Is a Terrestrial Neocloud Possible in China? — Part 1: A Primer

This is Part 1 — a primer. The words, the units, and how a number is checked. Part 2 is the 14 August analysis: the hangar tour and the Singapore appendix.

This piece follows one structure. What a neocloud is, how big it is, and what sits inside it. Which data centers are comparable, and where. How a number is verified, what unit it is in, and which GPU it assumes. Then the two numbers that decide the economics: the power price and the compute that is actually usable. Then whose order bought the megawatts.

Liquid cooling versus air cooling is a later note, not the opening. Air-cooled halls show rooftop or side-wall fan rows. Liquid-cooled halls are the ones where a cooling tower is the tell. In many Chinese halls the generators sit underground. The absence of a US-style tower is not proof of an empty shell.

1. What a neocloud is, how big, and what is inside

A neocloud is a company, not a building. It sells GPU-hours to someone else: on-demand or reserved, for training and inference. CoreWeave, Nebius (NASDAQ: NBIS), Lambda, and Crusoe are this category. AWS, Azure, Google Cloud, and Alibaba Cloud are hyperscalers. GDS, VNET, and Runze are landlords. They sell cabinets and megawatts. The landlord owns the airframe. The neocloud is the airline.

What sits inside a Western neocloud is three businesses stacked under one word.

The first is a developer cloud: a few cards to a few hundred, paid by the hour, a credit card can open it. Lambda and RunPod live here. The revenue is real. It is not CoreWeave’s backlog.

The second is reserved training fabric: InfiniBand or NVLink, thousands to tens of thousands of GPUs, take-or-pay for about three years. This is the meat. The customers are Microsoft, Meta, OpenAI, Anthropic. NVIDIA is a shareholder, an allocator of cards, and sometimes the residual-value story behind the debt. The loan is longer than the contract. The gap is a bet that the used cards still rent.

The third is a dedicated factory: Crusoe’s Abilene campus for Oracle / OpenAI, Fluidstack for Anthropic. The offtaker is one lab. The catalog is not public.

Scale is spoken in power, not in press-release GPU counts.

CoreWeave, Q2 2026: more than 1.5 GW active, about 4.2 GW contracted by 11 August 2026, year-end active guide above 1.85 GW, 51 sites. Capex guide $35–39 billion. Backlog about $104 billion.

Nebius: connected-power target for 2026 of 800 MW to 1 GW. Contracted-power target raised toward 5 GW by year-end 2026. Mäntsälä Finland is 75 MW live. A second Mäntsälä hall is up to 70 MW from 2027. Lappeenranta is up to 310 MW. Finland across three sites is planned at 455 MW. Independence, Missouri, is approved as a campus with potential capacity up to 1.2 GW. EMEA contracted power was already above 750 MW when Lappeenranta was announced (31 March 2026).

Crusoe, 9 June 2026: 4.9 GW contracted, including a 1.2 GW Abilene campus for Oracle and a 900 MW Abilene campus for Microsoft.

One card is not one megawatt. An H100 or H200 SXM is about 700 W. A B200 is about 1,000 W, configurable to 1,200 W. A GB200 NVL72 rack is about 120–132 kW for 72 GPUs. One thousand H100s are 0.7 MW of GPU silicon alone. Add CPUs, NICs, and storage, then multiply by PUE (western China often claims about 1.2). A rough working number: about 1 MW of campus power holds on the order of one thousand H100s. A B200 campus holds fewer cards per megawatt. A company’s “5 GW contracted” is interconnection and offtake, not five million cards already seated.

Four tests, all required. Frontier GPUs in training quantities. A fabric of about 10,000 cards, not a lab. Reserved cluster hours sold to a third party at high ACV. GPUs financed like aircraft, residual risk on the balance sheet. Fail one and the firm is something else. Fail all four and it is a landlord with a liquid-cooling brochure.

2. Comparable campuses, and the structure used here

The comparable object in China is not a ticker. It is a campus that already has halls, power, and a named offtaker. The hubs that reporters can enter, and that public satellite can at least frame, are these.

Ulanqab (Wulanchabu), Inner Mongolia: Yiwutang (Huawei Ave / Apple Ave / Ali Blvd), Bayin, and the Chayouqianqi Jingmeng pad. City figures at year-end 2025: 84 signed data centers, about 330,000 operating racks, 66% occupancy (Cailian, 26 June 2026). UCloud’s existing park: cabinet utilization above 95% and GPU 91.29% at 30 June 2026; Building D is about 75% intended power, not a signed take-or-pay (STAR inquiry, 31 July 2026).

Zhongwei, Ningxia: West Cloud Base on Fengyun Road. Rooftop PV and a cooling pond sit next to the halls. Unicom’s site visit copy is the offtake color, not the roof.

Qingyang, Gansu: China Computing Valley. The public mosaic over Wenquan is older than the park. Do not treat that frame as a Kingsoft hall.

Zhangjiakou / Huailai, Hebei: Chindata Donghuayuan and the Heying nameplate. Beijing overflow on landlord megawatts.

Langfang: Runze International Information Port (about 39.60N, 116.72E) and GDS Anci / Xianghe clusters. Runze sells cabinets. April 2026 IR had not confirmed the 200 MW all-liquid B-zone as live IT.

Shanghai Lingang: land AIDC on Lianggang Avenue, and the Highlander / Hai Lan Yun 2.3 MW underwater cabin 13.05 km east of Nanhui Cape, between offshore-wind phase I turbine 1-25# and phase II. The 14 m cylinder is about one Sentinel-2 pixel. Figures 17–18 in the illustrated pack are official sea-use planning maps, not satellite.

Johor–Batam is the ASEAN analog Semi and Cushman both write in public. Cushman’s 2026 Global Data Center Market Comparison ranks Johor first among APAC primary markets and fifth among global primary markets. Dallas is first globally. Shanghai is tenth on global market fundamentals and third among APAC primaries. Beijing is seventh in APAC. Ulanqab and Zhongwei do not appear on that public list.

Cushman & Wakefield’s public 2026 comparison covers 107 markets and 24 variables. What is free is the ranking and a few headline gigawatts (about 31.7 GW under construction globally in 2025). The city-level vacancy and megawatt table sits behind an email gate. The APAC construction-cost guide (90 clusters, 26 locations) is the same shape: Japan about $13.0–19.2 million per MW, Taiwan about $5.2–7.9 million, China mainland construction-cost inflation written as 0.4% in the public digest. Semi’s India teaser used about $5 million per MW and power under $0.10/kWh. That is shell capex, not a GPU-hour.

The structure of this essay keeps three tables unmerged: the shell, the power, and the order. A megawatt in the first table is not an offtake in the third.

3. How a number is verified, what the unit is, which card

Units first.

Do not add a city’s “1 GW” to a campus door. Do not add contracted GW to active GW.

Verification is a stack, not a photograph.

Satellite, public commercial mosaics, pulled 14 August 2026, archive, not live: can show a windowless hall, a rooftop AHU row, a side-wall outdoor-unit row, an empty pad, a cooling pond, a PV field. Yiwutang Ali Blvd at 40.9845N 113.2420E shows five north–south halls and rooftop AHU rows, plus a pad still under construction. Huailai Chindata shows cooling along the long sides of the halls. Langfang Runze shows rooftop fan arrays. Bayin is a strip, not a Zhongjin door pin. Qingyang’s mosaic predates the park. The Lingang cabin cannot be named at 10 m. There is no public live hall-level satellite. Starlink is communications, not Earth observation.

Satellite cannot show the SKU, the occupancy, the take-or-pay, the tenant, or whether the diesel sets are in the basement.

Filings and inquiry letters can show utilization and intended kilowatts (UCloud STAR). City releases can show rack counts and occupancy (Cailian on Ulanqab). Utility tariff sheets can show the door-to-door yuan per kWh. Company pricing pages can show a GPU-hour (CoreWeave, Nebius). Channel weeklies can show a China H100 monthly ask (SMM, 10 July 2026). None of these is a census of cards in a named hall.

Which card is in the hall is a separate claim. Export-control public record: H200 licenses were described as very few (BIS, July 2026). China west parks run a mix that public copy names as Ascend, Haiguang, and remaining NVIDIA inventory. An H100-hour price and an Ascend-hour price are not the same product. Convert power to card counts only after the SKU is named. Otherwise leave the number in megawatts.

4. Power price, and compute that is actually usable

Ulanqab door-to-door power is about 0.33–0.36 yuan/kWh. The Economic Observer, 14 August 2026: Inner Mongolia Power’s August 2026 agent-purchase price 0.3234 yuan/kWh; local officials said some compute centers land at about 0.358 yuan/kWh. Alibaba on site (21st Century Business Herald / Yicai, 13 August 2026) said about 0.3 yuan and claimed 90% green power. Zhongjin Bayin’s physical renewable substitution was 38.74% (Xinhua, 18 July 2025). That is a plant figure, not a satellite figure.

At 7.2 yuan to the dollar, 0.36 yuan is about $0.05/kWh. Semi’s India teaser was under $0.10/kWh. Texas industrial power is on the order of 6 cents. Western China is not expensive on electrons.

One H100-hour of electricity, at 700 W and PUE about 1.2, is about 0.3 yuan. Idle halls burn that too. East China cabinet power has been quoted around 6,000 yuan per unit per month in the SMM channel. West China idle is cheaper per kWh. City occupancy of 66% is a city number, not a hall meter. Ascend 910B2 was quoted at 13,500 yuan/month in Wenzhou and was hard to lease (SMM, 10 July 2026). A list price with no tenant is idle compute.

Usable compute is not nameplate megawatts. It is megawatts that have a card, a network, and a buyer. UCloud’s 91% GPU utilization is one park’s inquiry-letter number. Ulanqab’s 66% rack occupancy is the city’s. DeepSeek’s 1 GW is a city-level Bloomberg figure (31 July 2026), not a seated-card census. Do not multiply a city’s GW by 1,000 H100s per MW and call it a cluster.

GPU-hour, public, mid-2026

The west is cheap on power and expensive on the closed H100 lease. That gap is the export-control card, not the Inner Mongolia tariff. Building-shell capex is a fifth cell. Semi’s India $5 million/MW is that cell. It is not the cost of model compute.

5. Whose order

Bayin cannot be written as “Alibaba plus DeepSeek plus GDS behind one door.” The three public lines sit on three layers.

Alibaba is a field sighting in the Bayin core. 21st Century Business Herald / Yicai, 13 August 2026: Alibaba Cloud supernodes, Wang Chaoyang on tariff and green power. This is a walk-in, not a municipal memorandum.

DeepSeek is city-level Ulanqab. Bloomberg, 31 July 2026: about 1 GW, self-build plus lease. No public door GPS pins it to a Bayin hall.

GDS is a city memorandum. 3 June 2026: with Ulanqab, more than RMB 30 billion over five years, gigawatt-scale parks. GDS Q2 2026 said Ulanqab and Horinger are new markets, and that Changshu in Jiangsu is progressing. A MoU is not a Bayin take-or-pay. GDS is also a landlord. It can be a tenant of a park and a landlord to a hyperscaler. Do not collapse those roles.

Changshu itself is a real GDS campus in company copy (CS1 and CS2 delivered). Public Google satellite along Wangwan North Road and Penghu Road, checked 15 August 2026, showed warehouses and logistics sheds, not a named hall and not a cooling tower. The campus is not disproved. The door is not in that mosaic.

Western offtake, for comparison, is named in filings: Microsoft, Meta, OpenAI, Anthropic on CoreWeave and Nebius reserved hours. Crusoe’s 1.2 GW Abilene is written as Oracle. That is the order layer Semi sells in the model and teases in the newsletter.

6. What SemiAnalysis watches

Semi’s public datacenter writing is a method, not a China campus list.

It tracks more than 5,000 facilities through property records, permits, FOIA power, and satellite, with a vision model on rooftop chillers. A site moves from land, to ground broken, to shell and cooling plant, to energized, to live IT load. Announced gigawatts without site control, equipment deposits, or interconnection are treated as 2028-or-later air. Promised megawatts are scored against delivered megawatts each quarter.

It splits self-build, colo lease, and GPU-cloud rent, and asks who uses the megawatt: OpenAI, Anthropic, Microsoft, Meta. It prices the shell as capex excluding IT, broken into power, cooling, and building. It prices the card as a GPU-hour and a one-year contract. It prices the electron in cents per kWh. It asks whether the binding constraint is the chip, the hall, or the grid, including behind-the-meter gas.

Public case studies of delay are US doors: Nebius / DataOne New Jersey (rooftop chillers late), Core Scientific Denton for CoreWeave, STACK / Oracle New Mexico slipped toward 2029. Public China copy is thinner: China can build power and would lead gigawatt halls if the cards were allowed; they are not, so the US leads. Johor is written as Chinese operators internationalizing next to Singapore’s 900 MW and $0.23/kWh tariff. India, 15 August 2026, was a three-tweet teaser: under 10 GW by 2030, about $5 million/MW, power under $0.10/kWh, Microsoft $17.5 billion. The site list is gated.

What this essay can write that Semi’s public feed does not: the Bayin order split, the Ulanqab door-to-door yuan, the China H100 monthly ask versus the Western contract hour, and the 25 annotated frames that prove a shell and do not prove a contract.

The megawatts are on the ground. A terrestrial neocloud is a company that sells reserved cluster hours to a third party and finances the cards. That company is not in the taxi on Huawei Avenue.

Read next

• Part 2 — Is a Terrestrial Neocloud Possible in China? — hangar tour, Singapore / Asia appendix, four missing ingredients.