高出 40 到 60 倍：把整套数据中心算进 AI 电子垃圾

2026 年 9 月，环境非营利组织 Basel Action Network（BAN）发布了一份口径完全不同的分析：如果把 AI 数据中心的整套基础设施——服务器与加速器、配电、制冷、备电与储能、网络交换机与线缆，再加上被新算力要求倒逼淘汰的消费级电脑与手机、电信设备——全部计入，则在 2025 至 2050 年间，将有 3.95 亿至 6.17 亿吨 AI 驱动的电子设备退役。这个数字比此前的主流估计高出 40 到 60 倍。原因不是模型变了，而是统计边界变了。

统计边界变了，结论就变了

一、数字为什么突然变大：边界移动了，而不是模型变化了

此前多数估算只数服务器与加速器，BAN 的分析把整条基础设施链都算进来：配电设备、制冷设备、备电与储能、网络设备（交换机与线缆），以及因为新算力要求而被加速淘汰的消费级电脑、手机与电信设备。BAN 给出的换算规则很简单：每 1 吉瓦的数据中心容量，对应约 7 万吨电子垃圾。因此，当分母从「服务器」换成「整套基础设施」、分子又叠加了消费设备的替代潮时，结论自然跃升一个数量级。这也是这份报告最值得学习的地方——它的方法论差异，本身就是结论差异。

# 1. Why the number jumped: the boundary moved, not the model
BOUNDARY_OLD = ["AI servers", "accelerators"]
BOUNDARY_NEW = [
    "AI servers and accelerators",
    "power distribution equipment",
    "cooling equipment",
    "backup power and energy storage",
    "networking (switches, cabling)",
    "consumer computers and phones displaced by new compute requirements",
    "telecommunications equipment with the same displacement effect",
]

def estimate(boundary, gw_capacity, tonnes_per_gw=70_000):
    # BAN's rule of thumb: every gigawatt of data-centre capacity corresponds
    # to roughly 70,000 metric tonnes of e-waste. The headline changed because
    # the inventory changed.
    return gw_capacity * tonnes_per_gw

二、头条数字，按原文口径照抄

报告的公开口径是：2025–2050 年间，AI 驱动的电子设备退役量介于 3.95 亿至 6.17 亿吨之间；若装满标准集装箱首尾相接，可绕地球约六圈（约 2300 万个集装箱）；折合年退役量约 860 万至 1310 万吨。输入侧的关键假设是：每吉瓦容量对应约 7 万吨电子垃圾；容量参照 McKinsey 的预测，到 2030 年全球数据中心容量最高可达 219 吉瓦；美国在建数据中心管道中有 4,871 个。这些数字都应当连同假设一起引用——只引结论不引边界，是环境类数据最常见的失真方式。

// 2. The headline figures, exactly as published
const banFullInfrastructure = {
  published: "2026-09-15",
  organisation: "Basel Action Network (BAN)",
  scope: "first-of-its-kind full-infrastructure analysis",
  projection: {
    window: "2025-2050",
    retired_tonnes: "395,000,000 - 617,000,000",
    versus_previous_estimates: "40x - 60x higher",
    containers: "about 23 million standard shipping containers",
    laid_end_to_end: "would circle the Earth roughly six times",
    annual_rate: "8.6m - 13.1m metric tonnes retired per year",
  },
  inputs: {
    tonnes_per_gw: 70_000,
    capacity_2030_cited: "up to 219 GW (McKinsey projection)",
    us_datacenters_in_pipeline: 4871,
  },
};

退役量由更换周期决定

三、真正的驱动是更换周期，而这属于采购事实

报告中最「工程」的一段，是对设备物理参数的拆解：每机架服务器约重 1,360 公斤，交换机最高约 30 公斤，每个机架连接的铜缆重达数百公斤；而这些设备不会为维修与再利用设计，预计每 2 至 5 年就要整体更换。把这段话翻译成采购语言：你今天写进规格书的更换周期，直接决定了未来十年退役量的量级——甚至在任何业务负载跑起来之前。当设备不可维修时，「更换」在物理上就等于「废弃」。

# 3. The replacement clock is the real driver - and it is a procurement fact
ASSET_CYCLE = {
    "server_rack_weight_kg": 1360,
    "switch_weight_kg": "up to 30",
    "copper_cable_per_rack_kg": "hundreds of kilograms",
    "designed_for_repair_and_reuse": False,
    "replacement_interval_years": (2, 5),
}

def embodied_waste(fleet, horizon_years):
    low, high = ASSET_CYCLE["replacement_interval_years"]
    # A 2-5 year replacement interval means the hardware you specify today
    # determines waste volumes for the next decade, before any workload runs.
    return {
        "replacements_per_rack": horizon_years / high,
        "approx_replacements_at_fast_cycle": horizon_years / low,
        "note": "if equipment is not designed for repair, replacement equals disposal",
    }

四、规格书里真正能改的五个地方

任何团队都无法凭一己之力改变全球算力增长曲线，但可以在采购与运维规格里写进五条可执行要求：其一，寿命——要求供应商公布使用寿命与备件承诺周期；其二，可维修性——按现场可更换单元与维修文档给投标方打分；其三，再利用路径——在第一批设备到货前就指定下游责任方；其四，记账——像公布 PUE 那样公布「每吉瓦退役吨数」；其五，去向——追问退役设备实际流向哪个司法管辖区。报告的核心并非「停止 AI」，而是：退役硬件的账本目前无人认领，而规格书是修补它最便宜的位置。

# 4. What a team can actually change in a spec sheet
PROCUREMENT_LEVERS = {
    "longevity": "require a published service life and a spare-parts commitment",
    "repairability": "score bidders on field-replaceable units and documentation",
    "reuse_path": "name the downstream owner before the first shipment arrives",
    "accounting": "report e-waste per GW the same way you report PUE",
    "siting": "ask where retired equipment physically goes, and in whose jurisdiction",
}

def score(bid):
    return sum(1 for k, requirement in PROCUREMENT_LEVERS.items()
               if requirement in bid["commitments"])

# The report's point is not that AI should stop. It is that the retired-hardware
# ledger has no owner today, and a specification is the cheapest place to fix that.

规格书是修补退役账本最便宜的位置

五、一页纸的硬件退役计划

如果你需要一个能立刻用起来的模板，可以这样写：资产类别（推理机架，每架 1,360 公斤）；设计寿命（5 年）与真实更换周期（2.5 年）；下游责任人（具名再利用伙伴与认证回收商，并在合同中写入回收条款）；关注的物质（铅、汞、镉、PFAS）；度量口径（每年每吉瓦退役吨数，基准参照 BAN 的 7 万吨/吉瓦）；明确写出不覆盖的范围（消费设备替代、电信设备）；最后给这份计划一个复核时点——例如下一次容量扩张评审。把不可见的成本写成可见的条目，是这份报告带来的最直接启示。

{
  "hardware_retirement_plan": {
    "asset_class": "inference rack, 1360 kg per rack",
    "designed_service_life_years": 5,
    "actual_replacement_interval_years": 2.5,
    "downstream_owner": "named reuse partner + certified recycler, with a take-back clause in the contract",
    "materials_of_concern": ["lead", "mercury", "cadmium", "PFAS"],
    "reporting": {
      "metric": "metric tonnes retired per GW of deployed capacity per year",
      "baseline_reference": "BAN projects 70,000 t per GW across the full infrastructure boundary",
      "cadence": "quarterly, published internally"
    },
    "exclusions_not_covered": ["consumer device displacement", "telecom equipment"],
    "review_due": "at the next capacity expansion review"
  }
}