Everything About AI
NextFin News - Zankore, an Indonesian AI infrastructure platform backed by Nvidia Corp., signed a $3.1 billion loan deal on Wednesday to purchase advanced graphics chips, in one of Asia's largest artificial-intelligence infrastructure financings to date. The syndicated facility, underwritten by Citigroup Inc., ING, Natixis, Qatar National Bank and United Overseas Bank Ltd., with Citigroup as sole debt adviser, marks a turning point: the race to build sovereign AI capacity in Southeast Asia is now being funded by global bank balance sheets rather than only by state treasuries and venture capital.
The Deal: $3.1 Billion to Buy the Chips That Will Power Indonesia's AI Ambition
Zankore is not a start-up betting on an unproven idea. It is the AI infrastructure venture of Indosat Ooredoo Hutchison (IOH), Indonesia's second-largest mobile operator, launched in August with $800 million of sponsor equity from the Qatar-based Ooredoo Group, which holds a 49% stake. Nokia and Nvidia are the other named partners - Nokia supplying AI-native networking, Nvidia supplying the accelerated-computing hardware, software and GPU architecture that the platform is built around.
The platform's stated target is 1 gigawatt of Nvidia DSX AI Factory capacity across Southeast Asia, with an initial 200 megawatts deployed in the first half of 2027. To put that ambition in scale: a 200MW AI data center can house on the order of tens of thousands of high-end GPUs, and the full 1GW target would rank among the largest AI compute campuses in the Asia-Pacific region. The $3.1 billion loan is the debt engine for the first leg of that buildout.
The financing is notable for its structure as much as its size. Rather than requiring customers to buy GPUs outright - a capital barrier that has kept many enterprises and smaller cloud operators out of the AI race - the facility funds a platform that deploys chips on a revenue-sharing basis. That model, pioneered by Nvidia's DSX program, converts what would be a multi-billion-dollar upfront capex decision for a customer into an operating-expense contract, widening the pool of potential buyers for the compute Zankore is borrowing to build.
"As enterprises move from AI experimentation to mission-critical deployment, they need more than computing power; they need an integrated AI ecosystem," IOH President Director and CEO Vikram Sinha said in a statement.
The equity-debt mix is itself a signal. With $800 million of sponsor equity against a $3.1 billion facility, the deal implies roughly 80% debt funding of the initial capital stack - aggressive by traditional infrastructure standards, but consistent with a model in which the underlying assets are standardized, liquid chips rather than bespoke construction. Ooredoo Group CEO Aziz Aluthman Fakhroo framed the investment as a portfolio extension rather than a speculative bet: "With Zankore, we are taking the next logical step by adding AI compute to that portfolio, extending our approach into the high-growth Southeast Asian market - one of today's most dynamic AI infrastructure opportunities."
The demand Zankore is betting on is not hypothetical for its sponsor. IOH's neocloud services - the business line into which AI infrastructure will feed - generated about $35 million in revenue during 2025 and roughly $16 million in the first quarter of 2026 alone, with the company targeting approximately $170 million in contracted neocloud revenue over the following three years. That is the commercial bridge between a telecom operator and an AI infrastructure owner: existing cloud and compute revenue that can absorb the first wave of GPU capacity while enterprise offtake ramps.
Why Banks, Not Just States, Are Now Financing the AI Arms Race
The deeper story behind Wednesday's signing is a shift in who is willing to underwrite AI infrastructure risk. For the past two years, the largest GPU buildouts have been financed by a narrow set of actors: hyperscalers with investment-grade balance sheets, sovereign wealth funds, and a handful of well-capitalized crypto-mining firms that pivoted to AI. Zankore's facility - alongside comparable deals such as IREN's $2.4 billion private-credit package completed in late August to purchase Nvidia Blackwell Ultra GPUs - shows lenders increasingly comfortable treating GPU clusters as bankable collateral.
That comfort rests on a specific mechanism. A GPU cluster financed under a structure like this is not a speculative bet on a single tenant's success; it is an asset that generates revenue across a diversified pool of enterprise and government customers, with the underlying hardware retaining residual value even if one contract defaults. For banks, that transforms AI infrastructure from a technology bet into something closer to traditional project finance - toll roads and power plants, but with semiconductors instead of concrete.
The parallel with the crypto-mining cycle is instructive, and not entirely reassuring. In 2020-2021, lenders financed Bitcoin-mining rigs on the same premise: standardized hardware, liquid secondary markets, revenue-generating assets. When the cycle turned, the collateral value of the machines collapsed alongside the revenue stream, and the two risks compounded rather than offset. The difference this time is that AI compute demand is anchored by enterprise and government workloads with multi-year contracts, not by a single commodity price. But the structural lesson survives: hardware collateral is only as good as the cash flow it can generate, and cash flow depends on utilization.
The regional context sharpens the point. Indonesia is competing directly with Singapore, Malaysia and Thailand for the role of Southeast Asia's AI hub, and it has structural advantages that a bank syndicate can underwrite: a domestic market of 280 million people, a growing base of digital-native enterprises, and - critically - a telco operator that already owns the fiber, tower and data-center footprint into which AI capacity can be dropped. Indosat's digital-infrastructure spinoff, Infra Fiber Teknologi, controls roughly 86,000 kilometers of fiber connecting 5G towers, fiber-to-the-home networks and data centers. That is not an asset a greenfield start-up can replicate in three years.
Indonesia is not acting alone in the region. The broader buildout wave provides the competitive pressure that makes Zankore's timeline credible - and risky. Firmus Technologies, an Australian AI infrastructure firm backed by Nvidia, is building a 360MW Nvidia DSX AI Factory campus in Batam, Indonesia, just off the coast of Singapore, with the facility expected to be operational in the first quarter of 2027. Firmus and its Singapore-based partner DayOne are constructing the campus as a multi-tenant project aimed at AI-native customers, and the company has said it expects to win as much as $30 billion in committed offtake agreements in its first six years. That is the demand Zankore is betting against: a regional market in which two large, Nvidia-aligned platforms are racing to bring capacity online within months of each other.
Even so, the $3.1 billion facility is large relative to Indonesia's recent data-center financing. BDx Data Centers, another operator building an AI-focused campus in Jakarta, raised $320 million from a consortium of Indonesian banks earlier this year - less than a tenth of Zankore's facility. The step-change in size signals that international lenders, not just domestic ones, are now prepared to price Indonesian AI infrastructure risk.
The Counter-Thesis: What Could Break the Sovereign-AI Financing Model
The strongest argument against this wave of debt-financed GPU buildouts is simple: demand may not arrive as fast as the capacity is being built. The cyclical risk is real. AI infrastructure capex across the region has been running ahead of realized enterprise revenue, and if the transition from "experimentation to mission-critical deployment" that Sinha describes proves slower than the 2027 deployment timeline assumes, Zankore and its lenders face a classic overbuild scenario - fixed debt service against underutilized assets.
The competitive dimension of that risk is underappreciated. Zankore is not building into empty space. Firmus-DayOne's Batam campus targets the same AI-native customer base with a comparable Nvidia DSX architecture and a similar 2027 operational date. When two large platforms commission capacity into the same regional market on overlapping timelines, the first-order effect is lower utilization for both; the second-order effect is pricing pressure that compresses margins precisely when debt service is highest. In a leveraged structure, margin compression hits equity returns first, but refinancing risk follows if utilization does not recover before the facility's repricing dates.
There is also a technology-cycle risk that no loan covenant can fully hedge. GPUs depreciate faster than the buildings that house them. A Blackwell-class accelerator bought in 2026 will be economically obsolete well before a typical 7-to-10-year project-finance tenor expires, meaning the residual-value cushion that makes this collateral bankable today could erode if a materially more efficient chip generation arrives on schedule. Lenders are aware of this - the involvement of five international underwriters and a sole debt adviser suggests the risk was sliced and distributed rather than concentrated - but it remains the single largest underwriting assumption in the deal.
The counter-thesis is not that the model fails outright. It is that the economics compress: utilization rates fall short, pricing power shifts to the chip supplier rather than the infrastructure owner, and the equity returns that Ooredoo and IOH are underwriting come in at the low end of their range rather than the high end. In that world, the debt still gets serviced, but the "AI hub" narrative delivers modest utility returns rather than transformational ones.
Cyclical or Structural: This Is a Regime Shift With a Cyclical Overlay
The judgment this piece defends: the financing model is structural, but the timing is cyclical - and the two are being conflated in the market's enthusiasm.
The structural leg is clear. Emerging-market telcos and sovereign platforms are becoming permanent participants in AI infrastructure ownership, financed through syndicated debt and revenue-sharing arrangements with chipmakers rather than through state budgets alone. Once a lender has priced and syndicated a deal of this size, the playbook exists for the next one, and the next. The barrier to entry shifts from "who has $3 billion" to "who can originate and service the customer base" - a contest where incumbent telcos like IOH have a durable advantage.
Three structural forces support that call. First, the capital intensity of AI infrastructure now exceeds what most single corporations can fund from cash flow, forcing a permanent shift toward external finance. Second, the standardization of GPU-based compute - Nvidia's DSX architecture is effectively becoming the reference design for AI factories - makes the assets financeable in a way that custom AI accelerators never were. Third, the revenue-sharing model aligns the chip supplier, the infrastructure owner and the end customer on utilization, reducing the probability of a clean default even in a soft-demand scenario. Each of these forces is durable; none of them reverses on its own.
The cyclical leg is equally clear. The 200MW-by-first-half-2027 target rides a capex wave that has, in previous technology cycles, overshot realized demand. If enterprise AI adoption proves to be a step-function rather than a smooth curve, the capacity commissioned in 2026-2027 could sit partially idle until workloads catch up. That is not a failure of the model; it is the normal rhythm of infrastructure buildouts, and it is mean-reverting.
The practical implication: the structural shift means Indonesia will become a meaningful AI infrastructure node regardless of short-term utilization hiccups. The cyclical overlay means the first wave of projects - including Zankore's - may underperform their most optimistic return forecasts even as they succeed strategically.
What to Watch Next
Three signals will determine whether this deal is the first of a wave or a one-off that lenders regret:
- Utilization and offtake disclosure. If Zankore or IOH begins disclosing contracted capacity or named anchor tenants within the next two quarters, the demand thesis is validated. Silence on offtake through 2027 would support the overbuild counter-thesis.
- The second deal. A follow-on facility - either for Zankore's scale-up to 1GW or for a competing Indonesian platform - within 12 months would confirm that lenders see a repeatable model rather than a single asset.
- Chip-generation cadence. If Nvidia or a competitor announces a successor architecture that materially improves performance-per-watt before Zankore's first 200MW is fully utilized, the residual-value assumption behind the collateral weakens, and refinancing terms for the whole sector will reprice.
The falsifying signal for the structural thesis is specific: if, by the end of 2027, no second syndicated AI-infrastructure facility of comparable size has closed in Indonesia or a peer Southeast Asian market, the "new financing regime" narrative is wrong, and this deal reverts to being a one-off supported by sponsor relationships rather than a bankable asset class.
For investors and policymakers watching from outside Indonesia, the takeaway is narrower than the headlines suggest. This is not evidence that AI demand is accelerating faster than expected; it is evidence that the financing plumbing for AI infrastructure has matured enough for banks to treat GPU clusters as collateral. The demand still has to arrive on schedule - and the 2027 deployment target is the clock against which that assumption will be tested.
The real story of Zankore's $3.1 billion loan is not that Indonesia is buying GPUs. It is that global banks have decided GPU clusters are now ordinary infrastructure - and ordinary infrastructure is exactly where overbuilds happen.
Explore more exclusive insights at nextfin.ai.
