返回首页
Source: Aren Technology
In a data center, the first things to come alive are not the interfaces of AI applications, but the indicator lights on power distribution cabinets. Layers of fan noise fill the air as server racks begin to draw massive amounts of electricity, network bandwidth and cooling resources. In this way, a computing power facility is fully activated.
Bloomberg has released a straightforward report: China plans to invest around 2 trillion yuan over the next five years to build data centers across the country. The initiative aims far beyond building a handful of demonstration parks — its goal is to weave a nationwide computing network. State-owned enterprises including China Mobile and China Telecom will undertake most of the operational responsibilities.
Though appearing to be traditional infrastructure construction, this move essentially marks a reshuffle of industrial sovereignty. The bottlenecks holding back AI development lie not merely in model parameters or chip manufacturing processes. The real constraints often exist in unglamorous yet fundamental links: power supply, data center facilities, optical fiber networks and scheduling systems. Simply put, AI models can be showcased at product launches, while computing power must be grounded in physical infrastructure. Demonstrations rely on presentation, but computing power deployment depends on reinforced concrete and long-term operational contracts.
Over the past decade and more, the internet industry has long focused on front-end innovations. Mobile applications, cloud services, recommendation algorithms and generative AI all sound sophisticated and lightweight. Yet every technological leap eventually boils down to practical questions: Where does computing power come from? Who supplies electricity? Who provides hosting and maintenance services?
In the early days, most data centers were built in-house by enterprises. The cloud era later ushered in a trend of centralized deployment. As large model training pushed power consumption and bandwidth usage to the limit, standalone data centers became less relevant, and the demand for cross-regional resource scheduling grew increasingly urgent.
This vision is not a sudden idea. The groundwork has been laid by the internet backbone network, cloud computing clusters and supercomputing centers over the years. Previously, however, many projects ended up as isolated computing silos. Local governments built data centers while enterprises purchased equipment, with weak interconnection and low resource utilization rates. The new blueprint targets a unified national computing network, which distributes computing resources just like power grids allocate electricity to enable free resource flow. In other words, computing power is evolving from a corporate asset into a public infrastructure.
The timing of this push is no coincidence. AI has moved beyond basic chatbots to a more operation-intensive stage. Physical AI, intelligent agents and real-time inference all drive up computing density. Model training is a capital-intensive sprint, while inference is a long-duration endurance test. The former hinges on peak computing performance, and the latter requires stable and sustained resource supply.
There is far more than software engineering separating a lab-based AI model from real industrial application. Practical economic factors come into play across every link: cabinet density, Power Usage Effectiveness (PUE), circuit redundancy, bandwidth costs and liquid cooling upgrades, all of which continuously consume cash flow. Ultimately, the winners in the AI business are not those skilled at storytelling, but professionals who can calculate every cost down to the smallest detail — from kilowatt-hours of electricity and gigabits per second of bandwidth to server depreciation.
Notably, internet platforms are no longer the main players. Instead, state-owned telecom giants such as China Mobile and China Telecom are taking charge. This is not a random personnel arrangement, but a redivision of industrial interests. Data center operation is not a one-time equipment sales business; it functions like a long-term infrastructure project featuring slow capital recovery, long contract cycles and high operational stability.
Led by state-owned enterprises, the whole sector can follow unified planning, realize centralized procurement, adopt long-term payment mechanisms and promote standardized supply chains. For the market, this sends a clear signal: speculative sectors chasing quick profits will fade out. Only companies capable of undertaking large-scale engineering projects, delivering products and providing long-term maintenance can secure steady cash flow in this new landscape.
The industrial chain stretches across multiple segments. The upstream covers power equipment, distribution systems and backup power supplies. The midstream includes civil engineering, server cabinets, cooling systems and optical fiber interconnection. The downstream consists of servers, storage devices and acceleration cards, followed by hosting, operation and maintenance, and network optimization services.
This industry differs from consumer electronics driven by hit products, or internet advertising subject to fluctuating traffic. It operates more like high-speed railways and power grids: infrastructure comes first, before data and computing power run on top of it. Companies with fast delivery, high energy efficiency and stable production capacity will gain priority access to bulk procurement orders.
History is replete with similar transformations. During the expansion of power grids, the biggest beneficiaries were not companies making empty claims, but equipment manufacturers and construction teams that delivered power to every household. When the railway network was built, profits first went to suppliers of steel rails, locomotives and signal systems, as well as engineering contractors rather than pundits grandly outlining visions. The nationwide computing network follows the same rules. It pulls AI away from flashy technological stunts and grounds it in industrial supply, shifting competition from algorithm showcases to the overall ledger of energy and infrastructure.
From a broader perspective, this is also a major national strategic investment. Computing power is more than just a pile of servers; it underpins data sovereignty, industrial independence and economic resilience. Once the national computing network is fully operational, cross-regional scheduling, local inference and unified resource pooling will become standard practices. This will lower the threshold for small and medium-sized teams to rent computing resources and reduce the overall costs of technological innovation.
Nevertheless, the success of this system hinges on three core pillars: power supply, resource scheduling and operational efficiency. A single weak link could turn the 2-trillion-yuan investment from capacity building into idle assets.
Therefore, it is time to look beyond product launch events. What truly matters are construction announcements, equipment bidding results, delivery schedules and energy efficiency upgrades. Enterprises that secure orders from state-owned clients and develop cost-effective solutions for green power, liquid cooling and optical interconnection will gain a firm foothold amid this industrial restructuring.
Admittedly, AI’s future lies in model innovation, but its long-term development is determined by unassuming data centers, network lines and power meters. When the first batch of data parks are fully interconnected, the lights that light up the night will represent an entirely new industrial order. In the end, the computing power business will return from conceptual hype to the most tangible reality: electricity bills.
2万亿元砸向全国算力网，先亮起来的不是AI界面，是机房里的配电柜和电表
来源：阿忍科技
机房里最先亮起来的，不是AI应用的界面，而是配电柜上的指示灯，风扇声一层层铺开，服务器柜体开始吞电、吞网、吞散热，一座算力工厂就这样被点亮了。现在，彭博社抛出的消息很直接，中国准备在未来5年投入约2万亿元人民币，在全国各地修建数据中心，目标不是搭几个示范园区，而是织出一张覆盖全国的算力网，中国移动、中国电信等国企，将扛起大部分运营责任。
这件事看上去像基建，实则更像一场产业主权的重排。因为AI的瓶颈，从来不只在模型参数里，也不只在芯片制程上，真正卡脖子的地方，往往藏在电力、机房、光纤、调度系统这些最不性感的环节里。说白了，模型可以在发布会上跑，算力必须在地上落地，前者靠演示，后者靠钢筋混凝土和长期合同。
往前倒推，过去十多年，互联网产业的叙事一直偏向前台。移动应用、云服务、推荐算法、生成式AI，听起来都足够聪明，也足够轻盈。可每一次技术跃迁，最后都要落回一件事，算力从哪里来，谁来供电，谁来托管，谁来运维。早年的数据中心更多是企业自建，后来进入云时代，集中化成了主旋律，再往后，训练大模型把电力和带宽推到极限，单点机房的意义开始弱化，跨区域调度的必要性被彻底放大。
这并不是突然冒出来的新想法。互联网骨干网、云计算集群、超算中心，这些路线早就给今天铺过路。只是过去，很多项目停留在“算力孤岛”阶段，地方建机房，企业买设备，彼此之间连接松散，利用率也常常不高。今天的蓝图变了，目标变成全国算力网，像电网调配电力一样调配算力，让资源能流动起来。换句话说，算力正在从资产，变成基础设施。
再往深处看，这个时间点并不偶然。AI已经从聊天机器人，走向更重的阶段，物理AI、智能体、实时推理，都在抬高计算密度。训练是烧钱，推理是长跑，前者拼峰值算力，后者拼持续供给。一个模型从实验室跑到产业现场，中间隔着的不只是软件工程，还有极其现实的单体经济模型，机柜密度、PUE、线路冗余、带宽成本、液冷改造，每一项都在吞现金流。你想想看，真正能把AI做成生意的，不是会讲故事的人，而是能把每千瓦时电、每Gbps带宽、每台服务器的折旧，算到骨头里的人。
更有意思的是，这次的主角不是互联网平台，而是中国移动、中国电信这样的国企。这背后不是偶然的组织选择，而是利益结构的重新分工。数据中心不是一次性卖设备的生意，它更像长期运营的基础设施账本，资金回收慢，合同周期长，稳定性高。国企主导，意味着项目更容易进入统一规划，采购更容易集中，支付更偏长期，供应链也更容易围绕标准化展开。对于市场来说，这种模式的信号非常明确，赚快钱的题材会退潮，能拿工程单、能做交付、能持续运维的公司，才有资格吃到真正的现金流。
这条链条很长。上游要电力设备、配电系统、备用电源，中游要土建、机柜、冷却、光纤互联，下游是服务器、存储、加速卡，再往后还有托管、运维、网络优化。它不像消费电子那样靠爆款拉动，也不像互联网广告那样靠流量起伏，它更接近高铁和电网，先把路铺出来，再让流量和算力在路上跑。结果呢，谁的交付快、能效高、产能稳，谁就更容易被纳入大单采购体系。
历史上，类似的转折并不少见。电网扩张时，真正受益的不是喊口号的公司，而是能把电送到千家万户的设备商和施工方。铁路建设时，先赚钱的也往往不是最会讲宏大叙事的人，而是钢轨、机车、信号、工程承包这些硬环节。今天的算力网，同样遵循这套冷规则。它把AI从“技术炫技”拽回到“工业供给”，把竞争从算法秀场拉回到能源与基础设施的总账本。
更冷一点看，这还是一场关于国家能力的下注。算力不是简单的服务器堆叠，它对应的是数据主权、产业自主和经济韧性。全国算力网一旦成形，跨区域调度、就近推理、统一资源池都会成为常态，中小团队租用算力的门槛会下降，创新的启动成本也会被压低。可这套系统能否真正跑起来，取决于三个词，供电、调度、效率。任何一个环节拖后腿，2万亿元就可能从“能力建设”滑向“资产沉淀”。
所以接下来别只盯着发布会。真正值得看的，是开工公告、设备中标、交付节奏、能效改造。谁能拿到国企订单，谁能在绿色电力、液冷、光互联上做出低成本方案，谁才可能在这轮产业重构里站稳。AI的未来当然属于模型，但决定它能走多远的，往往是那些沉默的机房、线路和电表。
等第一批园区连成网，夜里亮起来的就不只是灯，而是一整套新的产业秩序，算力这门生意，终究会从概念回到电费单上。
2 Trillion Yuan Invested in National Computing Network: Before AI Interfaces Light Up, Distribution Cabinets and Power Meters in Data Centers Take the Lead
Source: Aren Technology
In a data center, the first things to come alive are not the interfaces of AI applications, but the indicator lights on power distribution cabinets. Layers of fan noise fill the air as server racks begin to draw massive amounts of electricity, network bandwidth and cooling resources. In this way, a computing power facility is fully activated.
Bloomberg has released a straightforward report: China plans to invest around 2 trillion yuan over the next five years to build data centers across the country. The initiative aims far beyond building a handful of demonstration parks — its goal is to weave a nationwide computing network. State-owned enterprises including China Mobile and China Telecom will undertake most of the operational responsibilities.
Though appearing to be traditional infrastructure construction, this move essentially marks a reshuffle of industrial sovereignty. The bottlenecks holding back AI development lie not merely in model parameters or chip manufacturing processes. The real constraints often exist in unglamorous yet fundamental links: power supply, data center facilities, optical fiber networks and scheduling systems. Simply put, AI models can be showcased at product launches, while computing power must be grounded in physical infrastructure. Demonstrations rely on presentation, but computing power deployment depends on reinforced concrete and long-term operational contracts.
Over the past decade and more, the internet industry has long focused on front-end innovations. Mobile applications, cloud services, recommendation algorithms and generative AI all sound sophisticated and lightweight. Yet every technological leap eventually boils down to practical questions: Where does computing power come from? Who supplies electricity? Who provides hosting and maintenance services?
In the early days, most data centers were built in-house by enterprises. The cloud era later ushered in a trend of centralized deployment. As large model training pushed power consumption and bandwidth usage to the limit, standalone data centers became less relevant, and the demand for cross-regional resource scheduling grew increasingly urgent.
This vision is not a sudden idea. The groundwork has been laid by the internet backbone network, cloud computing clusters and supercomputing centers over the years. Previously, however, many projects ended up as isolated computing silos. Local governments built data centers while enterprises purchased equipment, with weak interconnection and low resource utilization rates. The new blueprint targets a unified national computing network, which distributes computing resources just like power grids allocate electricity to enable free resource flow. In other words, computing power is evolving from a corporate asset into a public infrastructure.
The timing of this push is no coincidence. AI has moved beyond basic chatbots to a more operation-intensive stage. Physical AI, intelligent agents and real-time inference all drive up computing density. Model training is a capital-intensive sprint, while inference is a long-duration endurance test. The former hinges on peak computing performance, and the latter requires stable and sustained resource supply.
There is far more than software engineering separating a lab-based AI model from real industrial application. Practical economic factors come into play across every link: cabinet density, Power Usage Effectiveness (PUE), circuit redundancy, bandwidth costs and liquid cooling upgrades, all of which continuously consume cash flow. Ultimately, the winners in the AI business are not those skilled at storytelling, but professionals who can calculate every cost down to the smallest detail — from kilowatt-hours of electricity and gigabits per second of bandwidth to server depreciation.
Notably, internet platforms are no longer the main players. Instead, state-owned telecom giants such as China Mobile and China Telecom are taking charge. This is not a random personnel arrangement, but a redivision of industrial interests. Data center operation is not a one-time equipment sales business; it functions like a long-term infrastructure project featuring slow capital recovery, long contract cycles and high operational stability.
Led by state-owned enterprises, the whole sector can follow unified planning, realize centralized procurement, adopt long-term payment mechanisms and promote standardized supply chains. For the market, this sends a clear signal: speculative sectors chasing quick profits will fade out. Only companies capable of undertaking large-scale engineering projects, delivering products and providing long-term maintenance can secure steady cash flow in this new landscape.
The industrial chain stretches across multiple segments. The upstream covers power equipment, distribution systems and backup power supplies. The midstream includes civil engineering, server cabinets, cooling systems and optical fiber interconnection. The downstream consists of servers, storage devices and acceleration cards, followed by hosting, operation and maintenance, and network optimization services.
This industry differs from consumer electronics driven by hit products, or internet advertising subject to fluctuating traffic. It operates more like high-speed railways and power grids: infrastructure comes first, before data and computing power run on top of it. Companies with fast delivery, high energy efficiency and stable production capacity will gain priority access to bulk procurement orders.
History is replete with similar transformations. During the expansion of power grids, the biggest beneficiaries were not companies making empty claims, but equipment manufacturers and construction teams that delivered power to every household. When the railway network was built, profits first went to suppliers of steel rails, locomotives and signal systems, as well as engineering contractors rather than pundits grandly outlining visions. The nationwide computing network follows the same rules. It pulls AI away from flashy technological stunts and grounds it in industrial supply, shifting competition from algorithm showcases to the overall ledger of energy and infrastructure.
From a broader perspective, this is also a major national strategic investment. Computing power is more than just a pile of servers; it underpins data sovereignty, industrial independence and economic resilience. Once the national computing network is fully operational, cross-regional scheduling, local inference and unified resource pooling will become standard practices. This will lower the threshold for small and medium-sized teams to rent computing resources and reduce the overall costs of technological innovation.
Nevertheless, the success of this system hinges on three core pillars: power supply, resource scheduling and operational efficiency. A single weak link could turn the 2-trillion-yuan investment from capacity building into idle assets.
Therefore, it is time to look beyond product launch events. What truly matters are construction announcements, equipment bidding results, delivery schedules and energy efficiency upgrades. Enterprises that secure orders from state-owned clients and develop cost-effective solutions for green power, liquid cooling and optical interconnection will gain a firm foothold amid this industrial restructuring.
Admittedly, AI’s future lies in model innovation, but its long-term development is determined by unassuming data centers, network lines and power meters. When the first batch of data parks are fully interconnected, the lights that light up the night will represent an entirely new industrial order. In the end, the computing power business will return from conceptual hype to the most tangible reality: electricity bills.
2万亿元砸向全国算力网，先亮起来的不是AI界面，是机房里的配电柜和电表
来源：阿忍科技
机房里最先亮起来的，不是AI应用的界面，而是配电柜上的指示灯，风扇声一层层铺开，服务器柜体开始吞电、吞网、吞散热，一座算力工厂就这样被点亮了。现在，彭博社抛出的消息很直接，中国准备在未来5年投入约2万亿元人民币，在全国各地修建数据中心，目标不是搭几个示范园区，而是织出一张覆盖全国的算力网，中国移动、中国电信等国企，将扛起大部分运营责任。
这件事看上去像基建，实则更像一场产业主权的重排。因为AI的瓶颈，从来不只在模型参数里，也不只在芯片制程上，真正卡脖子的地方，往往藏在电力、机房、光纤、调度系统这些最不性感的环节里。说白了，模型可以在发布会上跑，算力必须在地上落地，前者靠演示，后者靠钢筋混凝土和长期合同。
往前倒推，过去十多年，互联网产业的叙事一直偏向前台。移动应用、云服务、推荐算法、生成式AI，听起来都足够聪明，也足够轻盈。可每一次技术跃迁，最后都要落回一件事，算力从哪里来，谁来供电，谁来托管，谁来运维。早年的数据中心更多是企业自建，后来进入云时代，集中化成了主旋律，再往后，训练大模型把电力和带宽推到极限，单点机房的意义开始弱化，跨区域调度的必要性被彻底放大。
这并不是突然冒出来的新想法。互联网骨干网、云计算集群、超算中心，这些路线早就给今天铺过路。只是过去，很多项目停留在“算力孤岛”阶段，地方建机房，企业买设备，彼此之间连接松散，利用率也常常不高。今天的蓝图变了，目标变成全国算力网，像电网调配电力一样调配算力，让资源能流动起来。换句话说，算力正在从资产，变成基础设施。
再往深处看，这个时间点并不偶然。AI已经从聊天机器人，走向更重的阶段，物理AI、智能体、实时推理，都在抬高计算密度。训练是烧钱，推理是长跑，前者拼峰值算力，后者拼持续供给。一个模型从实验室跑到产业现场，中间隔着的不只是软件工程，还有极其现实的单体经济模型，机柜密度、PUE、线路冗余、带宽成本、液冷改造，每一项都在吞现金流。你想想看，真正能把AI做成生意的，不是会讲故事的人，而是能把每千瓦时电、每Gbps带宽、每台服务器的折旧，算到骨头里的人。
更有意思的是，这次的主角不是互联网平台，而是中国移动、中国电信这样的国企。这背后不是偶然的组织选择，而是利益结构的重新分工。数据中心不是一次性卖设备的生意，它更像长期运营的基础设施账本，资金回收慢，合同周期长，稳定性高。国企主导，意味着项目更容易进入统一规划，采购更容易集中，支付更偏长期，供应链也更容易围绕标准化展开。对于市场来说，这种模式的信号非常明确，赚快钱的题材会退潮，能拿工程单、能做交付、能持续运维的公司，才有资格吃到真正的现金流。
这条链条很长。上游要电力设备、配电系统、备用电源，中游要土建、机柜、冷却、光纤互联，下游是服务器、存储、加速卡，再往后还有托管、运维、网络优化。它不像消费电子那样靠爆款拉动，也不像互联网广告那样靠流量起伏，它更接近高铁和电网，先把路铺出来，再让流量和算力在路上跑。结果呢，谁的交付快、能效高、产能稳，谁就更容易被纳入大单采购体系。
历史上，类似的转折并不少见。电网扩张时，真正受益的不是喊口号的公司，而是能把电送到千家万户的设备商和施工方。铁路建设时，先赚钱的也往往不是最会讲宏大叙事的人，而是钢轨、机车、信号、工程承包这些硬环节。今天的算力网，同样遵循这套冷规则。它把AI从“技术炫技”拽回到“工业供给”，把竞争从算法秀场拉回到能源与基础设施的总账本。
更冷一点看，这还是一场关于国家能力的下注。算力不是简单的服务器堆叠，它对应的是数据主权、产业自主和经济韧性。全国算力网一旦成形，跨区域调度、就近推理、统一资源池都会成为常态，中小团队租用算力的门槛会下降，创新的启动成本也会被压低。可这套系统能否真正跑起来，取决于三个词，供电、调度、效率。任何一个环节拖后腿，2万亿元就可能从“能力建设”滑向“资产沉淀”。
所以接下来别只盯着发布会。真正值得看的，是开工公告、设备中标、交付节奏、能效改造。谁能拿到国企订单，谁能在绿色电力、液冷、光互联上做出低成本方案，谁才可能在这轮产业重构里站稳。AI的未来当然属于模型，但决定它能走多远的，往往是那些沉默的机房、线路和电表。
等第一批园区连成网，夜里亮起来的就不只是灯，而是一整套新的产业秩序，算力这门生意，终究会从概念回到电费单上。