South Korean startup Furiosa AI has officially confirmed ambitious plans to significantly expand the production of its specialized artificial intelligence processors. The company intends to increase the output of its second-generation Renegade (RNGD) chips from the current twenty thousand units this year to forty to fifty thousand units as early as next year. This step more than doubles current volumes and marks the first official confirmation of the company's production plans for the upcoming period. Furiosa AI positions itself as one of the leading domestic AI semiconductor manufacturers in South Korea, and the decision to scale is explained by the rapid and steady growth in demand for efficient AI inference. In particular, the company is actively developing a third-generation accelerator named Stork for the RNGD platform, built on advanced two-nanometer technology from Broadcom. Developers assure that the new product will be able to pose a serious challenge to NVIDIA's dominant solutions thanks to outstanding energy efficiency and the minimal cost of processing a single token. Among the already confirmed clients of the startup are technological giants such as Samsung SDS and LG AI Research, which further underscores the high level of trust in the company's products in the domestic market.

The architecture of the new generation of Furiosa AI chips is entirely focused on the AI inference segment, which continues to gain enormous popularity against the backdrop of the widespread adoption of agentic artificial intelligence. The Stork accelerator is designed to handle the most demanding workloads and ensure high-speed data movement within computing clusters. The platform utilizes an advanced compute die manufactured using a two-nanometer process and supports the latest HBM4/E memory standard. In a teaser image, the company demonstrated a chip configuration with twelve HBM4/E memory blocks, two massive compute chiplets, and two input/output controllers. When using twelve-layer modules of thirty-six gigabytes per stack, the total memory capacity reaches an impressive four hundred and thirty-two gigabytes. In addition to the compute architecture, Furiosa AI integrates Ethernet and PCIe blocks from Broadcom, which guarantees higher bandwidth and reliable rack-scale networking within massive AI computing clusters.

A key competitive advantage of the new solution is its optimization for real-world AI workloads, such as post-training model sampling, where high throughput becomes an absolute priority. Furiosa AI states that focusing specifically on throughput, rather than the complex stream management traditionally required by graphics processing units, will allow for unprecedented efficiency and greater token generation density. To simplify deployment, the company's software stack includes a universal compiler capable of automatically converting high-level PyTorch framework code into machine instructions for the chip. For developers needing finer control, a Virtual Instruction Set Architecture (Virtual ISA) is available, offering a declarative programming model without the nondeterministic complexity of traditional GPU coding. Co-founder and head of Furiosa, Jun Pyoek, emphasized that combining Broadcom's infrastructure capabilities with Furiosa's Tensor Contraction Processor architecture and its standard-setting software stack allows them to offer a comprehensive solution for the emerging era of "token factories." The first samples of the third-generation accelerator are expected to be delivered to clients in the first half of 2028.