能源已成为最直观的瓶颈：德州电网运营商 Ercot 收到到 2030 年超过 226 吉瓦（GW）的用电请求，几乎是其在 2022 年批准量的 100 倍。1 月 13 日，唐纳德·特朗普总统承诺公民不会“买单”，而是由微软等科技巨头承担。
为绕开电网拥堵，企业转向“自带电力”（BYO）的离网方案：电网接入如今最长可能需要 5 年，而 xAI 在 2024 年于田纳西州用创纪录的 4 个月建成大型 GPU 集群并用卡车运来燃气涡轮与发动机。Boom 宣布将向数据中心开发商 Crusoe 供应 29 台天然气涡轮；高盛估计未来 5 年美国新增数据中心容量中最多三分之一（约 25GW）将在离网状态下建设，从而加快落地速度。
资本也在收紧并催生 SPV 等结构：xAI 于 1 月上旬完成 200 亿美元融资后，将通过 Valor Equity Partners 设立的 SPV 租赁价值 54 亿美元的英伟达 GPU；Meta 为路易斯安那州名为 Hyperion 的数据中心筹集 300 亿美元；甲骨文据报通过 SPV 获得 660 亿美元表外融资以支持 OpenAI。摩根士丹利预计到 2030 年的 5 年内，涉及私募信贷的数据中心融资将达 8000 亿美元，约占该热潮预计总借款的一半；但若只有少数公司能稳定盈利，信用冲击可能外溢至更广泛的金融体系。
Energy is the most tangible bottleneck: Texas grid operator Ercot has received requests for more than 226 gigawatts (GW) by 2030, nearly 100 times what it approved in 2022. On January 13, President Donald Trump promised citizens would not “pick up the tab”, saying tech giants such as Microsoft would instead.
To bypass gridlock, firms are adopting off-grid “bring-your-own” (BYO) power: grid connections can now take as long as five years, while xAI built a large GPU cluster in Tennessee in a record four months in 2024 by trucking in gas turbines and engines. Boom will supply 29 natural-gas turbines to data-centre developer Crusoe, and Goldman Sachs estimates up to a third of incremental U.S. data-centre capacity—about 25GW—will be built off-grid over the next five years, speeding deployment.
Capital is tightening and driving SPV-style structures: after a $20bn raise in early January, xAI will lease $5.4bn of Nvidia GPUs via an SPV set up by Valor Equity Partners; Meta raised $30bn for the Hyperion data centre in Louisiana; and Oracle reportedly raised $66bn in off-balance-sheet SPV financing to support OpenAI. Morgan Stanley expects private-credit involvement in data-centre financing to reach $800bn in the five years to 2030—about half of total borrowing in the boom—yet if only a few firms generate reliable profits, the credit risks could spill into the broader financial system.
Source: Innovations in energy and finance are further inflating the AI bubble
Subtitle: How to bust bottlenecks
Dateline: 1月 15, 2026 09:35 上午