复仇者联盟

Envision, DeepSeek’s Ulanqab pivot taps China’s energy strengths, natl plan; ‘Clusters of AI computing hubs show institutional strengths’_我的网站

一 |     

    Envision Galaxy Campus, the world's largest single AI computing facility by China's green technology company Envision Group, enters into operation in Ulanqab, North China's Inner Mongolia Autonomous Region, on August 6, 2026. Photo: Envision GroupChina's largest artificial intelligence (AI) computing industrial park was completed and put into operation in Ulanqab, North China's Inner Mongolia Autonomous Region on Thursday. Meanwhile, Chinese AI start-up DeepSeek is reportedly also planning a massive AI data center there.  
As more Chinese companies choose to stack data centers in a lesser-known city in Northern China, seeking cheaper land, electricity and cooler climate, Chinese industry experts said on Friday that the country's institutional advantages in developing AI are being fully demonstrated, providing solid support for an emerging industry that is continuously promoting welfare globally as an international public good.
China's green technology company Envision Group announced that its Envision Galaxy Campus - a large-scale AI data center (AIDC)  - in Ulanqab officially entered operation on Thursday. It is also the single AIDC with the largest token output capacity in the world, Chinese green technology company Envision Group said in a statement sent to the Global Times on Friday.
Envision Group said its Envision Galaxy Campus in Ulanqab, with a planned total capacity of 2 gigawatts, integrates renewable energy, intelligent power management and high-density computing at unprecedented scale.
Envision is not the only Chinese company interested in building an AI data center in Ulanqab.
Recently, Chinese AI start-up DeepSeek posted a job opening for a senior data center operations and maintenance engineer, listing "Ulanqab" as the work location, which analysts said "hinted at a new data center of DeepSeek."
A recent report by Bloomberg also indicated that DeepSeek is also planning a massive AI data center there, citing people familiar with the matter. The report describes the project as "an ambitious move" in computing infrastructure from the start-up that's shown it can compete with the leading AI models of Silicon Valley.
DeepSeek is looking to add 1 gigawatt worth of compute in Ulanqab, about 350 kilometers northwest of Beijing, the report said. The announced and planned AI data centers are the latest addition to more than a dozen firms that have projects planned in the city.
Chinese experts said that these data centers located in Ulanqab, with its cool and dry climate, enjoy a natural advantage in heat dissipation that accounts for a major part of a data center's maintenance. But it is more than that. The city's status as a massive producer of clean energy as well as a hub of China's national computing power network, and a national development blueprint that goes back for years, are behind the companies' decisions, and these factors together are a vivid display of China's institutional advantage.
Unique advantage
Envision told the Global Times that its Envision Galaxy Campus features an area of 120,000 square meters, roughly equivalent to 20 standard football fields, and supports the computing capacity for 1 million graphics processing units (GPUs) in parallel, making it the AIDC with the highest token output capacity globally and setting a new benchmark for compute density in AI infrastructure.
With a purposely built proprietary AI Power System that enables stable, high-efficiency power transmission and energy storage just like a "power expressway and electricity granary", the facility can deliver up to ten times compute output per square meter more than conventional data centers.
Explaining why Ulanqab was chosen as the facility site, Envision said the city is one of the eight digital infrastructure hubs of the national "east data, west computing" project, allowing the center to have minimal latency for online computing tasks - a required feature as the number of GPUs multiplies at AI clusters. 
Some 67 percent of the area's electricity is generated from wind and solar, making the base an ideal place to tap green energy, the company noted, adding that it aims to develop 5 gigawatts of green AIDC capacity in desert and arid regions worldwide by 2030.
"Light travels at roughly 300,000 kilometers per second. But when it comes to supercomputing, latency becomes a matter so the physical distance between computing demand and computing capacity becomes a factor to be considered," Ma Jihua, a veteran technology industry analyst, told the Global Times on Friday.
"The AI computing clusters will effectively serve demand generated in Northern China, particular the Beijing-Tianjin-Hebei region, where the digital economy is growing fast," Ma said. 
Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute, said that companies' selection of Ulanqab is more than a simple geographic location balancing - it is an act that reflects the underlying shift in companies' site-selection logic, as the compute paradigms shift from storage to training and inference.
To support the synchronous training of a cluster equipped with as much as one million GPUs, Envision needs to build its own wind farms and dedicated transmission lines; this is not merely a matter of "finding a cool spot to house server racks" but rather integrating power system engineering directly into the AI compute infrastructure, Tian told the Global Times on Friday.
"Data centers in the storage era have been passive consumers of the power grid. Now, in the AI training era, they require the grid to be custom-designed for them," Tian said. "This way, the energy company can lock electricity costs at its AI centers within a controllable range, giving it teeth for competition in the upcoming token economy era." 
As the next frontier of AI is infrastructure, the cost of computing becomes a company's core competitiveness, Tian said.
By tapping rich renewable energy sources and direct power transmission from wind farm to data center, the electricity price at data centers in Hohhot, provincial capital of Inner Mongolia Autonomous Region and home to another data center cluster, is only one third of that in Beijing, domestic news portal bjnews.com.cn reported in June.
The unfolding story around Ulanqab also highlighted China's institutional strengths. 
"The government drew up a plan, laid down the backbone of a national computing power network, and private-sectors companies followed up with their own investment, forming a self-reinforcing positive loop," Ma said.
To underpin AI growth, China's Government Work Report this year unveiled plans to launch new infrastructure projects focused on hyperscale intelligent computing clusters, alongside the coordinated development of computing capacity and power supply, according to the Xinhua News Agency. 
Meanwhile, China's robust clean energy infrastructure and aggressive expansion plans for the upcoming five-year cycle are also poised to secure the critical power supply needed to meet the energy-intensive demands of AI computing.
The National Development and Reform Commission, China's top economic planner, said computing power network construction during the 15th Five-Year Plan period (2026-30) will lead to new direct investment of 4 trillion yuan ($592.68 billion), citing estimates by relevant institutions, the Securities Times reported on July 31.
Statistics show that by the first quarter of 2025, the total computing power of the eight data computing hubs of the "east data, west computing" project had reached 215.5 EFLOPS, with intelligent computing power accounting for 80.8 percent, according to Xinhua.
Computing power benefiting world
Yin Hejun, minister of Science and Technology, termed AI, robotics and innovative drugs as China's "new new three" at a recent policy briefing.
Ma said the expanding buildup of Chinese AI infrastructure reflects the strong growth of China's AI-related industries, which have emerged as new drivers of China's export growth, representing China's new core strategic industries and fresh economic growth engines.
As China's AI computing power grows, the country will be more well positioned to handle global demand, as it becomes the world's "token factory" by capitalizing on its energy strengths, Ma noted.
As China's technology exports now extend to digital services, with companies increasingly providing AI algorithms, cloud computing, and digital solutions to overseas markets, analysts said the competitively priced compute power from China will serve the world as an international public good in the emerging "token economy."
"In the future, computing clusters in Northern China will be ideal for handling computing needs generated from the Central Asian and Northeast Asian markets, due to its geographic proximity," Ma said.
Although some Western media outlets have recently targeted China's cost-effective AI products with a flurry of far-fetched skepticism in a bid to curb the growth of China's AI industry, Chinese AI products are already benefiting the world, analysts and industry insiders said. 
For instance, Chinese AI company iFLYTEK said that its AI-powered intelligent teaching system has helped expand Chinese language learning worldwide and supported the digital development of Chinese education in overseas markets. The AI-powered intelligent Chinese teaching system has greatly expanded students' practice opportunities: On average, learners now speak Chinese 12 times per session, compared with only twice in regular classes at a Thai school, the company said.
As China embraces the concept of technology for good and inclusive AI to be shared by the whole of mankind, more people in the world will benefit from this trend, analysts noted.
。    作者声明：该图片由AI生成          公司毛利率-24%。         作者｜刘俊群          编辑｜刘钦文          你有没有想过，每次在手机上和AI对话、让AI写一段代码或生成一张图片，背后都在消耗一种叫"词元"（Token）的东西——它是AI处理信息的最小计量单位。         有一家公司专门做这门生意：把底层算力变成Token，像自来水一样按量供应给开发者和企业。

二 | 它就是北京硅基流动科技股份有限公司（下称"硅基流动"），近日，公司向港交所主板递交上市申请，拟冲刺IPO。          作者声明：该图片由AI生成          来源：罐头图库          公司用了不到三年时间，营收从6000元增长到了5533万元，注册用户突破1000万，日均处理Token近5800亿次。

三 | 不过，2025年，公司毛利率只有-24%，全年亏损3.45亿元，经营现金流净流出1.72亿元。

四 | 卖得越多亏得越多——这家AI"卖铲人"的IPO故事，能讲得通吗？          01          租算力卖Token，三年亏损超4亿冲IPO          硅基流动从创立之初便确立了“不绑定任何单一云、芯片或大模型厂商”的中立定位，致力于打造连接上游算力、中游模型与下游应用的系统软件中间层。         截至2026年4月，其平台已适配超过170个模型，跨芯片及多模型适配能力据弗若斯特沙利文数据位列全球第一。         公司的业务模式分为两类。         第一类是公有云服务，它又分两种卖法：一种是面向中小开发者的"无服务器词元服务"——用户预充值，按实际消耗的Token量计费，类似于充话费。2025年这块收入1430万元，付费客户从2455人暴涨到71.6万人。         另一种是面向中大型企业的"专属实例服务"——给客户预留专用算力，按量计费。2025年收入1500万元，客户从7家增长到49家，单客户平均年花费从8.77万提升到30.5万元。2025年整体贡献收入2926万元，占总营收的52.9%，首次超过一半。          作者声明：该图片由AI生成          来源：罐头图库          第二类是本地部署解决方案，2025年该业务收入2607万元，占总营收的47.1%。主要面向金融、能源、政务等对数据安全要求高的机构。

五 | 简单说就是将AI系统部署在客户自己的服务器上，靠软件许可费、实施费和后续维保费赚钱。客户数量从28家降到20家，但单客户平均年花费从22.4万猛增到130.3万元。         两类业务的背后，都是同一个模式：向外租算力，加工成Token卖给客户。算力是成本大头，Token是计价单位，赚的是中间的"加工差价"。         正是凭借上述业务布局，硅基流动在2025年春节DeepSeek爆火时抓住了机遇，成为业内首家在国产芯片上部署DeepSeekR1/V3模型的公司，一战成名。         随后公司收入和用户规模均呈现爆发式增长：2023年成立4个月仅收入6000元，2024年营收735万元，2025年直接飙升至5533万元，同比增长653%；同期注册用户超1028万，服务企业客户超1.3万家，日均Token吞吐量5785亿次，单日最高突破1万亿次。          作者声明：该图片由AI生成          来源：罐头图库          按2025年词元年吞吐量计，它已是国内最大的独立生态词元供应商，在所有词元供应商中排名前五。

六 |          但收入增长的背面，是持续扩大的亏损。2023年至2025年，净亏损分别为0.12亿元、0.82亿元和3.45亿元。经调整净亏损分别为0.12亿元、0.54亿元和1.87亿元。         此外，2023年公司毛利率高达83.3%；2024年降至39.4%；2025年直接转负至-24%。也就是说，公司每提供1元Token服务，算力等成本要花掉1.24元。         亏损背后，2025年销售成本高达6863万元，较全年5533万元的总营收高出1330万元。仅免费Token代金券就消耗5421万元。再加上研发投入2.09亿元，仅研发与销售两项支出就近3亿元。          作者声明：该图片由AI生成          来源：罐头图库          《招股书》中提示，公司尚处于亏损状态、经营现金流持续为负，未来若无法改善盈利或及时获得外部融资，可能影响持续经营能力。         客户结构的变化也暗藏风险。公司前五大客户收入占比从2023年的100%降至2025年的45%，单一最大客户占比从83.3%降至13.6%。且存在客户与供应商重叠问题。《招股书》显示，硅基流动2025年第二大客户同时也是其第五大供应商。该客户贡献销售收入710万元（占总收入12.8%），同时又向硅基流动提供算力资源，采购额1260万元。         据《每日经济新闻》援引天使投资人、AI专家郭涛观点称，大厂虽有技术、资源和生态优势，但市场需求多样且复杂，难以覆盖所有场景。独立第三方（例如硅基流动）可聚焦特定垂直领域、提供定制化服务，部分企业出于数据安全和避免过度依赖单一厂商的考量，也会选择与独立第三方合作。         02          “清华博士”打造“Token工厂”，阿里、美团齐齐入股          公司创始人袁进辉，2008年取得清华大学计算机科学博士学位，师从中国AI奠基人张钹院士。         2013年至2016年间，他就职于微软中国，担任主管研究员，主导研发了当时全球速度最快的大规模主题模型训练系统LightLDA。

七 | 2017年，他创办了北京一流科技（OneFlow），专注于深度学习框架的开发。         2023年4月，美团联合创始人王慧文旗下的“光年之外”以换股方式收购了OneFlow，袁进辉随之成为光年之外的联合创始人。然而仅隔两个月，王慧文因病离岗，光年之外随后被美团整体接手。          作者声明：该图片由AI生成          来源：罐头图库          但面对美团的留任邀请，袁进辉选择了拒绝。

八 | 同年8月，他带领原OneFlow的核心团队重新出发，成立了硅基流动。回顾那段经历，他曾感叹“像坐过山车”。         颇具戏剧性的是，据新媒体“创投日报”披露，袁进辉在筹备硅基流动之初，DeepSeek创始人梁文锋曾考虑投资硅基流动，但因种种原因未能成行。事后他在朋友圈感叹"为自己（当时）没有远见感到后悔"。

九 |          尽管错失了梁文锋的资金加持，硅基流动的股东阵容依然堪称豪华。         目前，阿里巴巴持股7.42%，为第一大外部股东；华为旗下的哈勃科技持股4.07%；美团持股1.05%。

十 | 此外，商汤、360、创新工场、壁仞科技及其战投部门、晶科能源、蔚来资本、巨人网络、携程战投、联通系资本、纪源资本等二十余家机构也纷纷入局。          作者声明：该图片由AI生成          来源：罐头图库          袁进辉及四个员工激励平台合计控制44.48%的投票权。

十一 | 公司成立不到三年完成7轮融资，估值从天使轮2.28亿元飙升至B+轮77.4亿元，涨了超过33倍。         《招股书》显示，公司本次IPO募资净额约40%将用于研发投入，约25%用于租赁算力资源，约15%用于扩大客户覆盖，约10%用于海外扩张。

十二 |          如今，袁进辉带着不到三年的运营记录、-24%的毛利率，走到了港交所的门前。只是，一家靠补贴吸引用户、靠租赁算力维持运转的AI中间层企业，能否真正在人工智能浪潮中站稳脚跟呢？          你怎么看硅基流动港股IPO？评论区聊聊吧！。

Current article:http://lyv.minpinpengcemiuanjiuxuan.pics/ee5zwnd/20260826/4376.html

Published on:01:00:14