AI 能源算力全景图 · 一焦耳变成多少 token · 2026-07 · 反向约束Energy & Compute × AI · The Joule-to-Token Chain · Jul 2026 · The Reverse Constraint
首页说「AI 吞噬世界」——这张图是它的反面:世界的电、地、水、硅,正在吞噬 AI。算力不是凭空长出来的比特,是被物理供给曲线死死摁住的重工业;理解 AI 的天花板,不看参数,看一焦耳能变成多少 token
The master map says «AI eats the world» — this map is its inverse: the world’s power, land, water and silicon are eating AI. Compute is no weightless bit but heavy industry pinned by physical supply curves; to see AI’s ceiling, read not parameters but how many tokens one joule buys
本图相信:主约束已从「缺芯」结构性迁移到「缺电」——芯片扩产以 12–18 个月计,输电线审批 7–11 年、变压器交付 4–5 年、并网排队中位超 5 年;马斯克 2026 达沃斯直言「很快芯片会多于能通电运行的数量」(B)。This map believes: the binding constraint has structurally migrated from chips to power — fabs scale in 12–18 months while transmission permits take 7–11 years, transformers 4–5, and interconnection queues median 5+; Musk at Davos 2026: «soon we will make more chips than we can plug in» (B).
本图不相信:「AI 几乎不耗电」与「AI 正烤干地球」——两边都在夸张。真实趋势线是单位成本快速下降 × 总量爆炸式增长,净效应是总消耗持续攀升,且高度集中于少数电网与水源紧张的地理节点(杰文斯悖论)。It does not believe either «AI barely uses power» or «AI is baking the planet» — both exaggerate. The real line is unit cost falling fast × volume exploding faster: net consumption keeps climbing, concentrated on a few strained grids and watersheds (Jevons).
主脊:算力供应链生命周期七环节(芯片→整机→选址→电力→散热→调度→折旧),②④标红为约束深水区;三条结构带:瓶颈迁移史与核电复兴、能耗水耗的诚实层、capex 与收入缺口;判断层:电与地 ＞ 芯片(时间尺度错配),效率进步是总消耗的催化剂不是解药。姊妹分工:芯片与代码侧→code,制造业底盘→factory,建筑工程→build,加密算力前史→crypto。
The spine: seven links of the compute supply chain (chips → racks → siting → power → cooling → scheduling → depreciation), with ② and ④ flagged deep water; three bands: the bottleneck migration and nuclear revival, the honest layer of energy and water, capex vs the revenue gap; the judgment: power and land ＞ chips (a mismatch of clocks), and efficiency is the accelerant of total burn, not its cure. Siblings: chips and code → code, manufacturing → factory, construction → build, crypto’s compute prehistory → crypto.
2.6 TW
美国电网并网排队总容量(≈现有总装机 2 倍),中位等待超 5 年;2000–2020 年申请容量最终仅 13% 投运(LBNL A)——电不是想接就能接The US interconnection queue (~2× installed capacity), median wait 5+ years; only 13% of 2000–2020 applications ever energised (LBNL, A) — power is not on demand
0.24–0.4 Wh
一次典型文本查询的能耗(一厂商首份一手全栈测量 0.24 中位;独立估算 0.3;比旧共识 2.9Wh 低约一个数量级)——不含训练摊销与重负载(10 万 token 输入约 40Wh)One typical text query (0.24 Wh median in the first full-stack first-party measurement; 0.3 independent; an order below the old 2.9 Wh) — excluding training amortisation and heavy loads (~40 Wh at 100k input tokens)
7250 亿$
四大云厂商 2026 年 capex 指引(+77%;2025 约 4100 亿);高盛测算 2025–2030 合计 5.3 万亿美元——对面站着 Bain 测算的 8000 亿美元收入缺口(B)Big-four 2026 capex guidance ($725B, +77%; ~$410B in 2025); Goldman sees $5.3T through 2030 — facing Bain’s $800B revenue shortfall (B)
~30%
中国智算中心平均算力利用率(官方媒体调研 B);部分国产芯片闲置率高达 70–80%——「东数西算」投资超 1 万亿元,沉睡的算力是另一种短缺Average utilisation of China’s AI compute centres (state-media survey, B); some domestic chips idle at 70–80% — past ¥1T invested, sleeping compute is scarcity of another kind
⚠️ 口径裁判(先读):① 单次能耗四口径并存——旧共识 2.9Wh(假设 1500 词长查询,高估)、独立估算 0.3Wh(自陈可能 10 倍误差)、厂商自报 0.34Wh(未同行评审)、一手全栈测量 0.24Wh 中位(A/B)——引用必须注明中位数、不含训练与重负载;② 用水四个数字最多相差近 2000 倍(每次对话 500ml→15ml→0.32ml→0.26ml):代际、统计边界、自报与独立估算的鸿沟,勿单引一个;③ 推理占比「63% 生命周期」与「80–90% AI 计算」两口径并存,仅方向可靠;④ 「每兆瓦吞吐提升 10 倍」为厂商引用第三方基准的营销口径;⑤ 负荷预测存在重复申报虚增(五年 166GW vs 分析师「不太可能超 65GW」);⑥ 循环融资中「1000 亿投资」截至 2026 初仍非最终协议;二手 GPU 价格、Burry 折旧测算为单源/做空立场,已降权;⑦ 四份深度研究交叉整理(一份全程带可核 URL 为主力);渗透%为编辑估值。
⚠️ Basis rulings (read first): ① four per-query bases coexist — the old 2.9 Wh (a 1,500-word query, inflated), 0.3 Wh independent (self-declared 10× error bars), 0.34 Wh vendor-reported (unreviewed), 0.24 Wh median first-party full-stack (A/B) — always cite as a median excluding training and heavy loads; ② four water figures span nearly 2,000× (500ml→15ml→0.32ml→0.26ml per chat): generations, boundaries, self-reports; never quote one alone; ③ inference share runs «63% of lifecycle» or «80–90% of AI compute» — direction only; ④ «10× throughput per MW» is vendor marketing citing a third-party bench; ⑤ load forecasts carry duplicate-application inflation (166GW five-year vs analysts’ «unlikely over 65GW»); ⑥ the «$100B investment» headline was not definitive as of early 2026; used-GPU prices and the short-seller depreciation math are single-source and down-weighted; ⑦ cross-compiled from four deep-research reports (one fully URL-verified as backbone); penetration %s are editorial.
◆ 中心装置 · 一焦耳 → 一FLOP → 一tokenThe central device · joule → FLOP → token
智能是一条能量转化链,每一步都有损耗与上限Intelligence is an energy-conversion chain, lossy and bounded at every step
GPU 单卡功耗五年翻近两倍(700W→1200W→整柜 130kW 实测),但单位算力能效同步改善——「耗电末日论」与「越来越省电论」都片面:功耗密度上升与能效改善同时为真。Per-card power nearly doubled in five years (700W→1200W; 130kW measured per rack), while efficiency per FLOP improved in parallel — both the doom and the thrift stories are half-true: density up and efficiency up, simultaneously.
▸ 单位在降(每 token 更便宜)Unit cost falling
一厂商 12 个月内单次查询能耗降 33 倍(一手 A);token 价格两年跌超 90%;两相浸没把 PUE 压到 1.01–1.03;新一代架构宣称每兆瓦吞吐 ×10(厂商口径 D)——省电的工程红利是真的。
One provider cut per-query energy 33× in 12 months (first-party, A); token prices fell 90%+ in two years; two-phase immersion pushes PUE to 1.01–1.03; the new architecture claims 10× throughput per MW (vendor, D) — the engineering dividend is real.
▸ 总量在爆(杰文斯悖论)Volume exploding (Jevons)
token 用量预计到 2030 年增 24 倍;全球数据中心用电 460TWh(2022)→945–1000TWh(2030,IEA);美国占比 4.4%→6.7–12%(LBNL 三情景);微软 CEO 直接援引杰文斯:「AI 越高效、越易得,用量越暴涨」(B)——效率进步是总消耗的催化剂。
Token volume is set to grow 24× by 2030; global data-centre power runs 460TWh (2022) → 945–1,000TWh (2030, IEA); the US share climbs 4.4% → 6.7–12% (LBNL’s three scenarios); the Microsoft CEO invoked Jevons outright: «the more efficient and available AI gets, the more it is used» (B) — efficiency is the accelerant.
◆ 判词Verdict低单价 × 大总量:两边都在夸张,净效应是总消耗持续攀升,且集中在少数电网/水源紧张的节点。PUE 行业平均已连续六年停滞在 1.54——「靠机房工程省电」的红利基本吃尽,往后只剩芯片能效与选址气候两张牌。Cheap units × huge volume: both camps exaggerate, and net burn keeps climbing, concentrated on a few strained nodes. Industry PUE has stalled at 1.54 for six straight years — the facility-engineering dividend is spent; what remains is chip efficiency and siting climate.