Noise at the first data center: what the neighbors and the staff would hear

Research note, October 30, 2026, Alexander Grabowski (B4L). Sources retrieved October 30, 2026; figures cited, our estimates marked.

This note covers the noise that the investor's first data center (DC1, on the Sikhio and Lam Takhong site in Nakhon Ratchasima) would make: the rules it has to meet, the sound power of each machine, the levels neighbors would hear by day, at night, during generator tests and in a grid outage, the exposure of the staff, and what a quieter plant would cost. B4L (buildfor.life) prepared it as consultant; the investor would build, own, and run the facility. The calculation is model/dc1_noise.py, with results in model/dc1_noise_results.md. It takes the position of every machine from the 3D layout of the first data center, the generator and cooling design of facility systems, and the C1 site of site selection. Money is in THB, with USD in parentheses at 33 THB per USD, the B4L planning rate.

In short

• The binding limit is the Thai annoyance rule at night. Between 22:00 and 06:00 a sound is rated as a 5-minute LAeq plus 3 dB(A), and it is a nuisance when that rating exceeds the background LA90 by more than 10 dB(A), so a steady plant must stay below about the background plus 7 dB(A) at the nearest home.[1546][1547] On a quiet rural night of 35 dB(A) (our estimate) that allows 42 dB(A); on the 40.5 dB(A) measured at the quietest house near a Chaiyaphum wind farm, 47.5 dB(A).[1548]
• The ten dry coolers make most of DC1's noise. At full speed they emit about 110 dB(A) of sound power, as much as the three generators together, because each 12 m unit is rated at 77 dB(A) at 1 m (our model).[1549] Fan sound falls by 50 lg of the speed ratio, so a 60% speed cap at night takes 11 dB off them at no cost.[1550]
• As drawn, DC1 meets the Thai night rule at a home 150 m outside the fence in its loudest direction, and 360 m away during a grid outage at night, on a 35 dB(A) background (our model). On the wind-farm background those distances shrink to 40 and 180 m. The nearest buildings in OpenStreetMap stand 1.6 to 1.9 km from the fence and would hear less than 26 dB(A), but the parcel is not chosen and rural homes are often missing from the map.
• Four measures costing about 2.9 to 9.4 million THB (USD 88,000 to 285,000) bring normal night running within the Thai rule at the fence itself, and cut the outage setback from 360 to 220 m (our estimate). They are a 60% night speed cap, low-noise dry coolers with larger and slower fans, the 40 dB(A) exhaust silencer that mtu offers for the reference set, and generator tests by day on a published schedule.[1551][1552]
• The generator enclosure grade decides more than any wall. The mtu Ultra Quiet Enclosure gives 76 to 77 dB(A) at 7 m for sets of this size, against 88 to 91.8 dB(A) for the weather-proof enclosure with a sound kit, which would add 6 to 9 dB at the fence (our model).[1553]
• Bangkok's only data-center noise complaint so far was about a generator tested at night, and operators there must now give notice of test schedules.[1554] At a home next to the Telehouse data center, PCD measured 61.5 dB(A) over 24 hours and an annoyance level of 2.3 to 6.1 dB(A) in September 2026, within every limit.[1555]
• Air-cooled hall B would sit near 91 dB(A) with the phase-1 fit-out, above the 85 dB(A) at which Thai rules require a hearing conservation program (our estimate).[1556][1557] No night background has been measured near Sikhio, so a survey of at least 48 hours at the candidate parcel and its nearest homes should come before the land is bought.[1558]

The rules: Thai law and the lenders' standards

The general standard of the National Environment Board (Notification No. 15, 1997) caps the 24-hour average at 70 dB(A) and the maximum at 115 dB(A).[1559] DC1 will not come near either; our model puts its own 24-hour level at the loudest fence point at about 54 dB(A). The binding limit is Notification No. 29 (2007), under which a sound whose annoyance level exceeds 10 dB(A) is a nuisance.[1547]

The Pollution Control Board's method of September 21, 2022 sets how that level is calculated.[1546] The background is the LA90 over at least 5 minutes with the source silent; for a source that never stops, it is measured at a similar place the source does not reach. Between 06:00 and 22:00 the source is rated as a 1-hour LAeq, so a 30-minute generator test is spread over the hour. From 22:00 to 06:00, or near temples, schools, and hospitals, it is rated as a 5-minute LAeq plus 3 dB(A), and impulsive, shrill, or vibration-type sound adds 5 dB(A).[1546] The method has no tonal penalty, but its vibration-type sound includes machinery whose sound comes with vibration, so an inspector could add 5 dB(A) to a generator's rumble (our reading).

Data centers are not classed as industrial facilities in Bangkok, and an environmental-law group says they fall outside the Factory Act and need no environmental impact assessment (EIA).[1554][1560] The EIA list starts at 10 MW for thermal power plants and has no data-center category, and in September 2026 ONEP was studying whether data centers should need one.[1561][1562] BOI's release of August 6, 2026 lists noise among its screening points, and on October 9, 2026 an MNRE subcommittee discussed a draft code of practice covering noise control, both without numbers.[1563][1564] We found no rule on generator test hours outside Bangkok.

Lenders will apply the IFC guideline: 55 dB(A) by day and 45 dB(A) at night as a 1-hour LAeq at homes, or no more than 3 dB above the existing background.[1558] WHO set 45 dB LAeq outside a bedroom at night in 1999 and a 40 dB Lnight target in 2009.[1565][1566] The British BS 4142 treats a rating 10 dB above background as a likely significant adverse impact and adds up to 6 dB for tones.[1567][1568]

Abroad, regulators are moving toward studies before approval. Virginia's legislative auditors recommended noise modeling and low-frequency limits, and Chandler, Arizona, requires a baseline sound study and a design that keeps a data center below the ambient level found.[1569][1570]

The sources and their sound power

The model takes the full build. Phase 1 has two generator sets instead of three, so its generator scenarios are up to 1.8 dB quieter on the generator part.

The dry cooler is made in Thailand by CIG BluSolutions, whose page gives "77 dB(A) @ 1m" for every model without saying whether it means one fan or the whole unit.[929][1549] Read per fan, the 12 m unit with twenty fans has 98.0 dB(A) of sound power; read per unit over a box 1 m from the casing, 99.7 dB(A). The model uses the higher reading and shows the other as a sensitivity (our arithmetic).

mtu gives 76.2 dB(A) at 7 m for the 12V4000 DS1500 and 77.2 dB(A) for the DS1750 in the Ultra Quiet Enclosure, at full load and without the exhaust.[1553] An unsilenced exhaust gives 120 to 130 dB(A) at 1 m, and silencers for this size cut 10, 30, or 40 dB(A) at mtu and 10, 20, or 35 dB at Caterpillar.[1571][1551][1572] Each set is three points: casing, radiator discharge, and exhaust. No data sheet in our evidence gives directivity, so every source radiates equally in all directions, and turning a set cannot be valued yet. The chillers are Carrier 30XA units rated at 89 dB(A), the central inverter measured 72 dB at 10 m, and a UK battery assessment used 96 dB(A) per power conversion unit and 82 dB(A) per container air conditioner.[1573][1574][1575] None of these sources gives octave bands, so each type gets a generic spectrum (our estimate), listed in the results file.

The method

The model applies ISO 9613-2 per octave band from 63 Hz to 8 kHz: spreading, air absorption, the general ground method, screening, and reflections. The standard predicts levels downwind or under a moderate night inversion, the conditions that carry sound best, with an accuracy of about plus or minus 3 dB between 100 and 1,000 m and no estimate beyond 1 km.[1576] Air absorption is computed for 30 °C and 75% humidity; at 70% the code reproduces the standard's table. Farmland is soft ground and paved yards near the sources are hard. The 7 m building screens sound over its roof and around its corners, proposed walls and berms screen it too, within the standard's caps of 20 and 25 dB, and the building's walls reflect it. Receptors stand 1.5 m above ground, as the Thai method and the IFC ask.

The receptors are the fence every 10 m, the public road on the south side, rings 100 to 1,000 m outside the fence, and the nearest buildings, villages, and quiet places in OpenStreetMap. The plot is centered on the C1 point, so every distance to a home is indicative until a parcel is chosen. As a check, one set in the Ultra Quiet Enclosure gives 41.5 dB(A) at 300 m from spreading alone, matching the research stream's hand figure of about 41 dB(A), and 38.7 dB(A) over farmland with air absorption.

The C1 weather series puts the 379 hours a year above the dry coolers' 34.5 °C limit between 09:00 and 19:00, so the trim chillers are a daytime source (our model). At night the air is below 29.2 °C in 99% of hours, about twice the design temperature difference, so the fans can slow down (our judgment). The base case caps them at 80% speed at night.

Results per scenario

The table gives the plant's own level at the loudest point of each set of receptors, and the Thai annoyance level at the loudest point 300 m outside the fence for the two backgrounds.

Levels are in dB(A). The monthly test is NFPA 110's 30 minutes at 30% of nameplate or more, as in the design, and its hourly rating at 300 m is 42.2 dB(A).[713] The fence maximum always falls on the east fence, which is 65 m from the cooling pad and 62 m from generator yard B.

Normal running at night (scenario 2), dry coolers at 80% speed: plant level in 5 dB(A) bands at 1.5 m, ISO 9613-2 downwind (our model). Dotted lines run 100, 300, and 500 m outside the fence.

Grid outage at night (scenario 6): three generators at full load, dry coolers at full speed, and the trim chillers on (our model).

Night with the recommended set: dry coolers capped at 60% speed and of a low-noise model (our model). The battery's power conversion becomes the loudest source.

What it means for the nearest homes

The table gives the smallest distance outside the fence at which the loudest direction meets each criterion, for the design as drawn and then with the recommended set (our model).

By day a home 200 m out meets the Thai rule even on a 35 dB(A) background, and the real daytime background near a highway is likely higher (our judgment). At night, normal running needs 150 m on the quiet background and 40 m on the wind-farm one. The IFC's +3 dB test is the strictest: on a 35 dB(A) background it lets the plant add no more than 35 dB(A), which takes 380 m at night. A louder background shrinks every distance.

A monthly test by day needs 210 m on the quiet background. A test at night would face the 5-minute rating and the 3 dB(A) night addition, and Bangkok's one complaint was exactly that case, so tests belong in the daytime (our judgment).[1577] An outage at night needs 360 m. The facility systems note expects one or two sustained grid interruptions a year, so this is rare, but we found no exception for it in the rule.

Direction matters as much as distance. The building and plant sit east of the plot's center, so at night, 100 m outside the fence, the east and south sides hear about 44 dB(A) and the west side 34 dB(A) (our model). Once the parcel is chosen, the plant should go to the end farthest from the homes. Moving the whole block to the middle of the plot would cut the fence maximum by 4.6 dB but only 1.3 to 1.7 dB at 300 m.

The nearest buildings in OpenStreetMap stand 1.55 to 1.9 km north of the fence along Highway 2, most of them unnamed, with the village of Ban Loet Nimit 1.7 km to the north-northwest and a school 2.5 km to the northeast. None would hear more than 26 dB(A), even in an outage, though ISO 9613-2 gives no accuracy estimate that far out. The map holds only six buildings within 2 km of the C1 point, which says more about the map than about the farmland (our judgment).

Two local conditions push the levels up. Flooded paddy fields in the rainy season turn the ground hard and raise the level at 300 m by about 6 dB, and a raised house or upper floor at 4 m adds about 2 dB (our model).

Tonal and low-frequency noise

Complaints about data centers abroad are more often about tones and hum than loudness. At a Compass data center in Loudoun County, Virginia, neighbors complained at levels of 25 to 35 dB, within the county's 55 dB limit, and the cause was a tone from the coil and fan support of one fan.[1578] Transformers hum at 100 and 200 Hz on a 50 Hz grid, fans produce a tone at their blade-passing frequency, and generators carry their energy in the lowest bands, which mtu says only massive barriers stop.[1579] WHO asks for a frequency analysis when dB(C) exceeds dB(A) by more than 20 dB.[1565] In our model the difference at 300 m is about 10 dB at night and 14 dB in an outage, growing to 13 and 18 dB at 1 km; the spectra are generic, so the trend matters more than the numbers.

The remedies belong in the purchase contracts (our judgment). Dry coolers and chillers should come with octave-band sound power at full and part speed, fans chosen for no tonal prominence, controls that skip any speed band where fans or supports resonate, and fan supports isolated from the coil frames, the fix at Compass. Siemens lists a 2,500 kVA low-loss transformer at 62 dB(A), 9 dB below the catalog's conventional 3,150 kVA unit.[1580][1581] The acceptance test at the fence should include a one-third-octave check against the BS 4142 tone test.[1568]

Staff inside: halls, yards, and hearing protection

Thai rules allow 85 dB(A) over 8 hours with a 3 dB exchange rate, so 88 dB(A) for 4 hours.[1556] At 85 dB(A) or more, the employer must keep a written hearing conservation program with hearing tests at the start and every year, and post noise maps and signs.[1557]

Hall B is the problem. NVIDIA rates one air-cooled DGX H100 at 98.7 dB(A) of sound power, says such systems can exceed 98 dB at 1 m, and makes hearing protection mandatory near them.[1582][1583][1584] A Korean field study of an air-cooled hall found staff exposure often above 85 dB(A).[1585] With the phase-1 reference fit-out of eight air-cooled GPU servers and 20 colocation racks, we estimate about 91 dB(A) in hall B, and about 95 dB(A) if it fills with GPU servers. Four hours a day there gives 88 dB(A) over 8 hours, and one hour behind the GPU servers 91 dB(A) (our estimates). Hall B therefore needs the full program from opening day: protection at the door, noise maps, and yearly audiometry.

Hall A, cooled by liquid, should stay near 75 dB(A) (our estimate); no maker publishes figures for liquid-cooled racks, and Supermicro's claim of about 50 dB states no metric.[1586] In the yards, half an hour in front of a running set during a test gives about 78 dB(A) over the day, and four hours near running sets in an outage about 82 dB(A); both are under the limit, but ear protection should be worn whenever a set runs (our estimates).

Mitigation and cost

The research found no price for silencers, enclosures, walls, berms, or low-noise coolers, so every cost is our estimate and needs a quote. Changes are in dB at the loudest point, at the fence and 300 m outside it.

The recommended set is the 60% night cap, low-noise dry coolers, the 40 dB(A) silencers, and generator tests by day only, with the schedule sent to the village head and the district. It costs about 2.9 to 9.4 million THB (USD 88,000 to 285,000), 0.2 to 1.1% of phase 1's facility capex of 0.86 to 1.27 billion THB (our estimate). Normal running at night then falls to 42.0 dB(A) at the fence and 29.6 dB(A) at 300 m, and an outage to 39.6 dB(A) at 300 m.

The silencer is insurance. If the unsilenced exhaust is 130 dB(A) at 1 m, the top of Cummins' range, an outage would be 4.3 dB louder at 300 m with the 30 dB(A) silencer and no louder with the 40 dB(A) one (our model). mtu warns that the last 10 dB(A) of generator quieting can be prohibitively expensive, and Cummins that silencer and enclosure grades are not standardized, so a specification that only names a grade does not bind the seller to the owner's limit.[1579][1571] The specification should state the limit at the fence. Walls around the yards must be absorptive inside, because one reflecting wall beside a set adds 3 dB(A), and the IFC asks for barriers without gaps of at least 10 kg per m².[1571][1558] Once the dry coolers are quiet, the battery's power conversion becomes the loudest source at night, so its suppliers should give a sound power figure at night load, and its yard should sit away from the homes (our judgment).

What to measure before buying the land

No night background has been measured near Sikhio. The nearest evidence comes from wind farms in the region: a 24-hour LA90 of 40.5 to 41.5 dB(A) at a house near Hanuman 1 in Chaiyaphum, a 24-hour Leq of 48.9 to 58.8 dB(A) at seven villages near EGCO's Chaiyaphum farm, and 43.0 to 47.8 dB(A) near a wind plant in Korat.[1548][1587][1588] None gives a night LA90, and the 5.5 dB between our two backgrounds moves every setback in this note. The survey should:

1. Log sound for at least 48 hours, day and night, at 1.5 m, at the candidate parcel and the nearest home in each direction, as the IFC suggests.[1558]
2. Record LA90 and LAeq with class 1 meters per the PCD method, in 5-minute periods at night; measuring at the homes before building starts settles the background before anyone argues about a "similar place."[1546]
3. Repeat in the other season, with dry fields and with flooded paddies, and note the wind direction.
4. List every home, temple, school, and clinic within 1 km; temples and schools get the 3 dB(A) addition by day too.
5. Note existing sources, such as Highway 2 and the Sikhio wind farms that site selection lists.

What this means for the investor

Noise will not stop DC1 on a rural Korat parcel if the plot leaves about 150 m between the plant and the nearest home at night, or about 360 m for the outage case on a quiet background. The recommended set brings the outage distance down to 220 m for under 10 million THB (our judgment). The background and the direction of the homes are the numbers to watch; the survey settles the first, and the layout can be turned to answer the second before the design is frozen.

The cheapest measures are decisions: a night speed cap on the dry coolers, generator tests by day with notice to the neighbors, the plant at the far end of the parcel, and noise limits written into every purchase specification. The Ultra Quiet enclosure class should stay in the generator specification. Inside, hall B needs a hearing conservation program from opening day, a running cost of training, audiometry, and protection, and moving GPU load into the liquid-cooled hall A would lower it.

Open questions