文 | 文武赵
2026年10月28日，央视新闻报道了中国八大国家算力枢纽的最新建设进展。
北京周边的张家口，单向网络时延已经控制在2毫秒以内；内蒙古和林格尔，绿电占比超过85%；甘肃庆阳，正成为国内大模型训练和词元计算的重要基地。
从张家口到芜湖，从贵州贵安到宁夏中卫，一张横跨中国东西部的算力网络逐渐形成。
一个值得讨论的问题是，为什么偏偏是这些地方？
如果按照科技产业实力排序，北京、上海、深圳、杭州显然更有优势。如果按照电力价格和土地成本排序，西部还有许多符合条件的城市。
实际上，中国算力枢纽的选址，涉及能源分布、产业结构、网络时延、气候条件，以及地方政府之间的竞争。
这项工程最初启动于2022年。国家发展改革委等部门确定了京津冀、长三角、粤港澳大湾区、成渝、内蒙古、贵州、甘肃、宁夏八大国家算力枢纽，并在其中规划了十大数据中心集群。
严格来说，八大枢纽是八个区域，具体落点包括张家口、芜湖、韶关、贵安、中卫、庆阳、和林格尔等地，以及长三角生态绿色一体化发展示范区、天府和重庆集群。
四年过去，生成式AI改变了中国算力市场。
原本围绕互联网、云计算和数据存储形成的区域布局，开始承接大模型训练、推理服务和人工智能应用需求。
这也使八大枢纽的选址逻辑，具有了新的商业意义。
一、算力为什么要离开北京、上海和深圳？
理解算力枢纽的选址，首先要理解数据中心是一门什么生意。
一家互联网公司在北京研发大模型，需要采购大量GPU，建设服务器集群，为设备提供稳定的供电、冷却和网络服务。
其中，芯片和服务器决定了初期资本投入，电力则是一项长期运营成本。
假设一座数据中心的平均IT负载为100兆瓦，全年运行8760小时，PUE为1.2，那么其全年总耗电量约为10.5亿千瓦时。
电价每降低0.1元/千瓦时，理论上每年就能节省约1.05亿元电费。
这是一个简化测算，没有考虑负载变化和具体电力合同，但足以说明数据中心为什么对电价敏感。
而在北京、上海、深圳这样的城市，土地、电力供应和建设指标都面临约束。
把大规模数据中心继续放在这些城市，意味着企业需要承担较高的资源成本。
问题在于，数据中心也不能无限远离客户。
金融交易、工业控制、视频会议等业务，对网络时延十分敏感。即使服务器成本更低，如果数据传输距离过长，也可能影响实际使用效果。
于是，张家口、芜湖、韶关获得了机会。
张家口靠近北京，能够承接首都地区的互联网、人工智能及政企客户需求，同时具备可再生能源和大型数据中心建设条件。
芜湖地处长三角，能够服务上海、南京、杭州、合肥等城市形成的产业集群。
韶关则处于广东省内，能够承接广州、深圳以及部分港澳相关的数据处理需求。
这三个地方的共同优势，是在建设成本和客户距离之间形成平衡。
它们既拥有建设大型数据中心所需要的空间，又能够相对便捷地连接中国数字经济最活跃的市场。
这一选址逻辑早在2022年就已经明确。
国家发展改革委提出，金融证券、工业互联网、远程医疗和人工智能推理等对网络要求较高的业务，应优先考虑东部枢纽；后台加工、离线分析、存储备份等业务，则可以向西部转移。
因此，东部算力枢纽的核心价值，来自客户密度和服务半径。
对于一家AI公司来说，能够稳定获得客户、保持较高算力利用率，有时比单纯节省电费更重要。
二、西部四个枢纽，争夺的是能源和产业投资
与东部相比，内蒙古、贵州、甘肃、宁夏的选址逻辑明显不同。
它们的共同特点，是拥有相对充裕的能源、土地或者自然冷却条件。
不过，四个地区各有优势。
内蒙古和林格尔拥有风电、光伏及电力供应基础，可以发展大规模绿色数据中心。
贵州贵安的优势，则来自较早形成的数据中心产业。
早在“东数西算”工程启动之前，贵州就开始承接云计算和政务数据中心项目，大型运营商和科技企业逐步在当地形成投资布局。
这是一种典型的产业先发优势。
数据中心建成后，需要稳定的电力、通信网络、运维团队、设备供应商和专业服务公司。
一家大型企业落地，往往会带动相关基础设施建设，降低后续项目进入的难度。
宁夏中卫的条件又有所不同。
中卫位于腾格里沙漠边缘，气候干燥，年平均气温约8.8℃。
对数据中心来说，较低的环境温度有利于利用自然冷源，减少部分机械制冷需求。
这项优势在高耗电的数据中心产业中具有实际价值。
甘肃庆阳则属于后来赶上人工智能产业周期的代表。
相较于较早发展云计算和存储业务的贵州，庆阳近年来积极引入智能计算项目，推动国产芯片、GPU集群和大模型训练产业发展。
这四个地方实际上代表着不同的竞争路径。
内蒙古侧重能源规模，贵州依靠产业积累，宁夏拥有气候及能源条件，甘肃则在智能算力需求增长期间加速布局。
其背后还有一个更深层次的能源经济问题。
长期以来，中国的电力资源与能源消费中心存在空间错配。
西北、华北部分地区拥有丰富的风能、太阳能及传统能源资源，但东部沿海集中了大量工业和服务业需求。
解决这种错配，可以把电力输送到东部，也可以把一部分高耗电产业放到能源产地。
数据中心恰好提供了后一种可能。
电力通过本地网络进入服务器，数据和计算结果则通过光纤传输。
不过，算力跨区域调度同样需要支付网络成本，还会遇到数据安全、时延以及集群协同问题。
所以，西部枢纽的竞争力还需要建立在算力能够稳定出售的基础上。
从商业角度看，电力资源决定了数据中心建设的成本下限，客户需求和算力利用率则决定了投资回报。
三、成都和重庆，为什么能够成为另一种枢纽？
八大枢纽中，成渝是一个很特殊的存在。
它既没有北京、上海那样高度集中的互联网总部，也不同于以能源资源为主要优势的内蒙古和宁夏。
成都和重庆能够入选，主要来自区域产业规模、计算需求和西南能源条件的结合。
成都拥有软件、电子信息、游戏、科研及人工智能产业基础。
重庆拥有汽车、电子制造、装备工业和工业互联网需求。
这两座城市本身就在产生大量数据和计算任务。
例如，一家汽车制造商需要利用AI开展质量检测、生产计划优化、供应链分析和自动驾驶仿真。
不同业务对算力的要求并不一样。
产线实时检测可能要求服务器靠近工厂，大规模仿真、历史数据分析则可以使用远距离算力。
成都和重庆因此有机会同时发展本地计算和跨区域计算服务。
更重要的是，西南地区还拥有相对丰富的水电资源，以及连接四川、重庆与西部其他省份的能源和通信基础设施。
央视新闻此次提到，重庆部分数据处理任务已经跨越约2000公里，利用新疆的风光能源开展大模型训练，再将结果传回产业端。
这说明成渝枢纽的价值，正在从本地数据中心建设延伸到跨区域算力协调。
它既是算力消费者，也可以是算力服务的组织者。
在全国算力版图中，成渝承担着连接西部能源资源与区域产业需求的角色。
这一特点决定了它不能简单套用内蒙古或宁夏的模式。
成渝未来的竞争优势，很大程度上取决于工业、汽车和软件企业能否真正形成规模化的AI需求。
四、八大枢纽最大的挑战，是算力能不能卖出去
八大枢纽的设立解决了一个重要问题：中国的大型数据中心应该优先建设在哪里。
如今，真正考验它们的，是运营效率。
国家数据局数据显示，截至2025年底，全国智能算力规模达到159万PFLOPS，八大国家算力枢纽及十大集群的智算规模占全国八成以上。
到2026年6月底，全国智能算力规模进一步增长至约218.5万PFLOPS。
算力规模迅速增长，意味着大量地方政府、运营商、互联网企业和投资机构正在投入资金建设基础设施。
但建设规模不等于商业收入。
对于一家数据中心运营商来说，项目的商业模型通常需要考虑四项指标：建设投资、电力和运维成本、客户租赁或算力采购价格，以及实际利用率。
假设两个地区建设了同样规模的计算集群。
A地区电价较低，但利用率只有40%；B地区电价稍高，却拥有稳定的大客户，利用率能够达到80%。
在设备折旧等条件相近的情况下，B地区完全可能获得更好的投资回报。
智能算力比传统机架托管还要复杂。
GPU的采购价格较高，设备更新速度较快。如果算力在较长时间里闲置，折旧和资金成本仍然会持续发生。
由此产生了一个容易被忽略的问题：在各地争相建设智算中心的同时，谁来承担最终的算力需求？
过去，互联网和云计算企业通常是主要客户。
现在，大模型公司、科研机构、制造业企业和政务单位成为新的潜在购买者。
但不同客户的支付能力、使用周期和采购方式差异很大。
训练大模型可能需要长时间占用大型GPU集群，推理服务则更在意持续调用成本、软件优化和服务稳定性。
随着模型推理需求增长，算力竞争也将逐步延伸至Token生产成本。
庆阳已经提出围绕词元计算发展产业。芜湖则尝试把不同类型的算力资源放到平台上销售，推动通算、智算、超算等资源的统一交易。
这些探索指向同一个方向：数据中心需要进一步形成标准化、可计量、能够持续交付的服务。
而国家一体化算力网络建设，也在降低跨区域采购和调度的难度。
2026年8月，央视披露，全国超过六成算力已被纳入统一监测范围，全国一体化算力网监测调度试验验证平台进入试运行。
一旦算力可以更加透明地进行比较和采购，地方之间的竞争就可能进一步加剧。
拥有更低能源成本、更好网络条件、更高设备利用率的运营商，会在价格竞争中获得优势。
当然，市场暂时还难以做到完全统一。
不同芯片架构、通信网络、软件框架和数据安全要求，都会影响实际使用。
尤其是大型模型训练，任务能否拆分、能否跨地域运行，远比普通云资源采购复杂。
但总体趋势已经清楚：算力逐渐从地方基础设施项目，进入跨区域竞争的服务市场。
这会改变地方政府的招商逻辑。
过去，一座大型数据中心落地，能够带来可观的固定资产投资，包括土建、供配电系统、服务器和网络设备采购。
但数据中心投入运营之后，直接创造的就业岗位通常有限。
如果当地缺少AI软件、模型服务以及行业应用企业，数据中心对经济发展的长期带动作用也可能低于最初预期。
因此，地方政府未来需要争取的，除了服务器和投资额，还包括稳定的算力采购订单，以及愿意围绕算力设施开展业务的企业。
从这个意义上看，八大枢纽的竞争才刚刚进入第二阶段。
张家口、芜湖和韶关，依靠靠近东部需求市场的区位优势；成渝依靠产业基础和区域计算需求；内蒙古、贵州、甘肃、宁夏则需要进一步发挥各自的能源、气候、网络和产业积累优势。
这些优势最终都需要接受商业市场的检验。
2022年，“东数西算”首先要解决的是算力建设布局问题。
2026年，当AI训练、模型推理和企业应用开始持续消耗大量算力时，中国已经建立起一个具有相当规模的计算基础设施体系。
接下来，投资者和地方政府需要更加关注算力利用率、客户结构、实际收入以及投资回报。
八大枢纽的未来地位，最终取决于谁能够以更低的综合成本，向全国客户稳定交付有价值的计算服务。
这也是中国算力产业从大规模建设走向商业化运营必须回答的问题。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen Wen Wu Zhao
On 28 October 2026, the news coverage reported the latest developments in the construction of China ' s eight national computing hubs.
The one-way network latency has been contained within 2 milliseconds; in Inner Mongolia and Linger, the share of green power exceeds 85 per cent; and Gansu Qingyang is becoming an important base for LLM training and word counting in the country.
From the mouth of Zhang to the lake, from Guian to Ningxia, Guizhou, to Ningxia, a computational network has evolved across the East and West of China.
One question that deserves to be discussed is why are these places?
It is clear that Beijing, Shanghai, Shenzhen and Hangzhou are better placed to rank according to the power of the technology industry. If ranked according to electricity prices and land costs, there are many more eligible cities in the west.
In fact, the location of China ' s computing hub involves energy distribution, industrial structure, network latency, climatic conditions and competition among local governments.
This work was initially started in 2022. Departments such as the National Development and Reform Commission (NDC) have identified and planned 10 major data centre clusters in the eight major national computing hubs of Kyosin, Long Triangle, Hong Kong Bay, Chengsu, Inner Mongolia, Guizhou, Gansu and Ningxia.
The eight hubs are, strictly speaking, eight regions, with specific locations such as Zhang Jiakou, Joo-Lu, Xuan Guan, Gui An, Middleguard, Qingyang and Linger, as well as the Long Triangle Eco-Green Integrated Development Model Zone, the Tsar and the Chongqing Cluster.
Four years later, Generating AI changed China's calculator market.
The regional layout, originally built around the Internet, cloud computing and data storage, began to take on the need for LLM training, reasoning services and artificial intelligence applications.
This has also given new commercial significance to the location logic of the eight hubs.
I. Why did you leave Beijing, Shanghai and Shenzhen?
Understanding the location of the computing hub begins with understanding what the data centre is.
An Internet company that developed LLM in Beijing needed to purchase large quantities of GPUs, build server clusters and provide stable power, cooling and network services for equipment.
Of these, chips and servers determine initial capital inputs, while electricity is a long-term operating cost.
Assuming an average IT load of 100 MW, 876 hours of operation and 1.2 hours of PUE for the year, the total annual electricity consumption is about 1.05 billion kilowatt-hours.
Every 0.1 kWh reduction in electricity prices would theoretically save about $105 million per year.
This is a simplified measurement that does not take into account load changes and specific power contracts, but is sufficient to explain why the data centre is sensitive to electricity prices.
In cities such as Beijing, Shanghai and Shenzhen, land, electricity supply and construction indicators are constrained.
The continuation of large-scale data centres in these cities means that enterprises bear higher resource costs.
The problem is that data centres cannot remain indefinitely away from customers.
Financial transactions, industrial controls, videoconferencing, etc. are very sensitive to the network latency. Even if the server costs are lower, if the data are transmitted too far, it may affect the effectiveness of the actual use.
As a result, Zhang ' s mouth, Zhang ' s lake, and Zheng Guan were given opportunities.
The Zhang family is close to Beijing and has access to the Internet, artificial intelligence and the needs of clients in the capital, as well as to renewable energy and large data centres.
The Joo-Huk Triang, which serves the industrial clusters formed in cities such as Shanghai, Nanjing, Hangzhou and Chai Fat.
It is located in Guangdong Province and is able to cover data-processing needs related to Guangzhou, Shenzhen and some Hong Kong/Australia.
The common advantage of these three places is to strike a balance between construction costs and customer distance.
They have both the space needed to build large data centres and the relative ease with which to connect China ' s most dynamic digital economy.
The logic of this location was established as early as 2022.
The National Commission for Development and Reform (CNDP) has proposed that the Eastern hub should be given priority for more demanding Internet-based operations such as financial securities, the industrial Internet, telemedicine and artificial intelligence reasoning, and that back-office processing, offline analysis and backup storage could be transferred to the West.
Thus, the core value of the East Calculator Hub is derived from the client density and service radius.
For an AI company, stable access to customers and high computing power utilization are sometimes more important than simply saving electricity.
Four hubs in the west, competing for investment in energy and industry
In comparison with the east, the locations of Inner Mongolia, Guizhou, Gansu and Ningxia are clearly different in logic.
Their common feature is that they have relatively ample energy, land or natural cooling conditions.
However, each of the four regions has advantages.
Inner Mongolia and Linger have a wind, photovoltaic and power supply base and can develop large-scale green data centres.
Guiang, Guizhou, has the advantage of having an earlier data centre industry.
The cloud computing and government data centre project began in Guizhou well before the start of the East China East Count project, with large operators and science and technology enterprises gradually developing local investment patterns.
This is a typical industry pre-emptive advantage.
Once the data centre is completed, there will be a need for stable power, communications networks, transport teams, equipment suppliers and professional services.
The landing of a large enterprise often leads to the construction of relevant infrastructure and makes it more difficult for follow-up projects to enter.
Ningxia's conditions are different.
It is located on the edge of the Tengri desert and has a dry climate with an average annual temperature of about 8.8°C.
For data centres, lower ambient temperatures facilitate the use of natural sources of cold and reduce some mechanical refrigeration requirements.
This advantage has practical value in the high-power-consuming data centre industry.
Gansu Qingyang, on the other hand, belongs to representatives who later catch up with the AI industry cycle.
Compared to Guizhou, where cloud computing and storage operations were developed earlier, Qingyang has been active in recent years in introducing smart computing projects to promote the development of the national production chips, GPU clusters and LLM training industries.
These four places actually represent different paths to competition.
Inner Mongolia, which focuses on the size of energy, Guizhou relies on industrial accumulation, Ningxia has a climate and energy conditions, and Gansu accelerates the layout during the growth of intelligent computing demand.
There is also a deeper energy economy behind it.
There has long been a space mismatch between China ' s electricity resources and its energy consumption centres.
While parts of the north-west and north-east China are rich in wind, solar and traditional energy resources, the east coast is concentrated in large demand for industry and services.
This mismatch can be addressed by sending electricity to the east or by placing a portion of the high-volume-consuming electricity industry in energy sources.
Data centres happen to offer the latter possibility.
Electricity enters the server through local networks, and data and calculations are transmitted via fibre optics.
However, calculation of cross-regional movements will also have to cover network costs, as will data security, latency and cluster synergies.
Thus, the competitiveness of the Western hub also needs to be based on the ability to stabilize sales.
From a commercial point of view, power resources determine the lower cost limits for the construction of data centres and the return on investment is determined by client demand and power utilization.
Why can Chengdu and Chongqing become another hub?
Of the eight hubs, Sensu is a very special presence.
It does not have a highly centralized Internet headquarters like Beijing, Shanghai, nor does it have an internal Mongolia and Ningxia, where energy resources are the main advantage.
Chengdu and Chongqing can be selected mainly from a combination of regional industrial size, calculated demand and south-west energy conditions.
Chengdu has software, electronic information, games, scientific research and an artificial intelligence industry base.
Chongqing has automobile, electronic manufacturing, equipment industry and industrial Internet demand.
Both cities themselves produce a wealth of data and computational tasks.
For example, a car manufacturer needs to use AI for quality testing, production plan optimization, supply chain analysis and autopilot simulation.
The requirements for computing differ from operation to operation.
Real-time testing of production lines may require servers to approach the plant, and large-scale simulations and historical data analysis can use remote computing.
Chengdu and Chongqing thus have the opportunity to develop both local and cross-regional computing services.
More importantly, the south-west region also has relatively rich hydropower resources, as well as energy and communications infrastructure linking Sichuan, Chongqing and other western provinces.
As mentioned in the news this time, some of the Chongqing data-processing tasks have already crossed about 2,000 kilometres, using Xinjiang's wind energy for LLM training, and then back to industry.
This demonstrates the value of the hub, which is extending from local data centre construction to cross-regional calculus coordination.
It is both a consumer and an organizer of computing services.
In the national run-up, Sung-su assumed the role of connecting energy resources to regional industrial needs in the west.
This characteristic dictates that it cannot simply apply the Inner Mongolia or Ningxia model.
The future competitive advantage will depend to a large extent on whether industry, automobiles and software enterprises can truly form a scalable AI demand.
The biggest challenge for four or eight hubs is whether or not they can sell it.
The establishment of the eight hubs addressed an important issue: where should China ' s large data centres be built as a matter of priority?
Today, the real test is operational efficiency.
According to data from the National Data Agency, by the end of 2025, the country had an intelligent calculus of 1.59 million PFLOPS, with eight major national computing hubs and ten major clusters accounting for over 80 per cent of the country ' s total.
By the end of June 2026, the size of the country ' s intelligent computing capacity had further increased to about 2,185 000 PFLOPS.
Rapidly growing calculators mean that a large number of local governments, operators, Internet enterprises and investment agencies are investing in infrastructure.
But the scale of construction does not equal business income.
For a data centre operator, the project ' s business model usually needs to take into account four indicators: construction investment, electricity and transportation costs, customer rental or computing prices, and actual utilization.
It is assumed that the same size of the computational clusters were built in both regions.
Electricity prices are lower in Area A, but only 40 per cent are used; in Area B, where electricity prices are slightly higher, there are large and stable customers, with a utilization rate of 80 per cent.
Where conditions, such as depreciation of equipment, are similar, Area B may well have a better return on investment.
Smart algorithms are more complex than traditional hangars.
GPU procurement costs are higher and equipment upgrades are faster. Depreciation and financial costs will continue to occur if computing power is idle over a longer period of time.
As a result, there is an easy problem to ignore: At the same time that local competitions are building centres of wisdom, who will bear the ultimate need for capacity?
In the past, Internet and cloud computing enterprises were usually the main customers.
Now, LLM companies, research institutions, manufacturing enterprises and government units are potential new buyers.
However, the ability to pay, the life cycle and the manner of procurement varied greatly from client to client.
Training LLM may require long periods of time to occupy large GPU clusters, while reasoning services are more concerned with the cost of continuous calls, software optimization and service stability.
As demand for model reasoning grows, computing competition will gradually extend to Token production costs.
Qingyang has proposed the development of industry around the word count. For its part, Shaan Lake attempts to market different types of computing resources on the platform, promoting uniform trade in such resources as computing, intelligence, overcaling, etc.
These explorations point in the same direction: data centres need to be further standardized, measurable and deliverable.
The development of integrated national computing networks is also making cross-regional procurement and movement less difficult.
In August 2026, Omnibus disclosed that more than 60 per cent of the national calculus had been included in the unified monitoring and that the National Integrated Calculus Network monitoring and movement control pilot validation platform had been put into commission.
Local competition is likely to increase further once computing power allows for more transparent comparison and procurement.
Operators with lower energy costs, better network conditions and higher equipment utilization would gain advantage in price competition.
Of course, for the time being, full market integration is difficult.
Different chip structures, communication networks, software frameworks and data security requirements affect actual use.
In particular, large-scale model training, the ability of missions to split and to operate across geographical areas is much more complex than the acquisition of conventional cloud resources.
However, the overall trend is clear: the economy is gradually moving from local infrastructure projects to competitive services markets across regions.
This will change the business logic of local government.
In the past, the landing of a large data centre has resulted in considerable investment in fixed assets, including the procurement of civil, electrical distribution systems, servers and network equipment.
However, when data centres are operational, the number of jobs directly created is usually limited.
In the absence of local AI software, modelling services and industry applications, the long-term driving effect of data centres on economic development may be lower than originally expected.
Therefore, what local governments need to do in the future includes, in addition to servers and investments, stable computing purchase orders, as well as businesses willing to operate around computing facilities.
In this sense, the eight hubs have only just entered the second phase of competition.
Zhang Jia-kwan, Zhang Shui-kwan and Zhui-kwan, relying on their locational advantages close to the eastern demand market; Zhang Zhang Zhang Zhongxia Zhang Zhang Zhang Zhang Zhui, Zhang Zhuang Zhuang Zhui, Zhang Zhuang Zhang Zhui, Zhang Zhang Zhuang Zhang Zhui, Zhang Zhang Zhang Zhu, Zhang Zhang Zhu Zhu, Zhang Zhu, Zhang Zhang Zhu, Zhang Zhu Zhu, Zhang Zhang Zhu, Zhang Zhang Zhu, Zhang Zhang Zhang Zhu, Zhang Zhang Zhu, Zhang Zhu, Zhang Zhang Zhu, Zhu, Zhang Zhang Zhu, Zhang Zhu, Z
Ultimately, these advantages need to be tested in commercial markets.
In 2022, “East China and West” was the first to address the problem of the configuration of power construction.
In 2026, when AI training, model reasoning and business applications began to consume significant amounts of computing power on a continuous basis, China had established a system of computing infrastructure of considerable scale.
Next, investors and local governments need to pay more attention to the use of computing power, client structure, real income and return on investment.
The future status of the eight hubs ultimately depends on who is able to steadily deliver valuable computing services to clients across the country at lower combined costs.
This is also a question to be answered by China’s computing industry from large-scale construction to commercial operations.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.