NightCafe: the bootstrapped AI-art site that beat Midjourney to market (interview)

NightCafe is an AI art platform in Cairns, Australia: it aggregates third-party image models, adds community features, and charges via credits and subscriptions.

Entirely bootstrapped — no outside investment; funded by the credit system and $4.79–$50/month subscriptions that undercut Midjourney and Civitai.

In 2019, Angus Russell — a designer and repeat startup founder newly moved into a Sydney semi-detached house with bare walls — couldn't find prints he liked online, so he built a side hustle: a marketplace to buy and sell AI-generated art, named after Van Gogh's 'The Night Café'. It failed: creation was free and fun, but nobody bought wall prints.

Then a hosting bill came in a few hundred dollars higher than usual — one user had generated thousands of images in days. Angus added a credit system to stop it; his inbox flooded with requests to buy more credits, and overnight the site broke even. Elle Russell joined to run the business side. In mid-2021, after OpenAI announced DALL-E, Angus put the open-source alternative VQGAN+CLIP on NightCafe and bought hundreds of GPUs; the images blew up on Reddit, NightCafe made $17,000 in a single day, and Angus quit his Atlassian job. Today it aggregates models from OpenAI, Google and Black Forest Labs, runs Stable Diffusion and Ideogram on its own servers, and still sells prints — as it has since 2019.

The accidental credit system was demand discovery: users asked to pay before the founder ever tried charging them.

Aggregation turns the model wars into someone else's capital expense — NightCafe ships every new model without training any.

Staying bootstrapped forced unit economics that venture-backed rivals skipped, which is why it profits 'most months' while they burn.

Pricing can be discovered by accident: a runaway user's hosting bill forces a limit, the limit becomes a product, and the users themselves tell you what to charge.

As of August 2024, NightCafe was adding video models like Stable Video Diffusion, keeping fine-tuning behind human moderation to blunt deepfake risk, and holding pattern on strategy: no enterprise offering, no outside money, with a court ruling on AI training data the main existential risk it tracks.