FuriosaAI raises $620M Series D to scale AI inference silicon

FuriosaAI, a South Korean AI chip developer, said on October 28 that it closed a $620 million Series D round. The company framed the raise around two goals: meeting surging enterprise demand for AI inference and accelerating development of its next-generation silicon.

The size of the round says something about where AI capital is moving. Spending was once concentrated on training clusters, but as models move into production the recurring bill lands on inference — every call, every step an agent takes, consumes compute.

Inference workloads behave differently from training. They are judged on cost per unit of work, power draw and latency, and they lean heavily on a software stack that can keep up with changing model architectures. That is the argument for dedicated accelerators: trade generality for a lower cost per inference.

FuriosaAI's lineup is built around that premise. RNGD is its inference accelerator card, positioned as a power-efficient option for serving models; NXT is a multi-card datacenter system. The company describes the approach as AI-native silicon designed for inference rather than general-purpose GPUs adapted from graphics work.

The competitive picture is well understood. Nvidia's grip on training is hard to challenge, but inference is widely seen as the more contestable layer: workloads are more standardized, buyers care more about total cost of ownership, and deployments range from hyperscale racks to edge facilities.

A Series D of this size shifts what FuriosaAI has to prove. Benchmark numbers matter less than volume production, stability under real customer workloads and a software stack that does not bottleneck deployments. Later-stage chip investors tend to look at wafer capacity, ecosystem maturity and whether one-off orders turn into repeat purchases.

For buyers, a credible second source of inference silicon has value on its own: it creates pricing leverage and supply redundancy. But switching costs are real, because moving workloads means rewriting kernels, retuning performance and revalidating accuracy.

Three things to watch: the timing and specifications of the next chip, how quickly RNGD lands in large-scale customer deployments, and whether the software stack can track the release cadence of mainstream open models. The round answers the ammunition question; delivery decides whether the company can hold ground outside Nvidia's shadow.