The $150K week that started it all
In 2022, Pieter Levels was fine-tuning Stable Diffusion to generate better interior designs — a side experiment that became Interior AI. For fun, he uploaded photos of himself and trained the model on his own face. It worked. The model learned him as a concept, and suddenly he could render himself as a medieval warrior, an astronaut, anything. As he later told the Lex Fridman Podcast: this is another startup.
He launched avatarai.me. It went viral immediately — roughly $150K in a single week, the most money he had ever made in that span. Then came the pattern every indie hacker dreads: Lensa, a VC-backed company far better at iOS than a solo PHP developer, cloned the concept and rode it to roughly $30 million, becoming a top-grossing app. Levels' reaction on the podcast was disarming: he thought it was amazing, honestly — the technology was open, and someone was going to build it.
The clone didn't kill the idea. Levels kept building on the same insight — people will pay for photorealistic AI images of themselves — and avatarai.me eventually pivoted into Photo AI: an AI photographer you train on your selfies, generating professional headshots, couple photos, and social content on an ongoing basis.
One podcast, 3x revenue
On August 20, 2024, Levels appeared on the Lex Fridman Podcast (#440) and told his entire life story. A month later he published the numbers on levels.io: a new $420,000/month revenue record across his portfolio — Photo AI at $161K/mo, his book MAKE at $93K/mo, Nomad List at $61K/mo, Interior AI at $43K/mo, merch and X at $34K/mo, RemoteOK at $29K/mo. The podcast had lifted total revenue 2.5x; Photo AI alone jumped 3x.
He was characteristically honest about what the record meant: it had already dropped way down, but it was a nice record for like one day a few days prior. The spike was real money, but it was a spike — and that was the point. It demonstrated what a single deep media appearance could do versus traditional press: the power of podcasts versus traditional media, because comparably, going on national TV gets him about three signups.
The cost structure is the other half of the story. At roughly 80% profit, his biggest expense by far was GPU compute — about $60K a month for Photo AI and Interior AI combined. Everything else was, in his words, quite cheap: servers at $500 a month and Cloudflare. One person, one laptop, $420K months.
70 projects, 4 winners
None of this was overnight. In 2014, broke and living cheap in Asia on about $500 a month, Levels set himself a public challenge: 12 startups in 12 months. Nomad List — born from loneliness, a spreadsheet of good cities for remote workers — became the first real hit. RemoteOK followed. By his own accounting in 2021, he had shipped more than 70 projects and only four ever made money.
The method never changed: build in public on levels.io, ship fast, let the audience decide. Photo AI is what happens when that method meets a technology wave — the same fine-tuning experiments he blogged about became the product. An independent Indie Hackers case study tracked Photo AI at $132K MRR in November 2025, still built solo with zero employees.
A 2026 documentary tallying his public revenue posts put the one-man business at about $3.1 million a year in 2025 — and the playbook keeps evolving. He later vibe-coded a 3D flight simulator, fly.pieter.com, that crossed $1M in 17 days after an Elon Musk retweet. The throughline across a decade is embarrassingly simple: ship the experiment, publish the numbers, repeat.
The moat is the audience
Photo AI's real defensibility was never the model — Stable Diffusion was open, and Lensa proved anyone could copy the trick. The moat was distribution: a decade of building in public had given Levels an audience of hundreds of thousands who trust his launches, plus the SEO gravity of years of tutorials and revenue posts.
The product also matured past novelty. Where avatarai.me sold cheesy one-off avatar packs, Photo AI became a standing AI photographer — headshots for LinkedIn, couple photos, ongoing social content — sold as a subscription. That is what turns a viral week into $132K MRR: recurring revenue from a use case people return to, not a filter they try once.
And the cost discipline compounds the advantage. A VC-backed competitor needs to justify its burn; Levels needs $500 a month in servers plus GPUs. At 80% margins, he can outlast hype cycles and keep the product alive through the quiet months — which is exactly what happened between the 2022 avatar craze and the 2024 revenue record.
There is a final, quieter advantage: time. Of the 70-plus projects Levels shipped, only four ever made money — most died. Photo AI survived the post-viral trough because its owner could afford patience: no investors demanding growth, no payroll forcing a shutdown. The avatar craze of late 2022 faded; the product kept improving; and when the Lex Fridman audience arrived in 2024, a mature product was waiting to catch the wave. Most startups die in the trough between hype cycles. A solo founder with 80% margins simply waits it out.
The playbook
- Ship the experiment, not the plan — avatarai.me came from fine-tuning his own face for fun. The billion-dollar insight arrived disguised as a joke.
- Build in public for a decade — levels.io posts compound into an audience; every launch starts with distribution most founders would pay millions for.
- Run a portfolio, not a company — 70+ projects shipped, 4 made money. Photo AI is one winner among many — the portfolio is the strategy.
- Podcasts beat TV for indie products — One Lex Fridman appearance: 2.5x revenue. National TV: about three signups. Depth of attention wins.
- Keep the cost floor near zero — $500/mo servers, GPUs as the only real cost. At 80% margins you can outlast every hype cycle.