AI inference startup Runware raises $50M to make AI run faster – News

Artificial intelligence startup Runware Ltd. wants to make high-performance inference accessible to every company and application developer after raising $50 million in an early-stage funding round.

Announced today, the Series A round is backed by some significant investors, including Dawn Capital, which led it, plus Comcast Ventures, Insight Partners, a16z Speedrun and Speedinvest. It brings Runware’s total amount raised to date to $66 million, underscoring the potential of its high-performance inference platform that specializes in real-time AI image, video and audio generation.

The startup was founded in 2023 by the Romanian developer duo Flaviu Radulescu and Ioana Hreninciuc during the early days of the AI boom. They immediately saw the potential of the first AI image and video generation tools, but they also became frustrated by how slowly they took to generate their outputs, and consequently decided to do something about it.

In a blog post, Radulescu said slow inference is a major impediment to teams trying to ship AI at large scale, as it tends to break the user experience. He also pointed to other problems, including fragmented access to the leading large language models and high costs.

“We built Runware to remove those constraints completely,” he wrote. “Our approach combines custom AI inference hardware with an optimized software stack that reaches up to ten times lower pricing and faster performance than traditional data-center deployments.”

Runware’s inference platform is powered by the Sonic Inference Engine and custom hardware. It can be integrated into any app via the company’s application programming interface, meaning that customers don’t have to worry about the underlying infrastructure of the LLMs they choose or maintain separate integrations.

The startup has carefully customized its inference infrastructure for open-source models, and provides zero-day access to new releases, meaning customers will always be able to use the latest versions of popular models. The startup sources its infrastructure from cloud providers and says it can reroute the most demanding workloads when more memory is needed to ensure they can be processed rapidly.

“On the software side, we heavily optimize model loading and offloading, which lets us support over 400,000 models and make any of them available for inference in real time,” Hreninciuc told News in an interview.

In essence, Runware can speed up inference for basically any open source model and get them to spit out their outputs in real time, and for that reason it has proved to be immensely popular. The company said it has powered more than 10 billion generations for over 200,000 developers in its first two years, supporting more than 300 million end users globally. Its customers include the AI startup Together Computer Inc., the question-and-answer forum Quora Inc., the image sharing website Freepik Inc. and the website builder Wix.com Inc.

Runware faces a number of competitors in trying to optimize AI inference, some of the most notable being Fal.ai and Replicate Inc. Fal.ai has raised far more money, recently closing on a $140 million round that valued it at $4.5 billion, but it’s much more focused on the scope of its model range rather than on inference speed.

One of the main differences between Runware and its rivals is its pricing. Fal.ai and Replicate both charge customers based on compute time, whereas Runware charges a simple cost-per-image to its customers, meaning they pay only for what they need.

Hreninciuc told News the company will use the funds from today’s round to enhance its Sonic Inference Engine and make it even faster, while expanding it to support more than 2 million AI models. In addition, it’s going to look at supporting new modalities besides image, video and audio generation, with its ultimate goal to become “the API for all AI,” she said.

Image: Runware