There is no inbox...

Good morning. This is Link, with your week in review for Monday, September twenty first through Sunday, September twenty seventh. Thirty eight newsletters came in this week, down again from last week's forty six, so whatever you're doing to that subscription list, it's working. Six things survived. That's about seven minutes: three headlines, then the learning corner, then one thing worth posting.

Headline one. The fight over who gets the click just produced real numbers, and they are ugly for anyone who runs a website. A controlled study of eleven hundred Google users found that switching on AI Mode cut the share of searches ending in an outside click by nearly nineteen points. News sites lost about twelve and a half points of that. Reddit lost twenty one. Over at Microsoft, internal Bing data shows Copilot has cut publisher click through rates by as much as ninety four percent. Ninety four. That's not a decline, that's a format change.

Here's the twist. The platforms starving publishers are now quietly testing paying some of them back. Google is piloting a program that pays select publishers when their work shapes an AI answer. And OpenAI just wired ChatGPT ads directly into HubSpot's CRM, so small businesses can buy back some of the attention AI search took from them for free. Take with one hand, sell with the other. Tidy business, if you're the platform.

Headline two. Nobody told the model labs there's a slowdown debate happening. Anthropic shipped Claude Opus five point five this week, matching its bigger sibling on most work for forty percent less to run. OpenAI answered with two cheaper GPT six variants. Grok four point seven landed with a bigger base model. And on the consumer side, Meta's Muse became the most downloaded personal AI agent in the country within two weeks, and immediately signed shopping deals with Walmart, Sephora, and Best Buy.

Here's the catch. Sixty four percent of adults say they don't trust Facebook, and ChatGPT still beats Meta on trust by a wide margin. So users get to pick their pill here. The agent everyone already trusts a little, or the one that just cut deals with half the mall. That's either genius or a very public faceplant, and we won't know which for a few more months.

Headline three. If there's a theme this week, it's control. Salesforce announced Koa, its first reasoning model, built so the company owns the weights and never has to send customer data across someone else's trust boundary. Tech press called it everything the big labs should fear. A Chinese lab named Z AI says its flagship model now runs entirely on Chinese made chips, with throughput that tripled in under two weeks. And a data center startup called Crusoe raised almost four billion dollars to build modular data centers in a factory and truck them wherever the power is.

All of that landed the same week the fight over how fast AI should even move went fully political, with the current administration pushing growth and two former presidents and a sitting governor pushing the brakes. Nobody agrees on speed. Everybody agrees they want the steering wheel.

And that brings us to the learning corner.

There's a genuinely good idea buried in an otherwise dense trade newsletter this week. Everybody worries about AI vendor lock in at the model level, and almost nobody worries about it at the memory level, which is where it actually lives now. Your assistant remembers your preferences, your project history, everything it's learned about how you work, and on most business plans that memory belongs to the individual seat, not the company, and it can't be exported in any real format. Switch tools, or lose the person who built up that context, and you start over from nothing.

The fix is almost embarrassingly simple. Write your own context down. Who you are, how you like work handed back to you, the standing facts you're tired of repeating, in one document you control, outside any chatbot. Anyone running more than one assistant, which, yes, describes this operation, should already be doing this.

Two smaller housekeeping moves while we're here. First, go check whether your chat tools are set to train on what you type, because more than one of them defaults to yes. Second, if you're paying frontier prices for repetitive, low stakes work, run the same batch of tasks through a cheap open model and see how close it gets before assuming you need the expensive one. The gap is usually smaller than the invoice.

Now, one to post.

If you post one thing this week, make it the argument that half the AI industry's spending might just be sitting in a warehouse. The case, laid out in detail this week, goes like this. Microsoft has told investors for three straight quarters that it's adding roughly a gigawatt of capacity every quarter. Dig into the actual numbers, and only about two gigawatts of that, total, ever, is AI specific capacity. The rest is warehoused chips waiting for buildings that don't exist yet, or power that isn't connected. Add up construction still in progress across the big cloud players, and you get north of three hundred and seventy billion dollars sitting half finished. The estimate for how much silicon is bought and paid for but not plugged in anywhere: somewhere around two hundred billion dollars, maybe more.

Here's the angle that makes it postable. Everyone's been told this is the dot com bubble again, cable in the ground that eventually got used. This argument says it's worse, because a chip sitting in a box for three years doesn't get cheaper to install and doesn't get more useful. It just sits there depreciating, waiting on demand that's mostly two customers signing very large contracts with each other. That's not a supply chain. That's an expensive game of hot potato, and somebody has to be holding it when the music stops.

So that's the week. Everybody wants to own something right now: the weights, the memory, the chips, the customer's attention, and the thing they all have in common is that ownership is expensive and slower to pay off than anyone's admitting out loud. That's Link. Talk to you next week.