Firmus picks VAST Data OS to power APAC AI factories
Firmus has signed an agreement with VAST Data for VAST's AI Operating System, a move the Australian AI infrastructure provider says will support its planned expansion of sovereign and energy-efficient AI data centres across Asia Pacific.
The arrangement focuses on the data layer inside Firmus facilities being designed for large-scale AI training and inference. Firmus operates across Australia and other regional markets, including Singapore, and is building new sites it describes as "AI factories".
Firmus is also an NVIDIA Cloud Partner, and its broader infrastructure strategy follows the NVIDIA Cloud Partner reference design. It describes the VAST software as a foundational component within that design, alongside large GPU deployments.
Regional build-out
Firmus has positioned Australia as a core market in its regional footprint, while also pointing to Singapore and other Asia-Pacific hubs as part of its expansion plan. The facilities are intended for public-sector and regulated workloads that require local control of data.
A key project is Southgate in Tasmania, which Firmus has described as a multi-billion-dollar facility. It says the site will be renewable powered and designed for sovereign AI workloads, with a focus on operating under energy constraints.
Firmus says its existing deployments run at the scale of thousands of GPUs. Its roadmap targets tens of thousands and, over time, hundreds of thousands of GPUs across the region. Those targets imply significant growth in storage and networking needs, as well as the ability to move large datasets through training and inference pipelines.
Data layer focus
The deal highlights the role of data infrastructure in AI systems built around large GPU clusters. As GPU fleets scale, the system must supply data fast enough to keep accelerators busy. It also needs controls for multi-tenant use, where different customers and agencies share the same underlying platform.
Firmus said it selected VAST because it needed a high-throughput, high-capacity data foundation. It also cited the ability to unify data access across environments and support secure multi-tenant workloads.
VAST positions its product as an operating system for AI infrastructure. In this deployment, the software will provide the data layer for Firmus sites, with a focus on sustaining GPU efficiency across large, disaggregated clusters as models and datasets grow.
Firmus CTO Daniel Kearney linked the choice to how inefficiencies scale in large GPU environments.
"The data layer has to scale in lockstep with compute and energy. At the scale of thousands of GPUs, small inefficiencies compound quickly. We selected the VAST AI Operating System because it is architected for that reality: high throughput, disaggregated, aligned with NVIDIA Cloud Partner reference design environments, and built to sustain GPU efficiency as we expand sovereign AI capacity across Asia," said Daniel Kearney, Chief Technology Officer, Firmus.
Sovereign requirements
Sovereign AI has become a central theme for infrastructure providers and governments in the region. The term is often used to describe AI systems where data residency, operational control, and compliance requirements keep sensitive data and workloads within national borders.
Firmus said its in-country deployments are designed for public-sector and regulated workloads. It also said the projects align with regional energy and sustainability priorities, an increasingly important constraint as electricity prices and grid capacity shape data centre development.
Alongside sovereignty, Firmus is positioning energy use as a core design parameter. It describes its facilities as vertically integrated systems that consider compute, cooling, networking, and data movement together. It has also outlined a "Model-to-Grid" architecture that factors in model behaviour, GPU performance, thermal dynamics, and grid conditions.
Kearney described the approach in a separate statement focused on responsiveness to grid conditions.
"Firmus was founded on the belief that AI infrastructure must be engineered as a single, vertically integrated system. Our Model-to-Grid architecture integrates model behaviour, GPU performance, thermal dynamics, and grid conditions into one optimisation framework, making our AI factories both model-aware and grid-aware, with real-time responsiveness to energy pricing and broader grid signals," said Kearney.
Supplier perspective
For VAST Data, the deal reflects a broader focus on storage and data movement as limiting factors for AI systems. Suppliers are working to reduce bottlenecks as GPU deployments grow, with an emphasis on keeping systems balanced as clusters become larger and more distributed.
Jeff Denworth, co-founder at VAST Data, said the economics of large AI deployments depend on the relationship between energy consumption and data flow.
"At this scale, eliminating inefficiency isn't a rounding error, it's the business model. When you're designing AI factories to operate across tens or hundreds of thousands of GPUs, energy, data movement, and compute efficiency become inseparable. Firmus understands that the economics of AI are defined by how efficiently data flows through the system per watt of power consumed. The VAST AI OS was built for exactly that reality: sustaining GPU efficiency, eliminating bottlenecks, and enabling large, disaggregated AI systems to operate as a single, coherent platform," said Jeff Denworth, co-founder, VAST Data.
Firmus said VAST will underpin the data layer across its regional footprint as it adds AI capacity and shifts workloads between training and inference across its facilities.