Zankore’s $3.1B Nvidia GPU Loan Makes AI’s Next Bubble a Bank Problem
Thirty-four days after launching, Zankore lined up up to US$3.1 billion in debt for Nvidia GPUs. AI is no longer a tech story. It’s a bank-risk story.
Thirty-four days after launching, Zankore secured an up to US$3.1 billion loan facility to buy Nvidia GPUs. AI is no longer a software story. It is a debt market with server racks bolted to the collateral. ([zankore.com](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia?utm_source=openai))
The core story: US$3.1 billion is being wagered on rented compute
On September 9, Indonesian AI-infrastructure platform Zankore announced a senior term-loan facility of up to US$3.1 billion. Citi, ING, Natixis CIB, Qatar National Bank and UOB are underwriting it. The money is earmarked for advanced Nvidia GPU infrastructure and cloud-computing deployment across Indonesia and Southeast Asia. Citi acted as exclusive debt adviser. ([zankore.com](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia?utm_source=openai))
That is not a venture round. It is lenders putting serious debt behind hardware that depreciates, needs enormous power and cooling, and can become yesterday’s kit faster than a founder can update a pitch deck.
Zankore launched on August 6, 2026, with Ooredoo Group, Indosat Ooredoo Hutchison, Nvidia and Nokia behind it. Ooredoo is the founding shareholder and lead investor, with a 49% stake. Zankore’s stated ambition is to build toward 1 gigawatt of Nvidia DSX AI Factory capacity. At launch, it said it targeted roughly 200 megawatts in the first half of 2027; its September announcement describes an initial 100-megawatt Nvidia buildout. ([zankore.com](https://zankore.com/press-release-eng?utm_source=openai))
For perspective, one gigawatt is not a cute little AI side project. It is an industrial-scale bet that Southeast Asia will keep needing more training, inference and agentic-AI capacity — and will be willing to pay enough for it to service the debt.
The clever bit is the structure. Nvidia is not merely selling chips into this buildout. Zankore says the arrangement includes revenue-sharing and credit support designed to align deployment with customer demand. Bloomberg reported that Nvidia’s service order helps make the financing bankable by easing lenders’ concerns around repayment and credit risk. ([zankore.com](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia?utm_source=openai))
Read that again. The world’s most valuable AI hardware supplier is helping turn its own product sales into financeable infrastructure.
That is where this gets interesting.
Why this is bigger than an Indonesian data-centre story
For the past few years, the AI conversation has been painfully simplistic: who has the best model, who has the most GPUs, who is spending the most billions.
That misses the real game. Models are improving. Chips are getting bought. But somebody still has to finance the steel, fibre, substations, cooling systems, networking gear, land, contracts and operational teams needed to turn racks of GPUs into usable compute.
Zankore is an early marker that the financing model is evolving. Banks have traditionally been cautious about lending against GPUs because the residual value is uncertain and the technology cycle is vicious. A GPU can be indispensable today and merely adequate tomorrow. Add geopolitical risk around advanced chips and you can see why Asian bank-led GPU deals have been rare. ([theedgesingapore.com](https://www.theedgesingapore.com/amp/news/tech/nvidia-backed-firm-indonesia-signs-us31-bil-gpu-loan?utm_source=openai))
The answer here is not that the risk has disappeared. The answer is that Nvidia’s involvement, contracted demand and a revenue-linked structure may make the risk easier to package.
This is what mature booms do. First, founders sell a vision. Then equity investors finance the vision. Eventually, debt investors finance the physical machinery behind it.
Railways did it. Telecommunications did it. Data centres did it. Now AI compute is joining the club.
And when banks enter a market at this scale, the question changes from, “Is this exciting?” to, “What happens if utilisation misses plan?”
That’s a far healthier question, even if it ruins the party a bit.
The background: Zankore is selling more than GPUs
Zankore is positioning itself as a “neocloud” platform: effectively, GPU capacity delivered as a service rather than boxes sold to customers outright. Its target customers include AI-native businesses, enterprises, startups, developers and institutions across Indonesia and Southeast Asia. ([zankore.com](https://zankore.com/zankore-expands-southeast-asias-ai-compute-infrastructure-with-nvidia?utm_source=openai))
That matters because the winning asset may not be the GPU. Nvidia owns the GPU. The winner for Zankore must be the ability to keep those GPUs busy, price them intelligently, deliver reliable capacity and match customers to infrastructure in the right jurisdiction.
Ooredoo’s argument is that Indonesia is the launchpad rather than the final market. The company has pointed to demand across Southeast Asia, particularly as other markets confront power constraints. Fortune reported that Zankore had already secured blue-chip demand around its planned 200-megawatt initial capacity and that the platform aims to contract one gigawatt over the next three years. ([fortune.com](https://fortune.com/2026/08/06/qatar-ooredoo-nvidia-nokia-unveil-multi-billion-dollar-ai-compute-platformand-southeast-asia-is-their-target/?utm_source=openai))
There is a commercial logic here. Companies increasingly care about where their data lives, where their inference runs and whether capacity will be available when they need it. American cloud giants will remain giants, obviously. But regional capacity, local relationships, telecom distribution and data-residency options can be a real wedge — if the operator executes.
Zankore also begins with an existing base rather than a blank sheet of paper. Fortune reported that Indosat Ooredoo Hutchison’s Nvidia-powered AI cloud business expected to operate about 28 megawatts of GPU capacity in the third quarter of 2026, after generating US$33 million in the first half of the year. ([fortune.com](https://fortune.com/2026/08/06/qatar-ooredoo-nvidia-nokia-unveil-multi-billion-dollar-ai-compute-platformand-southeast-asia-is-their-target/?utm_source=openai))
That is still a long way from one gigawatt. But it is more credible than a bunch of slides and a render of a shiny building in a jungle.
The second-order implication: Nvidia is becoming part vendor, part infrastructure financier
Most people will see this headline and conclude that Zankore has bought a truckload of Nvidia chips.
Too shallow.
The more important signal is that Nvidia is helping make an enormous GPU purchase financeable. That extends Nvidia’s influence beyond product performance and into the capital structure of the AI economy.
If Nvidia can help partners obtain debt, support revenue models and bring smaller customers into the market, it can expand the pool of buyers who can afford serious compute. That is good for Nvidia. It is also good for banks because they are not underwriting a random startup buying speculative hardware with a prayer and a Notion page.
But it creates concentration risk. Zankore depends on Nvidia hardware. The lending proposition is strengthened by Nvidia’s support. Zankore’s customer proposition is built around Nvidia capacity. One company becomes supplier, ecosystem partner and a key ingredient in the financing logic.
There is nothing automatically wrong with that. Plenty of industries have dominant vendors sitting at the centre of financing arrangements. But operators and investors should call it what it is: a tightly coupled ecosystem, not a cleanly diversified infrastructure play.
The other implication is that AI infrastructure will increasingly be judged like infrastructure. Cost of capital, utilisation, power availability, contract duration, credit quality and uptime will matter as much as benchmark scores.
Frankly, good. The AI world could use fewer demos and more boring questions asked by sceptical adults with spreadsheets.
The overlooked angle: this is not US$3.1 billion of proof
Here is the bit everyone should keep in their pocket before getting carried away: Zankore has signed an up to US$3.1 billion facility. That is not the same thing as having spent US$3.1 billion, installed US$3.1 billion of GPUs or earned US$3.1 billion in revenue.
Debt facilities are commitments with conditions, drawdowns, covenants and execution risk. The capital only becomes productive when the company turns it into operational capacity and then turns that capacity into paying usage.
Zankore’s early forecasts are ambitious. Fortune reported projections of about US$13 billion in revenue and US$9 billion in cumulative EBITDA over five years for the platform. Those are company forecasts, not bank statements from the future. Treat them accordingly. ([fortune.com](https://fortune.com/2026/08/06/qatar-ooredoo-nvidia-nokia-unveil-multi-billion-dollar-ai-compute-platformand-southeast-asia-is-their-target/?utm_source=openai))
The bullish case is straightforward: demand stays hot, Southeast Asian compute remains constrained, Zankore locks in quality customers and Nvidia-backed financing lets it scale ahead of slower rivals.
The bear case is just as straightforward: GPU supply improves, pricing falls, customers build their own capacity or rent from hyperscalers, and expensive capacity sits underused while debt keeps charging interest.
This is why “AI infrastructure” is not a magic phrase. It is a brutally operational business. You win by keeping expensive machines fully utilised without destroying your pricing.
What this means for you
If you are a founder, stop treating compute as a vague line item. Ask whether your workload truly needs premium GPUs, whether it needs them continuously, and whether a long-term reservation makes commercial sense. AI costs can wreck a good product faster than bad hiring.
If you run an AI company, negotiate compute like a grown-up. Know your utilisation, peak demand, failover needs, model mix and gross margin per customer. Do not sign capacity commitments because someone told you AI demand only goes one way. Nothing only goes one way.
If you are an investor, separate chip demand from durable infrastructure economics. A massive GPU order can be evidence of growth — or evidence that someone has found a more expensive way to be wrong. Ask who the end customers are, how long their contracts run, whether revenue is committed or merely forecast, and who wears the downside if GPU prices fall.
And if you are building any capital-heavy business, nick the useful lesson: find ways to make your assets legible to lenders. Revenue contracts, credible counterparties, clear operating data and sensible risk-sharing beat charisma every day of the week.
Zankore’s US$3.1 billion financing is not proof that every AI infrastructure bet will work. It is proof that the next phase of AI is being financed like an industrial buildout.
That means the winners won’t just have the flashiest model or the loudest founder. They will be the ones who can turn expensive hardware, power and debt into reliable cash flow.
That’s less sexy than an AI demo. It is also where the real money usually gets made.