As media attention focuses on the performance of foundation models, more and more startups are focusing on the issue of inference, namely the actual execution of models in production, at scale, in applications used daily by enterprises and millions of users, and of course its cost.
As models stabilize and become standardized, value shifts towards their operational exploitation. For businesses, it’s no longer just about what a model can do, but how much each query costs, how quickly it runs, and under what conditions it can be integrated into existing products.
A specialization in media generation
Runware positions itself as an inference platform specializing in generative media. Images, creative content, marketing or social uses. Faced with hyperscalers who offer general solutions, the British startup claims an approach built around the specific constraints of media generation. It provides developers with a single API to integrate visual content generation capabilities, without having to manage graphics models or the corresponding compute infrastructure.
This API-first approach relies on a proprietary inference engine, designed to optimize both performance and costs. An economic trade-off that is now central while inference represents the main expenditure item for AI applications. Media generation imposes high requirements in terms of latency and stability, which are not compatible with architectures that are too generic or insufficiently optimized.
Established customers and rapid adoption
Since its launch in 2023, Runware has counted among its clients established content and creation platforms such as Wix or Quora, as well as players more specialized in AI-assisted creation. According to Citi Research, the volume of tokens processed during the training and especially inference phases could increase by more than 400 over the next five years. Growth driven less by research than by the multiplication of concrete uses. Media generation is among the main drivers of this explosion, both for content platforms and for marketing and creative tools.
A competitive field that is becoming more dense
The competitive field of inference is quickly structured around several actor profiles, with distinct but partially convergent positions.
Fireworks AI is establishing itself as one of the most visible players in high-performance inference, with a promise focused on execution speed and cost optimization for open source and proprietary foundation models.
Together AI occupies an intermediate position between research and production. Very present in the open source ecosystem, the company offers a shared training and inference infrastructure, relying on hardware partnerships and strong proximity with research communities. However, its approach remains more general, less specialized in specific uses such as media generation.
Modal adopts a different logic, more geared towards the orchestration and execution of complex AI workloads. The platform is aimed at advanced technical teams, capable of composing their own inference pipelines, but it requires a finer mastery of the underlying architecture.
Baseten, for its part, explicitly targets the putting into production of models with high requirements in terms of performance and reliability. Here again, the positioning remains broad, covering several types of models and uses.
Finally, hyperscalers remain structural competitors. AWS, Google Cloud, and Microsoft Azure offer integrated, powerful, and deeply interconnected inference services to their cloud ecosystems. Their strength lies in the depth of infrastructure, but this generalization can become a hindrance for specific use cases such as media generation, where economic trade-offs and latency require more advanced optimizations.
Runware was created in 2023 by Flaviu R. and Ioana Hreninciuc, and has just raised $50 million in Series A, or approximately €42.5 million, in a round led by Dawn Capital, with the participation of Speedinvest and Comcast Ventures, as well as existing investors including Insight Partners, a16z speedrun, Zero Prime Ventures and Begin Capital.
