The complexity of AI infra should never stand in the way of human creativity.

None of it began as a plan to build data centers. Co-founder Flaviu Rădulescu had built PicFinder, an image generation app released shortly after Stable Diffusion that made image generation as fast as search, returning results in under a second while comparable apps took thirty. It reached millions of users in less than two months. His co-founder Ioana Hreninciuc had spent fifteen years in product, almost all of it in gaming, where you learn to think in margin per user and per session long before you think about hardware.

Keeping PicFinder up meant renting GPUs from every provider they could find, and identical specs delivered wildly different throughput. They traced it to CPU clock speed: providers sell GPU-hours, so nothing around the GPU gets optimized, and for inference that is exactly what matters. Rented GPUs were never going to get cheap enough to support the product they wanted, and renting meant they could not tune the parts that mattered. Hardware had stopped being a cost question and become a control question, so they built their own.

In 2023 that meant hand-assembling three liquid-cooled servers on Ioana’s kitchen table, then spending the summer swapping components and tuning the software until the same GPUs gave about 50% more performance. When a raise for the consumer app failed, the conclusion was clear: the app was not the business. The engine underneath it was.

The software platform had to exist first. Most of the first half of 2024 went into building Runware itself: the inference engine, the model catalogue and the API around them. It launched in October 2024 and scaled quickly, and the GPU bill grew with it, which settled any remaining argument about owning the hardware.

Today Runware is a trusted backbone for over 1 million developers and enterprise innovators worldwide, including global platforms like Wix and Higgsfield, processing more than 20 billion AI requests for 500 million end users across 400,000+ models. The two people who built those first servers still run the company.

Our mission promises infrastructure that runs on renewable power, consumes no water, and doesn’t draw on the resources communities already depend on. Evaporative cooling consumes around 550 billion liters of water a year globally, and that is the cost of one cooling method, not of running AI. Owning the hardware let us choose differently, and it shapes where our pods can go and what they ask of the places they go.