Meta Books Nvidia AI Factory Capacity With Firmus
Meta has signed agreements for GPU computing capacity at Firmus AI factories in Southeast Asia built on Nvidia’s DSX platform. The Australian infrastructure company says the arrangements cover contracted capacity and options to expand, supporting Meta’s AI research, model development and training. They extend a relationship in which Meta already uses Nvidia GB300 NVL72 systems at Firmus’s Melbourne facility, described by the partners as the Southern Hemisphere’s largest Blackwell Ultra installation. Firmus is also an Nvidia investment and strategic infrastructure partner. The named buyer and the choice of a Nvidia-based factory design give this agreement a clearer route to Nvidia systems demand than a generic regional data center plan. Firmus and Meta did not disclose the capacity or financial terms of the new agreements.
Firmus designs, manufactures and operates AI factories with computing, cooling and power engineered together. Its HyperCube modules, co-designed with Nvidia, put liquid cooling, mechanical equipment and electrical systems into factory-built units. The current generation supports GB300 and Vera Rubin NVL72 systems along with Spectrum-X networking. Those details matter because Meta is contracting for large-scale research and training capacity, where the entire rack, network and facility design influence performance and electricity use. Gaya Nagarajan, Meta’s vice president for engineering infrastructure, called the regional buildout a “foundation of personal superintelligence.” The agreement therefore links an established hyperscale customer to a regional operator with a design already aligned to multiple generations of Nvidia equipment.
The two companies have different roles in the transaction. Meta is buying computing capacity from Firmus, while Firmus develops and operates the factories; Nvidia supplies technology and holds an economic interest in the operator. Nvidia’s direct product opportunity arises as Firmus equips sites using the DSX stack, including GPUs, networking and system designs. Any benefit to the value of Nvidia’s Firmus investment is a separate channel. The distinction is particularly relevant as Firmus prepares an Australian flotation intended to raise about $5 billion for further AI projects, according to Reuters. A publicly traded Firmus would make the value and financing needs of this infrastructure model more visible, while customer contracts provide a stronger basis for raising capital than speculative construction alone.
Firmus already had a large Asia-Pacific development program before this Meta expansion. In September it announced a multiyear OpenAI agreement for dedicated capacity at two Malaysian AI factories and said its contracted portfolio exceeded 900 megawatts across customers. It described seven sites in Australia, Singapore, Indonesia and Malaysia, with two then operating and five in development. A separate Nvidia partnership covers a 360-megawatt DSX campus in Batam, Indonesia, with plans for as many as 170,000 Nvidia accelerators across several architectures through 2027 and 2028. The Meta agreement adds a second leading model developer to Firmus’s regional customer story. Its new Southeast Asian facilities have a commercial use in view, rather than relying only on a general expectation of future AI demand.
Firmus’s Melbourne installation gives Meta a working reference for the expansion into Southeast Asia. The new arrangements could place Nvidia systems in several more factories as Firmus uses repeatable HyperCube modules and the DSX stack. For Meta, the capacity purchase diversifies the geography and operating counterparties supporting its model programs without requiring it to own each building. For Firmus, a repeat hyperscale buyer can support project finance and improve the credibility of its planned flotation. Nvidia may benefit from systems and networking orders as the operator equips sites, and from its stake if Firmus earns an attractive return on those facilities. The economics hinge on the capacity Meta takes and the cost of building and operating it.
Analysis
Meta adds another leading model developer to Firmus’s regional customer base alongside an earlier OpenAI agreement, improving the case for financing repeatable DSX factories. Nvidia has two ways to participate: systems and networking as Firmus equips sites, and the value of its stake if the operator earns attractive returns on the capacity Meta buys. HyperCube’s standardized cooling and electrical design may shorten successive builds, but Meta’s expansion options let the customer scale purchases with its own model needs. That preserves Meta’s flexibility while placing the execution and utilization burden on Firmus, whose economics determine the strength of Nvidia’s equity channel.