The partnership directly tackles AI cost economics by maximizing GPU efficiency and energy utilization, a critical factor for scaling sovereign AI services across APAC.
Firmus is accelerating its presence in the Asia‑Pacific region by pairing with data‑storage specialist Vast Data. The partnership centers on deploying Vast’s AI Operating System as the backbone for Firmus’s next‑generation AI factories, including the flagship Project Southgate in Tasmania. Executives cite the need for a data foundation that can keep pace with extreme throughput and capacity demands as the company scales to thousands of GPUs. By leveraging a unified data layer, Firmus aims to eliminate the latency and inefficiency that typically bottleneck large‑scale AI workloads.
Vast’s AI OS distinguishes itself through a model‑to‑grid architecture that fuses model behavior, GPU performance, thermal dynamics, and real‑time grid conditions into a single optimization framework. This design enables the platform to react instantly to energy pricing signals and broader grid constraints, preserving GPU efficiency even in disaggregated, high‑throughput clusters. Integrated with NVIDIA’s Cloud Partner (NCP) reference design, the system aligns compute, storage, and power consumption, ensuring that each watt delivers maximum AI output. For Firmus, this translates into measurable cost savings at scale, where marginal inefficiencies quickly compound.
The timing aligns with Firmus’s recent $10 billion debt facility, providing the capital needed to broaden sovereign AI capacity across APAC markets. By offering a secure, multi‑tenant data layer, the collaboration positions Firmus to attract both anchor tenants and government‑backed projects, a strategic advantage in a region where data sovereignty and energy efficiency are increasingly regulated. As competitors race to build comparable AI factories, Firmus’s integrated stack—combining Vast’s OS, NVIDIA’s reference architecture, and robust financing—could set a new efficiency benchmark for large‑scale AI deployments.
Comments
Want to join the conversation?
Loading comments...