision sets the route it takes, so a small call never queues behind a large one.
2. 2The evidence gets assembled while the decision happens, because a record built afterwards is worthless in a room where somebody is asking.
3. 3Past a threshold, a named person signs, and the limit of that authority sits written down next to the number.

And a firm of a hundred thousand people runs on the same mechanics at a different scale, and how much of the structure is written down is what separates the ones that move from the ones that stall. The Seven Dimensions of Enterprise AI is the longer version of how those seven decisions fit together, and the authority half of it has its own post: the Agentic Authority Escalation Model.

What survives the review

Every bubble argument lands in the same place eventually: which of these things keeps working once money gets expensive.

But projects that survive a bad quarter share three properties. They are cheap to run in a way somebody can point at. They are clear about who may do what. And they can show their work, so the person who signed for one can answer a question six months later without reconstructing anything.

And that test is about design. Most projects turn out to fail it before a model is even chosen.

What I built instead

So I stopped making that argument after the money was spent, and put the test into a product.

TikSense measures each creator's ability to convert sales for your specific product, across commerce data, content fit, audience intent, and visual selling quality. Every analysis gives you a verdict and the outreach copy to send. Built for TikTok sellers and businesses deciding which creators to work with.

What matters here is what travels with the verdict. The evidence, a cost per analysis I can state out loud, and a clear line about what the system decides and what a person still has to sign.