From Models to Users

The AI industry in early 2026 was shaped like an upside-down pyramid. The news cycle was dominated by model companies across the top, holding most of the capital and all of the hype, tapering to a narrow point where coders did magical things with it.

Then Q2 came and the conversation shifted to orchestration, open source models, and cost optimization... and in the last 2 weeks, things accelerated.

Ramp released the tool it uses to keep its own AI bills down, free to anyone. Stripe agreed to pay more than seven billion dollars for OpenRouter, a company most people outside software have never heard of. Martha Stewart co-founded a "homeowner management" app that never says the word "AI." And 2 months after Anthropic announced Claude in Slack, Slack announced AI coworkers into shared project channels - well beyond just Claude.

Four stories. All four are about the same missing layer.

Why the picture looked so simple

Something has to sit between a raw input and a person who benefits from it. Something wires it into how the work gets done. Something wraps it in a product with a login, a bill, and someone to call. In most industries that middle is where the work lives, and most of the money.

Developers are the only customer who don't need anyone else to build that middle for them.

Sell a developer access to a model and they'll assemble the rest over a weekend. They pick the model. They wire it in. They build the thing on top. And when they do buy tools, they buy tools shaped like themselves. That's what got built: Cursor, Vercel, LangChain, a whole industry of software for people who write software.

So the middle was never missing. Models ran rampant for exactly one kind of customer, and that customer happened to be the only one who could have done without the packaged middle.

Which means the chains aren't getting longer so much as re-shaping around whoever is buying, and a finance chief needs a different middle than a marketing chief, who needs a different one again than a homeowner. Somebody has to build each of them.

Right now, everybody is building, and everybody extends far beyond just Anthropic and OpenAI juggernauts.

The enterprise, past engineering

Stripe is the one that needs explaining, because seven billion dollars is a lot of money.

The boring explanation covers most of it: OpenRouter meters usage, handles prepaid credits, and settles payments across dozens of AI suppliers, which makes a payments company buying payment plumbing for AI no mystery at all.

What that plumbing turned into is the interesting part. When Google launched a new low-cost model this month, it came in at half the old price - an introductory rate that doubles on the first of January. OpenRouter stacked another fifty percent on top of that, exclusive to its own platform. A developer who wired that model in during August is not going to revisit the decision in January.

That's not a piece of software routing traffic. That's a shelf, with suppliers discounting for eye-level placement and a customer who forms a habit and moves on. Whatever Stripe thinks it's buying, the shelf is what it's getting.

Ramp made the same move from the opposite direction, spending three years on an internal tool that kept its own AI features pointed at whichever model was cheapest and good enough, then releasing it to everyone this month. Free. Customers report their AI bills dropping by about forty percent.

Free is the tell. If picking between models were a business, Ramp would charge for it. Instead it gives finance departments a reason to choose Ramp, then spend more time and money inside Ramp. That's the business.

Then Slack, this week, put AI coding assistants into shared channels, where the team can see the work and review it like any other work. Slack owns the workplace, so that's where the assistants went.

None of those three is an AI company. Each one already owned a customer and reached down to claim the layers after the models. The labs tried reaching the other way up the chain: OpenAI killed Operator. Google killed Project Mariner.

The financial picture underneath rhymes with all of it. OpenAI grew 18% last quarter while its operating loss widened to $12.3 billion and it froze its newest frontier model, and Anthropic more than doubled, passing it on revenue for the first time. Anthropic sells mostly to developers and enterprises who build their own middle; OpenAI is building every layer itself, all the way down to a consumer. Two different distances to the customer, and one of them is enormously expensive to cover. (Note: that's not to say OpenAI is wrong, capturing the entire chain obviously has outsized value, but requires enormous capital and a lot of time).

The consumer, barely started

Everything we've seen this year has largely been centered on the enterprise user - starting with the CTO's office, now a race for the other buying centers. But the consumer user (and every cohort therein) is where I'm most fascinated.

I've spent the last few months watching agents shop, and the gap between what they'll research and what they'll actually buy isn't subtle. Americans research with agents constantly and buy with them almost never. A quarter of new-car buyers used an AI tool somewhere in their shopping last year. The share of purchases an agent completed is a rounding error.

Walmart found out why. It partnered with OpenAI last fall to let people buy through ChatGPT directly, and the checkout underperformed badly - wrong product details, conversion well below what Walmart does on its own site. So Walmart pulled it. Then it put its own assistant, Sparky, inside ChatGPT and Gemini instead, and OpenAI now sends shoppers back to the retailer's site to finish.

That's the same move Stripe and Ramp and Slack made, arriving at the consumer end. A company that owns the customer declined to let somebody else own the layer between them. Walmart kept the relationship and used the AI apps for reach.

Where that layer already exists, the numbers look nothing like ours. In China, one app tends to own identity, payment, catalogue and delivery together, so nothing has to be handed off mid-purchase. Alipay's agent payments passed 120 million transactions in a single week. Alibaba's assistant reached 300 million monthly users across Taobao and Tmall, against a catalogue of four billion products. Compare that to the US -

That level of adoption leans heavily on the super-app structure - mobile-first, a different regulatory history, no legacy card rails to work around - and it doesn't port cleanly to America, but the principle does. Everyone has access to the same models. Getting a purchase all the way finished isn't a model problem. It's an app and integration problem.

“Not one customer have I seen in our voice-of-customer data say, I really want to check out on ChatGPT… We want to solve customer problems.”

— Joe Cano, SVP of Digital at Lowe’s, to Jason Del Rey of The Aisle, April 2026

When the iPhone launched, the phone was the story, and it's still a huge business at $209 billion last year. The economy that grew on top of it now runs roughly seven times that. Apple's own study puts sales through the App Store at $1.4 trillion in 2025. That number deserves salt - Apple commissioned it, and it counts groceries and car rides ordered through apps, not Apple's revenue. Which is the useful part. On more than ninety percent of it, Apple collected nothing.

Analogies don't predict anything. But money is already moving as though people expect this shape.

Andreessen Horowitz raised a $1.7 billion fund in January aimed at applications, and said it had moved past infrastructure. Y Combinator has funded about 150 assistant startups.

And then there's Hint, the home app Martha Stewart co-founded in July. It tracks your maintenance schedule, your energy use, your insurance questions, your documents. You can ask it when the air conditioner was last serviced.

The homepage never uses the word "AI," or the word "agent." The headline is Homeownership without the guesswork. Human expertise. An app that knows your home. The word "model" appears once, in a promise never to train one on your data.

The technology runs all the way through it and never gets mentioned once, and what's actually for sale is judgment about your house.

What this means if you're not a tech company

Customers buy outcomes in a setting they recognize. Nobody has ever chosen a bank for its database, or a car for its steel supplier, and eventually nobody will choose a home app for its model.

For a couple of years AI was the exception, because its first customer could turn a raw ingredient into an outcome without help from anyone. That customer isn't alone in the room anymore, and the ordinary rules came back.

So the pyramid has flipped. It used to be the models on top of the conversation. Now it's (rightly) the customer and what has to sit between them and the machine. If you already have customers - a store, a dealer network, a patient list, a route - somebody is going to build that middle layer for them. The only real question is whether it's you, a competitor, or a company you've never heard of that decides your customer is theirs now.

What would prove me wrong is easy to watch for. If two years from now the model companies' own apps are where most people do this, the middle never mattered and the top of the pyramid kept everything.

I don't think that's where it goes. The last time the picture looked like this, the device sold enormously well and everybody else built the economy.