Three full-time employees. One 26-year-old self-taught founder who never finished his degree. An initial build measured in weeks. And today: 2.5 million people log in every single day.
Janitor AI is the AI romantic-fantasy roleplay platform you've probably scrolled past, dismissed, or quietly judged. Forbes profiled it in May 2026 and the numbers deserve your attention: 2.5 million daily active users, more than 15 million total users, roughly 100 million monthly visits, and a top-10 spot among US consumer AI apps according to Similarweb. Top AI roleplay apps pull average sessions of 78 minutes a day — right up there with TikTok.
And here's the stat that breaks everyone's mental model: 70–80% of Janitor AI's users are women.
If you're building an MVP in 2026, this story is a masterclass — not because it's clean, but because it's survivable. The company almost died at 21 days old. Its founder got scammed on a handshake deal before his first real round. Its moderation system imploded in public. And it's still standing, Khosla Ventures-backed, with a burn rate its investors describe as embarrassingly low. This is the survivor guide from 0 to 1, written the way it actually happened.
Jan Zoltkowski — "Jan" — grew up in Brisbane, taught himself engineering, dropped out of university in his first year, and went to work at crypto exchanges Bitfinex and Bittrex. Not the typical YC founder résumé. That matters, and we'll come back to it.
In early 2023 he hit the EthDenver hackathon and started playing with early GPT models. Then he became a power user of Character AI — the category king — until Character AI banned adult content and triggered what its own community still calls The Migration: a mass user exodus.
Zoltkowski was one of the leavers. So he built what he wanted to use. With AI assistance, he wrote the initial platform code in three weeks. He picked the name Janitor from the Latin ianitor — the doorkeeper who holds the keys to many gates. On June 16, 2023, he posted a single message on Reddit's Character AI community: "New website, try it out."
Days later it caught fire on BookTok — TikTok's enormous romance-reader subculture. By most counts the site crossed one million users in about three weeks (some timelines say 17 days), and roughly 3 million by September 2023.
No marketing budget. No launch video. No waitlist. One post, into exactly the right crowd, at exactly the right moment.
Survival lesson #1: Your MVP ships in weeks, not quarters. Zoltkowski's three-week build wasn't laziness — it was a bet that distribution and community would teach him what features to build. He was right. Early users openly acknowledged the bugs and stayed anyway, because the product solved a need nobody else would touch.
Look closer at that launch. Janitor AI didn't "go viral" through luck. Three deliberate, repeatable choices created the explosion:
1. He fished where the anger was. Character AI had just tightened its filters and alienated its most engaged creators. Zoltkowski launched inside that exact community, offering the thing they'd just lost. Every angry ex-user was pre-qualified demand. Most founders launch into a void. He launched into a crowd already primed to move — his product just had to be standing there with the lights on.
2. He rode a validated demand wave. The "romantasy" genre — fantasy-plus-romance — was already detonating in publishing: genre sales up 50% in 2024 per Circana, with Rebecca Yarros' Onyx Storm setting the fastest adult-fiction sales record since 2004. US romance print sales were still up 24% year-over-year in mid-2025. Zoltkowski didn't invent demand; he redirected an ocean of it into interactive form. His own framing: it's "an evolution of the amateur novelist" — millions of people who would never write a book can now co-author living, interactive stories, far faster.
3. He served the audience Silicon Valley ignores. VCs had filed AI romance under "vice" — alongside cannabis and porn — with many funds contractually barred from touching it. And a layer beneath that bias: consumer AI's romance products were assumed to be for lonely men. The data said otherwise. The biggest audience for AI romantic fantasy was women, historically the core market of the romance publishing industry itself. Mercedes Bent, the Premise VC partner who did invest, put the pattern bluntly: products serving women's emotional lives get chronically underestimated and unfairly stigmatized — the same bias that shaped early Tumblr and Reddit.
Competition stayed away because of the stigma. The bias was the moat.
Survival lesson #2: The best go-to-market in 2026 is an underserved, emotionally invested audience that bigger players are too squeamish to serve. Bias arbitrage is real, and it's available in every industry — find the customer segment your better-funded competitors consider beneath them.
Twenty-one days after launch, success nearly killed the company.
Traffic exploded, and the infrastructure bills exploded with it — including a Cloudflare bill that hit $180,000, an existential sum for a three-person team with no revenue. Into the panic slid one of the strangest characters in tech: Martin Shkreli, of pharma-price-gouging fame, who slid into Zoltkowski's Discord promising GPUs and cash for equity. The money never arrived. The deal collapsed. What actually saved Janitor AI was an angel check from Sky9 Capital.
Then came the second structural crisis, and it's the one every AI wrapper founder should tattoo on their forearm. In mid-2023, OpenAI began mass-terminating user accounts connected to Janitor AI through its API — the NSFW roleplay violated OpenAI's usage policies. Janitor AI was, at that point, a frontend renting intelligence from a provider that could — and did — pull the plug.
The company rebuilt its stack on open-source models it controlled. That forced reinvention turned out to be the foundation of its entire cost advantage. In May 2025, Khosla Ventures led its Series A.
Survival lesson #3: Model your success-case cost curve before you launch. Every founder stress-tests for nobody showing up. The killer is everybody showing up while your COGS scale linearly with usage. And if your MVP runs on rented infrastructure — an API, a marketplace, a platform's distribution — assume that tap can be revoked the day you become inconvenient. Janitor AI's wrapper phase was a legitimate MVP strategy. Staying a wrapper forever would have been suicide.
Now for the number that should reorganize how you think about AI product margins.
Janitor AI serves about 13 trillion tokens per month for roughly $130,000 in compute costs. Run that same traffic on the cheapest base model on AWS and you'd pay at least $1 million a month. On Anthropic-class frontier models, Forbes' math puts it at $50 million or more.
That's not a discount. That's a different species of business — the difference between a margin structure that can survive free access and one that dies the first month. How'd they get there?
Zoltkowski's line about burn rate — that it's "absurdly low" and should embarrass most VC-funded startups — isn't founder theater. Three people, living and working in one San Francisco loft, running a top-10 consumer AI app. The COO is 22.
Survival lesson #4: In AI, your unit economics are your product strategy. Roleplay usage is unbounded — long sessions, regenerated replies, re-reads. Every extra minute of engagement is simultaneously retention and cost. If you price like a SaaS seat while your costs scale like a utility, you've built a treadmill that speeds up the faster you run. Design cost architecture on day one, not after the invoice arrives.
Here's the part most technical founders will underweight, and it might be the single biggest reason this product retained users while better-funded rivals churned them.
Around 400,000 users sit in Janitor AI's official Discord, organized into self-forming subgroups that function as writers' rooms for specific fandoms and story universes. The most devoted creators draft thousands of words of character backstories in Google Docs before publishing a bot. They've taught themselves scripting and CSS to design profile pages. They write hundreds of pages of tutorials teaching each other the craft — unpaid, unassigned, because the platform is their creative home.
Forbes' profile follows "Cookie," a stay-at-home mom in North Carolina who discovered the platform while isolated in a new city, caring for a toddler while her husband traveled. She now co-plots characters with five other creators in a private Discord. Her words: it's her outlet, her escape, her creative practice. That's not a chatbot user. That's a community member with a body of work.
The readers became the writers. The writers became the moderators, the marketers, the onboarding department, the tutorial division. A three-person company scaled its human operations through its community's intrinsic motivation — until, as we'll see, it couldn't.
Survival lesson #5: Community does the retention work your roadmap hasn't gotten to yet. Janitor AI's code was mediocre in June 2023; its users said so publicly and stayed anyway. If your MVP has a pulse but an empty room, you don't have a retention problem, you have a belonging problem. Build the gathering place early.
Now the honest part of the survivor guide — the chapter where the model breaks.
By January 2026, the volunteer moderation system that had carried the platform caught fire in public. A moderator resigned with a scorching open letter: the volunteer team had shrunk to two or three people processing a backlog of 125,000 violation reports. Standards were inconsistent, tooling was primitive, and — most damaging — the dev team allegedly reversed bans on extreme content when popular creators appealed through back channels.
The community revolted. Zoltkowski — beloved on Discord under the handle "Shep," borderline deified by power users — had his halo cracked. The COO published an apology conceding the moderation system needed professionalizing. And Zoltkowski, to his credit, said the thing founders almost never say: part of the accusations were false, but most of the criticism stung because it was true.
The volunteer corps was scrapped entirely, replaced by professional AI-assisted and human moderation. The context makes the stakes obvious: Character AI, the category's poster child, spent 2025–2026 drowning in state lawsuits and teen-safety settlements. Janitor AI — which bans under-18 users and, per Forbes, faces no active US litigation — is trying to professionalize before its courtroom moment, not after.
There's a money trail behind all of it, too. A year earlier, Zoltkowski had published an open letter explaining why he was tightening image rules: payment processors won't touch a platform that can't keep content inside policy lines. Monetization forces governance. The freemium subscription the team is now planning has age verification, moderation capacity, and app-store compliance as prerequisites — not nice-to-haves.
Survival lesson #6: Governance debt compounds exactly like technical debt — silently, then catastrophically. Every UGC platform inherits this. The workaround that got you through month three (volunteers, manual reviews, "we'll figure out trust & safety later") is a liability with a fuse. Professionalize your trust layer before the open letter, not after it.
Strip away the salacious headlines and this is the playbook:
Mentor's honesty, before you copy-paste this playbook: Janitor AI's story isn't finished. It has never charged a dollar — the freemium tier is still ahead of it, along with age verification at scale and the fight to keep creators while adding rules they came here to escape. OpenAI is opening adult content in ChatGPT this December, which puts the world's biggest AI company directly into the arena. And the whole category lives one bad news cycle from the kind of regulatory earthquake Character AI's lawsuits are already producing.
The 0-to-1 was a masterpiece of asymmetry: tiny team, tiny burn, huge ignored demand, perfect timing. The 1-to-10 — turning a free community into a governed business — is the harder game, and it's being played right now. Watch it. The next chapter of the survivor guide is being written in real time.
Building your own 0-to-1? Take rule #1 seriously this week: pick the angry crowd, build the three-week version, and post it where they already are.
