Listen to the “podcast” style version of this issue below!
Generated with Google’s NotebookLM.
Your vendors have read the same headlines you have
Buying software used to feel like signing a lease. You picked the building, you moved in, and the price went up a little every year.
That’s not the market anymore.
Whether or not AI actually hollows out subscription software, the company on the other side of your next renewal knows the question is being asked. That’s leverage you didn’t have last time.
Bending Spoons agreed to buy Airtable at an enterprise value of $1.285 billion.
In 2021 the company was worth over $11 billion.
Bending Spoons buys software companies past their prime and cuts costs. It also owns AOL and Vimeo.
That isn’t a story about what AI can do. It’s a story about a tool your team might use, and what somebody just paid for it.
Airtable isn’t obscure. It’s the kind of tool a department head puts on a company card without asking anyone. When a vendor moves to an owner whose business model is cost extraction, the thing worth protecting at your next renewal isn’t the price. It’s whether you can get your data out.
Atlassian’s chief executive said he’ll spend $250 million of his own money buying company shares. The thesis he’s arguing against now has a name in the market: the SaaSpocalypse.
On one hand, a category leader sold for a tenth of its peak.
On the other, a CEO is putting $250 million of personal money on the opposite side of that bet.
Both are real, which is why the move this quarter is a conversation about renewal terms rather than a project to replace anything.
What to do this week: Pull your software renewal calendar for the next two quarters. For each vendor on it, find out who owns the company now and how you’d get your data out if you left.
The number that justifies the spend has to exist before you spend it
You’ve already spent money on AI this year.
Ask what it bought and the honest answer is usually a story rather than a number. Not because nothing happened. Because nobody wrote down what the work took before the tool showed up.
That’s your baseline, and it only exists if you catch it first.
404 Media reported that Uber went through its entire annual AI budget in four months, then capped what employees could spend on tools like Claude Code and Cursor.
Accenture went looking for what was driving its own token bill. Leaked audio has its agentic AI strategy lead saying it wasn’t the engineers.
It was ordinary staff doing ordinary things. Turning PDFs into slides.
Accenture had already told senior staff to use AI or risk their promotions. Then it went looking for why the bill was so high. Nobody had shown those people what a page costs.
Gartner’s first-half 2026 CIO research, reported by InformationWeek, found 83% of CEOs are increasing AI investment.
The same research found 59% of AI initiatives never reach production.
Put those two next to each other and the median organization is funding a second round of AI work before the first one shipped anything. That isn’t a spending problem. It’s a sequencing problem.
DeepSeek plans to raise what it charges. Its most recent model shipped at fourteen cents per million tokens of input, which is roughly the floor every other vendor prices against.
When the cheapest provider moves up, everyone’s negotiating position moves with it. Price your next AI case at what it costs today, not at a number nobody has quoted you.
What to do this week: Pick the one AI tool you’re most likely to expand next quarter. Before you expand it, write down the number it’s supposed to move and what that number is today.
The specification decides the result, not the model
How well an agent performs is mostly decided by how well the job was described.
Not by which model ran it.
A pair of researchers and an operator publishing his own numbers arrived there from opposite directions, which is more interesting than either one arriving alone.
Jennifer Sloan at UCL and Vern Glaser at the University of Alberta drew a line between prompting an AI and directing one.
Prompting is bounded by the questions you think to ask. Directing means configuring a system to work through the whole set on its own.
They name three things to specify:
Context: what data it can reach
Capabilities: what it can do
Orientation: what it prioritizes
Those three words are the usable part. They’re a brief you could hand someone this afternoon, and they happen to be the same three things a security review would ask about.
Azeem Azhar’s team has been running agents for six months. He audited one week of it: sixty-two substantial tasks, about $800 of agent cost, against roughly $19,000 of human equivalent.
His first lesson is to write the finish line before the goal. A test that says it’s done, not a description of what you want.
The ratio is what gets quoted. The first lesson is what’s usable.
“Make this look more organized” is not a finish line. “Every face of the cube is one color” is. If you can’t write the second kind of sentence about a task, the agent doesn’t have a target either.
What to do this week: Take the next task you were going to hand to AI and write the finish line first — the specific test that says it’s done. If you can’t write that sentence, that’s the finding.
Something is booking tables without seeing your website
Somewhere, someone is booking a table without ever looking at the restaurant’s page.
An agent is doing it for them. Software is being built for that quickly now, and it adds up to a channel into your business that most companies haven’t looked at yet.
Shopify said AI-driven traffic and orders to its merchants tripled year over year in the second quarter. Its read is that AI search isn’t pulling traffic away from its stores the way it has elsewhere.
Publishers have had the opposite experience with the same technology.
The difference doesn’t look like luck. A publisher’s product is the page, and AI can summarize a page. A merchant’s product is on the far side of a checkout, and something still has to complete it.
Google Maps added food ordering and hotel booking, moving it from something that tells you where a place is to something that transacts with it. Cloudflare, which sits in front of a large share of the web, launched a browser built for AI agents rather than people.
If customers find you through Maps, an agent is now the thing filling in your booking form.
Cloudflare building infrastructure specifically for agent traffic is a forecast about volume. Companies don’t build for a trickle.
What to do this week: Have someone on your team book or order from your company the way an agent would, through Maps or an assistant rather than your website. Then find out who owns that channel. The useful part is usually discovering there’s no name.
Of note
Four labs, one testing firm. OpenAI and Meta both named the same outside testing company, Irregular, in model-containment incidents four days apart. Moonshot’s Kimi made it four labs. The testing supply chain looks more concentrated than the individual incident reports suggest.
ChatGPT’s free tier goes unlimited. Free and Go users are getting unlimited text chats. Whatever your AI policy says, everyone in your organization now has unmetered access on a personal account. The rate limit had been doing quiet governance work.
Don’t be a meat proxy. Simon Willison surfaced a line worth stealing. If you’re forwarding AI output you haven’t read, understood, and rewritten in your own words, you’ve made yourself the least valuable part of the chain. That’s a standard you can say out loud in a team meeting.
A renewal, a baseline, a finish line, and a channel nobody owns
There’s a version of AI news that’s about what the machines can now do.
Today, none of the four things worth your attention are that. They’re a renewal, a baseline, a finish line, and a channel nobody owns.
All of them already in the building.