The Power Behind AI

01

Executive summary

The first wave of Indonesian data centres was a story about land and fibre. The next one is being decided by who can get megawatts connected, on time, in West Java.

1. 1

   The constraint has moved from floor space to megawatts.

   Jakarta’s capacity under construction more than doubled to 395 MW in the first half of 2026, inside a development pipeline of 1,699 MW. PLN already serves 159 data-centre customers with about 1,990 MVA connected, and 99% of that sits in the Java-Madura-Bali system.

   Source: Cushman & Wakefield source; PLN source

2. 2

   PLN’s own projection runs far ahead of today’s load.

   PLN puts installed data-centre IT capacity at about 580 MW in 2026 and projects a power need of 25,297 MW by 2034. Allowing for cooling and other overhead at an average PUE of 1.54, today’s IT capacity implies about 893 MW of facility load; the 2034 figure is roughly 28 times that.

   Source: PLN source; Uptime Institute source; Longhand calculation

3. 3

   Electricity is cheap per kWh; the scarce thing is the connection.

   The I-4 high-voltage tariff was held at about Rp997 per kWh for the third quarter of 2026. At that price a 100 MW hall running at 70% with a PUE of 1.54 pays about Rp941 billion a year for power. The harder problem is lines and substations: the 2025-2034 plan calls for about 48,000 kms of transmission and 108,000 MVA of substations.

   Source: CNBC Indonesia source; RUPTL source; Longhand calculation

4. 4

   Listed exposure is narrow and uneven.

   Colocation was about 94% of DCI Indonesia’s revenue in the first half of 2026. NeutraDC, Telkom’s data-centre arm, was about 1.1% of Telkom’s. The clearest second-order exposure is land: data centres took about 75% of the 85 hectares of industrial land demand at Kota Deltamas in the same half.

   Source: Bisnis Indonesia source; Selular.ID source; Kontan source

5. 5

   The build-out is being paid for with capex and new debt.

   Indosat raised capex 33.6% to Rp13.28 trillion in 2025. DCI Indonesia agreed a Rp17 trillion investment credit facility in May 2026, against total liabilities of Rp2.64 trillion at the end of 2025. Returns now depend on how fast new halls fill.

   Source: The Jakarta Post source; Halo Indonesia source

02

The AI infrastructure chain

AI spending does not land in one place. It passes down a chain, from the model to the chip, the building, the electricity, the plant that makes it, the wires that carry it, the land underneath and the fibre that connects it. Each link has a different owner, a different bottleneck and a different listed company, or none at all. Select a link to see its role.

03

Indonesia’s data-centre market

There is no official census of data-centre capacity in Indonesia, so the numbers below come from the grid operator and from a property consultant, and they measure different things. PLN counts installed IT capacity and connected MVA. Cushman & Wakefield counts Jakarta’s capacity by stage of development. I keep them apart rather than add them.

The pipeline grew about 140% in a year, from 709 MW to 1,699 MW, while operating capacity barely moved. It is now more than five times the capacity actually running. Most of what is planned is not yet a building, which is why the map and the timeline below separate what exists from what has only been announced.

The map makes the concentration plain. Almost every operating hall and most of the construction sit in one corridor, from central Jakarta east through Bekasi and Cikarang to Purwakarta, inside the Java-Madura-Bali system that carries 99% of PLN’s data-centre load. Batam and Bintan are the only other cluster, built on their fibre links to Singapore, and Surabaya is a single outpost. The largest numbers on the map are also the least certain: the 1,000 MW design at Bintan has no timetable, and the two biggest AI commitments, CoreWeave and Zankore, have not named a site.

Geographically the market is Greater Jakarta first, with Batam as the second hub because of its fibre routes to Singapore. The hyperscale cloud regions of Google (2020), Amazon (2021) and Microsoft (2025) all sit around Jakarta. The newer AI-oriented announcements, CoreWeave’s 360 MW for 2028 and the Zankore platform’s first 200 MW for 2027, are not yet tied to a disclosed site.

04

The power bottleneck

Why does AI need so much power? A chip turns electricity into computation and heat. An AI accelerator draws far more than an ordinary server, it runs close to flat out, and every watt of heat has to be removed by cooling that uses more electricity again.

1. AI computeAccelerators and servers; the IEA expects their electricity use to grow about 30% a year to 2030 source
2. Data-centre loadIT load plus cooling and losses, captured by the PUE ratio
3. MWDemand at a moment: what the connection must carry
4. MWhEnergy over time: MW multiplied by hours in use
5. Power generationPlants that must produce those MWh around the clock
6. Grid capacityLines and substations that deliver them to the site

Globally, data centres used about 415 TWh in 2024, about 1.5% of the world’s electricity, and the IEA expects about 945 TWh by 2030 source source. Indonesia is small in that total, but not small against its own grid. PLN sold 317.69 TWh in 2025, and it named data centres among the drivers of 5.4% growth in business-sector sales source. The calculator below turns a hall’s size into energy and cost, with every assumption on view.

Three paths to 2030

The scenarios below start from PLN’s 580 MW for 2026. The base case follows PLN’s own path to 3.5 GW by 2030, which works out to about 56.7% a year. The bear and bull cases are mine, and each lever is shown so it can be argued with.

On PLN’s own path, data centres would use about 32 TWh a year by 2030. That is about a tenth of everything PLN sold in 2025, from an estimated 1.7% today (Longhand calculation, 580 MW at 70% utilisation and a PUE of 1.54). The grid question is not whether Indonesia can generate that much in total. It is whether it can deliver it to a few districts east of Jakarta within four years.

05

Where the money flows

Nobody publishes how data-centre spending in Indonesia splits between power, grid, land, connectivity, cooling and the building. I did not find a reliable breakdown, so the diagram below shows where the money goes, not how much.

What can be observed is how much capital a megawatt attracts, and what a megawatt pays for power each year. Announced investment per MW ranges from about US$8 million to US$16 million across three projects, a spread that depends on whether the budget includes IT equipment, which the announcements do not say. At the sourced tariff, one MW of IT capacity at 70% utilisation and a PUE of 1.54 buys about Rp9.4 billion of electricity a year, around US$0.5 million.

06

Listed company exposure

This is a map, not a ranking. The companies are listed in chain order, from the data hall outward. No scores are given and nothing here is a view on the shares. Where a figure is not in the sources I reviewed, the table says so. Select a company for its financial history.

Source: company reports and press as cited in each company panel. DSSA, also named by BRI Danareksa as directly exposed source, is included from its 2026 public expose.

07

Financial analysis

The charts compare the same measure across companies. Units differ (Pertamina Geothermal reports in US dollars), so each company has its own small chart rather than a shared axis. Years without a verified figure are marked n/a.

Source: company reports and press as listed in section 13. Growth and margins are Longhand calculations from reported figures. The five-year view shows FY2021 to FY2025; FY2021 and FY2022 come from each company’s annual reports, and years a company did not report read n/a.

08

Valuation

Different businesses call for different yardsticks. A developer with land is read on book value, a tower company on EV/EBITDA, and a company still building capacity on a discounted cash flow. Table 2 names the method used for each company and why.

Prices in the table are closing prices on 24 September 2026, adjusted for share splits, and each can be changed. Share counts are shares issued at 31 December 2025; for Mitratel and XLSMART, treasury shares are not deducted. Figure 9 puts the same multiples on each year-end since FY2021, using the year-end close and that year’s reported figures.

Source: reported FY2025 figures and share counts as cited in section 13; closing prices on 24 September 2026 (Yahoo Finance), editable. EV uses net debt from the audited statements. Pertamina Geothermal reports in US dollars, so its market value is converted at Rp17,916 per US dollar (24 September 2026) before comparing it with its accounts. Puradelta and Jababeka do not report EBITDA, so they are read on P/B and P/E. Multiples are Longhand calculations.

A discounted cash flow you can change

A DCF suits businesses whose cash flows follow from capacity and contracts: DCI Indonesia, Mitratel and Pertamina Geothermal. The model is kept simple so each step can be checked. It projects five years, then a terminal value.

09

Scenario analysis

Three sets of assumptions, applied to the same three companies. Each lever maps to something in the chain: data-centre growth, utilisation, power cost, capex and margins. The outputs are what the model says under those assumptions, for the year 2030. They are not forecasts, and they are not a view on which shares to own.

10

Monte Carlo

Three scenarios hide how wide the outcomes are. The simulation below draws each input from a range, runs the same DCF 10,000 times and plots the spread of the model’s enterprise value. It uses a fixed seed, so the same inputs always give the same result.

The dashed line marks today’s market enterprise value, so each run can be read against the price. With the default ranges and seed, the three companies land in very different places. For Mitratel, the market value sits almost exactly on the median: about half the runs are higher, so the price assumes roughly the base case. For Pertamina Geothermal, about one run in six is higher, so the price already leans toward the upper end of the ranges. For DCI Indonesia, the market value of about Rp481 trillion lies above every run, about 34 times the model’s median. The price assumes growth far faster, or lasting far longer, than even the bull preset. That gap is the main thing a reader has to take a view on.

Simulation outputs are conditional on model assumptions and should not be interpreted as forecasts.

11

Risks

The risks are mapped to the links of the chain they act on, not scored. A score would suggest a precision the evidence does not have. Select a risk for the evidence, the mechanism and what to watch.

12

What to watch

Ten indicators decide whether the scenarios above drift toward the bear or the bull case. The latest reading is shown where a sourced one exists.

13

Sources and methodology

Data collection

The research was carried out up to 24 September 2026. Priority went to official and company sources: PLN, the Ministry of Energy, company releases and filings. Company financials for FY2021 to FY2025 come from annual reports, audited statements and the June 2026 interim statements (for balance sheets at 31 December 2025). Where those were not available, I used press reports and checked each material figure against at least one other outlet where one existed. Figures that could not be confirmed, or that conflicted across sources, are marked unavailable rather than chosen.

Financial calculations

Figures are as reported, in the company’s reporting currency. Growth rates and margins are calculated from the reported figures. Where a figure is derived rather than reported (for example Mitratel’s EBITDA from its reported EBITDA margin), the company panel says so and shows the method. Free cash flow is operating cash flow less capex, and is shown only where both were found. ROE and ROIC are not shown: consistent equity and invested-capital figures were not available for enough years to calculate them properly.

Valuation assumptions

Multiples use FY2025 reported figures and the share counts cited. Market prices are closing prices on 24 September 2026, adjusted for splits; the reader can change them. The DCF uses a five-year projection, free cash flow of (EBITDA − capex) × (1 − 22%), a Gordon-growth terminal value, and the reader’s WACC. Net debt enters only where it was reported or can be derived from a reported ratio; elsewhere the model shows enterprise value only.

Scenario methodology

The power scenarios compound PLN’s 580 MW for 2026 to 2030 at the stated growth rate, then apply utilisation and PUE to get energy, and the tariff to get cost. The company scenarios set the DCF inputs to three coherent sets of assumptions. All scenario inputs are Longhand assumptions and are shown in full.

Monte Carlo methodology

Each run draws revenue growth, EBITDA margin, WACC, terminal growth, a utilisation index and a power-cost change from triangular distributions (low, most likely, high) set by the reader. Runs where WACC does not exceed terminal growth by at least one percentage point are redrawn. The generator is a seeded Mulberry32, so results are reproducible for a given seed and set of inputs. The page runs 10,000 draws in the browser; the percentiles are read from the sorted results.

Limitations

• Market and project figures were mostly read through press coverage rather than the original documents.
• Market capacity figures come from different organisations using different definitions (IT MW, MVA connected, development stages) and are not additive.
• No reliable split of data-centre spending by component exists for Indonesia, so section 05 is a framework, not a measurement.
• Prices come from a market-data provider (Yahoo Finance), not an exchange closing file, and a single day’s close can move far from the average.
• The DCF ignores working capital, leases and minority interests, and treats depreciation as equal to capex. It shows how the inputs move the result; it is not a valuation opinion.
• Some five-year series still have gaps: capex, operating cash flow and net debt are not reported in every year for every company.

Source registry

Accessed . Each entry states what this report takes from it.