China AI Datacenter Buildout

China’s construction of AI datacenter capacity at a pace and scale second only to the United States — roughly 24 GW of operational capacity as of 2026 (against ~56 GW in the US, ~15 GW in the rest of Asia, ~14 GW in EMEA), with another ~50 GW under construction or announced — organized around a distinctive industrial-policy geography (the “Eastern Data, Western Compute” hub system), a construction-speed advantage (100 MW facilities routinely delivered in under 12 months, versus 18–24 months in the US), and a hard external constraint (US export controls on advanced chips) that makes the buildout chip-gated rather than power-gated. ^{1 2}

The buildout matters for three separate reasons that are easy to conflate: it is the physical foundation of China’s AI model ecosystem (ByteDance’s Doubao, DeepSeek, Moonshot’s Kimi, GLM — every one runs on this fleet); it is one of the largest directed capital-formation programs in the world right now (BAT capex hit $20B in 2Q26 alone, with all three of Alibaba/Tencent/Baidu posting negative free cash flow for the first time on record; the 15th Five-Year Plan layers grid capex to over $746B); and it is a live test of whether Chinese-style directed buildout produces durable infrastructure or another round of ghost-asset overcapacity — the same question Nathan flagged in the margin of the source (“State-Funded Real Estate Bubble v2.0?”). ³

Scale and demand structure

SemiAnalysis’s building-level tracking (1,000+ facilities across 60+ operators) puts China’s operational fleet at 24 GW as of 2026, excluding ~20 GW of dated pipeline and ~30 GW of announced projects. The demand side is concentrated: ByteDance alone occupies roughly a fifth of delivered national capacity and rents nearly all of it, making its tender calendar the order book for every wholesale colocation operator in the country. Combined ByteDance/Alibaba/Tencent/Baidu capex went from ~$35B (2024) to over $50B (2025) on a trajectory toward $100B (2026). ⁴

The structural context: China’s consumer internet has saturated (~80% penetration growing 1.5%/year; WeChat MAU +2% YoY; Baidu revenue falling; Alibaba customer-management revenue −7% YoY in the June 2026 quarter), and Chinese cloud revenue is almost entirely domestic (Alibaba Cloud earns ~20% overseas; carrier clouds effectively 0%), so AI is the only remaining growth engine for the platform titans. The hyperscalers have also built an investment web across the GenAI labs — Alibaba into five LLM unicorns, Tencent into six including the largest external ticket in DeepSeek’s Series A — and the labs in turn train and serve on hyperscaler clouds, making the datacenter expansion the compute foundation for the whole domestic ecosystem. ⁵

Four eras of the industry

SemiAnalysis divides the market’s history into four phases, which explain the otherwise-confusing present (headline ~50% utilization coexisting with an AI capacity shortage): ⁶

1. Carrier hosting (pre-2015) — datacenters as telecom real estate: the three state carriers held 60–70% share; ~90% of facilities ran below 2 kW retail racks bundled with bandwidth. This legacy stock is precisely what’s vacant now — physically unsuitable for H20/Ascend-class densities and impractical to retrofit. The headline vacancy number is largely a monument to this era.
2. Cloud land grab (2015–2021) — Alibaba and Tencent opened a winner-take-all capex race (Alibaba Cloud ~110% CAGR 2015–19, 17 consecutive price cuts in 12 months); third-party colos (GDS, VNET, Chindata) scaled alongside.
3. Digestion (2022–2023) — policy headwinds and macro slowdown stopped the expansion; Alibaba Cloud growth went to effectively zero; Tencent halved capex. Power-exclusive market rates were cut in half from ~$80/kW/month; GDS estimated in August 2026 that another 18 months remained to digest the price reset.
4. AI supercycle (2024–) — DeepSeek plus the hyperscalers’ own GenAI ambitions revived wholesale demand. GDS and VNET signed ~1.3 GW of wholesale in 1H26 with under 10 MW of retail orders — and those two listed names capture only ~1/3 of ByteDance/Alibaba orders, so listed-market data badly understates total demand. The wholesale/retail split produces the two-speed market: wholesale buildings back above 70% utilization, legacy retail racks near 60%.

Eastern Data, Western Compute: mechanism over slogan

The EDWC policy (东数西算), approved by the NDRC and three other central departments in February 2022, designates eight national computing hubs — Beijing-Tianjin-Hebei, Yangtze River Delta, Greater Bay Area, Chengdu-Chongqing, Inner Mongolia, Guizhou, Gansu, Ningxia — implemented as ten clusters. The official framing is cross-regional resource optimization in the lineage of the South-to-North Water Diversion Project and the west-east power transmission program: move compute toward the resource-rich west to relieve land and energy scarcity in the east. ⁷

What made EDWC bite, per SemiAnalysis, was not the slogan but a pre-existing transmission mechanism: the energy-consumption review. Since 2010, any project consuming over 10,000 tons of coal equivalent (>10 MW) cannot legally begin construction without a quota from the local Development and Reform Commission. When EDWC arrived in 2022, the same offices that issue quotas became the offices implementing EDWC, so quota issuance simply stopped flowing to non-hub locations. Guangdong halted datacenter energy quotas entirely in 2021–22 and reopened them only for the Shaoguan hub cluster. In mid-2023, Guangdong’s Energy Bureau found Tencent’s two Qingyuan campuses running at 13–14x approved scale (“approved small, built big”) and held their power supply to the approved quota; Tencent’s subsequent capacity went to Shaoguan. ⁸

Two enablers turned the quota stick into a viable migration: telecom-led fiber buildout (Shaoguan now runs 1.3 ms one-way latency to Guangzhou, 1.66 ms to Shenzhen, under 3 ms to Hong Kong — functionally a Shenzhen rack for all but the most latency-critical workloads), and AI demand itself (training workloads are latency-insensitive but power-price-sensitive). Inner Mongolia emerged as “China’s Johor”: the Mengxi grid (China’s only provincial grid outside State Grid/Southern Grid, with independent pricing), power at roughly half Tier-1 prices, land in square kilometers. The FT’s ground report from Ulanqab — 89 datacenters built or planned, ~15 GW committed, Rmb0.358/kWh versus Rmb0.80 in Beijing, an abandoned high school being converted into racks — is the micro-level picture of the same migration. ^{9 10}

The speed and cost advantage

Chinese 100 MW delivery time compressed from ~18 months (cloud era) to ~12 months (AI era), against a US best case of 12–18 months of construction sitting behind 12–13 months of non-overlappable permitting. The mechanisms: speculative shells pre-built at developer risk (hyperscaler tenders demand T+6-month delivery — GDS signed a 152 MW order in early 2025 and delivered in six months); low-rise long-span steel halls replacing multi-story concrete on greenfield western sites; prefabrication moving MEP work off-site (Alibaba’s CUBE 5.0 “100-day datacenter”: 30 days factory production, 50 days installation, 20 days commissioning). Full-modular container campuses remain rare — Chinese site labor is cheap enough that the marginal speed gain is small, and banks appraise a 20-year building as loan collateral far better than “a yard of containers.” Construction costs in Ulanqab run about 20% below big-city levels. ^{11 12}

The chip gate

The binding constraint on the buildout is not power, land, labor, or permits — it is advanced processors, by deliberate US policy. The October 7, 2022 BIS interim final rule (87 FR 62186) added advanced computing ICs (ECCN 3A090) and computers containing them (4A090) to the Commerce Control List with a presumption-of-denial license policy for the PRC; created new Foreign Direct Product rules extending EAR jurisdiction to foreign-produced chips made with US technology; imposed a “supercomputer” end-use control (defined as ≥100 double-precision petaflops in a 41,600 ft³ envelope); expanded Entity List licensing for 28 Chinese supercomputing/IC entities; and restricted “US person” support for advanced IC production in China. The stated rationale is military-civil fusion: advanced computing enabling WMD development, military decision-making, and surveillance. ¹³

The rule explains the otherwise-puzzling co-occurrence of vacancy and shortage that Nathan flagged in the margins: some fraction of “empty” datacenters are not speculative mistakes but powered shells waiting on chips they cannot legally obtain. Beijing’s counter-incentives are visible at the facility level — datacenters using domestic processors instead of Nvidia silicon receive better tax treatment and electricity/water discounts, and Chinese labs (DeepSeek, Z.ai) are siting facilities in Ulanqab alongside the hyperscalers. Jefferies’ Edison Lee’s summary: “There is no risk of overbuilding. Token consumption is going through the roof… The question is whether the data centres can get enough chips to put into the buildings.” ^{14 15}

The two open debates

Debate 1 — real demand vs directed malinvestment. The bull case (SemiAnalysis, Jefferies): wholesale utilization above 70%, gigawatt-scale tender calendars, capex doubling YoY, demand driven by token consumption. The bear case, best articulated by ChinaTalk’s Irene Zhang and Nick Corvino: provincial governments treated intelligent-compute parks with the same Field-of-Dreams logic they used for real estate — a Science and Technology Daily report found one western compute center at below-50% rack-up with sub-30% server utilization burning RMB 30M/year, ~150 intelligent-compute centers operating by late 2024 and ~400 more under construction or planned, in the state press’s own words “blindly constructed systems divorced from actual needs” (脱离实际需求的盲目建设). Both can be true at different layers of the stack: the wholesale AI segment can be genuinely tight while provincial general-purpose compute parks rot. Nathan’s prior — “China has a history of overbuilding to smooth economic downturns; how much of this is natural vs artificial” — is exactly the unresolved question. ^{16 17}

Debate 2 — did EDWC actually move compute west? SemiAnalysis: yes — the quota mechanism genuinely redirected construction into the hub map, and ByteDance’s self-build program (Datong, Wuhu, twin Inner Mongolia sites, Zhongwei) proves the hyperscalers put their own money there; EDWC concentrated the buildout and the market picked winners within it (Inner Mongolia won, Chengdu/Chongqing lost on power price, Gui’an became the cold-storage hub with Apple’s iCloud China as anchor tenant). ChinaTalk: mostly no — 2025 CAICT data shows chip location concentrated in Hebei, Guangdong, Jiangsu and the eastern metros’ hinterlands; only four of the eight hubs are genuinely remote; the NDRC justified the policy with electricity as 70% of operating costs when construction and chips dominate total cost (electricity ~5% of three-year cost for a 400 MW facility); the real pattern is “Eastern Data, 45 Minutes Outside the City Compute.” The disagreement is partly definitional — SemiAnalysis measures capacity built within the designated hub map (much of which is exurban-east), ChinaTalk measures chips moved to the remote west — but the policy-evaluation stakes are real: poorer western provinces may be the policy’s biggest losers, holding debt-financed shells. ^{18 19}

A third, smaller note: the energy story is coal as much as wind. Inner Mongolia pairs China’s largest wind fleet (117 GW, nearly 4x the UK’s entire capacity) with ~130 GW of mostly-coal fossil capacity, and the “green development” framing in the original NDRC announcement coexists with data centers explicitly valued as a soak for stranded fossil generation. ^{20 21}

Cross-domain connections

• The energy-quota mechanism is a textbook limited-access-order rent: construction rights flow through administrative rationing by the same offices that implement industrial policy, and incumbents learn to build where the rents are allocated rather than where demand is densest — see open-access-order.
• The US comparison is the abundance-agenda mirror image: China’s facility side has no permitting friction (3–6 months, overlapping), so its bottleneck moved entirely upstream to chips; the US has no chip constraint, so its bottleneck moved entirely downstream to power and permits (data-center-backlash, interconnection queues). Each system’s constraint is the other’s surplus.
• The Mengxi grid oddity (one provincial grid with independent pricing inside a national system) rhymes with japan-grid-frequency-divide: both are legacy grid partitions that became economically load-bearing long after their origins stopped mattering.
• The bank-collateral explanation for why full-modular didn’t win (“borrow against the asset, lease it to ByteDance”) is an asset-financing determinant of construction technology — the same shape of argument as ship-finance shaping hull design, and a reminder that build form follows collateral form.
• llm-inference-provider-landscape describes the Western three-layer API supply chain; this page is the physical-layer counterpart for the Chinese ecosystem — the labs in that supply chain (DeepSeek, Moonshot, Zhipu/GLM) train and serve on the hyperscaler clouds built on these campuses.

Sources

• Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model
• Financial Times 2026 — China races to build data centres in bid for AI supremacy
• 2022 — China approves mega project for greater computing power, digital future
• Irene Zhang and Nick Corvino 2026 — “Eastern Data, Western Compute” is Fake
• Bureau of Industry and Security, US Department of Commerce 2022 — Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification (87 FR 62186, preamble extract)

Footnotes

1. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

2. Financial Times 2026 — China races to build data centres in bid for AI supremacy ↩

3. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

4. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

5. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

6. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

7. 2022 — China approves mega project for greater computing power, digital future ↩

8. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

9. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

10. Financial Times 2026 — China races to build data centres in bid for AI supremacy ↩

11. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

12. Financial Times 2026 — China races to build data centres in bid for AI supremacy ↩

13. Bureau of Industry and Security, US Department of Commerce 2022 — Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification (87 FR 62186, preamble extract) ↩

14. Financial Times 2026 — China races to build data centres in bid for AI supremacy ↩

15. Bureau of Industry and Security, US Department of Commerce 2022 — Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification (87 FR 62186, preamble extract) ↩

16. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

17. Irene Zhang and Nick Corvino 2026 — “Eastern Data, Western Compute” is Fake ↩

18. Irene Zhang and Nick Corvino 2026 — “Eastern Data, Western Compute” is Fake ↩

19. Everlyn, Dylan Patel, Patrick Schaabi 2026 — The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model ↩

20. Financial Times 2026 — China races to build data centres in bid for AI supremacy ↩

21. 2022 — China approves mega project for greater computing power, digital future ↩