I built these maps mostly so I could keep track of all this myself, and to learn China geopolitics. They are three layers of one system, not three separate charts: the firms sit on the compute, and the compute sits on the grid.
01 · Companies
The industry, by headquarters
104 companies across 19 cities. One dab per company, coloured by where it sits in the stack and sized by how big the company is. Hover a dab to read it, click it for the profile; use the key to isolate a layer.
Industry layer
Company size
- Giant
- Large
- Start-up
Company geography
Explore the companies behind China's AI stack
Select a company on the map for its profile, or isolate one layer of the industry with the key in the corner of the map.
Runs it , as of early 2026
Trained at
Founded by
Ships
Computes insee the compute map ↓
The pattern to notice
Every company in this dataset is headquartered on the eastern side of the Hu Huanyong line of 1935, the side where roughly 94% of China's population lives on well under half the land. That crowding is why Shanghai, Hangzhou, Shenzhen and Guangzhou are drawn beside their true positions, with the small ring at the end of the leader line marking where each one actually sits.
| Company | 中文 | City | Layer | Founded | Leads it |
|---|---|---|---|---|---|
| Baidu | 百度 | Beijing | Platform | 2000 | Robin Li |
| ByteDance | 字节跳动 | Beijing | Platform | 2012 | Liang Rubo |
| Didi | 滴滴出行 | Beijing | Platform | 2012 | Cheng Wei |
| JD.com | 京东 | Beijing | Platform | 1998 | Sandy Xu |
| Kingsoft | 金山软件 | Beijing | Platform | 1988 | Zou Tao |
| Kuaishou | 快手 | Beijing | Platform | 2011 | Cheng Yixiao |
| Meituan | 美团 | Beijing | Platform | 2010 | Wang Xing |
| 微博 | Beijing | Platform | 2009 | Wang Gaofei | |
| Zhihu | 知乎 | Beijing | Platform | 2010 | Zhou Yuan |
| 01.AI | 零一万物 | Beijing | Model lab | 2023 | Kai-Fu Lee |
| BAAI | 北京智源人工智能研究院 | Beijing | Model lab | 2018 | Wang Zhongyuan |
| Baichuan Intelligence | 百川智能 | Beijing | Model lab | 2023 | Wang Xiaochuan |
| Langboat | 澜舟科技 | Beijing | Model lab | 2021 | Zhou Ming |
| ModelBest | 面壁智能 | Beijing | Model lab | 2022 | Li Dahai |
| Moonshot AI | 月之暗面 | Beijing | Model lab | 2023 | Yang Zhilin |
| Zhipu AI | 智谱 | Beijing | Model lab | 2019 | Zhang Peng |
| Cambricon | 寒武纪 | Beijing | Silicon | 2016 | Chen Tianshi |
| GigaDevice | 兆易创新 | Beijing | Silicon | 2005 | Zhu Yiming |
| Horizon Robotics | 地平线 | Beijing | Silicon | 2015 | Yu Kai |
| Hygon | 海光信息 | Beijing | Silicon | 2014 | Sha Chaoqun |
| Loongson | 龙芯中科 | Beijing | Silicon | 2010 | Hu Weiwu |
| Moore Threads | 摩尔线程 | Beijing | Silicon | 2020 | Zhang Jianzhong |
| NAURA | 北方华创 | Beijing | Silicon | 2001 | Zhao Jinrong |
| 4Paradigm | 第四范式 | Beijing | Applied | 2014 | Dai Wenyuan |
| Aibee | 爱笔智能 | Beijing | Applied | 2017 | Lin Yuanqing |
| Kunlun Tech | 昆仑万维 | Beijing | Applied | 2008 | Fang Han |
| Megvii | 旷视科技 | Beijing | Applied | 2011 | Yin Qi |
| Qihoo 360 | 奇虎360 | Beijing | Applied | 2005 | Zhou Hongyi |
| AUBO Robotics | 遨博智能 | Beijing | Machines | 2015 | Wei Hongxing |
| Galbot | 银河通用 | Beijing | Machines | 2023 | Wang He |
| Lenovo | 联想 | Beijing | Machines | 1984 | Yang Yuanqing |
| Li Auto | 理想汽车 | Beijing | Machines | 2015 | Li Xiang |
| Roborock | 石头科技 | Beijing | Machines | 2014 | Chang Jing |
| Xiaomi | 小米 | Beijing | Machines | 2010 | Lei Jun |
| Jingjia Micro | 景嘉微 | Changsha | Silicon | 2006 | Liu Wei |
| Changan Automobile | 长安汽车 | Chongqing | Machines | 1862 | Zhu Huarong |
| Midea | 美的集团 | Foshan | Machines | 1968 | Fang Hongbo |
| Rockchip | 瑞芯微 | Fuzhou | Silicon | 2001 | Li Min |
| Shein | 希音 | Guangzhou | Platform | 2008 | Chris Xu |
| CloudWalk | 云从科技 | Guangzhou | Applied | 2015 | Zhou Xi |
| XPeng | 小鹏汽车 | Guangzhou | Machines | 2014 | He Xiaopeng |
| Alibaba | 阿里巴巴 | Hangzhou | Platform | 1999 | Eddie Wu |
| Alibaba Cloud | 阿里云 | Hangzhou | Platform | 2009 | Eddie Wu |
| Ant Group | 蚂蚁集团 | Hangzhou | Platform | 2014 | Eric Jing |
| Cainiao | 菜鸟 | Hangzhou | Platform | 2013 | Wan Lin |
| DingTalk | 钉钉 | Hangzhou | Platform | 2014 | Ye Jun |
| Game Science | 游戏科学 | Hangzhou | Platform | 2014 | Feng Ji |
| NetEase | 网易 | Hangzhou | Platform | 1997 | William Ding |
| DeepSeek | 深度求索 | Hangzhou | Model lab | 2023 | Liang Wenfeng |
| Dahua Technology | 大华股份 | Hangzhou | Applied | 2001 | Fu Liquan |
| Hikvision | 海康威视 | Hangzhou | Applied | 2001 | Hu Yangzhong |
| Manycore Tech | 群核科技 | Hangzhou | Applied | 2011 | Huang Xiaohuang |
| BrainCo | 强脑科技 | Hangzhou | Machines | 2015 | Han Bicheng |
| Deep Robotics | 云深处科技 | Hangzhou | Machines | 2017 | Zhu Qiuguo |
| Geely | 吉利汽车 | Hangzhou | Machines | 1986 | Li Shufu |
| Leapmotor | 零跑汽车 | Hangzhou | Machines | 2015 | Zhu Jiangming |
| Unitree Robotics | 宇树科技 | Hangzhou | Machines | 2016 | Wang Xingxing |
| Zeekr | 极氪 | Hangzhou | Machines | 2021 | An Conghui |
| CXMT | 长鑫存储 | Hefei | Silicon | 2016 | Zhu Yiming |
| iFlytek | 科大讯飞 | Hefei | Applied | 1999 | Liu Qingfeng |
| Gotion High-Tech | 国轩高科 | Hefei | Machines | 2006 | Li Zhen |
| EVE Energy | 亿纬锂能 | Huizhou | Machines | 2001 | Liu Jincheng |
| Inspur | 浪潮 | Jinan | Platform | 1989 | Zou Qingzhong |
| Estun | 埃斯顿 | Nanjing | Machines | 1993 | Wu Bo |
| CATL | 宁德时代 | Ningde | Machines | 2011 | Robin Zeng |
| Bilibili | 哔哩哔哩 | Shanghai | Platform | 2009 | Chen Rui |
| PDD Holdings | 拼多多 | Shanghai | Platform | 2015 | Chen Lei and Zhao Jiazhen |
| Trip.com Group | 携程 | Shanghai | Platform | 1999 | Jane Sun |
| MiniMax | 稀宇科技 | Shanghai | Model lab | 2021 | Yan Junjie |
| Shanghai AI Laboratory | 上海人工智能实验室 | Shanghai | Model lab | 2020 | Zhou Bowen |
| StepFun | 阶跃星辰 | Shanghai | Model lab | 2023 | Jiang Daxin |
| ACM Research | 盛美上海 | Shanghai | Silicon | 1998 | David Wang |
| AMEC | 中微公司 | Shanghai | Silicon | 2004 | Gerald Yin |
| Biren Technology | 壁仞科技 | Shanghai | Silicon | 2019 | Zhang Wen |
| Enflame | 燧原科技 | Shanghai | Silicon | 2018 | Zhao Lidong |
| Hua Hong Semiconductor | 华虹半导体 | Shanghai | Silicon | 2005 | Bai Peng |
| MetaX | 沐曦 | Shanghai | Silicon | 2020 | Chen Weiliang |
| SMIC | 中芯国际 | Shanghai | Silicon | 2000 | Zhao Haijun |
| UNISOC | 紫光展锐 | Shanghai | Silicon | 2018 | Wu Shengwu |
| Zhaoxin | 兆芯 | Shanghai | Silicon | 2013 | Luo Yu |
| Hesai | 禾赛科技 | Shanghai | Applied | 2014 | Li Yifan |
| SenseTime | 商汤科技 | Shanghai | Applied | 2014 | Xu Li |
| Yitu | 依图科技 | Shanghai | Applied | 2012 | Zhu Long |
| AgiBot | 智元机器人 | Shanghai | Machines | 2023 | Deng Taihua |
| Fourier Intelligence | 傅利叶智能 | Shanghai | Machines | 2015 | Gu Jie |
| NIO | 蔚来 | Shanghai | Machines | 2014 | William Li |
| SAIC Motor | 上汽集团 | Shanghai | Machines | 1984 | Wang Xiaoqiu |
| Siasun | 新松机器人 | Shenyang | Machines | 2000 | Qu Daokui |
| Huawei | 华为 | Shenzhen | Platform | 1987 | Ren Zhengfei |
| Tencent | 腾讯 | Shenzhen | Platform | 1998 | Pony Ma |
| ZTE | 中兴通讯 | Shenzhen | Platform | 1985 | Li Zixue |
| HiSilicon | 海思半导体 | Shenzhen | Silicon | 2004 | He Tingbo |
| Mindray | 迈瑞医疗 | Shenzhen | Applied | 1991 | Li Xiting |
| Ping An Technology | 平安科技 | Shenzhen | Applied | 2008 | Xie Yonglin |
| RoboSense | 速腾聚创 | Shenzhen | Applied | 2014 | Qiu Chunxin |
| BYD | 比亚迪 | Shenzhen | Machines | 1995 | Wang Chuanfu |
| DJI | 大疆创新 | Shenzhen | Machines | 2006 | Frank Wang |
| Dobot | 越疆 | Shenzhen | Machines | 2015 | Liu Peichao |
| Inovance | 汇川技术 | Shenzhen | Machines | 2003 | Zhu Xingming |
| UBTech | 优必选 | Shenzhen | Machines | 2012 | Zhou Jian |
| Ecovacs | 科沃斯 | Suzhou | Machines | 1998 | Qian Dongqi |
| YMTC | 长江存储 | Wuhan | Silicon | 2016 | Chen Nanxiang |
| Meitu | 美图 | Xiamen | Applied | 2008 | Wu Xinhong |
| Allwinner Technology | 全志科技 | Zhuhai | Silicon | 2007 | Zhang Jianhui |
Every company on that map sits east of the Hu Huanyong line. That is where the trouble starts. Training and serving models takes data centres that draw power around the clock and need land and cooling, and the eastern cities are where all three are scarcest. Beijing has been closed to new large data centres for years. A handful of the platforms already compute far from home.
Open one of them on the map to see where. The next map is the state’s attempt to make that the rule rather than the exception.
02 · Compute
East Data, West Computing · 东数西算
8 national computing hub nodes and the 10 data centre clusters inside them, approved in February 2022. Four hubs hold the demand; four hold the power, the cold air and the land.
8 hub nodes10 data centre clusters4 · 4 demand and supply
- Demand hub · where the data is
- Supply hub · cheap power and land
- Eastern city · where the work starts
How to use it
- Follow a line from its ring, an eastern city, to the diamond meant to take its work.
- Hover a diamond to light up its hub and the line into it.
Demand hubs
Beijing–Tianjin–Hebei 京津冀
The research capital and its overflow. Beijing itself has been closed to new large data centres for years, so the capacity moved to the hills north-west of it.
- Zhangjiakou 张家口About 200 km from Beijing, cold most of the year, and sitting inside one of the country's oldest wind-power bases. Alibaba's northern cloud region is next door at Zhangbei.
Yangtze River Delta 长三角
The densest concentration of demand in the country: Shanghai's finance and silicon, Hangzhou's platforms, and the manufacturing belt between them.
- Yangtze Delta Demonstration Zone 长三角生态绿色一体化发展示范区Straddles the Shanghai, Jiangsu and Zhejiang borders on purpose — a cross-jurisdiction zone built so latency-sensitive workloads can stay close to the city without sitting inside it.
- Wuhu 芜湖The delta's cheaper hinterland, roughly 300 km up the Yangtze from Shanghai. Huawei, China Telecom and Alibaba campuses.
Greater Bay Area 粤港澳大湾区
Shenzhen, Guangzhou, Hong Kong and Macau as one labour market. Hardware demand, and the most land-constrained of the four demand hubs.
- Shaoguan 韶关The delta's designated relief valve: far enough north of Guangzhou for land and power to be affordable, close enough for round-trip latency to stay tolerable.
Chengdu–Chongqing 成渝
The interior's own megacity pair, and the reason the map is not a clean east-west split. Sichuan hydropower makes it a supply region too, and the cleanest grid in the country at roughly 265 g CO₂ per kWh — in the wet season. The 2022 drought shut factories for weeks.
- Tianfu 天府Chengdu's new-area cluster, drawing on Sichuan hydropower that would otherwise be curtailed each summer.
- Chongqing 重庆Paired with Tianfu across the Sichuan basin; automotive and industrial workloads rather than consumer platforms.
Supply hubs
Inner Mongolia 内蒙古
Cold, flat, windy, and already wired to the capital — the most built-out of the western hubs before the policy existed. Also the dirtiest grid in the country, at roughly 947 g CO₂ per kWh against Sichuan's 265. The wind is real; so is the coal underneath it.
- Horinger 和林格尔South of Hohhot, where all three state carriers put their northern bases. Alibaba, Huawei and Apple's second China data centre are a little further east at Ulanqab.
Guizhou 贵州
The original experiment. Mild year-round temperatures, hydropower, stable karst geology, and a decade of provincial policy aimed at exactly this. The one supply hub where the clean-power argument mostly holds — though Guizhou is a coal province with hydro on top, not the reverse.
- Gui'an 贵安Next to Guiyang, which already holds Apple's China iCloud data under its local partner, Tencent's tunnel data centre, and the three carriers' southern bases.
Gansu 甘肃
The newest of the four, and the most explicitly energy-led: built alongside the wind and solar bases of the Hexi Corridor. Whether a workload lands on that wind or on the coal beside it depends on the hour, not on the address.
- Qingyang 庆阳On the Loess Plateau in eastern Gansu, developed with China Mobile alongside new renewable capacity rather than after it.
Ningxia 宁夏
Small, dry, sunny and cheap, with the shortest political distance between a provincial government and a signed land deal. Sunny is not the same as clean: Ningxia sits in the coal-heavy northern half of the grid, well above the national average carbon intensity.
- Zhongwei 中卫Desert edge on the Yellow River. The AWS China (Ningxia) region operated by its local partner sits here, along with Meituan's own campus.
Not all work can travel
The arrows on that map run west, but a workload can only follow them if nobody is waiting on the answer.
- Training runs
- Overnight batch jobs
- Backups
- Rendering
- Models answering chat
- Search queries
- Payments
- Games and video calls
That is why the plan did not send everything to the desert: half the hubs are demand hubs, and clusters like Wuhu and Shaoguan sit a few hundred kilometres from the cities they serve, near enough to answer quickly and far enough to find land.
It also means the programme suits training better than everyday use, because the more people use a model, the more of its computing has to happen near them. And even the work that can move west only helps if there is power to run it, power that already has somewhere to go.
03 · Power
West-to-East Power Transmission · 西电东送
The grid the compute policy is a footnote to. Three corridors carry electricity from where it is generated to the seven provinces that consume close to 40% of it — and they run the opposite way to the data.
Power source
HydroWind and solarCoal
Northern corridor 北通道
→ Beijing–Tianjin–Hebei and Shandong
Coal, wind and solar from the northern and north-western bases into the capital region. This is the corridor that overlaps East Data, West Computing most directly: three of the four supply hubs sit on it.
- Hami 哈密Wind and solar
- Jiuquan 酒泉Wind and solar
- Ordos 鄂尔多斯Coal
- Yulin 榆林Coal
- Shanxi bases 山西煤电基地Coal
- Xilingol 锡林郭勒Wind and solar
Central corridor 中通道
→ Yangtze River Delta
Yangtze hydropower into Shanghai and its neighbours. The oldest and least controversial leg, and the one where the resource is genuinely enormous.
- Baihetan 白鹤滩Hydro
- Xiangjiaba 向家坝Hydro
- Three Gorges 三峡Hydro
Southern corridor 南通道
→ Pearl River Delta
Yunnan and Guizhou generation into Guangdong's factories. Run end to end by the Southern Grid, which was created to run it.
- Nuozhadu 糯扎渡Hydro
- Wudongde 乌东德Hydro
- Bijie 毕节Coal
What the corridor map leaves out
Westward does not automatically mean cleaner.
Electricity can move west to east while its carbon burden moves in the opposite direction. China's computing policy sends workloads across a grid whose emissions vary by more than a factor of three.
Grid carbon intensityThree of the four supply hubs sit in the coal-heavy north, not the hydro south-west.
265 g CO₂ / kWh · Sichuan947 g CO₂ / kWh · Inner Mongolia
3.6× higher at the northern endpoint
Four limits to a clean-grid reading
- 01Hydro is seasonal. Sichuan and Yunnan are abundant in the wet season and tight in the dry one. The 2022 Sichuan drought shut factories and rationed power for weeks.
- 02“Hydro province” is too simple. Guizhou combines coal with hydro. That mixed system is part of why its power remains cheap and steady enough for data centres.
- 03Curtailment is no longer the whole case. It supported the original argument for building west, but fell sharply after 2017 as transmission expanded. Renewables growth is now pushing it upward again.
- 04A green origin does not guarantee green delivery. Several ultra-high-voltage links include a guaranteed coal share for stability. A line leaving a wind base is not a line carrying only wind.
The implicationMoving a workload from Shanghai to Horinger changes its geography. Without hourly generation and delivery data, it does not prove that the workload became cleaner.
The constraint no map shows
Land, power and cooling are the inputs a map can draw. The one that has mattered most since 2022 is the chips.
- United States export controls cut China’s access to the most advanced AI accelerators and to the equipment needed to make them.
- The controls are tightened.
A data centre in Gansu with cheap power still has to fill its racks, and what it can fill them with is a limited stock of imported chips or domestic ones: the accelerators from Huawei, Cambricon and the Shanghai start-ups on the first map, made by foundries such as SMIC working without the most advanced tools.
Sources and method
All three maps are drawn on the same projection and the same frame, from Natural Earth admin-1 land polygons, so they can be laid over one another. No maritime boundary of any kind is encoded in the source or in this rendering. The dashed diagonal on each is the Hu Huanyong line of 1935, east of which roughly 94% of China’s population lives on well under half the land.
The companies. Headquarters as registered, which is not always where the work happens. Founders are historical; the person running a company is not, and leadership here is current as of early 2026. Where a leader trained is recorded for 52 of the 104 companies; the rest are left blank rather than guessed.
Dab size is company size in three steps, not a continuous scale, because private valuations and listed market caps do not compare closely enough to plot as numbers. Giant (17) are the household platforms and national champions, with tens of billions of dollars in revenue or value. Large (56) are listed or late-stage firms of real scale. Start-up (31) covers young companies, research institutes and small caps. The steps are my own judgement, and they measure size, not influence: DeepSeek is a small dab.
The compute. The eight hubs and ten clusters are the ones approved in February 2022 by the National Development and Reform Commission, the Cyberspace Administration of China, the Ministry of Industry and Information Technology and the National Energy Administration. The arcs show the direction the programme is named after, not designated routes: no hub is formally paired with any other. Reading the programme against the Hu Huanyong line is an idea taken from ChinaTalk’s Eastern Data, Western Compute is Fake. Company data-centre locations are recorded only where the cluster notes name them, so a company without that line is unrecorded rather than without data centres.
The electricity. Corridor structure follows the Wilson Center’s mapping of the West–East Electricity Transfer project, and service territories are as published by the two operators. Named generation sites are the major bases on each corridor, not a complete list — there are now dozens of ultra-high-voltage links. Carbon-intensity figures come from provincial grid carbon-footprint work covering 2020–2022; because estimates vary between studies, only the two well-attested endpoints are quoted.
