AI 新能源制造全景图 · 一度电的成本 · 2026-07 · 制造侧New Energy Manufacturing × AI · The Cost of One kWh · Jul 2026 · The making side

最便宜的电,最惨烈的行业 The cheapest power, the bloodiest industry

中国用二十年把光伏度电成本打到 0.043 美元/kWh(累计降约 90%),让 91% 的新建可再生项目比最便宜的化石能源还便宜——这是人类应对气候变化最重要的单一贡献,没有之一。代价是光伏全行业 2025 年预亏超 600 亿元、数十家企业破产退市。2024 年,LCOE 曲线十余年来第一次抬头(+0.6%):制造业能给的都给完了,下一段曲线的主战场不在硅料车间,在电网调度室。而 AI 对这个行业最大的贡献,不是优化它的产线,是成为它最大的新客户。 In twenty years China drove solar’s LCOE to $0.043/kWh (−90%), making 91% of new renewables cheaper than the cheapest fossil alternative — humanity’s single largest climate contribution, full stop. The price: the PV industry pre-announced 60B+ yuan of losses for 2025 amid a wave of bankruptcies. In 2024 the LCOE curve ticked up (+0.6%) for the first time in over a decade: manufacturing has given all it can, and the next battleground is the dispatch room, not the ingot shop. AI’s greatest gift to this industry is not optimising its lines — it is becoming its biggest new customer.

本图相信:周期已经分叉——锂电这一部翻篇了(碳酸锂年内 11.7→20.9 万元/吨,储能电芯加急单溢价 10%,散单价高于集采价=教科书式卖方市场);AI 最扎实的战绩在锂电视觉质检(宁德时代溧阳灯塔:产量 +320%、缺陷 −99%,WEF 认证)与需求侧(科技巨头买走全球 43% 清洁能源 PPA)(A/B)。This map believes: the cycle has forked — the battery act is over (lithium carbonate ran ¥117k→209k/t within the year; rush-order cells carry a 10% premium; spot above framework price is the textbook seller’s market); AI’s hardest wins sit in battery visual inspection (CATL Liyang lighthouse: output +320%, defects −99%, WEF-certified) and on the demand side (big tech buys 43% of global clean-energy PPAs) (A/B).

本图不相信:「AI 发现新材料」。Nature 2026-01 对 A-Lab 论文发出更正(「新材料」未必是新的);MatterGen 旗舰案例被证与 1971 年已知化合物相同且在其训练集里;MIT Technology Review:截至 2026 年初,零个 AI 预测材料进入规模化生产。德勤:91% 的中国制造业 AI 项目未达预期,老师傅的缺陷检出率常高于 AI(A/B)。It does not believe «AI discovers new materials». Nature corrected the A-Lab paper (its «new» materials may not be new); MatterGen’s flagship case matches a compound known since 1971 — present in its own training set; MIT Technology Review: as of early 2026, zero AI-predicted materials have reached commercial production. Deloitte: 91% of Chinese manufacturing AI projects miss expectations, and the veteran inspector often beats the model (A/B).

主脊:一块光伏板与一颗电芯的一生(路线之争→上游材料→制造扩产→装机应用→并网消纳→贸易壁垒→退役回收),③⑤标红;三条结构带:死法图鉴、规则替代补贴、AI 渗透评分卡;判断层:现金流 > 技术代际 > 产能规模——过剩周期里,产能是负债不是资产。三图分工:AI 要电与电网→energy,动力电池装车→car,化工材料→chem,工厂内卷→factory。 The spine: the life of one panel and one cell (route wars → upstream → manufacturing → installation → curtailment → trade walls → recycling), ③ and ⑤ in red; three bands: the book of deaths, rules replacing subsidies, the AI scorecard; the judgment layer: cash flow > technology generation > capacity — in a glut, capacity is a liability. Sisters: energy, car, chem, factory.

$0.043

2024 年全球公用事业级光伏 LCOE(美元/kWh),二十年累计 −90%;同年 +0.6%——十余年来首次抬头(A)Global utility-scale solar LCOE in 2024, down 90% in twenty years — and up 0.6%, the first rise in over a decade (A)

600 亿+

光伏全行业 2025 财年预亏(元,创历史新高);9 家龙头合计预亏 435–500 亿,通威一家 90–100 亿(A/B)PV industry pre-announced FY2025 losses (¥, a record); nine leaders alone guide ¥43.5–50B, Tongwei ¥9–10B (A/B)

+205% / −54%

2026Q1 独立储能 vs 新能源配储增速——136 号文取消强制配储:把「被迫装的」清零,把「算得过账的」放大三倍(A)Q1-2026 standalone storage vs mandated-pairing growth — Document 136 zeroed the forced and tripled the bankable (A)

91%

德勤调研:中国制造业 AI 项目未达预期的比例;失败主因是需求错位与数据混乱,不是算法(B)Share of Chinese manufacturing AI projects missing expectations (Deloitte); the killers are mis-scoped needs and messy data, not algorithms (B)

口径裁判:2025 光伏新增以能源局 317GW 为准(媒体常引 315);隆基 BC 产能 80GW vs 100GW 系「自有 vs 含合作」口径;储能出货排名两套口径对华为/海博思创位次差异巨大,引用必须注明统计机构;「中印 LCOE 0.33/0.38 美元」为小数点排版错误(应为 0.033/0.038);中国功率预测「准确率」=1−归一化均方根误差,与国际 RMSE 口径不可直接比较;灯塔工厂数据是精选样本——德勤的 91% 才是分母。Refereeing the numbers: use the NEA’s 317GW for 2025 additions; LONGi’s 80 vs 100GW BC is an owned-vs-partnered basis; storage-shipment rankings disagree wildly by agency — always cite the counter; the «$0.33/0.38 LCOE» for China/India is a decimal typo ($0.033/0.038); China’s forecast «accuracy» is 1−nRMSE, incomparable with RMSE; lighthouse factories are hand-picked numerators — Deloitte’s 91% is the denominator.

中心装置 · 抬头的成本曲线The central device · The curve that looked up

这个行业二十年的全部进步浓缩成一条单调下降的曲线;2024 年它第一次不再单调。分子(制造成本)与分母(可卖出的电)在两头同时出事。Twenty years of progress compressed into one monotonic curve — until 2024, when it stopped being monotonic. Numerator (cost) and denominator (sellable power) broke at the same time.

分子:制造端把成本打穿了Numerator: manufacturing hit the floor
组件 0.69 元/W(2024Q4 低点)、储能系统 0.52 元/Wh、陆风 1600 元/kW——全部逼近或跌破现金成本;硅料一度 3.4 万元/吨而行业平均成本 4.37 万。这些数字再往下压,压出来的不是更低的度电成本,是倒塔事故(哈尔滨半月两起)和破产清算。2026 年的反内卷武器不是限价(违反反垄断法),是能耗强制国标——用环保门槛做产能出清的合法外衣,并首次把组件衰减率写进强制标准。Modules at ¥0.69/W (the Q4-2024 floor), storage systems ¥0.52/Wh, onshore turbines ¥1,600/kW — all at or under cash cost; polysilicon once ¥34k/t against a ¥43.7k industry cost. Pressed further, these numbers produce collapsed towers (two in Harbin in a fortnight) and bankruptcy filings, not cheaper power. The 2026 anti-involution weapon is not price-fixing (illegal) but mandatory energy-intensity standards — environmental thresholds as the lawful cloak of capacity exit, with degradation rates written into compulsory standards for the first time.

分母:发出来的电,送不出去Denominator: the power cannot leave
2026Q1 弃光率 9.4%、弃风 8.6%,双双击穿 95% 红线;西藏光伏利用率仅 60.8%;风电装机新增 1577 万千瓦而发电量同比只增 0.4%——装机去了送不出去的地方。项目级证据:新疆北塔山连续三年弃风 12.5%/21.3%/18.3%;28 省机制电量完成率仅 75.7%;风电 REITs 可比 IRR 低至 0.03%——公开市场已把弃电风险打进资产价格。Q1-2026 solar curtailment hit 9.4% and wind 8.6%, breaching the 95% line; Tibet’s solar utilisation is 60.8%; wind added 15.8GW while generation grew 0.4% — capacity went where power cannot exit. Project-level: Beitashan curtailed 12.5/21.3/18.3% three years running; provinces filled only 75.7% of guaranteed volumes; wind REITs price at IRRs as low as 0.03% — the public market has already priced the curtailment in.

判词VerdictLCOE 的下一段主战场,不在硅料车间,而在电网调度室和交易大厅——这正是本图与 energy 图的交接点。弃电的根因是省间壁垒与地方保护(制度),不是调度算法(技术):长三角用交易机制而非 AI 减少了弃风弃光。The next stretch of the LCOE curve will be fought in dispatch rooms and trading halls — the hand-off to the energy map. Curtailment is rooted in inter-provincial walls and local protection (institutions), not dispatch algorithms: the Yangtze Delta cut curtailment with a trading mechanism, not with AI.

Reading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本图为新能源制造业的结构分析,不构成任何投资建议;所有价格、装机与亏损数据保留原始口径,口径冲突处已标注;厂商与灯塔工厂数据是精选样本(分母见德勤 91%),按 D 级证据看待厂商自述;市场地图 ≠ 推荐背书;仅供研究与批判性参考。Stance: a structural read of new-energy manufacturing — not investment advice; prices, installations and losses keep original bases with conflicts flagged; vendor and lighthouse numbers are curated numerators (Deloitte’s 91% is the denominator) and self-reported claims are grade-D; a market map is not an endorsement; for research and critical reference only.