CCTV footage in Thailand: storage on disk and tape, and video AI at DC1

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

This note covers what Thailand has and has not decided about keeping CCTV footage in Thai data centers, how much footage the country's state cameras produce, who stores footage today and at what price, and how the investor's first data center (DC1, on the Sikhio and Lam Takhong site in Nakhon Ratchasima) could keep that footage on disk and tape and run video recognition and model training on it. B4L (buildfor.life) prepared it as consultant; the investor would build, own, and run the facility and sell the services. The arithmetic is in model/cctv_storage.py, with results in model/cctv_results.md. Money is in THB, with USD in parentheses at 33 THB per USD, the B4L planning rate. A platform to manage the cameras themselves, and a research program on whether they call home, are in one system for Thailand's state cameras.

In short

• No law or cabinet resolution requires CCTV footage to be stored in Thailand. The reports of October 2026 describe a voluntary label that TCT, an industry association, plans to launch in December 2026 with AIS, True, and NT, and we found none of the binding guidance it expects from PDPC, NCSA, and NBTC.[1282][1283] State agencies are bound only by a 2020 guideline that keeps a central copy of 20% of state cameras in Thailand for at least 60 days, and by budget rules that send cloud spending to the government cloud GDCC first.[1284][1285]
• A central 30-day copy of Bangkok's cameras alone would take 43 PB of disk, far more than BMA stores today. BMA records on recorders spread across the city for 7 to 30 days, and its own data centers are close to full; its newest holds 640 TB.[1286] The 75,476 cameras chosen in 2022 as the 20% imply about 377,000 state cameras, and the Bangkok Metropolitan Administration (BMA) runs 65,015 (our arithmetic).[1287][1288] A camera at 2,048 kbps writes 22.1 GB a day.[1289]
• On a shared Ceph cluster at DC1, 30 days of footage cost about 88 to 107 THB (USD 2.7 to 3.2) per camera a month (our model). All of BMA's cameras fit in 14 racks, 23% of hall B, for 232 to 254 million THB (USD 7.0 to 7.7 million) of equipment. KIRZ sells 30 days at 367 THB, and BMA spends about 1,025 THB per camera a month on its whole operation.[1290][1288]
• Tape costs less than disk for footage kept longer than about 36 to 46 days. A year on 14 days of disk plus LTO-9 tape costs 155 to 251 THB per camera a month, against 880 to 1,030 THB on disk alone (our model).
• The investor may run video AI for agencies but may not train its own models on their footage. A processor that goes beyond its instructions becomes a controller, so training runs inside the agency's tenancy for the agency (our judgment).[1291][1292] A heavy detector costs 932 to 1,299 THB per stream a month on the investor's GPUs, against about 5,783 THB on AWS in Bangkok (our model).[1293][1294]
• The likeliest first customers are the telcos in TCT's pilot and private video services (our judgment). Direct sales to agencies wait for DGA's certification of cloud providers and compete with GDCC, which NT runs on 4,980 million THB (USD 151 million) for FY2027.[1295][76]
• DC1 opens between the second half of 2028 and the first half of 2029, so a pilot in a rented Bangkok rack would come first (our judgment). The offer would price 30 days of 2MP footage at 170 THB per camera a month, 1.5 times the high-case cost (our model).

What Thailand has decided and what is only proposed

The reports of October 2026 trace back to one initiative of TCT, a non-profit industry association, which wants personal data from CCTV and IoT devices stored on cloud infrastructure in Thailand.[1282] It plans to launch in December 2026 with AIS, True, and NT; makers and distributors would declare for themselves where data is stored, on the model of Thailand's No. 5 energy label, and pay about 2 to 3 THB per barcode sticker plus 20,000 THB per camera model a year.[1282] TCT calls the scheme voluntary, says direct legislation is hard, and expects binding guidance from PDPC, NCSA, and NBTC; we found none from any of them.[1282][1296] TCT's news page carried no notice of the scheme on October 30, 2026, and we found no certified product.[1297] A May 2026 headline said the state was "about to issue a rule", which its own article does not support.[1298]

The guideline is the only instrument we found that puts state footage in Thailand, and it binds through budget approval, since CCTV projects above 10 million THB need the national sub-committee.[1284] It names the store's location but not its operator. We could not confirm whether the central store was built, who runs it, or whether the guideline still applies; the copy we read is a provincial upload of 2024.

What this means for who may store state footage

A private Thai operator may store state footage today when an agency buys the service through an allowed route, because no rule reserves it for NT or GDCC (our reading). Agencies spending state budget must use GDCC first and may buy elsewhere only where GDCC cannot yet serve, for security projects, or for large or complex systems.[76] CCTV is not named in DGA's data classes; identifiable street footage probably falls in class 2 or 3 and police footage in class 3 or 4, and classes 3 and 4 "should" stay in Thailand (our reading).[1300] The PDPA does not forbid storage abroad.[1310] If the draft bill passes as written, storing agency data would also need a DE ministry license held by a Thai company.[1303]

How many cameras and how much footage

We found no current official count of Thailand's CCTV cameras, state or private. The counts we found are dated and partial:

The two Phuket sources disagree on the existing count (223 or 233) and on whether 503 is the new total or the number of new cameras. Broken and offline cameras are common, so the footage actually produced is below what the counts suggest (our reading).

We plan on 2,048 kbps, Hikvision's recommended H.265 maximum for 2MP at 25 fps or 4MP at 15 fps; the guideline's floor of 1080p at 10 fps is 768 kbps in the same table.[1289][1284] THAI DATA CLOUD sizes a 2MP H.265 camera at about 2 Mbps and 21.6 GB a day, within 3% of our planning case.[1319] Axis claims its smart codec saves 50% or more on average, with savings from 25% to 99.7% depending on the scene, so measured bitrates should replace these maxima before anything is ordered.[1320]

Who stores footage today and at what price

The state cloud and the telcos

GDCC, the government's central cloud, received 4,980 million THB (USD 151 million) for FY2027, up from 2,264 million THB.[76] NT operates it for more than 4,000 government systems and is adding GPU capacity.[1321] NT also sells "CCTV ON CLOUD" to large agencies without a published price, and Pakkret City Municipality awarded it 3 million THB including VAT for the circuits to its 287 cameras.[1322][1323] NT therefore runs the state cloud, sells cloud CCTV, rents the circuits, and sits in TCT's pilot.

NT adds a cloud camera with 7 days of playback to its home fiber for about 85 to 90 THB a month, AIS charges 99 THB for a 3MP camera with 7 days, and True bundles two cameras with 7 days into a 699 THB fiber package.[1324][1325][1326] None of these pages says where the footage is stored. The only telco offer that claims Thai hosting is AIS MEC with Huawei Cloud, which processes CCTV and face recognition with "all data hosted in Thailand" (vendor claim, no price).[1327]

Thai video and storage providers

KIRZ, a Bangkok fiber operator, publishes the clearest price list. Its Standard tier records on AWS and charges 367 THB a month for a 2MP camera recorded around the clock for 30 days, 481 THB for 60 days, 486 THB for 90 days, and 683 THB for a year.[1290] Its private-cloud service costs 310 THB for 14 days and 370 THB for 30 days, on KIRZ's own storage in Thai data centers (vendor claim).[1328][1329] THAI DATA CLOUD sells object storage for CCTV at 699 THB per TB a month and a T4 GPU server for video AI at 33,210 THB a month.[1319] NIPA Cloud charges 0.77 THB per GB a month plus 2.31 THB per GiB of egress.[413]

Consumer brands store abroad

The consumer camera brands sold in Thailand list no data center in the country. Hik-Connect names the United States, Singapore, Brazil, and Ireland and says some staff in mainland China have limited remote access.[1330] EZVIZ adds Frankfurt, Imou lists the United States, Germany, and Singapore, and Xiaomi keeps "other" regions on Alibaba Cloud in Singapore.[1331][1332][1333] TP-Link's Tapo Care stores clips on AWS without naming a region and costs 99 THB a month.[1334]

Integrators and reference prices

Large state projects go through integrators such as Forth's Genius Traffic System and Samart Comtech (see the Bangkok section). Agencies price equipment against the DE ministry's reference list of June 2021, still its newest published CCTV list, which sets a 32-channel recorder with at least 32 TB of disk at 120,000 THB.[1335][1336] A draft of a new list went to public hearing on October 20, 2026 (search result only, not verified).[1335] The National Anti-Corruption Commission opened an inquiry after a subdistrict paid the reference price of 16,000 THB for a 2MP camera whose market price it puts at 5,000 THB or less.[1337]

A design for footage at DC1

The design leaves the agency's recorders in place and makes DC1 the home of the central copy, the long copy, the evidence copy, and the analytics. The guideline already makes agencies keep 30 days themselves and says the center need not pull every camera.[1284]

Ingest and transport

Cameras keep recording locally, and the agency's command center forwards the selected streams; ONVIF Profile G and Genetec's city-wide pattern let a center pull footage from the edge later and back-fill gaps after an outage.[1338][1339] Nokia's reference design carries camera traffic on fiber rings to city hubs and over an IP/MPLS core to the data center.[1340] Provincial traffic can run to Sikhio directly or through the investor's point of presence at Telehouse Bangkok (connectivity); the last mile stays with the agency.

The streams are modest for a data center: BMA's 65,015 cameras send about 133 Gbps (our model). A reseller's nationwide MPLS list prices 10 Gbps at 45,000 THB and 100 Gbps at 300,000 THB a month, while best-effort business fiber of 1000/1000 Mbps costs 2,499 THB at True and 2,199 THB at AIS without a committed rate.[1341][1342][1343] MediaMTX records streams as fMP4 or MPEG-TS segments, and it and go2rtc are both MIT-licensed.[1344][1345][1346] We size one ingest gateway per 2 Gbps, about 1,000 cameras, plus a spare (our estimate).

Recording software

Agencies keep their video management systems (VMS). Milestone XProtect cannot write directly to S3 because it "expects to write to a file system", so an S3 archive needs a separately licensed bridge such as Tiger Surveillance Bridge.[1347][1348] Genetec stores footage as small G64 files and sells its own cloud archive, and it ends support for its legacy Cloud Archives on January 8, 2027.[1349][1350] For small-business products the investor could run open recorders: Frigate (MIT) keeps all video for a few days and then only the motion around detections, and Kerberos Agents write recordings to the customer's own storage, Ceph included.[1351][1352]

Ceph for the hot tier, tape for the long copy

The hot tier is Ceph's S3 gateway, as in the open stack note, with erasure coding 8+3, 32 TB drives in 30-bay nodes, and an 80% fill (our design).[596] Large segment objects suit object storage better than many small ones, which also waste space on an erasure-coded pool, so the gateways write segments of several minutes (our judgment).[1353][596] Ceph supports S3 Object Lock in governance and compliance modes for evidence.[1354]

Drive workload sets a floor under the hot window. Drives of this class are rated for 550 TB of reads and writes a year.[1355][1356] With erasure-coding overhead and 10% for reads (our estimate), a drive holding fewer than about 19 days of footage exceeds that rating, so a 7-day window costs as many drives as 19 days (our model).

Tape keeps the long copy. An LTO-9 cartridge holds 18 TB and costs 4,477 THB in Thailand, a half-height LTO-9 drive writes 300 MB/s, and IBM's TS4300 with 80 slots and four drives costs USD 54,500.[628][1357][631] A drive finds a clip in about 60 seconds, so one clip comes back within minutes once a drive is free.[1357] The newest 30 days of one copy stay in library slots, older cartridges go to shelves, a second copy goes off site, and cartridges are reused as footage expires (our design). LTFS lets a tape be mounted like a disk, so the archive does not depend on one vendor's backup format.[1358] Above about 36 days of retention with one tape copy, or 46 days with two, 14 days on disk plus tape costs less than disk alone (our model).

Evidence that holds up

A Thai court may not reject electronic data because it is electronic and weighs it by how reliably it was created, stored, and kept unaltered (Electronic Transactions Act, section 11).[1359][1360] Four measures protect that weight (our judgment):

• Cameras and servers take their time from a Thai stratum-1 source, as the guideline requires of state cameras.[1284]
• Exports use the ONVIF export file format, adopted by IEC as IEC 62676-2-32 and signed with RSASSA-PSS and SHA-256; NIST recommends signing at export and again as a file is shared.[1361][1362]
• Every segment gets a hash at ingest, and an incident's segments are copied into a compliance-mode Object Lock bucket and onto WORM tape in two copies.[1354]
• A chain-of-custody log records who exported what and when, which TCSG makes a "should" at its E and H levels.[1306]

Security and certification

State footage would be handled at TCSG level H (our judgment). TCSG warns that storage in Thailand alone is not sovereignty if administration, keys, or support can be reached from abroad, and asks buyers to know who controls the keys.[1306] The agency should hold the keys, and DC1's staff should administer the platform only from Thailand (our judgment). For high-impact systems the NCSC cloud standard, in force since September 10, 2026, asks for ISO/IEC 27017 or CSA STAR Level 2, which the investor would add to DC1's ISO 27001 scope, along with a section 40 processing agreement with every agency.[1306][1291]

Where it sits

All of BMA's cameras at 30 days take 14 racks and 78 kW, 23% of hall B's racks, while every state camera would take 75 racks, more than hall B holds (our model). The core room already holds the Ceph cluster and a tape room (first data center), so a large contract would spread across both (our judgment).

Bangkok as the first case

Bangkok is the largest and best-documented camera owner, so it is the lead case. BMA runs 65,015 cameras at 16,623 locations.[1288] It spent 2.3 billion THB on CCTV in fiscal years 2023 to 2025, more than 80% of it, about 800 million THB a year, on maintenance, data management, communications, and circuit rental.[1288] That is about 12,300 THB per camera a year, or 1,025 THB a month, for the whole operation; we use it as a ceiling for comparison, since storage is only one item in it (our arithmetic).

BMA does not keep its footage in one place. Its cameras record on equipment spread across the city: the fiscal 2027 maintenance tenders describe a mix of DVRs and NVRs with 4, 8, and 16 channels, network switches, viewing stations, and recording servers, all set to the traffic department's master clock.[1363] Recorders that the contractor replaces must keep at least 30 days at the camera's highest resolution and at least 15 frames per second, a clause that recurs in the tender of each zone.[1364][1365] That is the only retention rule we found in BMA's own documents. A district councilor told the city council in July 2025 that the cameras record only 7 days, which is a claim made in a council debate rather than a written policy.[1366] The palace-area system, one of the few whose tender names central storage, has a storage-management server with at least 20 TB of external storage.[1367]

Residents reach the footage through the city. An older district page lists 13 CCTV centers.[1368] Under the request system that replaced the old one around November 28, 2025, a request needs a police report, online requests are filled within 24 hours for at most 6 cameras and 3 hours in total, and longer requests or original files must be collected at one of "all 7" CCTV centers.[1369] In fiscal 2025, 86% of requests were made online.[1370] The record has gaps: the State Audit Office found that BMA had no footage for 2,614 of 8,977 camera points requested between January 2017 and September 2018, 2,367 of them because recorders had failed.[1371]

The new command center had more than 36,000 cameras connected in February 2026, of about 60,000 to be connected, and invites agencies and private owners of cameras facing public space to join.[1372] BMA's new traffic and security control center has a budget of 194.6 million THB for phase 1, and phase 3 is planned at 65.5 million THB.[1373][1374] In 2022 BMA linked 14,968 cameras, its 20% share, to the Interior Ministry's national system, and it warns residents that older footage may be overwritten.[1287][1375]

BMA plans to move recording closer to the cameras. In July 2025 the deputy governor said that about 33,000 of its cameras are analog and more than 18 years old and about 19,000 are digital and more than 8 years old.[1376] He said BMA would move from DVRs and server storage to NVRs or SD cards in the camera, which cost less to run, and keep server storage only where AI processing is needed: crowded places, key and high-risk areas, and events such as Songkran.[1376] The system was built around 2010 on H.264, and operators still work in three viewing programs (Mirasys, Nice Vision, and Gfin).[1376] BMA's own data centers are close to full. An August 2026 strategy deck shows disk use of 93.8%, 81.2%, and 95.4% in DC1, DC2, and DC3, marked on the slide as a crisis level, and a fourth center opened in May 2026 with 640 TB.[1286] The same deck plans an "AI CCTV & Centralized Data Hub" with a single CCTV gateway and AI plate recognition.[1286] BMA's digital plan for 2024 to 2027 names its own DR site and central data center as the core infrastructure and mentions no outside cloud, and the fiscal 2026 budget holds 91.8 million THB to expand the BMA Data Center; we found no budget line for cloud services or for footage storage.[1377][1378]

The camera money goes to upkeep and circuits. BMA's open-contracting data give about 703 million THB of camera items for the traffic department in fiscal 2026: 287 million for maintenance, 165 million for circuit rental, and 251 million for cameras and other items (a sum of the budget lines).[1378] One directly contracted line rents circuits for the standalone cameras' status monitoring at 126 million THB a year.[1378] The fiscal 2027 maintenance tenders, announced between October 8 and 15, 2026, range from 7.3 million THB for the palace-area cameras to 45.6 million THB for the East zone of 9 districts.[1379][1364] Genius Traffic System, 99.9997% owned by Forth, won at least 29 BMA camera projects worth 1,273.31 million THB in fiscal 2015 to 2022.[1380] Governor Chadchart's policy P203 aims for 100,000 BMA and private cameras, with 30,000 private cameras linked to the city's network, a target with no result reported yet.[1381][1381]

Bangkok already runs AI on its cameras. BMA and the police Central Investigation Bureau (CIB) ran face recognition against the CIB's warrant database on Khao San and Silom roads at Songkran 2026; the 2025 trial flagged 256 warrants and led to 7 arrests among more than 725,000 visitors.[1382] BMA has sent images and plates of motorcyclists riding on sidewalks to district offices since August 2024, and BMA and the CIB run a National CCTV Center under a prime minister's directive, piloted in Bangkok.[1383][1384] The city's police cameras are a separate fleet of 9,138 standalone 4G cameras that keep 15 days on a memory card.[1313] Genius Traffic System, a Forth affiliate, held BMA's maintenance contracts of 2019 worth 251.9 million THB and built its 12 management centers for 295.65 million THB, and Samart Comtech installed cameras in parks and public buildings for 359.72 million THB in 2012.[1385][1386]

BMA's AI runs mostly at the edge or on a few servers. In July 2025 it had 557 cameras that detect red-light and crossing violations on the camera and 268 that collect plates, speeds, and traffic counts, and fiscal 2024 funded 100 licenses of server-side AI software.[1376][1376] Because replacing analog cameras is costly, BMA chose software that analyzes existing feeds in the highest-risk areas, and it depends on the police to prosecute what the cameras find.[1370] The follow-through is thin: of the pavement riders caught by AI cameras, only 5.58% had been acted on by August 2026.[1286]

What this means for a central copy

BMA does not hold anything like 43 PB today. It keeps 7 to 30 days on recorders spread across the city, and its newest data center holds 640 TB, about 1.5% of what a 30-day copy of every camera would take (our arithmetic). The table below therefore sizes a central copy that BMA does not run and has not budgeted. The openings we see are narrower and more likely: storage behind the AI sites and the Command Center's connected cameras, an evidence archive that outlasts the 30 days, overflow capacity for BMA's full data centers, and GPU capacity for its software-based analytics (our judgment). None of them has a budget line yet.

All cameras send about 133 Gbps, which two 100 Gbps circuits on the reseller's list carry for 9 to 18 THB per camera a month (our model).[1341] At 90 days, 14 days on disk plus two tape copies cut the cost to 128 to 151 THB per camera a month, but the tape side needs 222 drives; at that size a frame library is the better buy, and we found no price for one (our model and judgment).

Analytics for 10% to 20% of BMA's cameras, the intersections and entry points (our estimate), would need 12 to 24 servers for a light detector, 41 to 82 for license plates, or 34 to 68 for a heavier detector, at 2.1 to 20.5 million THB a month (our model). The larger fleets draw up to 558 kW and belong in hall A, which is built for 1.4 MW of GPUs (our judgment).

What it costs and what it could sell for

The cost covers drives, nodes, ingest gateways, and network at the study's 8% cost of capital over 5 years with 3% to 5% a year for spares, plus power and the facility charge of 4,662 to 8,392 THB per kW a month from server packages. Drives are Seagate's Exos M 32 TB at 58,270 THB including VAT, the price a Thai reseller listed on October 30, 2026.[1387] At 30 days DC1's storage costs about 134 to 163 THB per usable TB a month, against 144 THB in the study's 1 PB example with two tape copies, which left out space, ingest gateways, and the 2026 memory prices (our model).

Four results matter for the business.

DC1 is cheaper than every Thai list price for footage kept 30 days or longer, at scale. At 30 days the rule price of 170 THB leaves a margin of 37% to 48% and is less than half of KIRZ's 367 THB, which leaves a telco reselling DC1's storage about 197 THB per camera a month for transport, its platform, sales, and the camera (our model).[1290] Thai object storage at 699 THB per TB or 0.77 THB per GB costs four to six times DC1's cost for the same footage.[1319][413]

DC1 is not cheaper for small, isolated customers. A municipality of 1,000 cameras on its own cluster pays 352 to 451 THB per camera a month, because erasure coding 8+3 needs 12 nodes however little footage there is, and its circuit adds 45 to 90 THB (our model). Small agencies make sense only as tenants of a shared cluster.

Tape carries long retention. A year of footage costs 155 to 181 THB per camera a month with one tape copy and 219 to 251 THB with two, against 880 to 1,030 THB on disk (our model). At 10 PB a tape copy costs about 12 to 13 THB per TB a month, below AWS Glacier Deep Archive's 59 THB in Bangkok and far below S3 Standard's 743 THB.[638][392]

Evidence is cheap to keep. If 0.5% to 2% of camera-hours become evidence kept 5 to 10 years on WORM tape in two copies (our estimate), the archive costs 3.4 to 23.3 THB per camera a month for BMA and 8.3 to 28.1 THB for a city of 10,000 cameras (our model).

Hard-disk prices are the largest uncertainty. VDURA's index has a 30 TB drive rising from USD 495 in the third quarter of 2025 to USD 1,216 in August 2026; Seagate and Western Digital say their nearline output is sold out through 2027, and an analyst puts the spot premium at 30% to 40%.[1388] At the 2025 index price, 30 days would cost 58 to 75 THB per camera a month; with a 35% spot premium, 104 to 123 THB (our model).

Video AI: recognition as a service and training

What agencies run today

Thai agencies already run video AI, including Bangkok's face recognition with the CIB described above. Phuket Eye received 98 million THB for 503 cameras with face detection, plate recognition, and AI at junctions, and a second phase of more than 1,200 cameras needs another 198 million THB.[1315] The Tourist Police signed a five-year program with Gorilla Technology estimated at USD 50 to 60 million, on Intel Gaudi accelerators, and Gorilla announced a 200 MW "Korat AI Campus" in DC1's province on June 24, 2026.[1389][1390] Airports of Thailand verifies faces at check-in and boarding and logged about 7.89 million vehicle movements by plate recognition in 2025, and the Department of Highways runs incident detection on Motorway M81.[1391][1392] An advocacy analysis counts 914 Tourist Police arrests from AI face and plate recognition without published retention limits or independent oversight, and it reports that the PDPC ordered 1.2 million iris records deleted in November 2025.[1393]

What the law allows

A state agency usually records CCTV under section 24(4) of the PDPA, its public task, and section 4 exempts state security and criminal-justice work by function.[1394][1309] Face templates used to identify people are biometric data under section 26 and need explicit consent unless an exception applies, such as a substantial public interest with safeguards.[1395][1396] Section 21 allows a new purpose only with notice and consent or another law's permission, and model training is likely a new purpose (our reading).[1397] A processor acts only on the controller's written instructions, becomes a controller when it goes beyond them, and faces fines of up to 3 million THB.[1291]

The PDPC's non-binding AI guideline of October 12, 2026, which we read through law-firm summaries, asks for at least a new notice before data collected under section 24 or 26 is reused for training and suggests anonymizing first.[1398][1399] It holds that a model trained on personal data is not automatically anonymous and treats a provider that trains its own models on customer data as a separate controller.[1399][1292] The draft AI Act, open for comment until August 14, 2026, and still a draft on October 19, 2026, would let sandbox participants train on state agencies' personal data only where anonymized or synthetic data is not enough, and could require AI that matters to the country to process data in Thailand.[1400][1401][1402] TCSG AI says the ability to track people is no reason to switch tracking on, and it is silent on training with an agency's footage.[1403] Article 5 of the EU AI Act, the outside reference, bans real-time remote biometric identification in public spaces for policing with narrow exceptions.[1404]

How the investor would offer it

These rules set the product (our judgment):

• Training runs inside the agency's tenancy on the agency's instructions, and the agency owns the data, labels, and weights.
• The investor never pools state footage into its own models or reuses one agency's footage for another.
• Footage is anonymized or replaced by synthetic data where the task allows, using tools such as NVIDIA's synthetic traffic video and the Apache-2.0 federated-learning frameworks NVFlare and Flower.[1405][1406][1407]
• Face recognition is offered only to agencies with a legal basis, at TCSG level H, with templates kept in the agency's tenancy.
• Video, metadata, and AI output stay in DC1 by default; NVIDIA's video search pipeline can run with no data egress.[1408]

Open models and licenses

Ultralytics YOLO is AGPL-3.0, and Ultralytics requires a paid Enterprise license for cloud services and for fine-tuned models used commercially.[1409][1410] RT-DETR, D-FINE, and RF-DETR are Apache-2.0, and so is YOLOX, last updated in June 2025.[1411][1412][1413] Qwen3-VL's 8B and 235B models carry Apache-2.0 weights for video search and description.[1414][1415] CVAT (MIT) and Label Studio (Apache-2.0) label video.[1416][1417] For blurring, deface (MIT) handles faces only, and Meta's EgoBlur (Apache-2.0) handles faces and plates with models trained on footage from its Aria glasses.[1418][1419]

GPU sizing

The planned counts keep NVIDIA's advised 15% headroom.[1293] An 8x RTX PRO 6000 server costs 5.84 to 7.30 million THB with freight and duty (hardware) and 179,000 to 249,000 THB a month with power and space, against USD 5.76 an hour per GPU for AWS's g7e.2xlarge in Bangkok (our model).[20][1294] AWS Rekognition Video at USD 0.135 a minute would cost about 195,000 THB for one camera analyzed around the clock, and a cloud VMS sells plate reading at USD 20 (660 THB) per camera a month.[1421][1422] Our plate-reading cost is above that because it applies the small L4 card's benchmark to a larger card, so throughput should be measured before plate reading is priced (our judgment). Training is small by comparison: RF-DETR Nano trains on the full COCO dataset in about 5 GPU-hours on an RTX PRO 6000.[1423]

Thai license plates: a data gap

Hugging Face held no Thai license-plate dataset and only four small models on October 30, 2026.[1424][1425] NECTEC's test standard NTS 1005.3.1-2564 asks for at least 20,000 plates for real-time tests and 100,000 for offline tests.[1426] A sanctioned project in which agencies label plates inside their tenancies to NECTEC's standard would build what the market lacks; the commercial alternative is an API such as iApp's, from 1,500 THB a month (our judgment).[1427]

Who would buy and how to get in

We rank the routes by how realistic they are (our judgment):

1. The most realistic route is wholesale capacity behind NT, AIS, and True, which run TCT's pilot because TCT says they have verifiable domestic clouds.[1282] A telco reselling at KIRZ's price would keep about half of it (our model). The three run their own data centers, so the investor would compete on price, tape retention, and analytics.
2. The second is a Thai region for camera brands that store abroad, since a brand that wants TCT's label needs a Thai home for its cloud.[1330][1282] TCT speaks of "tens of millions" of devices sending data abroad, an estimate it does not quantify.[1282]
3. The third is private recording sold through partners to condominiums, retailers, and factories, on the open recorders above; the PDPC consulted in April 2026 on a CCTV guideline for housing estates and condominiums.[1428]
4. The fourth is selling to agencies directly, which becomes practical once DGA certifies cloud providers, and where GDCC-first allows exceptions such as security projects and systems GDCC cannot yet serve.[1295][76] Municipalities such as Udon Thani, with 972 cameras recording offline, are candidates for a central copy.[1316]
5. The last is the central 20% store, if it is ever tendered or tendered again. We could not establish whether it exists.

The competitors are GDCC and NT, True IDC, AIS and CSL, THAI DATA CLOUD, KIRZ, and Gorilla's planned Korat campus.[1321][1429][1390] Footage with its analytics fits hall B's offer of data and AI agents in one building (business hosting).

Risks

• There is no mandate: the label is voluntary, and the binding guidance TCT expects does not exist yet.[1282]
• State budgets go to GDCC first, and NT runs GDCC, sells CCTV clouds, rents the circuits to towns, and sits in TCT's pilot.[76][1323]
• Hard-disk supply is tight. Nearline output is sold out through 2027, and spot buyers pay 30% to 40% more.[1388]
• Transport can cost as much as storage for small towns, and the last mile fails: more than half of Pattaya's cameras were unusable in 2020 because their fiber links kept burning.[1317]
• Face recognition by the police draws criticism, and the PDPC has ordered biometric records deleted.[1393] A data center known for state surveillance carries a reputational risk for its other tenants (our judgment).
• A processor answers for the security of what it holds and faces administrative fines of up to 3 million THB under section 40.[1291]
• Thailand paused 166 data-center projects in September 2026 and is setting new rules: seven tiers, minimum requirements, and a pitch on benefits for projects without a construction permit.[315][1305][1430] Applications open in early 2027.[497]
• DC1's first IT load comes in the second half of 2028 at the earliest (first data center), so earlier contracts must start elsewhere.
• If the draft bill passes as drafted, it would require a license held by a Thai company and let officials order a provider to suspend access to a customer's data within 12 hours.[1303][1304]

A staged plan

A pilot cluster costs more per camera than DC1's shared cluster, so its job is to prove the service and win the customer (our judgment). Ceph suggests 4+2 or 6+3 coding when unsure, which needs fewer nodes than 8+3.[596]

What this means for the investor

Thailand has not decided to move state footage into domestic data centers; it has a voluntary label, a 2020 guideline that keeps a 20% central copy in Thailand, and budget rules that favor GDCC. The opportunity is real but smaller and slower than the headlines suggest (our judgment).

Footage is a good fit for DC1's design. Bangkok's 65,015 cameras at 30 days take 14 racks and 232 to 254 million THB (USD 7.0 to 7.7 million) of equipment, and at the rule price of 170 THB per camera a month the same contract would bring about 11 million THB (USD 0.33 million) a month (our arithmetic). The cost of 88 to 107 THB per camera a month is less than a third of KIRZ's retail price and about a tenth of what Bangkok spends per camera today. BMA has not budgeted for such a copy and plans to record at the cameras instead, so in Bangkok the nearer openings are storage for the AI sites, evidence kept beyond 30 days, and overflow from its full data centers (see Bangkok as the first case).

The margin lies in what small providers cannot do cheaply: long retention on tape, evidence on WORM tape, and analytics next to the footage. Plain 30-day storage at scale is a wholesale product for a telco partner (our judgment).

Drives should be bought per contract, because the halls need no change for footage and drives bought early would carry 2026's peak prices (our judgment).

The legal model should be settled before any agency signs. The investor acts as processor, trains only for the agency inside the agency's tenancy, holds no keys to state footage, and offers face recognition only where the agency has a legal basis (our judgment).

Open questions