Skip to content
MakerGain
← Back to project

He Could Not Figure Out the GPU and Hired Someone: Photo AI Self-Reports \$105,000 in Monthly Revenue

Edited by MakerGain

Photo AI lets users upload selfies, train a personal model, and then generate photos and videos with sound by scene; Pieter Levels says Python and GPUs were too hard, so he hired someone to cover the model layer. The site lists plans from $9/month billed annually, with generations charged in points; by his own account, monthly revenue is $105,000, and he cut Photo AI's monthly GPU bill from $36,000 to $7,000.

Photo AI reduces taking a photo that's good enough to post to a few actions: upload a few selfies, let the system learn your face, then pick a scene — old money, nightlife, street style, wedding — and generate a photo with you in it that you never posed for (Photo AI FAQ[1]). It says it can change the scene, clothing, pose, and expression, and even make you speak to the camera in your own voice.

Pieter Levels, who makes the product, admitted on a podcast that the bottom layer of his AI startup — the model — required him to get help. He said Python and GPUs were too hard, so he hired an AI developer to handle the model; he tried a lot himself but just couldn't get it, and merely understanding the tooling would take six months: “That shows my ceiling.” He used to think he could do everything alone, but in that interview he saw his limits (The Bootstrapped Founder[2]).

From signing up to getting an image, users don't have to do much. After registering and entering Studio, they can apply shoot preset templates, write their own prompts, or use someone else's photo as a reference (FAQ[1]). Scenes listed on the homepage include old money, nightlife, street style, luxury, music festival, founder, and wedding (Photo AI[3]). The page displays a cumulative figure: 31,106,552 photos. The page doesn't say what date that count starts from, whether it includes videos, or whether it covers all users; it shows that production volume is large, but not who stays or how often they come back.

Image: source Photo AI · Original AI portrait sample from the official website[4], retrieved 2026-11-08; original image content preserved.

Payment comes through subscriptions and also point top-ups: plans start at $9/month, are billed annually, and can be canceled at any time (Photo AI[3]). Generations are deducted in points: one Hyper Realism photo costs 2 points, and a video with sound costs 100 points — fifty times the former (Points FAQ[5]). Members receive points each month, subscription points expire monthly, and extra purchased points expire after two years (same source).

Compare that with the task of “taking a usable professional headshot”: Aragon's pricing page describes delivery in as fast as 30 minutes, up to 100 photos, 20-plus outfits, and a choice of background; in the same table, hiring a photographer takes 2–3 business days, yields 5–10 photos per person, 1–2 outfits, and 1 background (Aragon pricing page[6]). Photo AI deducts subscription points per generation; Aragon's page does not list specific plan prices, but it spells out the refund terms: if you haven't downloaded and don't plan to download, you can get a full refund; if you have downloaded, contact customer service to handle it case by case, and a free redo is offered (same source). The two purchase models differ: the former buys generation points that expire, while the latter puts the number of delivered photos and refund terms up front.

It also connects its entry point into other AI apps: Photo AI can connect to ChatGPT, Claude, Cursor, and other apps that support MCP, and users can say in a conversation, “Give me four photos in Paris,” with the results appearing directly in the chat (Photo AI[3]). The page doesn't say how much account data an app can read after authorization; here we can only say “authorized connection.”

The GPU and other model bills Pieter disclosed land on Photo AI's books; revenue comes from subscriptions and points. Who actually pays — the person, a team, or someone paying on their behalf — is not explained on the public page; Photo AI's customer count, paid-user distribution, and refund rate are also not disclosed. A more practical point before paying is this: its own refund terms are not clearly stated on the homepage or in either of the two FAQs, on workflows and points — a gap buyers will encounter.

Pieter also changed his habits around customer acquisition. He said he used to rely mainly on organic traffic and SEO, was learning TikTok marketing at the time, and had hired someone to help. He recalled that an influencer posted content about photoai.com, and monthly recurring revenue (MRR) rose from 12k to about 40 or 50k, then stayed at that level (The Bootstrapped Founder[2]). The interview transcript doesn't specify the currency, give the date it was circulated or the promotion spend, or clarify that the influencer was a TikTok creator. Another remark of his is clear in direction: media coverage almost never brings link clicks and is far less effective than influencer spread (same source). The product also has a referral mechanism: after logging in, users can get a referral link, and after someone signs up through the link, both sides get free points (Referral FAQ[7]). The materials don't say how many signups and how much revenue referrals bring.

His description of the difficulties is more direct: AI startups have high user churn and lack a moat; everyone uses the same set of models and technology, and a smart person can clone a similar product in 3 months, or a slower one in 6 months to 1 year; GPU costs will quickly squeeze margins (The Bootstrapped Founder[2]).

This isn't empty talk. In another post he wrote out the bills: Photo AI's GPU spending fell from $36,000/month to $7,000/month by switching to a cheaper video model, moving some image tasks to Google Cloud where he had credits, and handing some tasks to Sume, which sells models at market prices; together, those items saved about $29,000/month. The monthly Cloudflare R2 bill fell from $1,200 to $800 by having photos automatically move to the cheaper Infrequent Access tier after 90 days (levels.io[8]). The same post says that adding Photo AI's $7,000 GPU and $800 R2 costs to the expenses of his other websites/apps brings total business spending to about $8,000/month; the “$300/month” in the title excludes Photo AI. The post doesn't break down how much each of the three items — switching the video model, moving to Google Cloud, and handing work to Sume — contributed to the GPU bill, or what task volume was during the same period.

He put revenue into a post: photoai.com has $105,000 in monthly revenue and $80,000 in monthly profit; the post doesn't give the settlement month, a definition of profit, or a cost breakdown (levels.io[9]). These two figures and the GPU and R2 bills above don't add up to the same income statement, and how profit is calculated remains unclear.

Put the podcast and the bills together: he got someone to fill in the model layer, while the costs still sit on Photo AI's books. He estimates that cloners will catch up in 3 months to 1 year; the GPU bill has been brought down, but the moat question still has no answer.

Continue reading

AI video
Free video ideas, finished videos by the minute — how Subscribr runs a YouTube workflow

Subscribr handles the long-form YouTube workflow from ideas and scripts to thumbnails and finished videos. Video ideas and scheduling are free, and finished video is priced by the minute. Founder Gil Hildebrand left a funded startup and bootstrapped Subscribr, which launched in 2024. He sold lifetime licenses before the product was built, with delivery promised within 60 days and refunds available through two weeks after delivery. The tool later grew from ideation and scripting to the full production workflow.

AI video
Zero monthly fees, with pricing by the second: VOSU unbundles AI video, then asks who will deliver it

VOSU aims to take a campaign from concept to an editable workflow: users describe the requirement, and agents plan concepts, build prompts, select models and generate variants in bulk. Its three full-time creator founders chose pay-per-use billing and non-expiring balances, while betting growth on a marketplace connecting brands with AI creative talent.

AI images
A thumbnail non-designers can make: AIThumbnail turns “style” and “face” into two upload boxes

Sacha Dumay got stuck on thumbnails while running a YouTube channel. He described himself as having little marketing and design experience but stronger backend and AI-development skills, so he built AIThumbnail to turn non-designers' video ideas, faces and preferred thumbnail styles into images. Friends gave the first version good but brutal feedback, and he rebuilt it. He upgraded on the day OpenAI released image-1, then reached Product Hunt's daily Top 8 with zero spending on paid marketing. Payment notifications convinced him to commit fully. In that interview, he identified steady traffic as his challenge and described trying multiple channels.

AI video
Break down the script before opening AI: how Layer8 sells films

Layer8 is an AI film studio co-founded by Ali, Kevin, and James. It breaks down scripts before opening AI tools, with directors, writers, artists, and AI experts working together while filmmakers retain visual judgment. After the team sent early samples to contacts, California Highway Patrol became its first paying client. An earlier attempt at templated ads failed to take off, and a trademark dispute prompted the Layer8 name. Agencies, production companies, and brands buy finished films.