Close to the compute.

For twenty years, the expensive thing about a software company was building the software. You hired engineers, they wrote code for two years, they shipped a workflow — a CRM, an email tool, a spreadsheet with opinions — and you charged ninety-nine dollars a seat. Hosting was a rounding error against the cost of building. That's why SaaS existed as a category: the workflow was the hard part, and the infrastructure underneath was cheap enough to hide.

That's over. Building software is cheap now. I can vibe-code a CRM in a week. A good engineer with an AI assistant can clone most of a big platform's feature surface in a month. The code isn't the moat. The workflow isn't the moat. The UI isn't the moat. If the thing you were selling was a spreadsheet with opinions, you're about to find out what it's like to compete with infinite engineers working for free.

What's expensive now is running the intelligence. Tokens cost money. Compute costs money. Every meaningful thing a software company does from here runs through a model at some point, and every one of those runs is real money — not rounding-error money, money that shapes your margin and decides whether you have a business at all. The center of gravity moved from building the software to running the intelligence.

So the companies that matter next sit close to the compute. Not on top of it — close to it. Close enough that what the customer buys is your operational advantage over raw inference, not the workflow wrapped around it. The pitch stops being we built this clever thing, pay a subscription and becomes you could wire this up yourself — stand up the models, the vector store, the prompts, the rate limits, the glue — or you could let us run it for a thin margin, and we carry all of it. That's the business. Thin margin over raw cost, in exchange for carrying the operational weight of the AI layer.

To make that shape work, you need a domain where the AI does most of the work, most of the time, across a lot of surface, on a recurring cadence. A calendar app has one AI moment and a lot of plumbing; the economics don't work. Marketing is the opposite. It's the most token-hungry function any business has — blog posts, landing pages, emails, product descriptions, FAQs, customer replies, endlessly, on a cadence, and the output gets better with context about the business. That's the exact shape AI is good at, and the exact shape a founder could wire up herself and never will.

That gap — between could and will — is where the business lives. And the pricing has to follow the cost: consumption, thin margin over compute, the meter running with the AI. Seats would be a lie, because the cost doesn't scale with logins; it scales with how much the AI does.

There's a tailwind under this, and it's the reason I'm building here and not somewhere safer. The cost of a unit of intelligence has been falling fast — roughly an order of magnitude a year at constant capability, for two years running. Every time the underlying compute gets cheaper, the math of a thin-margin, compute-denominated business gets better, not worse. I route every AI call through a layer that can swap between providers — Anthropic, Google, OpenAI, and the widening field of open models — and use the best model for each step: a cheap one to classify, a stronger one to write, a vision model to check the result.

And the honest exposure, because the honest version is the one worth trusting: I run on top of models other companies own. A provider could change pricing, deprecate a model, or build a marketing platform themselves. Every AI company alive has that exposure; it's the ambient condition of the industry, not a unique flaw of this one. If a frontier lab launched a marketing platform tomorrow, it wouldn't be fatal — it would force me to be clear about what I offer that they don't: eight integrated channels, a meter you can trust, and a willingness to operate at margins a big model company would never tolerate. I built this alone, on that bet. I'm not speculating about the new shape of software. I'm standing on it.