Three delivery models dominate China-sourced data center procurement in 2026: hyperscale self-build on a greenfield campus, carrier-tenant colocation inside China Telecom facilities, and prefabricated modular halls shipped from Shenzhen or Suzhou [S4][S5]. The selection call is driven by IT load (5 MW vs 100 MW), latency to end users, and whether the operator wants a Chinese operating license or a foreign-owned build.
Site geography clusters around four provinces: Inner Mongolia (Ulanqab) and Guizhou for cool-air free cooling, the Yangtze River Delta for latency to Shanghai finance, and the Greater Bay Area (Guangdong) for cross-border to Hong Kong PoPs [S3][S4]. Apple publicly confirmed the Ulanqab Inner Mongolia site in 2018 with a 2020 operation target, an early signal that the cold-climate corridor has been bankable for foreign tenants for almost a decade [S3].
Delivery model comparison: self-build vs colocation vs modular
Self-build on a 10+ hectare campus suits 30 MW and up, with a 24-36 month delivery and a CAPEX-heavy profile [S4]. China Telecom Americas (CTA) colocation is the practical choice for 1-10 MW tenants, since the parent carrier is one of China's largest fixed-line and data center providers and reaches 100+ countries through a Tier-1 backbone [S5]. Prefab modular halls are a faster path, typically 6-12 months from PO to energization, but they cap at around 20 MW per site before grid upgrade risk reappears.
Decision criteria line up as: (1) power scale in MW, (2) months to first load, (3) capex vs opex mix, and (4) the operator's tolerance for ICP filing (the Chinese operating license for internet content). Hyperscale tenants with their own license gravitate to self-build; foreign SaaS firms without an ICP partner usually start in carrier colocation and migrate later.
Site selection: climate, latency, grid carbon intensity
Inner Mongolia and Guizhou hold average annual temperatures near 5-10 deg C, which lets operators hold annual PUE in the 1.15-1.25 band with air-side economizers for 7,000+ hours per year [S4]. The Greater Bay Area trades that free-cooling headroom for sub-10 ms fiber to Hong Kong IXPs, which is why Hong Kong carrier-neutral buildings (iTech Tower, MEGA-i, Equinix HK1) still see Western tenants using them as the front door to mainland networks.
Grid carbon intensity is now a board-level spec, not a CSR line item. Inner Mongolia and Ningxia sit on wind+coal mixes above 500 g CO2/kWh; Yunnan and Guizhou hydropower brings that down below 200 g CO2/kWh in good water years. Buyers with RE100 or SBTi commitments should pin a g CO2/kWh number on the RFP, since 2026 ESG audits in the EU are pushing Chinese landlords to publish facility-level figures for the first time.
Certifications, standards, and what to put on the data sheet
Two certification pairs cover almost every 2026 RFP: Uptime Institute Tier III or Tier IV (concurrent maintainability / fault tolerance), and TIA-942 Rated 3 or Rated 4 for the structured cabling and topology backbone [S4]. Tier IV commands roughly a 25-40% premium over Tier III for the same IT load because of the 2N+1 electrical and mechanical paths. Chinese operators also reference GB 50174-2017, the national data center design code, on top of the Uptime and TIA specs.
For power, spec dual 10-12 kV utility feeds from separate substations, transformer N+1, and UPS in 2N with 10-15 minute battery plus diesel rotary UPS (DRUPS) or genset N+1 for Tier IV. Cooling should call out chilled-water or DX row-oriented units with CRAH redundancy, leaving hot-aisle containment optional only when the hall is below 8 kW per rack. Don't accept an undefined "high-density" label on the data sheet: pin a kW/rack number, typical 12-30 kW/rack for AI training, 6-10 kW/rack for general compute.
Industrial hardware: where the China-sourced parts actually live
China supplies the bulk of the electromechanical layer: switchgear, busways, modular UPS frames, Li-ion battery cabinets, CRAH units, and the linear guide rails and crossed-roller guide bearings inside server slide rails and robotic storage retrievers. Prefab skid manufacturers in Suzhou and Dongguan buy these motion components in volume, so a buyer's RFQ to a skid vendor can pull a useful leverage signal on unit cost for the wider project. [S3]
For the facility monitoring layer, BMS front ends increasingly bundle a pressure transmitter on chilled-water headers, a flow meter on the condenser loop, and a multi-channel data logger at each row to trend inlet temperature, humidity, and door state. Cooling loop control valves are usually industrial valve assemblies sized to the loop's GPM and dp, not generic utility valves. Buyers who write the data sheet to the component (not to a black-box "monitoring system") get cleaner warranty splits when a sensor fails.
Cross-border bandwidth, ICP filing, and data residency
CTA's network reaches 30+ cloud IXPs globally, and the operator's IP transit and CDN services ride ChinaNet, which covers more than 60% of China's wireline and wireless internet infrastructure [S5]. For an enterprise buyer, that translates into a managed cross-border option: SD-WAN, IP VPN, or wavelength between a Hong Kong PoP and a mainland IDC. Latency from Hong Kong to a Guangzhou or Shenzhen data center is typically 5-10 ms round trip; to Beijing, expect 35-50 ms.
Three regulatory gates must be on the project plan from day one: (1) the ICP filing or ICP license (Bei'an for websites, Xu'uke for commercial hosting), (2) the Cybersecurity Law / DSL / PIPL data classification for any workload touching personal information of Chinese residents, and (3) the multi-tier protection scheme (MLPS 2.0) rating for the system, usually Level 3 for SaaS and Level 4 for finance and government. Foreign-owned data centers without a local partner are not legal to operate in China, so the colocation path often sidesteps that problem by renting from a Chinese carrier that already holds the license.
Cost bands and lead-time ranges to put in the model
Wholesale colocation in Tier 1 Chinese cities runs 700-1,200 RMB/kW/month inclusive of power in 2026, with Inner Mongolia and Guizhou at the low end and Shanghai, Beijing, Shenzhen at the high end [S4]. Carrier-neutral retail colocation in Hong Kong is roughly USD 180-300/kW/month and is often used as a staging point before the mainland cutover. Modular hall CAPEX sits around 35,000-55,000 RMB per kW IT, turnkey, for a Tier III hall, with Tier IV pushing 60,000-90,000 RMB per kW.
Lead times vary sharply: shell-and-core plus fitout 18-30 months for a 30 MW campus, 6-12 months for a 5-10 MW prefab modular hall, and 4-8 weeks to deploy a rack inside an existing carrier-tenant room. Power availability from the grid is the single biggest schedule risk; Inner Mongolia and Guizhou have land and wind capacity but weaker transmission, while the Greater Bay Area has transmission but a multi-year queue for new HV connections. Pin a grid energization date with a defined substation name on the RFP, not a generic "utility power available" clause.
Limits, failure modes, and the spec gaps that bite first
The three failure modes that hit 2024-2026 China builds hardest: (1) grid curtailment in Inner Mongolia during winter heating season, which forces operators to lean on diesel and breaks carbon math; (2) cross-border link instability between Hong Kong and the mainland, which makes a single-carrier SD-WAN a poor primary path; (3) liquid-cooling retrofit, where the original air-cooled CRAH plant was not piped for warm-water loops above 40 deg C inlet, capping AI rack density at the 30-40 kW/rack range. Each of these is a paper-stage fix: write the curtailment backup plan into the power purchase agreement, demand dual-carrier cross-border with BGP failover, and oversize the cooling plant header for 45 deg C supply from day one. [S5]
Buyers who treat the data sheet as a binder of component specs (Uptime tier, kW/rack, g CO2/kWh, ms latency, MPa working pressure on the chilled-water loop) rather than a marketing brochure get fewer surprises at commissioning. For paired guidance on AI-era capacity planning, see the Data Center Procurement Strategy: Spec-First Sourcing for AI-Era Builds walkthrough, and for the cabinet-level motion and control hardware that often ships with a prefab hall, the VFD Suppliers in 2026: Specs, Brands, and Sourcing Map rounds out the bill of materials. Track the China Telecom Americas product-promotions page and any Uptime Institute Tier IV award announcements in Inner Mongolia, Guizhou, and Ningxia as the two near-term signals that new capacity is genuinely online, not just announced.