NextFin News — Ulanqab是距离北京数百公里的北部城市，已经成为中国最密集的数据中心建设集群之一。 地方当局签署了数十个项目。 Alibaba，华威，Kuaishou和专家操作员都占据了空间。 一所被废弃的学校被改建为计算设施。 吸引人的不是软件人才。 它是相对廉价的电力、可用的土地和缓解冷却负荷的气候。
建筑繁荣与模型价格的急剧下降同时发生。 5月，DeepSeek将其V4-Pro API率永久削减75%. 到10月，由硅数据汇编、金融时报引用的Token价格指数显示，综合价格比5月峰值低52%以上。 并不是每个模型都同样下降；一些高价报价比较稳定。 然而，方向是明确的：托肯人出售的情报越来越便宜。
然而销售计算成本的公司却在花费更多而不是更少。 Alibaba在6月30日结束的季度的资本支出达到676.78亿人民币，逐年上升75%，主要受AI基础设施驱动。 同期，其AI云计算服务收入增长45%，达到484.37亿元。 这两个数字反映了该行业新的紧张状况。 模型提供者竞相降低列表价格。 云运营商必须为其偿还期持续多年的设施提供资金。
动力是最显眼的线条项目。 如果单价每千瓦时0.40元，假设10兆瓦的信息技术负载持续运行，功率为1.2，则每年的电费缺口约为4 200万元。 5年来，累计差额超过2亿元人民币。 真实的合同随利用率，费率和冷却设计而异，但灵敏度足够大，可以将场地选择拉入理事室。 能够提供稳固动力、直接连接发电和降低环境温度的区域成为战略资产。
地点仍然涉及权衡。 训练集群倾向于密集的高带宽织物。 对延迟敏感的推论倾向于接近用户和网络枢纽。 大规模培训、批量工作和冷藏可转向廉价电力；交互式产品不能忽视延迟。 因此，操作员分担工作量，而不是追逐一个理想的网站。
所有权结构新增另一层风险。 收购公司自己的GPU吸收使用风险和折旧。 向云提供商出租的一家公司为了灵活性而交易较高的长期单位成本。 GPU使计算进一步复杂化：实际生命可以超过经济寿命。 更新型的芯片生成器或更有效率的模型架构，即使硬件仍能运行，也能延缓早期的购买。 廉价电能提高运行保障；它不会消除技术过时。
模型侧的价格削减不将一对一映射到较低的任务成本上。 API清单价格反映了竞争地位和基本效率。 客户越来越频繁的路线常规分类，归纳和支持工作到更便宜的模型上，并储备昂贵的模型进行更硬的推理。 Agents放大效果。 一个单一的用户请求可以触发数十个模型呼叫，工具使用和重试。 更低的托肯价格扩展了一套可行的工作；它们不会自动缩小每个完成的工作流程消耗的总计算。
Alibaba的结果显示，AI云计算服务提供了不断增长的收入，并大幅提升了经调整的EBITA. AI实验室和应用部分包括Qwen等消费品，由于公司吸收了能力投资和较高的推论成本，因此出现了经过调整的大额EBITA损失。 同样的千瓦时可以出现在云分类账上作为收入，也可以出现在模型分类账上作为成本。 芯片销售商在服务器发货时的账面销售。 云运营商只能通过多年使用回收资本。 模范公司只有在支付客户到货量时才能收回。
排序是核心管理问题。 模型价格在几个月内可以下降四分之三。 数据中心的折旧计划持续数年。 需求增长可以填补能力——或者让经营者与不再确定成本界限的资产争夺价格。 算法和硬件融资在AI所能做的天花板上的进步；它也缩短了任何给定的一代设备的舒适的回报窗口。
中国的AI产业在过去两年里证明了它可以训练有竞争力的模型和安全芯片。 现在，它正在学习一个更传统的学科：如何定位、电力、金融以及忙碌的庞大资本基础，而在这个基础之上销售的产品却越来越便宜。 Ulanqab不断上升的服务器大厅天际线，更难的答案——无论这些大厅是否在下一轮提高效率后挣得钱——将用使用率和免费现金流动来写，而不是用Token价格表来写。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
NextFin News — Ulanqab, a northern city hundreds of kilometers from Beijing, has become one of China’s densest clusters of data-center construction. Local authorities have signed dozens of projects. Alibaba, Huawei, Kuaishou and specialist operators have all taken space. One abandoned school is being converted into a computing facility. The attraction is not software talent. It is relatively cheap electricity, available land and a climate that eases cooling loads.
The building boom coincides with a sharp decline in model prices. In May, DeepSeek made permanent a 75 percent cut on its V4-Pro API rates. By October, a Token price index compiled by Silicon Data and cited by the Financial Times showed composite prices more than 52 percent below their May peak. Not every model fell equally; some premium offerings held steadier. The direction, however, is clear: intelligence sold by the Token is getting cheaper.
Yet the companies that sell compute are spending more, not less. Alibaba’s capital expenditure for the quarter ended June 30 reached 67.678 billion yuan, up 75 percent year on year, driven largely by AI infrastructure. In the same period, its AI cloud and computing services revenue rose 45 percent to 48.437 billion yuan. The two figures capture the industry’s new tension. Model providers race to lower list prices. Cloud operators must finance facilities whose payback stretches over years.
Power is the most visible line item. A hypothetical 10-megawatt IT load running continuously at a power-usage effectiveness of 1.2 would generate an annual electricity-cost gap of roughly 42 million yuan if the unit price differed by 0.40 yuan per kilowatt-hour. Over five years the cumulative difference exceeds 200 million yuan. Real contracts vary with utilization, tariffs and cooling design, but the sensitivity is large enough to pull site selection into the boardroom. Regions that can offer firm power, direct connections to generation, and lower ambient temperatures become strategic assets.
Location still involves trade-offs. Training clusters favor dense, high-bandwidth fabrics. Latency-sensitive inference prefers proximity to users and network hubs. Bulk training, batch jobs and cold storage can migrate toward cheap power; interactive products cannot ignore delay. Operators therefore split workloads rather than chase a single ideal site.
Ownership structure adds another layer of risk. A firm that buys its own GPUs absorbs utilization risk and depreciation. A firm that rents from a cloud provider trades higher long-run unit costs for flexibility. GPUs complicate the calculation further: physical life can outlast economic life. A newer chip generation or a more efficient model architecture can strand earlier purchases even if the hardware still functions. Cheap electricity improves the operating margin; it does not erase technological obsolescence.
Price cuts on the model side do not map one-to-one onto lower task costs. API list prices reflect competitive positioning as well as underlying efficiency. Customers increasingly route routine classification, summarization and support work to cheaper models and reserve expensive ones for harder reasoning. Agents amplify the effect. A single user request can trigger dozens of model calls, tool uses and retries. Lower Token prices expand the set of viable jobs; they do not automatically shrink the total compute consumed per completed workflow.
Alibaba’s results illustrate the split inside a single group. AI cloud and computing services delivered rising revenue and sharply higher adjusted EBITA. The AI laboratory and applications segment, which includes consumer products such as Qwen, posted a large adjusted EBITA loss as the company absorbed capability investment and higher inference costs. The same kilowatt-hour can appear as revenue on a cloud ledger and as cost on a model ledger. Chip vendors book sales when servers ship. Cloud operators recover capital only through years of utilization. Model companies recover it only if paying customers arrive in volume.
That sequencing is the core management problem. Model prices can fall by three-quarters in months. A data center’s depreciation schedule runs for years. Demand growth can fill capacity—or leave operators competing on price with assets that no longer set the cost frontier. Progress in algorithms and hardware raises the ceiling on what AI can do; it also shortens the comfortable payback window for any given generation of equipment.
China’s AI industry spent the last two years proving it could train competitive models and secure chips. It is now learning a more traditional discipline: how to site, power, finance and keep busy a heavy capital base while the product sold on top of that base keeps getting cheaper. Ulanqab’s rising skyline of server halls is one answer to the power question. The harder answer—whether those halls earn their keep after the next round of efficiency gains—will be written in utilization rates and free cash flow, not in Token price lists.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.