An Australian data center operator and Nvidia have agreed to build a 360-megawatt AI computing campus on the Indonesian island of Batam, deploying as many as 170,000 graphics processors in a project valued at up to $30 billion in expected offtake.
The deal between Firmus Technologies and Nvidia is one of the largest AI infrastructure commitments disclosed this year. It would deliver liquid-cooled GPU clusters capable of supporting large-scale AI training and inference, with initial deployments scheduled to begin in the first quarter of 2027 and continue into early 2028.
Firmus, which funded its GPU purchases through a combination of bonds and loans, said the partnership would let smaller AI startups and researchers access computing power that has become increasingly concentrated among a handful of cloud providers. The campus is intended to serve as both a regional hub for Southeast Asia and a test case for alternative models of AI infrastructure ownership.
The Vera Rubin GPU architecture, Nvidia's next-generation chip expected to debut in 2026, would feature the company's first custom central processor design and deliver substantially higher token throughput than current systems, according to people familiar with the planning. The Batam facility would be among the first deployments of that architecture at full scale.
Global demand for AI capacity has strained power grids and semiconductor supply chains, prompting technology giants and sovereign wealth funds to compete for access to land, electricity, and cooling. By locating the campus in Indonesia, the partners are betting that Southeast Asia will become a major market for AI services even as the United States and Europe expand their own infrastructure.
The agreement still requires regulatory approvals and final financing arrangements. If completed on schedule, it would rank among the largest single-site AI deployments outside North America and underscore how quickly the business of AI computing is globalizing beyond traditional technology hubs.
Image source: v3b.fal.media