OpenAI's Malaysia Bet With Firmus Just Made the US Hyperscaler Capex Trade a Two-Region Story

Firmus Technologies said on September 8, 2026 that OpenAI has signed a multi-year agreement to take dedicated AI compute from two of its AI factory sites in Malaysia, making the ChatGPT maker the company's first publicly named anchor customer. The deal lifts Firmus's total contracted capacity across customers to more than 900 MW — a 15× jump from the 18,400-GPU Melbourne deal it disclosed in March 2026 — and hands the unlisted Australian builder a marquee reference customer just as bankers price what could be one of Australia's largest IPOs in years.

The two Malaysia sites are part of a seven-factory portfolio across Australia, Singapore, Indonesia and Malaysia. Two sites are operational today (Australia and Singapore) and five — including the two earmarked for OpenAI — are slated to come online over the next 24 months. Firmus's factories will run NVIDIA's DSX AI Factory platform integrated with its HyperCube liquid-cooled modular system, which is prefabricated in regional New South Wales.

Two macro numbers do the work here. The first is electricity. Tenaga Nasional's medium-voltage industrial tariff for data-center-eligible customers in Peninsular Malaysia sits at roughly 29.83 sen/kWh — about 6.7 US¢/kWh at recent exchange rates. Northern Virginia, the world's largest data-center cluster, is at 8.99¢/kWh. For a 100 MW AI factory running at a 70% load factor, that 2.3¢/kWh spread is roughly $13M of operating-cost relief per year before any cooling, water or land savings.

The second number is PJM's capacity market. The 2026/27 clearing price came in at $329.17/MW-day, almost 10× the $28.92/MW-day level of two auctions ago, and the U.S. Department of Energy had to issue emergency orders in July 2026 to let PJM curtail data-center load during a heat wave. PJM has now proposed a new rule that would force large data centers to fund their own power capacity or accept interruptible service. The message to hyperscalers is blunt: new US capacity will be slower, more expensive and less reliable than it looked two years ago.

The conventional read is that this is a NVIDIA win — more Vera Rubin systems shipping, more DSX factory slots filled. That is true but small relative to the NVIDIA order book. The bigger read is on the US hyperscaler build-out. Microsoft, Alphabet, Amazon and Meta collectively guided to roughly $725B of capex in 2026, up 77% year-on-year, and a large share of that spend is now being repriced by Asia-Pacific capacity that arrives in months rather than years.

Firmus's stack is a working model for what a US-anchored AI builder can do when it sources hardware, cooling and energy outside the US tariff net. Upstream, the chips, the rack-scale system and the factory blueprint all come from NVIDIA (DSX AI Factory + Vera Rubin NVL72 + up to 170,000 accelerators across Asia). Connectivity is being built out through the SUBCO Bernacchi-1 trans-Pacific cable announced June 2026 and a separate SUBCO subsea deal on September 3, 2026. Energy supply was locked in via a 600 MW South Australia agreement in June 2026 linked to 1.2 GW of new renewable generation and battery storage. The data hall itself is fabricated in regional New South Wales and shipped.

On the Malaysia side, the obvious local winner is YTL Power International, which has already completed a 500 MW Nvidia-powered data-center park in Kulai, Johor that hosts Microsoft's first Malaysian region, and is now doubling capacity to 2.4 GW by 2032. Grid power is supplied by Tenaga Nasional, the Peninsular Malaysia utility whose RP4 tariff structure sets the 6.7¢/kWh benchmark above. Downstream, the customer set is broadening: Fireworks signed on as a Firmus AI Cloud customer in August 2026, and Firmus has a separate Indonesian 360 MW AI factory with DayOne that pulls in sovereign-AI workloads.

Days to quarters: the Firmus ASX float is expected to price within weeks of the OpenAI announcement, and the IPO valuation (reportedly up to US$30B) will be the first market-clearing test of how much equity capital is willing to underwrite an Asia-Pacific AI infrastructure play that is not a US hyperscaler. The next operational milestone is the energisation of the two Malaysia sites; the OpenAI deal is not yet a paid contract but a contracted-capacity reservation that becomes revenue as the sites come online. Watch for: (1) the ASX prospectus and any disclosure of contract tenor, dollar value and capacity reservations vs. take-or-pay; (2) Malaysia site naming (the announcement did not specify the two locations); and (3) the 2027/27 PJM capacity auction in mid-2026, which will show whether US capacity scarcity deepens or stabilises.

One to three years: the structural risk is concentration. Both Malaysia sites run on the same NVIDIA Vera Rubin platform that Stargate runs on; if NVIDIA's supply prioritises Stargate and direct hyperscaler buyers, Firmus could face allocation risk on its 170,000-accelerator pipeline. Sovereign-data rules in Malaysia, Indonesia and the EU are tightening; the September 2025 Stargate UK pause and the April 2026 Stargate Norway handover to Microsoft show that OpenAI has already lost two non-US build-outs to regulatory friction, and Malaysia is not exempt from this risk. The Firmus floating-datacenter cost advantage also depends on the Malaysian ringgit staying weak and Tenaga Nasional's RP4 tariff not being repriced upward for large industrial users.

The asymmetry the Firmus deal exposes shows up clearly in the multiples. Microsoft is on 11.2× trailing sales and a 27.8× P/E with TTM operating cash flow of $183B and capex of $116B (capex/operating cash flow of 63%) — a balance sheet that can absorb 24 months of Asian diversification without strain. Equinix, the most rate-sensitive US data-center landlord, trades on 27.7× EV/EBITDA with a 51.6% gross margin, a valuation that implicitly assumes the grid stays the binding constraint. Constellation Energy, the US nuclear-backed power generator most exposed to data-center PPAs, is on 4.2× EV/sales but on 423.9× EV/FCF — a market paying for growth in a single buyer category. None of these valuations break immediately, but the Firmus deal introduces a competing supply curve that none of them were modelling at scale two quarters ago.