4 SEP 2026 — Batam has around 900 megawatts of installed generating capacity, on the last figure published. Firmus, Nvidia and DayOne are building a 360-megawatt AI factory campus at Kabil on the island, with up to 170,000 Nvidia accelerators through 2027 and 2028 and operations targeted for the first quarter of 2027. DayOne has signed a 450-megawatt power purchase agreement for the site. That is one campus contracting for about half an island's power.
What has been agreed
Firmus Technologies, an Australian company, has a compute partnership with Nvidia running to 2034, anchored on this campus. The accelerators span the Grace-Blackwell, Vera-Rubin and Vera platforms. Firmus contributes HyperCube, its liquid-cooled architecture, co-designed against Nvidia's DSX blueprints; DayOne, headquartered in Singapore, develops and operates the campus.
Firmus has said it expects $25 billion to $30 billion in committed offtake during the partnership's first six years. The Kabil Tech Park campus carries a 450-megawatt power purchase agreement with PT PLN Batam, phased across 2026 and 2027.
Batam is an island grid, not a national one
The ratio only makes sense once you know that PLN Batam is a separate system. It is not connected to the Java-Bali grid, so there is no large neighbouring network to lean on when demand spikes or a unit trips.
An island system has less margin by construction. Reserve is whatever is physically on the island, a single generating unit is a larger share of the total, and adding a load equal to half of installed capacity changes the dispatch problem rather than just the annual energy figure.
Batam's own numbers are older than the projects being planned against them. The last public installed capacity figure, roughly 900 megawatts, dates from 2021. Capacity has been added since, including a 100-megawatt rental gas plant commissioned in stages through 2025. An outsider cannot state the current total precisely, and that uncertainty is part of the difficulty.
A signed supply agreement is a firmer thing than a queue position
We reported this week on the 474 gigawatts of interconnection requests Texas has frozen and is auditing, and the lesson there was that a queue is a register of free options.
A power purchase agreement is not a queue position. It is a contract with a counterparty, a price and, usually, take-or-pay obligations, so it costs money to hold and much more to abandon. When DayOne signs for 450 megawatts it is committing, and PLN Batam is committing in the other direction.
That changes the risk. The Texan problem is phantom demand inflating a plan. The Batam problem is real demand arriving faster than an island grid can be built out, and those need opposite responses from a regulator.
There is one shared exposure, though. A contract signed for 2027 delivery assumes the generation exists by 2027, and a utility that cannot supply is in breach of an agreement rather than merely disappointing a queue. The commitment runs in both directions, which is what makes it worth watching.
The offtake number is the one to watch
Twenty-five to thirty billion dollars of committed offtake over six years is the figure that determines whether any of this gets built at the announced size. It is the customer commitment that justifies the accelerators, and the accelerators are what justifies the power.
It has not been broken down. No customer has been named. The split between AI-native firms, enterprises and software vendors has not been given. And committed carries a range of meanings, from a signed contract to a letter of intent.
That matters for the region rather than for Firmus. A campus that fills gives Southeast Asia a large block of accelerator capacity outside Singapore. A campus that half fills still leaves 450 megawatts of contracted power on an island grid, and PLN Batam's other customers share the consequences either way.
Why Batam and not Singapore or Johor
The geography is the whole proposition. Batam sits about 20 kilometres from Singapore, close enough for low-latency connectivity to the region's densest network hub, and outside Singapore's own constraints on data centre power.
Johor plays the same role from the other side, and we covered the first organised community objection to a data centre there. The Singapore-Johor-Batam corridor is now a single planning problem being handled by three regulators who do not share a grid.
Batam's advantage is that it is a free trade zone with its own utility and a national government that wants an AI hub. Its disadvantage is the same island system that makes it easy to approve things quickly.
Singapore's own moratorium on new data centre capacity, since replaced by a controlled allocation, is what created this corridor in the first place. Capacity did not stop being built when Singapore stopped approving it; it moved twenty kilometres and became somebody else's grid problem.
The questions worth asking now
Three of them are answerable and none has been answered. The first is what Batam's current installed capacity and peak demand actually are, since every ratio in this story depends on figures from 2021. The second is what new generation is contracted to serve the 450-megawatt agreement, because a supply agreement is a commitment to supply and not by itself a plant. The third is what happens to tariffs for everyone else on the island if that generation is built for one customer.
The fourth is water, which nobody has quantified. A liquid-cooled campus at this scale needs a supply, Batam depends on rainfall-fed reservoirs, and the dust-and-water arguments that surfaced in Johor arrived after construction started rather than before.
View author profile → · Editorial policy
