OpenAI Anchors Firmus’ Malaysian AI Factories: AI Infrastructure Enters the Power–Chip–Facility Integration Phase

Australian AI infrastructure company Firmus announced on September 8, 2026 that it had entered a multi-year strategic partnership with OpenAI. OpenAI will contract dedicated AI compute capacity from two Firmus AI Factory sites in Malaysia and will become an anchor customer. Firmus also said that its contracted capacity across all customers now exceeds 900 MW.

This is not a public purchase order with disclosed pricing, GPU counts, or a confirmed service date. Firmus did not disclose the contract value, and Reuters reported that OpenAI had not commented on the value. The defensible reading is that OpenAI has secured future regional capacity while Firmus has gained a customer anchor for construction and financing. The 900 MW figure should not be presented as OpenAI’s allocated capacity or as compute already delivered and running.

What happened?

• Scope: OpenAI will procure dedicated computing capacity from two Firmus data-center sites in Malaysia.
• Regional footprint: Firmus says its AI Factory platform spans Australia, Singapore, Indonesia, and Malaysia across seven sites. Two are operational and five are under development, with service targeted within 24 months.
• Hardware and systems: Firmus plans to deploy NVIDIA Vera Rubin at scale in Asia-Pacific and use the NVIDIA DSX AI Factory Platform to integrate Vera Rubin NVL72 systems with its HyperCube liquid-cooling, mechanical, and electrical design.
• Commercial status: “More than 900 MW” is Firmus’ contracted capacity across all customers. It is not the capacity of this OpenAI contract and is not the same as delivered effective compute.

Why does it matter?

1. OpenAI’s constraint is extending from model launches to regional supply

Model capability is only one part of an AI service. Reliable availability also depends on power, cooling, networking, GPU delivery, facility construction, and cross-border compliance. The Malaysian expansion shows Southeast Asia becoming a candidate region for AI training and inference infrastructure, not merely a data-center real-estate market.

2. AI clouds will compete on token economics

Firmus presents power, cooling, accelerated computing, networking, and software as one AI Factory, with lower token-production cost as a stated goal. For developers and small teams, model selection will increasingly require more than comparing API prices. Latency, concurrency, regional availability, failover, and cost per successful task will matter too.

3. Where can smaller teams participate?

• Compute selection services: compare the cost of a completed business task across regions, models, and providers—not just leaderboard scores.
• Inference cost observability: track tokens, latency, failure rate, and retries, then route workloads to the cheapest reliable model.
• Regional compliance deployment: provide data residency, logging, access control, and model-routing workflows for Southeast Asian businesses.
• Cooling and efficiency software: monitor GPU utilization, rack power, and workload scheduling.

Three signals to watch

1. Whether the two Malaysian sites disclose locations, capacity, and service dates;
2. Whether OpenAI discloses the actual contracted scale, service region, or products using the capacity;
3. Whether Vera Rubin NVL72 delivery, PUE, network connectivity, and real token-cost data become verifiable.

Bottom line

This is a capacity-locking and infrastructure-construction story, not evidence of delivered compute revenue. The confirmed facts are Firmus’ announcement of a multi-year OpenAI partnership, two Malaysian sites, and more than 900 MW of total contracted customer capacity. The contract value, OpenAI’s specific MW allocation, GPU count, and service date remain undisclosed. For AI entrepreneurs, the opportunity is less about hoarding GPUs and more about measurable services around compute cost, availability, regional compliance, and multi-cloud routing.

Sources

• Firmus official announcement
• Reuters via Investing.com
• Data Center Dynamics