Hello,

On 6 August, Indosat stopped treating AI infrastructure as somebody else’s business. Its new platform targets 1 GW of regional capacity.

Let’s dive in.

Indosat Ooredoo Hutchison launched Zankore by Indosat on 6 August with Ooredoo Group, Nokia and NVIDIA. The partners are targeting 1 gigawatt of NVIDIA DSX AI Factory capacity, with Southeast Asia as the launch market.

The first delivery target is concrete but still forward-looking. Zankore plans approximately 200 megawatts in the first half of 2027, using NVIDIA GB300 NVL72 systems. NVIDIA says its MaxLPS power software can provide up to 40% more compute inside the same facility power envelope. Those figures are supplier claims, not measured production results.

The ownership split matters. Ooredoo Group supplies capital and regional reach. Indosat brings Indonesian infrastructure, customers and execution. NVIDIA supplies accelerated computing and software. Nokia provides the data-centre and network fabric. Zankore is aimed at hyperscalers, governments and enterprises, rather than Indosat’s mobile customers alone.

That puts the operator closer to the centre of the AI value chain. It can sell compute, sovereign hosting, models, connectivity and edge distribution as one commercial proposition. Indosat already has a local-language model asset in Sahabat-AI; Zankore adds the infrastructure layer below it.

The risk is equally clear. A 1 GW target demands capital, power, cooling, utilisation and reliable customer demand. Telecom operators understand regulated infrastructure and national distribution, but GPU economics move faster than radio networks. Success will depend on phased commitments and contracted workloads, not the headline capacity.

What this means for you: Treat Zankore as an operating-model test, not a data-centre announcement. Track who owns customer contracts, capacity risk, orchestration and service assurance. Those control points will decide whether operators capture AI margin or merely provide sites, fibre and power.

Ericsson said on 4 August that it had become the sole global network vendor in SK Telecom’s government-backed Hyper-AI Network consortium. Korea’s Ministry of Science and ICT and National Information Society Agency support the programme, which will build and validate AI-RAN pilot networks.

The programme will demonstrate three physical-AI services through 2027. Its scope includes network slicing, 5G Standalone, distributed computing, service orchestration and AI-based network autonomy. Proposed use cases include patrol robots, autonomous transport and low-power modes for humanoid robots.

Ericsson will contribute its AI-RAN technology, Intelligent Automation Platform and AI-powered rApps. The second phase is expected to move into a real logistics setting at KG Mobility’s Pyeongtaek plant. That would take the work beyond laboratory validation, though it remains a pilot rather than a commercial deployment.

This changes what must be measured. Conventional RAN automation focuses on network outcomes such as energy, capacity and fault resolution. Physical AI adds application deadlines, uplink reliability, mobility continuity and the placement of compute across device, edge and data centre.

Korea is also using the programme to compare vendor architectures and computing configurations. That makes the testbed relevant to future standards and procurement. Operators will need a common way to express application intent without locking every industrial workload to one RAN stack.

The hard question is commercial ownership. A network can support a robot, but the operator still needs a contract model covering compute, latency, assurance and liability. Technical validation alone will not create revenue.

What this means for you: Add application-level service objectives to AI-RAN trials now. Measure task completion, compute placement and recovery behaviour alongside radio performance. If the commercial team cannot price those outcomes, the pilot is still a network experiment.

Globe Telecom’s second-quarter release, published on 4 August, shows how an emerging-market operator is funding the AI transition through its existing network economics. During the first half of 2026, Globe invested ₱26.3 billion in capital expenditure, up 39% year on year. About 91% went to data-related work.

The operator added 878 5G sites and 80,052 fibre-to-the-home lines. Its 5G network reached more than 98% outdoor population coverage in Metro Manila and key cities across the Visayas and Mindanao. It also added 43 non-key cities, bringing 215 towns to at least 80% population coverage.

Globe said it will invest in AI-enabled capabilities to improve customer experience and operating efficiency. It did not break out AI expenditure, workloads or returns. That distinction matters: the financial evidence supports a larger network build, while the AI outcome remains a management commitment.

The wider numbers give Globe room to invest. First-half service revenue reached a record level, EBITDA rose 6% to ₱44.9 billion and the EBITDA margin reached 52.6%. Mynt, the parent of GCash, contributed 28% of Globe’s pre-tax net income.

For operators in Southeast Asia, this is a more common route into AI than a purpose-built gigawatt platform. Revenue from connectivity and digital finance funds capacity first; AI is then layered into customer and operating platforms.

What this means for you: Separate AI-enabling capex from AI-generated value in board reporting. Globe’s disclosure shows the funding base clearly, but not the return from AI. Require named use cases, adoption and unit economics before calling ordinary data investment an AI programme.

Deutsche Telekom reported second-quarter adjusted EBITDA after leases of €11.8 billion on 6 August, up 7.3% organically. Free cash flow after leases reached €5 billion, and the group raised its full-year cash-flow guidance to about €20 billion. It also expanded its 2026 share-buyback ceiling to €5 billion.

At the same time, chief executive Tim Höttges said the operator’s Munich AI computing facility was fully utilised. Deutsche Telekom is discussing additional GPUs with NVIDIA and evaluating the European Commission’s €10 billion AI gigafactory tender, although no decision has been made. Reuters reported the capacity and tender comments after the results briefing.

The tension is useful. Deutsche Telekom can return more cash while considering another capital-heavy infrastructure role. Strong US earnings and European cash generation give it that option; few regional operators have the same balance-sheet support.

Demand has been proven at the current facility, but a European gigafactory is a different risk class. It needs long-term power, customers, public funding and a clear division of responsibility between operator, GPU supplier and software partners.

What this means for you: Do not copy the asset scale. Copy the sequencing: prove utilisation, secure customers, then expand. The next disclosure to watch is contracted demand, not installed GPU count.