Tune Talk Runs Nokia 5G Core on AWS: Cloud-Native MVNO Blueprint

Every platform lead who has run a carrier-grade workload knows the trade: you either own the metal and eat the capex, or you rent the cloud and pray the latency budget holds. Tune Talk just picked door number two, at national scale. The Malaysian operator is running Nokia's full 5G Core on AWS, with the user plane pushed out to Outposts at the edge, and an agentic AI layer minding the store.

What Happened

Tune Talk has moved from being a standard on-premises MVNO to a network architecture capable of full Mobile Network Operator duties, and it did so by putting Nokia's entire 5G Core software stack on AWS public cloud. As Telecoms Tech News reported, the control plane sits in the AWS Malaysia region while the user plane runs locally on AWS Outposts at the network edge. That split is the whole game: central scale for signalling and subscriber logic, local metal for the fast path.

The migration was phased, not big-bang. The engineering team first lifted legacy workloads onto a standard cloud footprint, using that as a stabilised baseline before rewriting for a hyper-cloud setup with Nokia and Mavenir. Only then did the cloud-native rebuild land. Deploying the first AWS RP gateway in Southeast Asia required deep engineering collaboration between Nokia and AWS to hit carrier-grade availability inside a shared public cloud.

On top of the core, Tune Talk added an agentic AI framework powered by AWS Agent Core Bedrock to run day-two operations. It executes automated root cause analysis across the network fabric, handles In-Service Software Upgrades without dropping active subscriber sessions, and expands infrastructure capacity on the fly during traffic surges.

The business signal is loud. In December 2024, Tune Talk was picking up 50,000 new subscribers a month. Once the cloud-native stack was steady, sign-ups reached as many as 15,000 per day. Whatever you think about public cloud telco economics, that curve is not ambiguous.

Technical Anatomy

The interesting part is not that Nokia's 5G Core runs in containers. That is table stakes now. The interesting part is where the containers live and what isolates them.

The edge deployment uses bare-metal instances running the AWS Nitro architecture. Nitro pulls virtualisation functions off the main CPU and dedicates hardware resources directly to the containerised network functions. That matters because 5G user-plane traffic is unforgiving: jitter, tail latency, and noisy-neighbour effects show up as dropped calls and degraded throughput, not as a Grafana warning you can triage on Monday. Dedicated hardware paths give the CNFs a predictable performance envelope inside what is otherwise a shared public cloud.

The control plane, session management, subscriber data management, and policy control functions sit in the AWS Malaysia region. Split-plane designs like this are documented in reference material such as the cloud architecture playbooks the hyperscalers all publish. What differs here is the carrier-grade SLA sitting on top: signalling can tolerate a regional round trip, user plane cannot, so the topology follows the latency budget instead of the org chart.

The agentic layer is where I'd raise both an eyebrow and a glass. Automated RCA across a multi-tenant network fabric is genuinely useful when it works. In-Service Software Upgrades that keep subscriber sessions alive are the kind of thing that used to require a change-freeze weekend and three vendors on a bridge call. Autoscaling capacity during traffic surges turns standard CI/CD pipelines into a self-healing operational layer.

My take: agentic RCA is powerful and terrifying in equal measure. Every production incident I've seen where automation drove remediation without a human in the loop had one common failure mode: the automation confidently did the wrong thing at 3am, faster than any human could have. The engineering discipline that keeps this safe is not the AI framework, it's the guardrails, the blast-radius controls, and the rollback paths. Those don't come in the Bedrock box.

Who Gets Burned

Three groups should be reading this deployment carefully.

First, incumbent MNOs with brownfield infrastructure. Tune Talk went from MVNO to MNO-capable without buying a national tower footprint. If the small cell plan lands, 80,000 to 100,000 self-installable units across Peninsular and East Malaysia, the densification arithmetic gets ugly for anyone still running traditional RAN rollouts. Small cells that end users mount with standard tools, plug into a domestic outlet, and connect via ethernet to a home router, managed by a centralised Self-Organising Network that configures neighbour relations and handles interference automatically, is a completely different capex curve. Field technician trucks don't roll. That is two engineers worth of budget saved per site, times tens of thousands of sites.

Second, regional MVNOs. The open-access model means other regional telecom providers can offload traffic onto Tune Talk's small cell layer. Small cell hosts earn passive income when subscribers from other networks connect. That flips a proprietary footprint into shared national utility infrastructure, and it flips MVNO economics with it. If you're a competing MVNO, you now have a choice: rent capacity from a competitor's cloud-native fabric, or explain to your board why you're building the same thing two years late.

Third, telco software vendors selling closed-appliance 5G cores. Nokia clearly saw the wind direction and shipped its stack into a public cloud shape. Vendors still selling sealed boxes with support contracts priced like enterprise storage in 2015 are going to have some awkward renewal conversations.

The uncomfortable read: greenfield operators unburdened by legacy technical debt can adopt public cloud from day one, wiping out huge upfront capital spend and tying running costs to active subscriber revenue. Brownfield operators cannot. That gap widens every quarter.

Playbook for Engineering Teams

If you run platform infrastructure adjacent to any of this, here is what deserves calendar time this week.

Audit your latency budget by tier. The Tune Talk split, control plane central, user plane at the edge, is a pattern that generalises. Fintech teams doing real-time fraud scoring, iGaming platforms serving live odds, ad-tech running RTB bidders: all of you have workloads that need edge metal and workloads that don't. Draw the line explicitly. Teams I've worked with that skipped this exercise ended up paying edge prices for control-plane work that would have been happy in a region.

Get honest about your migration sequencing. Tune Talk lifted legacy onto standard cloud first, then went cloud-native. That intermediate step is not glamorous, but it prevented the classic failure where you try to rewrite architecture and change infrastructure in the same quarter. Boring sequencing beats heroic sequencing.

Instrument before you automate. Agentic RCA is only as good as the signals it consumes. Standardised telemetry with something like OpenTelemetry across every service is the prerequisite. Skip this and your AI agent is guessing.

Define blast radius for every automated remediation. Which actions can the agent take unattended? Which need a human ack? What is the rollback path for an ISSU that goes sideways mid-upgrade? Write these down before production, not after the incident review.

Finally, treat infrastructure-as-a-service monetisation as a real product line, not a slide. Tune Talk's small cell hosts earn passive income. That is a billing system, a settlement system, and a fraud surface. Plan accordingly.

Key Takeaways

• Tune Talk runs Nokia's full 5G Core on AWS, with control plane in the Malaysia region and user plane on Outposts at the edge on Nitro bare metal.
• Subscriber growth went from 50,000 per month in December 2024 to as many as 15,000 per day after the cloud-native rebuild.
• An agentic AI layer on AWS Agent Core Bedrock handles automated RCA, In-Service Software Upgrades without session drops, and elastic capacity.
• The planned 80,000 to 100,000 self-installable small cells create an open-access densification layer and a passive-income model for hosts.
• Greenfield operators just got a working blueprint. Brownfield incumbents just got a problem.