OpenAI는 훈련 부하를 지원하기 위해 시드니 기반 공급업체로부터 말레이시아에서 컴퓨팅 용량을 조달하고 있다.

An Australian firm is building data centres in Malaysia, equipping them with American silicon, and selling the resulting compute to another U.S. company. Its co-chief executive describes the arrangement as the moment the Asia-Pacific region transitions from consuming artificial intelligence to producing it.

Firmus announced the partnership from Sydney on Tuesday. Under a multi-year agreement, OpenAI will purchase dedicated compute from two Firmus facilities in Malaysia, establishing itself as an anchor customer and pushing total contracted capacity beyond 900 megawatts. “This multi-year partnership marks the moment Asia-Pacific becomes a producer of intelligence, not just a consumer of it,” said Tim Rosenfield, co-founder and co-chief executive.

The statement is compelling and warrants closer examination. The underlying hardware is supplied by Nvidia. Firmus will deploy Vera Rubin NVL72 systems on Nvidia’s DSX platform. The end customer is American, and the generated tokens will serve OpenAI’s global user base, as the company explicitly states. Malaysia’s contribution consists of land, grid access, and power. While economically significant, this does not constitute the production of intelligence. Instead, it represents a strategic position within a foreign supply chain—a tangible economic benefit that diverges from the broader claim being advanced. The distinctly Australian component is narrower but more substantive than the marketing slogan. Firmus manufactures modular units it calls HyperCubes, integrating liquid cooling, mechanical systems, and electrical infrastructure. These modules are prefabricated in regional New South Wales and exported, representing genuine manufacturing activity.

Malaysia’s developments unfolded over a single day. Hours before Firmus’s announcement, The Next Web reported that Malaysia is considering Huawei chips for a sovereign artificial intelligence initiative, despite Washington’s warnings regarding U.S. export controls. Within twenty-four hours, the nation was courted from opposing directions: Chinese silicon for domestic government systems, and American silicon for a U.S. commercial client, deployed on identical territory. This is not a policy contradiction; it is deliberate strategy. By hosting both factions, Malaysia can sell compute capacity to each while remaining unaligned with either—a balancing act most governments struggle to achieve. The approach has a clear precedent. Armenia’s AI facility operates following a U.S. license approval, effectively making the site an extension of American policy alongside an Armenian national asset. Malaysia appears to be sidestepping that dependency by engaging both sides.

Contracted capacity does not equal operational capacity. The 900-megawatt figure requires important context omitted from the official release. Firmus currently manages seven AI data centre projects across four nations: Australia, Singapore, Indonesia, and Malaysia. Only two are fully operational; the remaining five are under construction, with commissioning targeted over the next 24 months. Consequently, the figure reflects contractual commitments rather than physical infrastructure. While standard practice for industry financing, contracted megawatts and delivered megawatts are fundamentally different metrics. The execution gap between the two is where data centre projects typically falter. Firmus’s Indonesian operations illustrate the project’s scale. In June, The Next Web reported that the company plans to construct a 360-megawatt campus in Batam, scheduled to become operational in the first quarter of 2027. Projected revenue from offtake agreements at that site is expected to reach $25 billion to $30 billion over six years.

A separate development followed shortly after: Firmus is preparing for a public listing. The Australian Financial Review noted that the company is scheduling investor meetings for the coming week, drafting a pathfinder prospectus, and appointing banking advisors, according to sources familiar with the matter. While no specific timeline has been set for an Australian Securities Exchange debut, the publication characterized the process as accelerated. Carmeli Argana filed a related report for Bloomberg. Securing the world’s most recognizable artificial intelligence company as an anchor customer, followed by investor roadshows the subsequent week, follows standard corporate sequencing. The arrangement is entirely proper, yet it underscores a straightforward reality: the OpenAI agreement represents the most critical asset in Firmus’s upcoming prospectus.

Firmus has already secured substantial private capital. In August, The Next Web reported that the company raised $2 billion at a valuation exceeding $10.5 billion, approximately double its equity value just four months prior. Investors included Coatue, Nvidia, Blackstone Tactical Opportunities, and Jane Street. The listing would place Firmus in a crowded pipeline. Anthropic is anticipated to go public within weeks, OpenAI has indicated a potential 2027 debut, and Moonshot AI is pursuing a Hong Kong exchange listing.

Firmus originated as a Bitcoin mining operation in Tasmania in 2019. That background informs the company’s current architecture. Cryptocurrency mining trained the founders to prioritize low-cost electricity, manage thermal loads at high density, and treat energy consumption as the core operational variable. Firmus now frames its model as ‘energy in, tokens out’—the same fundamental equation, merely serving a different purchaser. It also dictates the company’s geographic strategy. Long before the term ‘AI factory’ entered mainstream usage, Firmus scouted Tasmania for stranded renewable energy resources.

Alongside the OpenAI agreement, Firmus has announced intentions to launch an Australian AI Access Program designed to support researchers and organizations operating in science, education, agriculture, energy, and climate resilience. ‘Intends’ remains the operative term. The company provided no budgetary figures, capacity allocations, or implementation timelines. At present, the initiative exists as a declared objective rather than a binding commitment. A more complex dynamic underlies the program. The compute infrastructure resides in Malaysia, the primary customer is based in California, the modular units are manufactured in New South Wales, and the company is preparing for a Sydney listing. Australia’s role is essentially that of a manufacturer supplying the equipment that constructs the data centres—a strategically sound position, albeit distinct from directly generating artificial intelligence.

Key developments to monitor include whether the five facilities under construction achieve operational status within the projected 24-month window. This timeline represents the definitive validation of the 900-megawatt commitment. Observers will also track whether the AI Access Program secures formal funding and defined compute allocations. Finally, the actual financial terms of the OpenAI agreement remain undisclosed by both parties. Until pricing is revealed, the transaction’s significance will be measured in megawatts rather than revenue.