Jakarta’s JK1 expansion makes a case for local AI inference
Equinix is adding about 1,125 cabinets at JK1. For AI buyers in Indonesia, private cloud connections and local network access could matter as much as the space itself.
An Indonesian retailer choosing where to run its product-search model has two decisions to make. One is where it can secure GPU power and cooling. The other is how far each request travels between the storefront, the product database and the model. An inexpensive rack can be a costly choice if every search waits for several network round trips. Equinix’s expansion of JK1 in Jakarta gives buyers a reason to measure that second cost before signing for the first.
On 10 November, Equinix announced plans to add about 1,125 cabinets and introduce liquid cooling in the new phase. Equinix puts its total investment in JK1 at US$94 million. The part of the announcement worth a closer look is already in operation. Connected capacity on the internet exchange inside JK1 grew 46% from the preceding quarter. That measures the capacity participants have connected to the exchange. Actual traffic, cabinet occupancy and booked GPU demand each need separate evidence. Equinix’s Jakarta expansion announcement, 10 November 2026
What the extra cabinets buy
Equinix opened JK1 in May 2025 with 550 cabinets in Jakarta’s central business district. In August 2026 it announced expanded Microsoft Azure connectivity in Jakarta, alongside access to other cloud providers. An enterprise with its customer records in the cloud can therefore test a private route to colocated hardware before committing to a large installation. It can start with a small deployment and measure the path that its application actually uses. JK1 opening announcement · Equinix on Azure connectivity in Jakarta, 3 August 2026
The project also offers an unusually clear distinction between two investment figures. In its 2025 annual filing, Equinix listed 1,125 sellable cabinets, approximately US$39 million of capital expenditure and a fourth-quarter 2026 opening target for JK1 phase two. The November figure of US$94 million covers the facility as a whole. Dividing the phase budget by the sellable cabinets gives roughly US$34,700 per cabinet in the original investment plan. An investor can use that figure to ask how quickly the phase needs to fill. A tenant still needs a quote for power, cooling and connectivity, plus a handover date for its specific deployment. Equinix 2025 Form 10-K, expansion projects
For Equinix, the commercial advantage of extending an operating exchange is that a new customer can reach counterparties already there. For the customer, the value depends on which counterparties it needs. A cabinet beside an exchange offers little advantage if the application still retrieves its data from another region, or if its most important network partner connects elsewhere. Before pricing a cross-connect, identify the cloud region holding the data, the networks serving the users and the routes to both.
How much does a shorter route change a request?
Take a search service that makes three successive calls across the same metro link before it can start returning an answer. The table shows how much response time it could save if a different location shortened each round trip. These are illustrative inputs that buyers can replace with their own measurements.
| Reduction per round trip | Sequential trips per request | Network time saved per request |
|---|---|---|
| 4 ms | 3 | 12 ms |
| 8 ms | 3 | 24 ms |
| 12 ms | 3 | 36 ms |
The calculation is the reduction in round-trip time multiplied by the number of trips that must finish in sequence. Calls that run in parallel only contribute the delay on the slowest required path. GPU execution, queueing and application processing add their own time. A buyer can test the model by replaying the same workload from two sites and comparing time to first response at the 95th percentile during busy hours. If GPU queueing accounts for most of the delay, extra accelerator availability may be worth more than the shorter route. For overnight batch processing, the decision can tip even further towards power cost and the ability to expand.
The alternative east of Jakarta
DCI Indonesia illustrates why the shortlist should extend beyond the central business district. At its H1 campus in Cibitung, DCI lists five live data centres and 75 MW of total IT load. It also lists six global cloud providers at the campus. A buyer comparing H1 and JK1 can therefore test two connected locations. DCI’s campus-wide IT figure describes a different scope from the cabinets in Equinix’s next phase. Both operators would need to specify the power they can commit to the buyer, at what density and from which date. DCI Indonesia’s H1 campus disclosures
PowerlandMap’s public Indonesian inventory on 12 November contained 86 published, non-demo project records. Of those, 67 carried the explicit status “Live” or “Operational”, and 12 of the 67 had a positive recorded IT-MW figure. The unit is a database record, which can represent a building or a campus. The count describes what is documented in this sample. A missing figure is a reason to request an operator commitment when shortlisting a site. The Equinix JK1 record, for example, confirms an operating location while leaving its IT-MW field open. PowerlandMap market coverage
The decision before a lease
For a customer serving live Indonesian users, I would ask each shortlisted operator for two things in the same proposal. First, a priced allocation for the initial racks, including the power and cooling available at handover and the terms for adding the next racks. Second, a test circuit to the cloud and network endpoints the application uses. Run the same peak-hour traffic over both options, and put the measured response time next to the full monthly cost, including the links between sites.
If the central location consistently meets a customer response target that the alternative misses, the buyer has evidence for paying a premium. If both meet it, commissioning dates, expansion rights and total cost can decide the lease. That is the real opportunity in Jakarta’s expansion. The new cabinets give tenants more choice, and the existing networks give them a way to find out what that choice is worth.
*Cover illustration by PowerlandMap. It represents the relationship between computing and interconnection in Jakarta.*
The data behind the analysis
Follow Market Intelligence with sourced, dated evidence
Each analysis draws on qualified supply records, anonymised demand signals and evidence-based matching. The underlying records retain their sources, confidence levels and verification dates.
Related analysis
Norway’s data-centre bottleneck is becoming an allocation regime
Norway’s planned licensing regime and Statnett’s reservation pause turn grid access into a selection problem. The market must separate requested, reserved, connected and IT capacity.
Japan’s Data Centre Market After STTGDC’s ¥93bn Loan
STTGDC’s latest Tokyo financing highlights how Japan’s data centre competition is splitting between established customer locations and projects built around access to power.
Inside Helios: What $990 Million in Annual Revenue Reveals About AI Data Centre Leasing
Galaxy's Helios lease with CoreWeave shows how AI capacity is priced, phased and owned. Only 133 MW of the 526 MW contracted is delivered, and the US$990 million is a full-turnover run-rate, not current rent.
