South Korea's FuriosaAI closed a $620 million Series D round on 28 October at a post-money valuation of about $2.9 billion, four times its previous level. The money will fund mass production and international rollout of its RNGD accelerator, which is designed for running models rather than training them. The company is also developing a third-generation chip with Broadcom. Its main argument is lower power consumption in data centres.

What it means: The market for AI processors is gradually splitting in two: model training, where Nvidia dominates, and running models, known as inference, where many new players are arriving with specialised chips. Inference, meaning every time a user asks a chatbot a question or a system analyses documents, now accounts for most of the computing power consumed. There, electricity costs matter more than raw speed, which is why companies such as FuriosaAI emphasise energy efficiency. For Europe and Greece, where energy is expensive and grid capacity is limited, this trend matters for anyone planning data centres or AI services. More suppliers should mean, in the long run, more price competition and less dependence on a single manufacturer. However, new chips often come with less mature software, and moving away from Nvidia's tools is not always simple. For a business that rents computing power, the practical question is whether its cloud provider offers such alternatives at a lower price per request.

Read the original on FuriosaAI

Related service: Custom software

Related stories