Is the ground under Cikarang's data centers moving?

Cikarang, in Bekasi regency east of Jakarta, now carries most of Indonesia's new data center capacity. It sits on the same young alluvial plain that has made North Jakarta one of the fastest sinking cities on record. We processed two years of free Sentinel-1 radar over the cluster. The answer is that the ground under the halls is flat to within our noise, and that the first version of this page said something else because it sampled the wrong 400 metres.

The questionWhy a data center operator would care

A hyperscale hall does not fail because the whole site goes down by a centimetre. It fails because one part of it goes down faster than the part next to it. Differential settlement racks a slab, pulls a raised floor out of level, and puts a cooling pipe run in tension. The ground movement that matters is therefore a gradient across a few hundred metres, not an absolute number against sea level.

That is convenient, because a gradient is also the thing satellite radar measures best. An interferogram is a difference by construction. What it struggles with is the absolute rate, which is where the atmosphere and the choice of reference point get in.

What we processedTwo years of ascending track 98

Sentinel-1 ascending track 98, bursts 098_210452_IW2 and 098_210453_IW2, every 12-day pass from 1 September 2024 to 10 September 2026. That is 61 epochs. We ordered 119 two-burst interferograms from ASF HyP3, consecutive pairs plus skip-one pairs so the network stays connected if one pass decorrelates, at 10 by 2 looks, about 40 m on the ground. All 119 jobs came back and all 119 were usable.

The stack was inverted per pixel by coherence-weighted least squares over 186,656 coherent pixels. Every interferogram is referenced to the median of 23,553 coherent pixels in a strip south of latitude −6.377, away from the estates. Each site figure below is the median of coherent pixels within 400 m of a coordinate.

The mistakeAn estate centroid is not a building

The first pass sampled one point per industrial estate: MM2100, Jababeka and GIIC Kota Deltamas. Those points gave +4.3, −19.3 and +4.2 mm/yr, and we wrote up the Jababeka point as if the whole estate were sinking 23 mm/yr faster than its neighbours, with the estate's data center tenant named next to the number.

Then we went and found out where the buildings actually are. DCI Indonesia's H1 campus is at a published address with coordinates in PeeringDB. Seven more facilities in the area are mapped in OpenStreetMap with telecom=data_center. Checked against those, the estate points were nowhere near a data hall:

Kota Jababeka covers about 5,600 hectares. A 400 m circle dropped at its centre is a sample of township, not of anybody's slab, and the one we dropped landed on the flank of a subsidence lobe 6 km from the nearest hall. That is the whole error: 23 mm/yr of headline produced by a coordinate nobody had checked. The point says nothing about the estate as a whole, and how much of the lobe falls inside the estate boundary cannot be seen from outside.

The answerEight sites, all flat

Re-sampling the same inverted stack at the verified coordinates gives this. Rates are line of sight, relative to the reference strip, over the same 61 epochs.

The signalSomething is moving, 6 km away

The subsidence is real, it is simply not under the halls. Scene-wide the median rate is −2.1 mm/yr over 152,555 pixels with the 5th percentile at −31.2, so there is a genuine subsidence tail. It sits in a lobe covering the north east of the scene, which is the signature of regional groundwater drawdown rather than of anything built on top. The first-pass point in Kota Jababeka reads −19.3 ± 3.2 mm/yr on that flank. Where the lobe sits against the estate boundary is not something this data can show, because there is no boundary in it.

For an operator inside that lobe the number would matter a great deal. None of the eight verified sites is inside it. Whether a ninth site is depends on its coordinates, which is the point of the exercise.

What held upThe parts of the method that survived

Two internal checks came through the correction intact, and they are the reason we trust the flat answer above rather than assuming the whole stack is noise.

• Distant quiet ground agrees with itself. The MM2100 and Kota Deltamas estate points, 12.2 km apart, differ by +0.1 ± 2.8 mm/yr over two years. Two independent patches of ground at opposite ends of the scene landing that close together is a check on the inversion, not on the geology.
• Differencing removes the atmosphere. Single-site series scatter 8 to 15 mm per pass, most of it tropical water vapour. Difference two sites in the GIIC cluster, 0.5 to 2 km apart, and the residual scatter drops to 1.7 to 4.3 mm. That is the delay cancelling, and it is why a gradient between neighbouring assets is worth more than any single rate.

One caution on our own error bars. The sigmas quoted here come from the least-squares fit and assume each pass is an independent draw. Atmospheric delay is correlated in time, so the true uncertainty on every rate on this page is larger than the number next to it, by a factor we have not yet measured. The differences inside the GIIC cluster reach 6.9 mm/yr with a fitted sigma of 0.6, and we would not defend that as a 12 sigma result.

The limitsWhat this does not say

• The absolute rates are not calibrated. In a separate check this week over western Puerto Rico, tropical and vegetated on purpose, our own rates disagreed with 7 GNSS stations by 54 mm/yr RMS over a six-month window, an implied displacement error of about 45 mm. Spread across the two-year Cikarang window that is still of order 20 mm/yr. Nothing in the table above clears that bar, which is exactly why the table is read as a null result and not as eight precise measurements.
• The reference strip may be moving too. Every number here is relative to a strip of ground south of the estates. If that strip is sinking, every rate on this page is shifted by the same amount.
• One look direction. Ascending track only. Vertical settlement and horizontal movement cannot be separated without a descending stack over the same window. On this geometry a purely vertical motion reads at about 0.8 times its true size.
• No tropospheric correction. No ERA5 or GACOS delay model has been applied.
• A 400 m circle is still not a building. It is much better than an estate centroid, and it is still roofs, car parks and access roads averaged together. A slab-level answer needs the footprint, which needs the customer.
• Some of these sites were building sites. Princeton Digital JC3 broke ground in November 2025 and is due in service in Q4 2026, so for much of this window its coordinate covers earthworks rather than a finished structure. A settling new build and bare ground being graded look different, and two years is too short to tell them apart.
• The list is not complete. Eight sites are what PeeringDB and OpenStreetMap could confirm inside the scene. DAMAC Digital, formerly EDGNEX, has announced a 144 MW campus in GIIC that we could not place to a coordinate, and the larger cluster around 107.25 east sits outside the area we processed.

NextWhat would make this sellable

The correction changed the order of the list. Coordinates come first: this result was wrong for a day because an estate name was treated as a location, and the fix cost one lookup in two public databases. After that, a descending stack over the same two years, which is more free data and another day of processing and which turns line of sight into vertical and east-west. Then a tropospheric correction from ERA5. Then one GNSS tie or a corner reflector on a customer's site, the only item that costs real money and the only one that converts a relative gradient into an absolute rate anyone can audit.

The honest product here is a screening tool. It ranks assets against each other from free archive data in about a day, and it tells an operator which of their sites deserves an instrument. This run is what that looks like when the answer is no: eight sites, two years, nothing moving, and a subsidence lobe 6 km away that would have been worth knowing about if anyone had been building in it.

MethodReproducing this

Sentinel-1 SLC bursts from the Copernicus archive, interferograms from ASF HyP3 running ISCE2, and about 400 lines of Python for the network build, the weighted least-squares inversion and the figures. Coherence threshold 0.3 per pair, mean coherence threshold 0.4 per pixel. Displacements are in line of sight, negative away from the satellite. Site coordinates, their sources, the per-epoch series at both 250 m and 400 m, and every pairwise difference are in cikarang_sites.json in the repository, alongside the job list and urls. We will send the pipeline to anyone who wants to check the arithmetic. Contains modified Copernicus Sentinel data 2024 to 2026, processed by ASF HyP3.