AI算力扩张的下一道瓶颈：电力供应|AI's Next Bottleneck: Electricity Supply

产业分析 / Industry / 2026.09.04

从2030年用电展望切入，用四项关键指标甄别可落地的AI基础设施项目。 / Four metrics for investable AI infrastructure.

谈到AI基础设施，市场最易追踪的是GPU订单，最易被忽视的却是"何时真正接通电源"。劳伦斯伯克利国家实验室2026年6月的研究显示，美国数据中心到2030年的基准用电量为649太瓦时，约占全美总用电的11.8%；综合不确定性区间在521至843太瓦时之间。如此宽泛的区间表明，决定回报的不仅是芯片出货，还涉及供电、输电、散热、审批和设备实际利用率。

For AI infrastructure, GPU orders are the easiest metric to track, while the date a facility actually receives power is often overlooked. A June 2026 update from Lawrence Berkeley National Laboratory estimates a 2030 reference case of 649 terawatt-hours for U.S. data-center electricity use, about 11.8% of national consumption, with a compounded uncertainty range of 521 to 843 TWh. That wide range shows that returns depend on more than chip shipments: power supply, transmission, cooling, permitting and utilization also matter.

美国能源部2026年7月公布的国家输电需求研究报告草案提到，超大规模AI数据中心和制造业正推动负荷持续上升，新负荷接入与瓶颈缓解都亟需更多输电设施；意见征集截至9月7日。在我看来，AI资本开支已从"谁能抢到GPU"过渡到"谁能将兆瓦转化为稳定收入"的阶段。项目公布的规划容量，既不等于已接入电网的容量，也不等于被客户实际消化的算力。

The U.S. Department of Energy's draft National Transmission Needs Study, released in July 2026, says hyperscale AI data centers and manufacturing are driving load growth and that more transmission is needed for new load connections and congestion relief; its public-comment period runs through September 7. My view is that AI capital spending is moving from who can obtain GPUs to who can turn megawatts into durable revenue. Announced capacity is not the same as energized capacity, and energized capacity is not the same as compute that customers consistently use.

投资者可以搭建一个四列追踪表。第一列是并网节点：研究完成、协议签订、变电站建设与通电日期。第二列是电力质量：已锁定的电力占规划负荷的比重，以及其中稳定部分与可中断部分的比例。第三列是运营效率：除PUE之外，还需关注GPU利用率、冷却水耗和每兆瓦创造的收入。第四列是负荷弹性：训练任务能否错峰或跨区域调度，储能与本地发电能否在用电高峰期响应。

Investors can build a four-column tracker. First, interconnection milestones: study completion, signed connection agreements, substation construction and expected energization. Second, power quality: contracted power as a share of planned load, and how much is firm versus interruptible. Third, operating efficiency: PUE is only a starting point; also track GPU utilization, cooling-water use and revenue per megawatt. Fourth, load flexibility: whether training can shift across hours or regions, and whether storage and onsite generation can respond during peak periods.

伯克利实验室2026年5月发布的报告将数据中心灵活性归纳为计算迁移、设施调节、储能应用和现场发电四种机制，并指出灵活负荷可加速并网流程，但无法取代发电与输电领域的长期投入。对个股研究而言，这意味着应对新闻中的"规划吉瓦"打个折扣：缺少接入协议、交付时间表、客户合同与利用率路径的项目，其估值更接近于期权而非确定产能。

A May 2026 Berkeley Lab study groups data-center flexibility into four mechanisms: shifting compute, adjusting facility operations, using storage and deploying onsite generation. It says flexible load can accelerate interconnection but cannot replace long-term investment in generation and transmission. For company analysis, announced gigawatts deserve a discount: without an interconnection agreement, delivery schedule, customer contract and utilization path, a project should be valued more like an option than committed capacity.

乐观情形下，电网投资加速，灵活负荷获得更快接入，算力供给与客户需求同步攀升；基准情形下，部分地区排队现象持续，拥有现成电力合同和成熟园区的运营商将享受估值溢价；悲观情形下，需求预测偏高，电网为闲置负荷提前投资，相关成本最终由企业、用户或纳税人承担。三种情形下，最具价值的共性资产并非宣传口号，而是可核实的并网权、稳定电源与真实客户。

In the upside case, grid investment accelerates, flexible loads connect faster, and compute supply ramps with customer demand. In the base case, queues persist in some regions and operators with existing power contracts and mature campuses earn a premium. In the downside case, demand forecasts prove too high and grids invest ahead of idle loads, leaving companies, customers or taxpayers with the cost. Across all three cases, the valuable assets are verifiable grid access, reliable power and real customers—not slogans.

因此，研究AI产业链时，不妨调整研究顺序：先核实电力供应与并网进度，再确认设备交付情况，最后审视利用率与单位经济性。只有四项指标同步改善，资本开支才更有可能转化为现金流。若仅能观察到芯片采购，却看不到通电与客户消化的实际能力，就应当提高折现率、缩短预测期，并为延迟和闲置预留更宽的安全边际。

A practical research sequence is therefore: verify power and interconnection first, confirm equipment delivery second, and then test utilization and unit economics. Capital spending is more likely to become cash flow only when all four improve together. If chip purchases are visible but energization and customer absorption are not, use a higher discount rate, a shorter forecast horizon and a larger margin of safety for delays and idle capacity.

Risk notice / 风险提示 风险提示：用电预测、并网进度、电价波动、政策调整与客户需求均存在不确定性；本文情景分析并非确切预测，亦不构成投资建议。 / Risks: Power forecasts, grid timing, prices, policy and demand can change. Scenarios are not forecasts or investment advice.

关注EthanFang，持续追踪算力如何转化为可验证的现金流。 / Follow for measurable AI infrastructure analysis.