Table of Contents
The Nvidia-powered AI-RAN deal moves the partners’ year-long collaboration to deployment
In sum – what we know:
- From demos to deployment – The agreement is the first to anchor Nokia, Indosat, and Nvidia’s year-long AI-RAN work to a nationwide commercial contract with a firm timeline.
- A staged rollout – Field trials on Indosat’s live network are slated for end of 2026, with broad commercial AI-RAN deployment expected to begin in 2027.
- The economics are unproven – Analysts still question the capex, opex, and energy draw of GPU-accelerated cell sites in a tight-ARPU market like Indonesia, claims the trials are meant to test.
Nokia and Indosat have signed a strategic network modernization and AI-RAN rollout agreement that puts their Nvidia-powered collaboration on a path to actual commercial deployment across Indonesia. The deal makes Nokia the key 5G RAN supplier for Indosat’s low- and mid-band 5G rollout nationwide, and formally locks in NVIDIA’s GPU-accelerated AI-RAN platform as the compute foundation for both AI-enhanced RAN functions and edge AI workloads on Indosat’s network.
On the hardware side, the modernization brings Nokia’s latest radio portfolio into the network, including the next-generation Habrok and Pandion models, alongside the Levante baseband platform, centralized RAN (C-RAN) systems, and Nokia’s network management and automation tools. In other words, this is a full network refresh — one designed from the start to be AI-ready rather than retrofitted later.
The two companies have been working with Nvidia on AI-RAN for over a year, but until now the work has lived in labs and conference demos. This agreement is the first time it’s been anchored to a nationwide commercial contract with a firm timeline attached.
From proof-of-concept to live network
The June announcement extends a collaboration that started with an MoU at MWC 2025, when Indosat, Nokia, and Nvidia first agreed to develop and deploy AI-RAN together. That was followed in late 2025 by the launch of an AI-RAN Research Centre in Surabaya, positioned as a long-term R&D hub for AI-native wireless and edge AI in Indonesia. Then at MWC 2026, the partners demonstrated an AI-powered 5G call linking Barcelona and Jakarta in real time — cited as the first AI-powered 5G call in Southeast Asia.
Those were proofs of concept, but the new agreement is a deployment roadmap. It explicitly sets a timeline to move from lab conditions to field trials on Indosat’s live network in Indonesia by the end of 2026, with the trials focused on validating spectral efficiency gains, energy savings, and AI service performance in real-world conditions rather than controlled environments. Assuming those trials go well, broad commercial AI-RAN deployments are expected to begin across Indosat’s footprint in 2027 — a window that aligns with what Nokia and Nvidia have publicly pointed to for commercial AI-RAN globally.
AI-RAN implementation
The architecture combines three things on a single, software-defined infrastructure. AI-for-RAN uses machine learning to optimize how the network itself operates — dynamic resource allocation, traffic prediction, interference management. AI-on-RAN runs edge AI inferencing workloads, like video analytics or generative AI, on the same hardware that handles baseband processing. And AI-and-RAN treats that hardware as a multi-purpose platform, hosting connectivity and third-party AI workloads concurrently.
Technically, this means cloud-native RAN functions (DU and CU) running as Nokia anyRAN/Cloud RAN software, orchestrated via Kubernetes on GPU-accelerated servers. Nvidia supplies the AI Aerial platform — the accelerated computing stack and GPUs that host everything. Nokia and Nvidia argue this approach delivers better spectral efficiency and energy optimization than traditional purpose-built RAN hardware, though that’s precisely the claim the end-of-2026 field trials are meant to prove. Indosat, for its part, envisions the architecture as a single intelligent “AI mesh” layer — centralized AI processing combined with distributed AI-RAN nodes, distributing intelligence and connectivity to millions of devices and users.
The deal also slots neatly into national policy. Indonesia’s “Golden Indonesia 2045” vision calls for transforming the country into a high-income, digitally advanced economy by its centenary, and this rollout is explicitly framed as supporting that goal. There’s a practical dimension too. Indonesia is a large, geographically dispersed market where coverage outside Java remains uneven, and the low- and mid-band 5G focus is aimed at improving connectivity in underserved areas while managing dense urban capacity needs without proportionally increasing hardware and energy costs.
If the trials and 2027 rollout hold to schedule, the network would lay the foundation for low-latency consumer services like AR/VR and cloud gaming, along with enterprise applications including smart city traffic monitoring and AI-based video analytics in private 5G factory deployments. It would also put Indonesia among the earliest adopters of commercial AI-RAN globally — ahead of regional peers like Singapore, Malaysia, and Vietnam, which is not where you’d necessarily have predicted Southeast Asia’s first commercial AI-RAN network to land.
The future of the network?
Strategically, the deal is another data point in the broader Nokia-Nvidia alliance, which the two companies have positioned as a Western-aligned counterweight to Ericsson, Huawei, and ZTE in critical telecom infrastructure. NVIDIA has committed a $1 billion investment in Nokia as part of that partnership, and Indosat now joins a roster of early AI-RAN operators that includes T-Mobile US, SoftBank, and Vodafone. For Nvidia, every base station that becomes a GPU-powered micro-data center is an extension of its AI dominance into a market it didn’t previously touch.
That said, the skepticism around AI-RAN hasn’t gone away. Analysts have repeatedly questioned the high capex and opex of GPU hardware, particularly in a tight-ARPU market like Indonesia, where the economics of replacing purpose-built RAN silicon with general-purpose accelerators are far from settled. Energy consumption is a persistent open question too — GPU-accelerated cell sites need to demonstrate that their efficiency gains actually offset their power draw versus traditional RAN systems, and so far that case rests largely on vendor claims. There are also concerns about operators locking themselves into Nvidia’s specific AI ecosystem, a familiar worry for anyone who’s watched the data center market over the past few years.
And then there’s the gap between the lab and the field. AI algorithms trained in controlled conditions in Surabaya will need to perform reliably across Indonesia’s noisy, heterogeneous real-world environments — thousands of islands, wildly varying traffic patterns, and infrastructure quality that ranges from dense urban Jakarta to remote rural sites. The end-of-2026 field trials are designed to answer exactly these questions.
