Meta has agreed to rent AI chip capacity at data centers in Southeast Asia operated by Australian builder Firmus. Meta has reserved a certain amount of capacity with the option to add more if needed and will use the chips to research, build and train its AI models. The arrangement extends an existing cooperation in Australia and secures Meta additional regional compute for the next generation of AI across Asia-Pacific.
Reserved capacity and factory-built units
The companies did not disclose financial details and did not reveal which sites Meta will use. The structure is a reservation of capacity with room for expansion rather than a fixed one-off purchase. Firmus currently has seven AI sites in Australia, Singapore, Indonesia and Malaysia, and two of them are up and running. That footprint makes site selection important, since power availability, cooling and delivery dates differ by location, while Meta retains flexibility to scale.
The new sites will feature Nvidia DSX design inside Firmus HyperCube units. Each unit is made in a factory in rural New South Wales and comes with its own cooling and power equipment. The latest model can also support Nvidia upcoming Vera Rubin chips. In Melbourne, Meta already uses Nvidia GB300 systems, which Firmus describes as the largest group of Nvidia Blackwell Ultra chips south of the equator.
The timing is linked to Firmus market plans. The agreement was announced a week before Firmus starts the bookbuild for its share market debut on 6 October. The company plans to raise about $5 billion, which would make it Australia second-largest listing. Earlier this month, OpenAI became an anchor customer at two Malaysian sites, lifting Firmus contracted capacity above 900MW, while in June Firmus agreed with Nvidia to build a 360MW site on the Indonesian island of Batam with room for up to 170,000 Nvidia chips.
What this means for companies using AI
For businesses that buy or build on AI, the direct effect is regional availability of high-end compute rather than a new product. When Meta and OpenAI reserve large blocks from the same operator, remaining capacity, delivery slots and engineering attention become tighter for others. Large companies can respond by reserving capacity early and negotiating expansion options similar to Meta. Smaller firms are more likely to feel the change through cloud or colocation providers, in lead times, contract terms and the regions where Blackwell-class systems are actually available.
The announcement leaves several points that buyers need to verify before drawing conclusions. Price, volume and site choice were not disclosed, so the news alone does not show cost per workload or time to deployment. Only two of the seven Firmus sites are operating, which makes construction schedules, power delivery and acceptance criteria central to any contract. Gaya Nagarajan, Meta vice president of engineering for infrastructure, called Firmus a long-term strategic infrastructure partner, but that statement describes intent rather than proven performance at the new locations.
Confirmation will come from the 6 October bookbuild and what follows it: the confirmed fundraising amount, construction progress at the Malaysian and Batam sites, and whether Meta converts its expansion option into extra reservations. If contracted capacity keeps growing beyond 900MW and new sites move into operation, the model of factory-built units with dedicated power and cooling will look confirmed for business planning.
