What Firmus (Nvidia) Signing Two Data Centers in Malaysia with OpenAI Means — September 8, 2026
· CompaniesAutomation
Radar Flash Edition: Firmus, backed by Nvidia, signs a multi-year contract with OpenAI for two data centers in Malaysia, exceeding 900 MW contracted. Why capacity signed today won't exist until 2028, where your inference is already running, and three things to ask your AI provider in writing.
Flash edition. Firmus, the Australian data center company backed by Nvidia, has signed a multi-year contract with OpenAI to serve compute from two data centers in Malaysia. Complement to this morning's Radar.
What happened
Firmus announced it this Tuesday, September 8: OpenAI becomes the anchor customer for two "AI factories" in Malaysia, which will be built over the next 24 months using Nvidia's Vera Rubin chips (Firmus note). With the agreement, Firmus's contracted capacity exceeds 900 megawatts: it currently operates two centers (Australia and Singapore) and is building five more (Reuters). The firm, valued at over $10.5 billion, is preparing for an IPO this year. Neither party has disclosed the amount. "They will help us serve growing demand in the region and around the world," said Sachin Katti from OpenAI.
Why it matters
A week after launching GPT-6 Astra, OpenAI continues to buy capacity wherever it exists, and it exists in Asia: Malaysia is the fastest-growing data center market in Southeast Asia, with electricity and water already under scrutiny. The pattern is the same as we've seen with Lambda and CoreWeave: an intermediary that takes on debt, builds, and leases to a large laboratory. For you, there are two takeaways. The capacity being signed today won't exist until 2028, so usage limits and "promotional" prices will remain tight for a while. And your AI provider executes further and further away: the question is no longer which model you use, but where it runs.
For your company
Three things that cost nothing. One: ask in writing where your inference is processed. OpenAI offers data residency in Europe for API, Enterprise, and Edu, but only if you activate it; it is not in the plan your team uses by default. If you handle customer data, that checkbox is your GDPR. Two: design your flows for the limit, not the average. Every automation that depends on a single model needs a Plan B when a 429 error arrives—queue, retry, or a second provider—because new capacity takes two years. Three: if you negotiate an annual contract, ask for a reserved capacity clause with guaranteed limits; now that laboratories are signing multi-year anchors, you can ask for yours too.
Frequently Asked Questions
Does anything change today in my ChatGPT or API account?
Nothing. The Malaysia centers will take up to 24 months to be ready. The signal: OpenAI is reserving compute years in advance; those who don't reserve will depend on the leftovers.
Can my data end up being processed in Malaysia?
It depends on the plan. With European data residency activated, OpenAI commits to processing and storing in the EU; without it, there is no guarantee of where the inference runs. Review it before assuming.
What is Firmus and why does Nvidia matter here?
Firmus is an Australian operator in which Nvidia invests and to which it sells its chips: it finances the builder. It's the same circle that sustains Lambda or CoreWeave, and the one that decides how much compute is available.