The Last Subscription

Earlier today I published a forecast: one billion people paying $100 a month to OpenAI or Anthropic by 2030. An hour later, while working out how to post it to LinkedIn and X, I was offered a scheduling tool. It had a free tier. It would have taken thirty seconds. I said no, and the reason I said no is the second half of the argument.

I do not add subscriptions. I only remove them. That is not frugality; I will happily pay $100 a month for something that works while I sleep. It is a rule about leverage. Every tool between me and my audience is a party that can change its price, its terms, or its mind. A free tier is a subscription with the invoice deferred. So the post went out through a browser I already had, and nothing new got a hook in me.

That small refusal is the whole economic mechanism of the $100 bill. People will not pay $100 a month for intelligence in addition to everything else they pay for. They will pay it instead. The intelligence bill is the last subscription because it is the one that makes the others unnecessary.

Start with what people already spend. CNET’s 2026 survey puts the average American at $111 a month across their subscriptions. Narrower surveys that count only entertainment come in far lower, around $35, and broader ones that count everything with a recurring charge come in higher. The number itself is less important than its shape: a stack of small charges, each one a single capability, most of them forgotten until the card gets declined.

Now look at what those charges buy. A grammar checker. A scheduler. A transcription service. A note-taking app. A password manager. A PDF tool. A newsletter tool. A social media scheduler. A meeting summarizer. A habit tracker. Each is a thin capability wrapped in an account, a login, and a monthly charge. None of them is a business a person would miss if the capability arrived some other way.

The $100 is not a premium over the $10 app. It is the $10 app, ten times over, with no logins, no seams, and no one between you and the work.

When the capability arrives as a scheduled task, a plugin, or a skill inside the one system that already knows your calendar, your files, and your inbox, it becomes what I teach as ambient intelligence: invisible, always on, never prompted. The standalone version stops making sense. Not because it got worse. Because it became a feature of something you already pay for. A person paying $100 to Anthropic and $0 to nine other companies is not spending more than they were. They are spending about the same, for a great deal more, with one relationship instead of ten.

That is why the forecast is conservative. I did not assume a billion people would find a new $1,200 a year. I assumed they would redirect the $1,332 a year most of them are already spending, and that the thing they redirect it toward would be worth more than what it replaced. The first condition is arithmetic. The second is the product getting good, and that is happening on a schedule I can see.

If you want to know whether the market believes this, it already voted. Between 12 January 2026, when Anthropic released Claude Cowork, and 23 February, the S&P Software & Services Index lost about a quarter of its value. The sharpest move came after 30 January, when Anthropic open-sourced eleven plugins, each aimed at a job function: legal, finance, sales, marketing, support, data. Within two days, roughly $285 billion in software market capitalization was gone. Thomson Reuters, which owns Westlaw, fell almost 18 percent. RELX, which owns LexisNexis, lost more than 14 percent in one day.

Read step five carefully, because it is the shape of the future for the survivors. Docusign, Intuit, and Salesforce did not beat the agent. They became something the agent calls. That is a real business, but it is a wholesale business. The customer’s relationship, and the customer’s $100, sits with the stack. The software company gets a connector fee and a place in the menu.

In 2009 Apple ran a commercial with the line “there’s an app for that.” It described a world where every capability was a separate product with a separate icon and, soon after, a separate subscription. The line for the next decade is different, and shorter: your stack does that. Not an app for the thing. A scheduled task, a plugin, a skill, living inside the one system that already has your context.

Grammarly is the clearest example of what that does to a company, and I wrote the full case study last week. The interesting part is that Grammarly saw it coming. Grammarly was, for fifteen years, a feature with a subscription: check my writing. When every model could check writing for free, the company did not try to out-grammar the models. In October 2025 it renamed itself Superhuman, bought the Superhuman email client, bought Coda and Rows, and started selling a suite with an assistant in it. Whether that works is not settled. But the move itself is the confession: a standalone capability is not a company anymore. The only defensible positions are to own the stack, or to be a connector the stack calls.

That frames the economic impact on software more precisely than “SaaS is dead.” Gartner has worldwide software spending at about $1.44 trillion in 2026, still growing at double digits. The dollars are not disappearing. They are migrating, and they are migrating in three distinct directions depending on what a company actually owns.

1. Thin wrappers go to zero. A product that is a model plus a prompt plus a login is a feature with a burn rate. Its entire value can be reproduced as a skill file. The market has already priced these; the funding will follow within a couple of cycles.
2. Systems of record become connectors. Companies that own data, contracts, compliance, or a ledger that other things depend on survive, at wholesale margins, as a database and a server that people’s own AIs call. Their revenue compresses from seats to calls. The February 24 partnerships are the template.
3. The stack and the metal capture the surplus. The person’s $100, and what a company pays to keep its records reachable by its people’s AIs, flow to whoever owns the relationship and the compute underneath it. That is the labs, the clouds, and the chip and power supply chain. This is where the $1.2 trillion in the first piece comes from.

The unit of billing changes too. Seat-based pricing assumes one human uses one login. When one agent does the work of ten logins, the seat count falls even as the work rises, which is why Gartner expects a large share of SaaS spend to move to usage and outcome pricing by 2030. Software companies that survive will bill for things that happened, not for people who might use them.

I do not think most of those companies survive to 2030, and I do not want you inside one when it happens. The reason is deeper than “the platform will ship that feature,” though it will. The premise of an “AI for X” company is that there is a customer out there, a dental office, a property manager, a church, who will buy intelligence for a purpose. In the world the first piece describes, there is no such customer. Every person with the means to buy intelligence has exactly one customer: themselves. They pay one utility for raw intelligence. They say something as plain as “help me be a better me.” It starts working on their behalf, invisibly, with no dashboard except for diagnosis. There is no buying AI for this and that. There is only your stack. The dental office does not buy AI for dental offices. The dentist’s own AI, the hygienist’s own AI, and the front desk’s own AI do the work, and the office keeps its records somewhere those three can reach.

Nobody will buy AI for this and that. Every person will have one AI, and it will be theirs. The work is helping them set it up.

If that sounds radical, look at how a business actually works in 2026. Everyone in it, from the janitor to the CEO, carries their own smartphone. The company did not buy it, does not own it, and does not decide what is on it. Nobody would call the phone in the CEO’s pocket a business phone, even though the whole business runs through it. What the company provides is not the phone. It is the things the phone can reach: the email account, the shared calendar, the badge system, the payroll portal. The device is personal. The company is the set of records and permissions the device is allowed to touch.

That is exactly what “enterprise AI” becomes, and it is smaller and plainer than the market thinks. A company will not have an AI of its own that its people talk to, any more than it has a phone of its own that its people carry. A company will be a database, a server that its people’s AIs can call, and a governance layer that says which AI may create, read, update, or delete which record. That is the whole thing. The intelligence lives with the individual, on their bill, the way the phone does. The company keeps the ledger and the rules. The MCP servers that are already standard are the first draft of exactly this.

So here is where I would go instead, and it is not a startup. Be the person who helps individuals set up their stack: connectors, skills, plugins, scheduled tasks, one person at a time, the way there used to be someone who set up your home network or your first smartphone. Every one of the billion people in the first piece will need that at least once, and most will need it again each time their life changes. Apps die, except for a few. The one-to-one relationship between a person and their AI does not, and somebody has to stand it up. That is what we do at Charleston AI, one owner and one household at a time. It is the least glamorous job in the industry, which is exactly why it will still exist in 2030. Nobody’s plugin drives to Summerville.

Claim: the intelligence bill displaces the subscription stack rather than adding to it

Already true

The average American already spends about $111 a month across subscriptions. The market has repriced software once on the mere release of role-specific plugins. Systems-of-record companies have already accepted connector status. Grammarly has already reorganized itself into a suite to avoid being a feature.

What has to happen

The stack has to actually be good enough to cancel against. The scheduled task has to be as reliable as the app it replaces, and the person has to trust it with the same authority. Consolidation also has to run toward one or two bills rather than five: if Claude, ChatGPT, and Gemini each become a $100 line, the stack has not been consolidated, it has been rebuilt at a higher price.

Where I am probably wrong

Wrappers with real distribution survive longer than I think, because switching is annoying and most people are not me about subscriptions. A few “AI for X” companies survive by becoming the records-and-governance layer for an industry, the database every person’s own AI has to call, and those few are worth working for. And the incumbents, not the labs, might end up owning the stack: if Microsoft or Salesforce becomes the place the agents live, the $100 flows to them and my two-company forecast is wrong about the payee while right about the bill.

The rule I applied today is the rule the whole economy is about to apply. Not “is this tool good,” but “is this tool a separate relationship I need.” Increasingly the answer is no, because the relationship already exists, and it is the one that runs while you sleep.