One machine, walked backwards. Every tier, every flow, every control point — click a node, open it, then go one level deeper. The best place to fix an emission is the point where it enters the process, not the year it appears in a report. 把一台 AI 服务器倒着走一遍。每一层级、每一股流、每一个控制点——点开一个节点，再往下点一层。控制排放最好的位置，是它进入工艺的那个点，而不是它出现在报告里的那一年。
Read this as a study, not a claim. Every number tagged indicative is a public-source range used for benchmarking and teaching — not an audited figure for any company. Always verify the current legal text before acting. 这是研究与基准参考，不是承诺。凡标注「参考区间」的数字，均为公开来源的量级范围，用于对标与理解，并非任何公司的审定数据。执行前请核对现行法规原文。
Three rules run under every panel on this page. If a metric fails one of them, it is a slogan, not data.本页所有面板都建立在三条规则上。任何一个指标若不满足其中之一，它就只是口号，不是数据。
"Low carbon" is not a number. 1,860 kgCO₂e per 8-GPU node, cradle-to-gate, 2024 BOM v3.1 is. State the functional unit, the system boundary, the cut-off rule and the data year — or you cannot compare year to year. 「低碳」不是数字。1,860 kgCO₂e／8 卡节点，摇篮到大门，2024 年 BOM v3.1 才是。写清功能单位、系统边界、截断规则与数据年份——否则无法做年度对比。
A kilogram of CF₄ released today is still up there in the year 52025. Downstream offsetting cannot recall it. Find the mass-flow controller, the abatement unit, the drain — that is where the tonne is decided. 今天排放的一公斤 CF₄，到公元 52025 年仍在大气中，任何下游抵消都追不回来。找到质量流量控制器、尾气处理装置、排水口——吨数是在那里决定的。
Every figure needs an owner, a source document, a refresh date and a version. The auditor's real test is simple: pick one part, walk it back five tiers, and see whether the paper survives the walk. 每个数字都要有责任人、来源文件、更新日期与版本号。审核员真正的测试很简单：任选一个零件，向上追五层，看这条纸面链条是否走得通。
Tap a hotspot marker on the map, or open a tier below. Inside each tier: who sits there, how much data you can realistically get, what breaks, and one deep dive.点击图上的热点标记，或展开下方任一层级。每层内含：谁在这一层、现实中能拿到多少数据、通常在哪里断链，以及一个深入剖析。
Six hotspots — where risk concentrates六个热点 — 风险集中之处
You never contract a mine. You do, however, converge on a finite list of smelters and refiners — roughly 300 3TG facilities globally on the RMI list. That list is the only realistic hinge point for origin claims. 你不会直接与矿山签约，但全球钽、锡、钨、金冶炼厂在 RMI 名录上大约只有 300 家，所有供应链最终收敛于此。这份名录是产地追溯唯一现实的支点。
For an AI server, semiconductors typically dominate embodied carbon (indicative: 60–80% of cradle-to-gate). Inside a fab, purchased electricity leads, but process F-gases can add a further 10–25% of fab Scope 1+2 if abatement is weak. Both are controllable at specific, physical points — see section 04. 对 AI 服务器而言，半导体通常主导内含碳排（参考区间：摇篮到大门的 60–80%）。厂内以外购电力为首，但若尾气处理不足，工艺含氟气体可再占晶圆厂范围一＋二的 10–25%。二者都能在具体的物理点位上控制——见第 04 节。
Primary aluminium ingot averages roughly 16 kgCO₂e/kg globally including electricity (IAI); recycled secondary aluminium sits near 0.5–0.6 kgCO₂e/kg — a ~95% reduction for the same function. A 12 kg chassis therefore swings ~185 kgCO₂e on a single sourcing decision. 全球原铝锭含电力约 16 kgCO₂e/kg（IAI）；再生铝约 0.5–0.6 kgCO₂e/kg——同等功能下降幅约 95%。一个 12 公斤机箱，仅材料选择一项就相差约 185 kgCO₂e。
OSAT and HBM stacking add wet chemistry, ultrapure water and fluorinated photoresists/etchants on top of an already high-value die. Yield loss here multiplies every upstream impact: a 5-point yield drop on a stacked package can raise per-good-unit footprint by a similar order. 封测与 HBM 堆叠在高价值裸片之上叠加了湿法化学、超纯水与含氟光刻胶／蚀刻剂。此处的良率损失会放大上游全部影响：堆叠封装良率下降 5 个百分点，单位合格品足迹可能同量级上升。
Direct-to-chip and immersion cooling introduce fluoropolymer seals, hoses and in some designs fluorinated working fluids. Under the ECHA universal PFAS restriction proposal, that is a design decision with a market-access horizon — not just a chemistry footnote. 冷板式与浸没式液冷引入了含氟聚合物密封件、软管，部分方案还使用含氟工作液。在 ECHA 的 PFAS 通用限制提案下，这是关乎未来市场准入的设计决策，而不只是化学附注。
The tier you see best is the tier where labour peaks are easiest to hide inside seasonal ramps. Working-hour records, agency-labour ratio and packaging mass per rack are the three fastest audit reads at L11/L12 integration. 你看得最清楚的一层，也是最容易把用工高峰藏进季节性冲量里的一层。在 L11/L12 集成环节，工时记录、派遣工比例、每机架包装重量，是三项最快的审核切入点。
Open a tier — then open the deep dive inside it展开层级 — 再展开其中的深入剖析
Evidence: meter layout drawing · utility invoices · MES output records · allocation memo signed by the plant EHS manager.证据：计量布点图 · 能源账单 · MES 产出记录 · 由工厂 EHS 经理签署的分摊说明。
Four sentences, added at requalification, do more than three years of surveys: (1) disclosure of manufacturing sites for the parts we buy, (2) annual full material declaration in IPC-1752A class D, (3) right to audit extended to your suppliers for our parts, (4) 30-day notice of site or sub-supplier change. 在重新认证时加入的四句话，胜过三年问卷：（1）披露我方采购件的制造场址；（2）每年提交 IPC-1752A D 级全材料申报；（3）审核权延伸至你方为我方供货的供应商；（4）场址或次级供应商变更提前 30 天通知。
Where a supplier will not give primary data, a documented secondary value with stated uncertainty is auditable; an empty cell is not. Use published PCF sheets, ecoinvent/industry averages or IPCC defaults, and flag the data-quality score. Then set a date to replace it. 供应商不给一手数据时，一个注明不确定度的二手数值仍可审计；空白格不行。使用公开 PCF 文件、ecoinvent／行业均值或 IPCC 缺省值，并标注数据质量等级，同时设定替换的时间点。
This is where the physics lives: furnaces, plasma, ultrapure water, electrolysis. Visibility drops to 5–25%, yet this tier usually carries the majority of the product footprint. Everything in sections 04 and 05 happens here. 物理过程都在这一层：炉子、等离子体、超纯水、电解。可见度降至 5–25%，但产品足迹的大部分往往就在这里。第 04、05 节讲的全部发生在此层。
Direct visibility is usually below 5%. You get here through the smelter/refiner list, through certification schemes, and through targeted origin evidence for the handful of materials that carry human-rights and forced-labour risk: cobalt, 3TG, mica, and polysilicon feedstock. 直接可见度通常低于 5%。进入这一层要靠冶炼／精炼厂名录、认证体系，以及针对少数高人权与强迫劳动风险材料（钴、3TG、云母、多晶硅原料）的定向产地证据。
You do not have a data problem. You have a boundary problem that looks like a data problem.你面对的不是数据问题，而是一个看起来像数据问题的边界问题。
Each flow: definition, unit, what is in and what is deliberately out, the formula, a benchmark ladder, where the data lives, and the rule for when it must be included.每股流包含：定义、单位、纳入与刻意排除的内容、计算公式、对标阶梯、数据来源，以及何时必须纳入的规则。
Must include when: a material is >1% of product mass, is on a restricted list, or is a critical raw material. Can defer when: <0.5% of mass, no restricted substance, no high-risk geography — record the reason and revisit annually. 必须纳入：占产品质量 >1%、属受限清单、或属关键原材料。可暂缓：占质量 <0.5%、不含受限物质、不涉高风险地区——须记录理由并每年复核。
The distinction that decides your claim: post-consumer (PCR — material that finished a use life) versus pre-consumer (manufacturing scrap). Most programmes weight PCR far higher, and some count pre-consumer at zero. 决定声明成立与否的关键区分：消费后（PCR，已完成一次使用寿命的材料）与消费前（生产废料）。多数体系对 PCR 赋予高得多的权重，有些体系对消费前材料计为零。
Regulatory anchors: EPEAT / IEEE 1680.1 award points for verified recycled content; the EU Batteries Regulation sets recycled-content thresholds for Co, Pb, Li and Ni from 2031; ESPR digital product passports will require declaration at product level. 法规锚点：EPEAT / IEEE 1680.1 对经核证的再生含量给分；欧盟《电池法规》自 2031 年起对钴、铅、锂、镍设定再生含量门槛；ESPR 数字产品护照将要求在产品层级申报。
Datacenter advantage: racks ship to few destinations in large lots. Reusable transit packaging pays back faster here than in consumer PC channels — model it per trip, not per unit. 数据中心的优势：机架批量大、目的地少。可复用运输包装在此的回收期远快于消费 PC 渠道——按「每次运输」而非「每台」建模。
Two views must reconcile: the corporate view (Scope 1/2/3 by category) and the product view (PCF per unit). If your Scope 3 Category 1 divided by units shipped is wildly different from your PCF, one of them is wrong. 两个视角必须能对上：公司层（范围一／二／三分类）与产品层（单台 PCF）。若范围三类别 1 除以出货量与 PCF 相差悬殊，其中之一必然有误。
Indicative embodied carbon (cradle-to-gate, public PCF sheets vary widely): 8-GPU AI node ~1,500–3,500 kgCO₂e · 2U volume server ~1,000–1,800 · business notebook ~200–400. Semiconductors typically drive 60–80% of the AI node figure. 内含碳参考区间（摇篮到大门，公开 PCF 差异很大）：8 卡 AI 节点约 1,500–3,500 kgCO₂e · 2U 通用服务器约 1,000–1,800 · 商用笔电约 200–400。AI 节点中半导体通常占 60–80%。
Must include when: a site sits in WRI Aqueduct "high" (40–80%) or "extremely high" (>80%) baseline water stress — there, m³ matters more than tonnes. 必须纳入：场址位于 WRI Aqueduct「高」（40–80%）或「极高」（>80%）基准水压力区——在那里，立方米比吨更重要。
Two inventories, not one: product-embedded substances (what ships to the customer) and process chemicals (what the factory uses). RoHS and REACH govern the first; PFAS restrictions, worker safety and F-gas control reach into the second. 要建两份清单而非一份：产品内含物质（随货交付给客户）与工艺化学品（工厂使用）。RoHS 与 REACH 管前者；PFAS 限制、职业安全与含氟气体管控则伸入后者。
PFAS in an AI system hides in: fluoropolymer wire insulation (FEP/PTFE), connector seals, thermal interface materials, cold-plate gaskets, some dielectric fluids, and upstream in photoresists and etch chemistry. Map the process, not just the BOM. AI 系统中的 PFAS 藏在：含氟线缆绝缘（FEP/PTFE）、连接器密封、导热界面材料、冷板垫片、部分介电液，以及上游的光刻胶与蚀刻化学品中。要绘制工艺地图，而不只是 BOM。
Design connection: a rack that can be disassembled in under 10 minutes with common tools is not a sustainability slogan — it is a measurable driver of both the reuse rate and the service cost. 与设计的连接：一台用通用工具 10 分钟内可拆解的机架，不是可持续口号——它同时可量化地驱动回用率与服务成本。
You cannot manage it by reading the SDS. You manage it by knowing where the gas enters, how much is consumed in the chamber, what is created there, and what the abatement unit actually destroys — while it is running. Tap M1–M6.看安全数据表管不住它。要管住它，必须知道气体从哪里进入、腔体内消耗多少、又生成了什么、以及尾气处理装置在实际运行时到底销毁了多少。点击 M1–M6。
Cylinders in − cylinders returned − inventory change = gas consumed. This is the number an auditor can verify against invoices in ten minutes, and it must reconcile with the sum of your recipe-level data within a few percent. 进厂钢瓶 − 退回钢瓶 − 库存变动 = 气体消耗量。审核员十分钟内就能用发票核实这一数字，它必须与配方级数据之和在几个百分点内吻合。
Evidence: purchase invoices · cylinder weight logs · warehouse inventory report.证据：采购发票 · 钢瓶称重记录 · 仓库库存报表。
Flow × time × number of runs = feed per tool. This is where a process engineer can cut emissions before any abatement exists: shorter clean steps, remote plasma NF₃ instead of in-situ C₂F₆, optimised endpoint detection. 流量 × 时间 × 运行次数 = 单台设备用气量。这是工艺工程师在任何尾气处理之前就能减排的地方：缩短清洗步骤、以远程等离子 NF₃ 替代原位 C₂F₆、优化终点检测。
C₂F₆, C₃F₈ and CHF₃ partially convert to CF₄ — the longest-lived of the family. An inventory that counts only what you purchased will understate reality. IPCC Tier 2b handles this with a byproduct formation factor per gas and process. C₂F₆、C₃F₈ 与 CHF₃ 会部分转化为 CF₄——族中寿命最长者。只统计采购量的清单会低估实际排放。IPCC Tier 2b 用「按气体与工艺的副产物生成因子」处理这一项。
A unit rated 99% DRE that runs 90% of production hours delivers 89% — and a rated DRE is gas-specific. NF₃ and C₄F₆ destroy comparatively easily; CF₄ is the hardest, and a system tuned for NF₃ may barely touch it. 标称销毁率 99% 的装置，若只在 90% 的生产时段运行，实际只有 89%——而且标称销毁率是按气体区分的。NF₃ 与 C₄F₆ 相对易于销毁；CF₄ 最难，为 NF₃ 调校的系统可能几乎处理不了它。
Ask for: the DRE test report (which gas, which flow, which dilution), the interlock logic, and 12 months of uptime data. Those three documents are the whole conversation. 要三份文件：销毁率测试报告（何种气体、何种流量、何种稀释条件）、联锁逻辑、12 个月运行率数据。这三份就是全部对话。
Central wet scrubbers are designed for acid gases and particulates. Counting them as PFC abatement is one of the most common inventory errors — and one of the easiest for a verifier to catch. 中央湿式洗涤塔是为酸性气体与颗粒物设计的。把它算作全氟化物减排，是最常见的清单错误之一，也是核查员最容易发现的错误之一。
Beware the dilution trap: nitrogen purge makes stack ppm look tiny while mass flow is unchanged. Always convert to mass per unit time and compare with M1. If they disagree by more than ~10%, find out why before you publish. 警惕稀释陷阱：氮气吹扫让烟囱 ppm 看起来很小，但质量流量未变。务必换算为单位时间质量，并与 M1 比对。若偏差超过约 10%，发布前必须查清原因。
Atmospheric lifetime, years — bars are on a log scale. GWP₁₀₀ from IPCC AR6 in brackets.大气寿命（年），横条为对数刻度。括号内为 IPCC AR6 的 GWP₁₀₀。
Read it this way: one kilogram of SF₆ vented during a single maintenance event equals roughly 24 tonnes of CO₂ — about the annual footprint of two average EU households — and it is still there in the year 5225. 这样理解：一次维护中泄放的 1 公斤 SF₆，约相当于 24 吨 CO₂——大致等于两个欧盟家庭一年的排放——并且到公元 5225 年它仍在那里。
A fab's water story and a datacenter's water story are different problems with the same discipline: segregate at source, reclaim close to the tool, and weight every cubic metre by where it was taken.晶圆厂与数据中心的用水是两个不同问题，但纪律相同：源头分流、就近回用，并按取水地的水压力对每一立方米加权。
Screen every manufacturing site through WRI Aqueduct. 1,000 m³ from a low-stress basin and 1,000 m³ from an extremely-high-stress basin are not the same number. Report both raw and stress-weighted volumes.用 WRI Aqueduct 对每个制造场址做筛查。低压力流域的 1,000 m³ 与极高压力流域的 1,000 m³ 不是同一个数。原始量与压力加权量都要披露。
Making 1 m³ of ultrapure water can take well over 1.4 m³ of feed once RO reject is counted. Improving recovery here reduces intake without touching a single process recipe.计入反渗透浓水后，制取 1 m³ 超纯水常需 1.4 m³ 以上进水。提高此处回收率可在不改动任何工艺配方的前提下降低取水量。
Once fluoride rinse water joins general drain, you must treat the whole diluted volume and you can no longer reclaim the clean fraction. The engineering decision that determines your reclaim rate is made at the tool drain, years before anyone writes a water target. 一旦含氟漂洗水汇入总排水，你必须处理被稀释后的全部水量，也再无法回用其中洁净的部分。决定回用率的工程决策发生在设备排水口，比任何人写下用水目标早好几年。
Typical discharge permits for fluoride sit around 15–20 mg/L; leading fabs run well under 10 and treat the recovered calcium fluoride as a saleable by-product. Anchor your monitoring here: one inlet sample, one outlet sample, one number that everybody can argue about honestly. 氟化物排放许可通常在 15–20 mg/L 左右；领先晶圆厂可稳定低于 10，并把回收的氟化钙作为可售副产品。把监测锚定在此：一个进水样、一个出水样，一个所有人都能据实讨论的数字。
Withdrawal minus discharge equals consumption. For a datacenter using evaporative cooling, almost all withdrawal is consumption — which is why WUE, not withdrawal, is the honest datacenter metric.取水量减排水量等于耗水量。使用蒸发冷却的数据中心几乎全部取水都被消耗——所以对数据中心而言，诚实的指标是 WUE 而非取水量。
Most disputes about sustainability data are not about the arithmetic. They are about which line was skipped.关于可持续数据的争论，多数不在算术，而在哪一行被跳过了。
Requirements land on the supply chain as documents. Here is the translation — regulation → what you must actually collect. Verify current legal text before acting.法规最终落到供应链上都是「文件」。以下是这层翻译：法规 → 你实际必须收集什么。执行前请核对现行法规原文。
Ten regulations. One dataset. Build the dataset once, and the reports become queries.十部法规，一套数据。把数据建一次，报告就只是查询。
34 documents. This is what a validator actually opens. Nothing is stored — the count is yours, on this screen, right now.34 份文件，这是核查员真正会打开的东西。本页不保存任何数据——计数只存在于此刻你的屏幕上。
Evidence ready:已就绪证据： 0 / 34 · nothing leaves this page数据不会离开本页
A validator picks one part number at random and asks you to walk it back: supplier → site → material → mill → smelter. If the chain breaks at tier 3, your whole dataset is graded at tier 3.核查员随机抽取一个料号，要求你向上回溯：供应商 → 场址 → 材料 → 加工厂 → 冶炼厂。若链条在第三层断裂，你整套数据都按第三层评级。
Every file needs an issue date, a version, an owner and a next-review date. An undated PDF is not evidence — it is a rumour with a logo.每份文件都要有签发日期、版本、责任人与下次复核日期。没有日期的 PDF 不是证据，只是带 logo 的传闻。
"1,860 ± 15% kgCO₂e, 62% primary data" earns more trust than a bare number to three decimals. Verifiers reward stated limits; they punish false precision.「1,860 ± 15% kgCO₂e，一手数据占 62%」比精确到小数点后三位的裸数字更可信。核查员奖励坦白的边界，惩罚虚假的精确。
A programme that tries to include everything in year one collects nothing in year three. Deferring is legitimate — undocumented deferring is not.第一年就想把一切纳入的项目，第三年往往什么都没收到。暂缓是正当的；无记录的暂缓不是。
The hard cap: the cumulative deferred set must stay below 5% of the footprint, be listed by name, and be reviewed every year. Once it crosses 5%, it is not a gap — it is a finding.硬性上限：累计暂缓项合计须低于足迹的 5%，逐项列名，并每年复核。一旦超过 5%，它就不是缺口，而是审核发现。
Spreadsheets, annual questionnaires, compliance letters. Coverage claimed, not evidenced.表格、年度问卷、合规函。有覆盖率声明，无证据支撑。
Supplier master data linked to part numbers and sites. You can answer "which sites make this part?" in a day.供应商主数据与料号、场址关联。一天内能回答「这个零件由哪些场址生产」。
Tier-1 primary data, hotspot Tier 2–4 measured, uncertainty stated. Benchmarks in place.一级一手数据，二至四级热点实测，标注不确定度，已建立对标。
DRE, reclaim rate, PCR share and removal efficiency each carry a target and an owner. Third-party verified.销毁率、回用率、再生比例、去除率各有目标与责任人，并经第三方核证。
Machine-readable exchange (PACT / DPP-ready), automated evidence, proven year-on-year improvement per unit.机器可读交换（符合 PACT／DPP），证据自动化，单位产品逐年改善有据可查。
A realistic sequence: Year 1 map and baseline (Levels 1→2) · Year 2 primary data on the top-20 hotspots (→3) · Year 3 control points instrumented with targets (→4) · Year 4 verification and DPP-readiness · Year 5 automated exchange (→5). Moving one level per year is fast. Moving two is usually a rebuild. 现实的节奏：第 1 年测绘与基线（1→2 级）· 第 2 年前 20 个热点取得一手数据（→3）· 第 3 年控制点仪表化并设目标（→4）· 第 4 年核证与 DPP 就绪 · 第 5 年自动化数据交换（→5）。一年提升一级已属快速；一年两级通常意味着推倒重来。
Primary standards, regulators and open datasets. Read the current version — this page is a study, and the texts move.标准原文、监管机构与开放数据。请阅读现行版本——本页只是研究，而法规文本会变。