OpenAI just locked Southeast Asia into Nvidia's infrastructure playbook

Technicians monitoring liquid-cooled GPU racks in Malaysian AI data centre facility — Technicians monitor liquid-cooled GPU racks in a Malaysian AI data centre | Illustration: Indoneo

OpenAI and Australian infrastructure operator Firmus announced a multi-year partnership on September 8, 2026, giving OpenAI dedicated AI computing capacity from two Firmus data centre sites being developed in Malaysia. The deal makes OpenAI an anchor customer and lifts Firmus’s total contracted capacity across all customers above 900 megawatts.

It marks Firmus’s first entry into Malaysia, expanding its APAC footprint beyond Australia, Singapore, and Indonesia. The agreement runs on NVIDIA hardware, underscoring the chipmaker’s deepening role in the region’s AI infrastructure buildout.

The deal between Firmus and OpenAI—making the AI lab an anchor customer for two Malaysian data centres—signals how quickly the region’s compute expansion is coalescing around a single chip supplier. Firmus will deploy NVIDIA’s Vera Rubin processors and DSX AI Factory Platform at the Malaysian sites, integrating them with its prefabricated HyperCube units. The Australian operator, backed by NVIDIA, Coatue, Blackstone funds, and Jane Street, is scaling faster than most of its peers, using the chipmaker’s reference designs to cut through the complexity of power, cooling, and construction. The result is not just more capacity for OpenAI—it is a template for how AI infrastructure gets built in the region, and who controls the backbone.

The numbers that turn a partnership into a pipeline

With the OpenAI contract, Firmus’s total contracted capacity across all customers now exceeds 900 megawatts—a figure that puts the three-year-old operator in the same conversation as established hyperscalers. The company operates two AI data centres today, in Australia and Singapore, and is developing five more across the Asia-Pacific. The new Malaysian sites are expected to be ready for service within 24 months, though Firmus has not disclosed the individual capacity of each site.

Get the latest APAC news as it happens — follow Indoneo on Facebook

The technology stack at the Malaysian sites is entirely NVIDIA’s. Vera Rubin processors, the latest GPU architecture for large AI training workloads, will run on the DSX AI Factory Platform, an integrated design that ties together compute, networking, and infrastructure management. The HyperCube units—prefabricated, liquid-cooled modules manufactured in regional New South Wales—will house the hardware, compressing the usual construction timeline. The system is a closed loop: silicon, platform, and cooling all from one supplier’s playbook.

Malaysia’s policy environment is reinforcing the lock-in. In February 2026, Prime Minister Anwar Ibrahim told parliament that new data centre investments not linked to artificial intelligence would be restricted. The policy funnels new capacity toward AI-specific projects, which in turn demand the kind of high-end hardware that only a few chipmakers can supply. The table below maps the regulatory shift.

Malaysia’s AI-linked data centre policy timeline
Entity	Current rule	New rule	Effective date
Malaysia	Open to all data centre investments	Restricted to AI-linked data centres	February 2026 (PM Anwar’s parliamentary statement)
Malaysia	Personal data rules exist; no specific AI governance law	AI governance bill in early drafting	Drafting initiated 2026; no enactment date
Source: Malaysian government, parliamentary comments by Prime Minister Anwar Ibrahim, February 2026

Who’s winning the AI infrastructure race

Firmus is not the only operator racing to lock in capacity. Malaysia’s YTL is building AI-ready data centres, and regional players in Singapore, Australia, and Japan are all chasing the same combination of power, permits, and premier tenants. The differentiator is chip supply. Operators that secure NVIDIA’s Vera Rubin and DSX platforms can promise faster timelines and higher performance, making them the default choice for anchor customers like OpenAI.

Winning this race means converting a headline agreement into financed megawatts before energy and permitting constraints tighten. The operators that move first will define the latency and availability map for AI services across Southeast Asia. Losing means being left with generic colocation capacity in a market that increasingly privileges AI-specific infrastructure.

The building blocks of that speed advantage are easier to see than describe.

Firmus says it has seven AI factories across four countries, with two operational sites in Australia and Singapore and five under development. The real metric is not the 900 megawatts, but how quickly those megawatts convert to operational capacity that is NVIDIA-compatible. Every month of construction locks in a future where the region’s AI compute is a single-vendor pipeline. Firmus’s backers—NVIDIA, Coatue, Blackstone, and Jane Street—valued the company above $10.5 billion in its latest fundraising round, a bet that the pipeline will hold.

Firmus declined to disclose the financial terms of the OpenAI contract, and the specific power capacity of the two Malaysian sites remains undisclosed. The company’s five additional projects across APAC have not named customers or locations, leaving the full picture of the pipeline incomplete. For a climate researcher in Australia waiting for compute through the planned AI Access Program, the capacity is a distant promise. For a Malaysian startup, the anchor tenant’s contract already occupies the available slots, and the next wave of capacity is at least 24 months away.

The framework is in place. The question is whether the policy environment is accelerating the shift or simply clearing a path for a single supplier.

A policy funnel dressed as a market

Malaysia’s restriction on non-AI data centre investments is part of a broader pattern. According to market observers and infrastructure analysts, Southeast Asian governments are competing for AI infrastructure investment, with terms of entry increasingly favouring projects that can demonstrate high-end compute commitments. That means NVIDIA’s silicon, because no other chipmaker offers a comparable integrated platform at scale. The policy does not name NVIDIA, but it does not need to: the market is selecting the same default.

The regulatory environment is more permissive than the EU’s and less procedurally dense than the US’s, but Malaysian authorities retain the discretion to slow or redirect projects on water and electricity grounds. The result is a gatekeeper function that can accelerate AI-linked builds while filtering out generic colocation. For an operator like Firmus, with a named anchor tenant and a prefabricated, liquid-cooled design, the gate is open. For a conventional data centre without an AI story, it is not.

The Western parallel: a neocloud, not a hyperscaler

Firmus’s model most closely resembles a Western neocloud—a specialist compute provider building around a named anchor customer, rather than a hyperscaler’s broad consumer cloud. The difference is the delivery: prefabricated HyperCube units and NVIDIA’s DSX/Vera Rubin stack are scaling across Southeast Asia, while Western peers typically rely on larger, vertically integrated campuses in more mature regulatory environments.

The modular approach makes Southeast Asian capacity look more customer-committed and more directly tied to chip supply. That rigidity is a feature, not a bug, when the anchor tenant is OpenAI and the chip supplier is NVIDIA, but it also means the region’s infrastructure is hardening around a single supplier’s roadmap.

Firmus’s five additional projects are targeting ready-for-service within 24 months, and the next Malaysian policy cycle could clarify whether the AI-linked restriction becomes formal legislation. The speed at which NVIDIA-compatible capacity materialises will determine whether the current pipeline remains a single-vendor affair or opens to competition. For now, the region’s AI infrastructure buildout is not just using NVIDIA—it is building NVIDIA’s ecosystem into the grid.

Beyond the headline

The Money Trail

Industry analysts argue that the real financial beneficiary is NVIDIA’s ecosystem, not just OpenAI or Firmus. According to competitive infrastructure assessments, each new anchor-tenant deal increases the likelihood that subsequent rounds of AI infrastructure buying will be locked to NVIDIA-compatible power, cooling, and rack design choices, potentially narrowing procurement flexibility for operators and extending the chip supplier’s leverage over where expansion occurs.

The Bigger Picture

This is less a one-off Malaysian build than a sign that AI infrastructure is becoming a geopolitically distributed supply chain. Capacity is no longer concentrated in a few US cloud hubs; it is being split across jurisdictions that can offer power, permits, and political willingness to prioritise AI over ordinary data-centre demand.

The Reach

For Western cloud and infrastructure funds, the mechanism is simple: early commitment to Southeast Asian power and land can secure scarce future compute capacity before the market reprices it. The implication is that capital allocation decisions made now may shape who controls latency, availability, and regional service reach later.

Capital, supply chains, and the next buildout

With the OpenAI-Firmus deal closing, the next 12 to 18 months will determine whether the region’s AI infrastructure pipeline becomes a single-vendor standard or opens to broader competition. Four groups face distinct decisions.

Western investor in APAC digital infrastructure
Re-evaluate portfolio exposure to data centre operators that rely on generic colocation. The market is shifting toward AI-anchored, NVIDIA-compatible capacity. The next wave of value will belong to operators that can lock in anchor tenants and secure power approvals. Monitor Malaysia’s investment policy announcements through the Malaysian Investment Development Authority over the next quarter.
US-based AI service provider with APAC expansion plans
Assess your own infrastructure strategy. If you are not an anchor tenant with a dedicated build, your latency and availability will trail those who are. Consider partnerships with specialist compute providers like Firmus or direct investment in regional capacity to meet future demand. Check NVIDIA’s investor news releases for Vera Rubin and DSX developments to time your procurement.
Western semiconductor procurement manager for AI hardware
Monitor NVIDIA’s production and allocation strategies for Vera Rubin processors and DSX platforms. The Firmus deal signals that the chipmaker is prioritising its own ecosystem for the latest silicon. Verify your spot in the queue by reviewing NVIDIA’s investor news releases and contacting your supply chain representative.
European data center operator eyeing Southeast Asian markets
Analyze Firmus’s operational model closely. The prefabricated, liquid-cooled HyperCube approach is a replicable template for entering resource-constrained markets. The anchor-tenant strategy is not optional—it is the key to securing financing and regulatory approval. Begin your market entry planning with a study of Malaysia’s AI-linked investment policy.

Explainer

Vera Rubin: NVIDIA’s next-generation GPU architecture designed for large-scale AI training workloads. It succeeds the Blackwell architecture and is expected to power the company’s AI Factory reference designs. The architecture is named after the astronomer Vera Rubin, who provided evidence for dark matter.
DSX AI Factory Platform: NVIDIA’s integrated infrastructure design that combines compute, networking, and management software for AI data centres. It is the reference blueprint that operators like Firmus use to deploy large-scale AI compute clusters. The platform is designed to work with NVIDIA’s latest silicon, including Vera Rubin, to maximize throughput and energy efficiency.
HyperCube: Firmus’s prefabricated AI data centre module that integrates liquid cooling, mechanical systems, and electrification into a single unit. Manufactured in regional New South Wales, Australia, the HyperCube is designed to be deployed quickly and scaled modularly. It is the physical container for the NVIDIA hardware and DSX platform at Firmus’s AI factories.
Anchor customer: A large, creditworthy tenant that commits to a significant portion of a data centre’s capacity before construction is completed. The anchor customer’s contract provides the financial assurance needed to secure project financing and supply chain commitments. In the AI infrastructure market, an anchor customer like OpenAI signals that the capacity will be fully utilised, reducing investment risk for operators.
AI factory: NVIDIA’s term for a data centre purpose-built for AI workloads, using its integrated hardware and software stack. AI factories typically employ high-density, liquid-cooled compute and are designed to handle both training and inference at scale. The concept is central to NVIDIA’s strategy of selling entire infrastructure systems rather than individual chips.