Indosat Is Buying a Gigawatt of AI With a Telecom's Cash Flow — And Won't Book the Revenue
Q3's EBITDA dip was the price of a full-stack AI pivot: IDR 29T of capex, a token platform, and an accounting choice that keeps the top line clean and the story hard to verify
ISAT.JK · Earnings Call · 2026-10-28
The bill arrives before the business
Nine months into 2026, PT Indosat Ooredoo Hutchison delivered almost exactly what it promised at the top of the P&L: revenue up 11.1% year-on-year to IDR 45.8 trillion, EBITDA up 8.8% to IDR 21.1 trillion, normalized net profit up 10.6%. Then you look at the third quarter alone and the shape inverts — revenue down 2.2% sequentially, EBITDA down 12.4%, margin collapsing five points to 42.7%. The culprit was regulatory rather than commercial. Indonesia's biometric SIM registration regime, live for the full quarter, cut gross adds across the industry. Management's framing is that this is a cleanse, not a shock: “biometric registration is not a headwind, but a positive structural reform” — Syed Kazmi, Chief Operating Officer · 2026-10-28. The underlying evidence mostly supports them — ARPU rose to IDR 46,000 from IDR 39,000 a year earlier, data traffic grew 16%, and home broadband customers were up 21%. A lower-quality base earning more per user is a better business than a large base earning less. What is less clean is the cash. Capex for the nine months hit IDR 29 trillion — 63.4% of revenue — split between IDR 9.7 trillion of 5G and IDR 10.3 trillion of AI cloud spend. Third-quarter capex alone was IDR 19.5 trillion, up 265%. The tell is in the working capital: procurement payables jumped from IDR 9.8 trillion to IDR 20.46 trillion, which the CFO attributed simply to "the timing of payments as we committed to increase our investment into 5G." Net debt nevertheless fell 19.2% to IDR 10.3 trillion, or 0.36x EBITDA, after IDR 11.7 trillion of FiberCo monetization proceeds landed.What changed in the vocabulary
Go back three months. The company's own keyword trail for last quarter was a telco's vocabulary: cellular revenue, ARPU growth, data center only as an adjacency, MIDI revenue, the fiber asset. This quarter the top of the list is unrecognisable — Intelligence Suite, token platform, ZanKore, model builders, open-source hosting, megawatt capacity, the whole value chain. The CEO's framing is explicitly vertical integration: “we are not just providing connectivity, we are systematically building an end-to-end AI ecosystem from energy and data center through compute, model, usable applications” — Vikram Sinha, Chief Executive Officer · 2026-10-28. The stack, as presented: 86,000 km of terrestrial and submarine cable as the connectivity backbone; Indonesian power as a structural cost advantage; data centres via BDx and NexAI; compute via ZanKore with 200MW targeted for the first half of 2027 and 1GW by 2029, of which Indosat itself will take roughly 15MW; then 300-plus hosted models; and at the top the Intelligence Suite, aiming for USD 100 million of revenue by 2028 at a 70% EBITDA margin. Funding is secured for the first 100MW.The revenue that may never be revenue
Here is the detail that deserves the most scrutiny, and it came out only under questioning. When pressed on token economics and revenue share with model providers, the CEO said this: “we will have an asset-light model where we will make sure that we will be only booking margin after deducting all the infra cost. We will not be booking the top line. We will be only looking at margin.” — Vikram Sinha, Chief Executive Officer · 2026-10-28 That is a defensible choice for a capital-intensive compute business — it flatters margins and hides scale. It also means the AI cloud revenue number and the reported revenue line are two different objects, and investors will have to take the margin bridge on faith. Which makes the guidance history worth laying out. In October 2025 the company guided the GPU business to roughly USD 30–35 million for that year, with “more than 65 million to 70 million for full year next year” — Vikram Sinha, Chief Executive Officer · 2025-10-29. By February 2026 the framing was “the outlook for 2026 would be about $50 million to $60 million based on the current contract” — Chi Lee, Chief Financial Officer · 2026-02-09. Today: USD 50 million booked in the nine months, "on track" for the full-year aspiration. Consistent, but only after two re-basings, and the reward for scale keeps getting pushed right — the 2027 line in the presentation reads "more than USD 3 million in recurring revenue," a figure that sits so oddly against the USD 100 million 2028 target that it is either a transcription slip or a remarkably cautious bar.That is the entire bull case in one sentence: three-to-four-fold revenue per unit of the same silicon, if the customer mix skews to enterprises rather than model builders. It is also entirely unproven at Indosat's scale — which is why the company hedged with an adjusted earnings presentation. The CFO's defense: “our underlying 9 months '26 EBITDA would have been IDR 21.6 trillion, increasing by 12% year-on-year” — Chi Lee, Chief Financial Officer · 2026-10-28. Analysts pushed back hard that spectrum and FiberCo costs are recurring, not one-off. The CFO conceded the point — “this will be a continuing obligation for us to pay the annual fee every year” — Chi Lee, Chief Financial Officer · 2026-10-28 — and the maintenance line inside cost of services rose roughly 40% quarter-on-quarter to IDR 1.2 trillion on FiberCo formation. In February, the same CFO had told the same analysts the FiberCo deconsolidation impact would be “low single digit on EBITDA” — Chi Lee, Chief Financial Officer · 2026-02-09. It was low single digit. It just arrived as a cost, not a benefit.When you sell GPU as a service, you sell it on a long term at $6, $7 per hour. But when you are selling token... you can earn up to $15 to $20. And if you are able to convert that into application, if enterprise buys output, it goes as high as $30.