Cluster Interconnect and Data Centers

1. Introduction

1.1. Background

Cluster interconnect and data centers are the layer of AI Infra where "the machines are visible": inward, they carry the Scale-up domain of supernodes, while outward they provide the Scale-out network and physical hosting (power distribution, cooling, load-bearing, and floor space). This layer is undergoing two transformations at once: first, a multi-route competition in interconnect standards (NVLink, UALink, and ESUN/SUE running in parallel with the domestic EthLink/Eth-X/OISA/LingQu), and second, data-center constraints being elevated from a "supporting discipline" to a policy red line — the "Special Action Plan for Green and Low-Carbon Development of Data Centers" issued by the NDRC and three other ministries (2024-07) requires that by the end of 2025, newly built and expanded large/extra-large data centers have a PUE no higher than 1.25, projects at national hub nodes no higher than 1.2, and the green-power share of newly built data centers at national hubs exceeding 80% (A-level, policy original text).

The demand-side push is equally concrete: AI training workloads exhibit power oscillations of roughly 15 MW (versus roughly 1.5 MW for traditional non-AI workloads, B/C-level relayed framing), and in Llama 3 training, synchronised power changes across tens of thousands of GPUs produce oscillations on the order of tens of MW, approaching the power-supply limits of a data center. For the first time, interconnect and power supply have become first-order constraints on par with the chips themselves.

1.2. Definition and Scope

The Cluster Interconnect and Data Center track covers, on the operational side:

Boundary notes: for in-supernode (domain) implementations see 02-supernode.md; for the full text of the PUE policy and the energy statistics see 03-cluster.md in this group and the data-center-related content in Group 04 - Hardware Development; the GB300 power-shell capacitive peak-shaving parameters (about 65 J/GPU, grid peak reduction of up to 30%, A/B-level vendor framing) are outlined in Section 1.4.3.

1.3. Positioning in the AI Harness Architecture

The cluster network is the "bus" of the L2 execution layer, while power distribution and PUE are the physical projection of the L6 governance layer:

Bottleneck layer assessment: the focus of this track lies in L2 (bus capability) and L6 (energy governance). The combination of AI-load power oscillation plus battery buffering (e.g. Tesla Megapack) is a typical case of coordination between L3 orchestration and physical infrastructure — computing scheduling must be co-designed with power scheduling, a cross-layer problem unique to this track.

1.4. Current State of Development

1.4.1. The Three-Way Scale-up Race

Writing discipline: as of 2026-09 the native UALink switch chips were not yet in mass production, and AMD’s Helios racks carry the UALink protocol over an 800GbE physical layer (UALoE) as a transitional measure — it must be stated that "native silicon is not yet ready, bridged over Ethernet", and it is forbidden to write that "UALink is already deployed at scale commercially".

1.4.2. Domestic Interconnect Protocols

• LingQu (Unified Bus) 2.0: announced by Huawei at HC2025 to release the technical specification as an open standard; built-in reliability at every OSI layer, with 100 ns-class fault detection and protection switching on the optical path (A-level).
• EthLink (ByteDance): 1-4 Ethernet interfaces, up to 1024 GPUs per domain (B-level academic-survey framing).
• Eth-X (Tencent) and OISA (China Mobile): OISA advocates unified packets, multi-semantic fusion, and multi-level flow-control retransmission (B-level survey framing).
• Domestic IP progress: XinDong Technologies released a complete domestic UALink IP suite in 2026-05 (112G in production, 224G taped out); ODCC released a UALink test-and-verification service in 2026-04 (C-level; re-verify sources for item-by-item citation).

1.4.3. Scale-out, Optical Interconnect and Data Centers

• Scale-out dual stack: InfiniBand (Quantum-X800) and Ethernet (Spectrum-X800 and CPO switches, claiming 5x energy efficiency and 10x network resilience — vendor framing); ConnectX-9 SuperNIC at 1.6 Tb/s per GPU (GTC 2026).
• Optical interconnect evolution: NPO (optoelectronic chips moved down to the GPU board, meter-scale → centimeter-scale) is already deployed; CPO (co-packaged optics, centimeter-scale → millimeter-scale) is evolving; the Lightmatter LightSphere X supernode uses in-rack DAC plus cross-rack AEC (5-7 m) to reduce optical modules (B/C-level).
• Power distribution and oscillation governance: the GB300 NVL72 power shell has built-in electrolytic capacitors (about 65 J/GPU), and NVIDIA claims a peak grid-side power reduction of up to 30% at the same GPU output (A/B-level vendor framing); the "Implementation Plan for Facilitating the Synergistic Digital-Green Transformation (2026-2030)" quantitatively operationalizes for the first time 800V HVDC, over-100-kW single racks, liquid cooling plus waste-heat recovery, and "compute-power coordination" scheduling (B-level, relayed by Xinhua News Agency).
• Liquid cooling: cold-plate cooling has the highest maturity (roughly 65% share by industry estimate, C-level); liquid-cooling penetration is about 8% (2024) → about 35% by industry forecast (2026/2027, C-level); the Gui’an New Area spray-liquid-cooling pilot delivers +30% compute per rack, -40% energy, and a PUE persistently below 1.1 (B-level); Huawei Cloud Gui’an has an average annual PUE of 1.1, with iCooling AI tuning cutting energy by 8-15% (B-level).
• Energy overview: global data-center electricity use of 415 TWh (2024) → 945 TWh (2030 forecast, IEA, B-level); China’s compute-center electricity use of 196 billion kWh (2025) → 500 billion (2030E, CAICT, B-level).

Figure 3-1 | The three-way Scale-up interconnect-standards race and the domestic protocol landscape

Diagram: drawn from public information from the UALink Consortium, OCP, NVIDIA GTC 2026 and Huawei HC2025 (A/B-level) and the "Special Action Plan for Green and Low-Carbon Development of Data Centers" (A-level); the three states (in production / published / transitional) are this document’s analysis framing.

2. Glossary

3. Case Studies

3.1. UALink 1.0 to 2.0: The Advance of the Open Interconnect Route and Its Transitional State

Evidence level: B (publication facts of official specifications are cross-confirmed by multiple sources; some domestic progress is C-level)

3.1.1. Background

The closed nature of the NVLink ecosystem (it does not open up the Switch chip and PHY, and entrants must connect through NVIDIA products to join the network) prompted AMD, AWS, Google, Meta, and Microsoft to found the UALink Consortium in 2024-10, seeking an open Scale-up standard.

3.1.2. Solution

• 1.0 (2025-04): UALink 200G — up to 1024 accelerators per Pod and 200 GT/s per channel, supporting memory-semantic access, memory sharing, and Switch-based networking; a 64B flit/640B packet structure with average latency hundreds of ns lower than Ethernet (B-level).
• 2.0 (2026-04, including three companion specifications: the 200G physical layer, the management framework, and the Chiplet interface): adds In-Network Compute (all-reduce embedded in switch nodes); Chiplet 1.0 and UCIe 3.0 compatibility; over 85 members.
• Transitional state: native UALink switch chips were not yet in mass production as of 2026-09; AMD’s Helios racks in 2026 H2 carry the UALink protocol over an 800GbE physical layer (UALoE).
• Ecosystem moves: Broadcom left the UALink board in 2025-10, moving to OCP ESUN; ODCC released a UALink test-and-verification service in 2026-04; XinDong Technologies released a complete domestic UALink IP suite in 2026-05 (112G in production, 224G taped out, C-level).

3.1.3. Effects

• Specification-level effects: the combination of a 1024-accelerator single domain plus In-Network Compute gives non-NVIDIA ecosystems an open option on a par with NVLink 5/6; In-Network Compute sinks collective communication down to the switch layer, directly improving the synchronization overhead in L3 orchestration.
• Real-world constraint: as of the information cutoff, "the open standard is published" and "native silicon is not yet in mass production" coexist — the true usability of the open route depends on the first chip deliveries in 2026-2027 (first 1.0 chips expected in 2026, B-level).
• Harness interpretation: In-Network Compute marks the L2 execution layer being pushed down toward the network layer — the execution primitive (all-reduce) moves from GPUs to switches, changing L2’s implementation boundary rather than the six-layer structure itself.

3.2. Huawei LingQu (UnifiedBus) 2.0 Goes Open: Standard Output of Domestic Interconnect

Evidence level: A (official release at Huawei Connect 2025)

3.2.1. Background

After Huawei stated at HC2025 that "China’s semiconductor-manufacturing processes will lag for a long time, and sustainable compute can only be built on actually available process nodes", interconnect became the strategic fulcrum of the domestic route: with single chips constrained, push interconnect to the extreme and open the interconnect specification itself as a standard, expanding the boundary of the ecosystem.

3.2.2. Solution

• Release the UnifiedBus 2.0 technical specification as an open standard.
• Built-in reliability: reliability mechanisms built into every OSI layer; 100 ns-class fault detection and protection switching on the optical path.
• Linkage with the supernode product line: LingQu is the UB network-specification foundation of the CloudMatrix 384 and the Atlas 950/960 SuperPoD (supernode implementation see 02-supernode.md).

3.2.3. Effects

• It upgrades domestic Scale-up from a "private implementation" to a "specification others can comply with", providing an interface foundation for third parties to join the Ascend ecosystem (with the open CANN, Mind suite, and openPangu open-sourcing).
• The 100 ns-class optical-path fault detection directly serves L5 observability (fault-detection latency metric) and L3 fault-tolerant orchestration (protection switching).
• The effect data are all release framing; ecosystem adoption (the number of third-party chips/systems following suit) had not received independent statistics as of the information cutoff and is marked as a gap.

3.3. xAI Colossus: The Data-Center Engineering of a 100,000-GPU Cluster

Evidence level: A/B (Supermicro official case-study PDF); part of Colossus 2 is B/C (encyclopedia and media framing)

3.3.1. Background

A 100,000-GPU-scale training cluster pushes interconnect and data-center constraints to the limit: network selection, power-supply fluctuation, and cooling density must all be solved simultaneously under time pressure for build completion. Colossus 1 is the highest-publicity extreme case (100,000 H100s, built in 122 days, first 10,000 GPUs live in 19 days; then expanded to 200,000 GPUs over a further 92 days, about 300 MW).

3.3.2. Solution

• Network: dropped InfiniBand in favor of Spectrum-X 400G Ethernet; SN5600 switches (51.2 Tbps); nine 400GbE ports per server (8 BlueField-3 + 1 ConnectX-7), 3.6 Tbps in total; case-state network throughput efficiency of 95% (A/B-level official case study).
• Power: Tesla Megapack battery buffers (3.9 MWh each) handle 10-20 MW-class sudden power changes in training workloads; about 300 MW site scale.
• Cooling: Supermicro 4U liquid-cooled racks (8 4U servers + CDU), 64 GPUs per rack; EB-scale all-NVMe storage (VAST Data → DDN EXAScaler/Infinia in two stages).
• Colossus 2 (B/C-level framing): announced operational on 2026-01-17, the world’s first gigawatt-class AI training cluster; about 555,000 GB200/GB300 (equivalent to 1.4 million H100s), starting at 1 GW with a 2 GW target; 41 natural-gas turbines plus close to 1.2 GWh of Megapack; PUE 1.12 (all-liquid-cooled framing, C-level); accompanied by environmental controversy (EPA violation ruling, NAACP lawsuit).

3.3.3. Effects

• Network-route effects: demonstrates that RoCE-family Ethernet can replace IB in 100,000-GPU training scenarios, but at the cost of the operational complexity of "aggressive QoS and flow-level telemetry" (C-level operations framing).
• Framing conflicts presented side by side (must be retained): the Colossus 1 MFU of about 11% (C-level) and the "network utilization 95%+" framing (C-level) coexist — the two definitions differ (training efficiency vs network efficiency) and are easily conflated, so each metric definition must be noted separately; multiple PUE values exist (1.05/1.1 for Colossus 1, R03 framing, and 1.12 for the Colossus 2 all-liquid-cooled framing), and any PUE citation must note the site and generation; cross-generation comparison is forbidden.
• Harness interpretation: Colossus demonstrates the physicalization of the L6 governance layer — power-fluctuation governance (Megapack) and the energy metric (PUE) are essentially translating L6’s "budgets and guardrails" into power and cooling equipment; the 95% network-utilization framing is the direct mapping of L5 observability onto the Scale-out network.

4. Practice Standards

Nature statement: the following is a draft proposal distilled from publicly disclosed industry practices, not the original text of official industry standards. This track is based primarily on policy documents (the PUE red line) and de-facto standards (IB/RoCE configuration practices, interconnect specifications), with no IEEE/ISO/JEDEC standards specific to clusters. The group-level higher-order specification is in AGENTS.md in the same directory.

4.1. AGENTS.md Specification

4.1.1. AGENTS.md (AI Infra Group · Cluster Interconnect and Data Center Track)

# AGENTS.md —— AI Infra 组 · 集群互联与机房方向

> 本文为基于行业实践提炼的建议稿，非官方行业标准原文。
> 上位规范：01-AI-Infra组/AGENTS.md（不可逆动作禁止清单、变更窗口、可观测要求）。

## 角色与边界
- 角色：集群互联与机房运行侧智能体，负责网络拓扑与配置分析、
  Scale-up/Scale-out 选型比对、PUE 与能耗核验、功率波动治理方案评估。
- 不负责：网络全局配置（RoCE QoS/ECN/PFC、IB 子网管理器）的直接修改
  （属禁止清单）、供配电工程设计、互联标准组织的立场声明。
- PUE 等政策条文只引用有 A 级原文的条目；仅媒体转述的政策标注 [待核实]。

## 环境假设
执行前必须显式声明：
- Scale-up 互联体系与代际（NVLink/UB/UALink/其他）与单域规模。
- Scale-out 网络形态（InfiniBand / RoCE / Spectrum-X / 其他）、交换层级、
  跨机带宽、QoS/ECN/PFC 配置现状。
- 机房参数：单机柜功率上限、散热形态、设计 PUE 与实测 PUE、绿电占比。
- 训练负载画像：同步作业数、功率摆动幅度、储能/削峰设施。
- 未声明网络形态与机房参数时，禁止产出通信效率、PUE、扩展性结论。

## 上下文加载顺序（Context Budget）
1. 网络拓扑与配置现状（层级、带宽、QoS）。
2. 机房参数与能耗台账。
3. 目标负载的通信模式（集合通信规模、频率）。
4. 互联标准动态（区分已量产/已发布/过渡三态）。

## 工具契约
- 网络诊断通过受控工具（IB 探测、RoCE 遥测、nccl-tests）输出结构化结果。
- nccl-tests 输出必须含 busbw、buffer size、拓扑范围。
- 互联标准动态引用必须注明三态与来源级别。

## 任务执行流程（SOP）
- S1 需求解析：把负载需求换算为域内带宽、跨节点带宽、时延、功耗四类约束。
- S2 现状盘点：拓扑、QoS、机房参数、能耗台账。
- S3 候选生成：网络配置调整、拓扑调整、散热/供电改造候选方案。
- S4 受控验证：在维护分区或影子链路验证，记录 busbw 与时延。
- S5 合规核验：PUE 对照红线（1.25/1.2）、绿电占比对照要求。
- S6 证据打包：变更方案附回滚命令，进入审批队列。

## 验证与证据要求
- 通信结论附 nccl-tests 或遥测原始数据；网络利用率注明指标定义
  （吞吐效率 vs 链路利用率，禁止混用）。
- PUE 结论注明站点、代际、测量周期；禁止跨代/跨站点混比。
- 互联标准结论注明三态；过渡方案（如 UALoE）不得写成规模商用。

## 失败与升级策略
- RoCE 网络出现不可归因的拥塞或丢包 → 停止自动调优，升级网络工程师。
- 功率摆动接近供电上限 → 触发削峰预案并上报，不得自行调整作业调度策略。
- 实测 PUE 逼近红线 → 输出能耗报告并启动散热优化评估。

## 安全与合规红线
- 网络全局配置修改属禁止清单动作，只产出待审批方案。
- 政策引用限于 A 级原文；转述条文标 [待核实]。
- 出口管制语境下保持中性表述。
- 禁止残留非标准占位符。

## 输出格式
- 首行：可判定结论 + 阻塞项。
- 拓扑与配置表：层级、带宽、QoS 参数、来源。
- 能耗报告：PUE、实测功耗、红线差值、绿电占比。
- 标准三态表：各路线的量产/发布/过渡状态与来源级别。
- 数值规范：参数带单位；范围用 ~ 连接；百分比数值与 % 之间不留空格。

## 评估与自检
- 网络利用率与 MFU 等指标定义是否分清？
- PUE 是否注明站点与代际？
- 互联标准是否标注三态？
- 是否产出过禁止清单内的直接修改动作？（必须为否）
- 是否残留非标准占位符？

4.2. SKILL.md Specification

4.2.1. SKILL.md (AI Infra Group · Cluster Network Diagnosis and Data-Center Energy Verification)

---
name: cluster-network-and-facility-audit
description: AI Infra 组集群互联与机房方向标准技能——网络配置诊断、互联
  选型三态比对、PUE 与能耗红线核验。当需要诊断 RoCE/IB 网络问题、评估
  Scale-up/Scale-out 路线、核验 PUE 政策符合性、评估功率波动治理时使用。
  触发词：RoCE、InfiniBand、UALink、NVLink Fusion、PUE、液冷、功率摆动。
version: 1.0
created: 2026-09-12
---

# AI Infra 组 · 集群网络诊断与机房能耗核验

> 本文为基于行业实践提炼的建议稿，非官方行业标准原文。
> 上位规范：01-AI-Infra组/AGENTS.md。

## 适用场景
- 训练作业通信异常（掉速、超时、集合通信失败）的网络侧诊断。
- Scale-up 互联路线选型（NVLink / UALink / ESUN / 灵衢，三态比对）。
- Scale-out 网络选型（IB vs RoCE）与 QoS 配置评审。
- PUE 政策符合性核验（1.25/1.2 红线）与能耗预算评估。
- 功率摆动治理方案评估（削峰、储能缓冲、算电协同）。

## 前置条件
- 已声明网络形态、交换层级、QoS 现状与机房参数。
- nccl-tests、网络遥测工具可用；变更窗口机制可用。
- 能耗台账（PUE、分项功耗）可查询。

## 输入
| 输入项 | 说明 | 必需 |
|---|---|---|
| 网络拓扑与配置 | 层级、带宽、QoS/ECN/PFC、固件版本 | 是 |
| 负载通信画像 | 集合通信规模、频率、作业同步结构 | 是 |
| 机房参数 | 功率上限、散热形态、PUE 实测、绿电占比 | 是 |
| 故障现象 | 告警、日志路径、时间窗口 | 条件必需 |
| 互联标准动态 | 各路线三态与来源级别 | 否 |

## 输出
| 输出项 | 说明 | 必需 |
|---|---|---|
| 可判定结论 | 通过 / 不通过 / 部分通过 + 阻塞项 | 是 |
| 网络诊断报告 | busbw、时延、拥塞/丢包定位，附原始数据 | 条件必需 |
| 三态比对表 | 各互联路线量产/发布/过渡状态 | 条件必需 |
| 能耗报告 | PUE、实测功耗、红线差值、绿电占比 | 是 |
| 变更方案 | 配置调整建议 + 影响面 + 回滚命令 | 条件必需 |
| 证据表 | 工具、版本、命令、日志路径、数值、时间窗口 | 是 |

## 执行步骤
1. 环境校验：核对网络形态、QoS 现状、机房参数；确认不触及禁止清单。
2. 基线确认：nccl-tests 基准 busbw、能耗台账基线。
3. 问题定位（诊断场景）：分层定位——链路（光模块/线缆）、交换
   （拥塞/ECN）、配置（QoS/PFC）、软件（NCCL 超时）。
4. 选型比对（选型场景）：按三态（已量产/已发布规范/规划过渡）比对
   候选路线，标注来源级别。
5. 能耗核验：PUE 对照红线，分项功耗对照预算，绿电占比对照要求。
6. 受控验证：配置类建议先在维护分区验证并记录 busbw 差异。
7. 交付：诊断/比对/核验结论 + 变更方案（含回滚）+ 待人工确认项。

## 质量标准（DoD）
- 网络结论附 busbw 等原始数据；指标定义明确（吞吐效率 vs 链路利用率）。
- PUE 注明站点与代际，无跨代混比；红线核验有计算过程。
- 互联路线标注三态；过渡方案不写成规模商用。
- 变更方案含回滚命令；未执行禁止清单动作。
- 参照口径（不得直接套用）：PUE 红线 1.25/1.2（A级政策）；八大枢纽平均
  PUE 约 1.3、最先进 1.04（B级，国家能源局）；Tomahawk 6 = 102 Tbps /
  2048 端口 / 小于 400ns（B级）；灵衢 2.0 光路径 100ns 级检测（A级发布）。

## 常见失败与处理
| 失败模式 | 现象 | 处理 |
|---|---|---|
| 指标混用 | 把网络利用率 95% 当训练效率引用 | 分别注明指标定义与来源级别 |
| 过渡态误判 | 把 UALoE 桥接写成 UALink 规模商用 | 标注三态与过渡事实 |
| QoS 盲调 | 未测基线直接改 ECN/PFC | 先测基线，变更走审批窗口 |
| PUE 混比 | 跨站点/跨代比较 PUE | 注明站点与代际 |
| 红线误引 | 引用仅媒体转述的政策条文 | 标 [待核实] 或仅引 A 级原文 |
| 波动治理越权 | 自行调整全局作业调度应对功率摆动 | 升级人工，输出预案建议 |

## 示例
任务：某 8 千卡训练集群作业频繁出现集合通信超时，怀疑 RoCE 网络配置。
1. 环境校验：确认 Spectrum-X 类以太网、交换层级、QoS/ECN/PFC 现状、
   固件版本；确认网络全局配置修改不在本轮直接执行。
2. 基线确认：nccl-tests 分层实测 busbw（节点内/跨机/跨柜），记录时间窗口。
3. 分层定位：跨柜 busbw 显著低于节点内，ECN 标记率在跨柜段异常升高。
4. 候选方案：ECN 阈值调整、PFC 策略收敛、故障光模块替换，各附影响面
   与回滚命令。
5. 受控验证：维护分区验证 ECN 调整，busbw 恢复至基线的 97% 以上。
6. 能耗核验：同步核对 PUE 实测与红线差值（本例无变化）。
7. 交付：结论"部分通过，跨柜段需按方案调整 ECN 并更换 2 个光模块"；
   变更方案进入审批队列；证据表含命令、日志路径、时间窗口。

4.3. Landing Checklist

5. Summary

The Cluster Interconnect and Data Center track covers the L2 "bus" and the L6 "physical projection" in the six-layer Harness model. Three conclusions:

1. Interconnect standards have entered a long-term pattern of multiple routes coexisting. The three-way race of NVLink (closed + semi-open), UALink (open, In-Network Compute), and ESUN/SUE (Ethernet-based), plus the domestic EthLink/Eth-X/OISA/LingQu 2.0, leaves no single winner in the short term; the selection discipline is to assess against the three states of "in production / published specification / planned or transitional", and transitional schemes such as UALoE must not be written as commercially deployed at scale.
2. Data-center constraints have been elevated from supporting to red-line. A PUE of 1.25/1.2 and an 80% green-power share are hard constraints with A-level policy original text; the 15 MW-class power oscillation of AI workloads makes power-supply governance (battery buffering, peak shaving, compute-power coordination) the second constraint after interconnect, and L3 computing scheduling must be co-designed with power scheduling.
3. Metric discipline matters especially in this track. Network utilization versus training efficiency, PUE across different sites and generations, and deployment data at release versus dissemination framings are all high-conflation points in this track; AGENTS.md’s separation of metric definitions and site/generation labeling are mandatory requirements.

Information-Gap Statement

The following items have not been confirmed by A/B-level sources and are marked [To be verified] in the body text:

1. This track found no IEEE/ISO/JEDEC standards dedicated to AI cluster interconnect and data centers; it is expressed in terms of policy documents + interconnect specifications + de-facto standards, and no standard numbers may be fabricated.
2. The "Action Plan for Facilitating the Two-Way Empowerment of AI and Energy" (four ministries, 2026-04, "100% liquid cooling for newly built large AI data centers, retrofitting existing air-cooled facilities by 2028") has only been seen in media relays; the original text must be verified before citation.
3. Colossus 1’s MFU of about 11% and a network utilization of 95%+ are both C-level and have different definitions; they have been separately noted.
4. Colossus 2’s exact scale (about 555,000 cards), PUE 1.12, 41 gas turbines, and 1.2 GWh of Megapack are all encyclopedia/media framing (B/C-level).
5. XinDong Technologies’ UALink IP and Biren’s cooperation in the ODCC interoperability test are C-level; sources must be re-checked for item-by-item citation.
6. The EthLink/Eth-X/OISA parameters (1024-GPU domain, unified packets, etc.) come from relays in academic surveys (B-level); no first-hand specification text from the companies was obtained.
7. Liquid-cooling penetration of 8% → 35% and a 65% cold-plate share are C-level industry observations, rewritten in the body text as "by industry estimate/forecast".
8. The GB300 power-shell 65 J/GPU and grid-peak -30% figures are vendor framing (A/B-level) with no independent retesting.
9. No public AGENTS.md/SKILL.md standard templates were found for this track; this document is a draft proposal.

6. References

1. Special Action Plan for Green and Low-Carbon Development of Data Centers — NDRC and three other ministries, 2024-07. https://www.ndrc.gov.cn/
2. Leading a New Paradigm of AI Infrastructure with Pioneering Supernode Interconnect Technology (LingQu 2.0 open) — Huawei, 2025-09-18. https://www.huawei.com/cn/news/2025/9/hc-xu-keynote-speech
3. UALink 1.0/2.0 specification release — UALink Consortium, 2025-04 / 2026-04. https://ualinkconsortium.org/
4. Multiple AI Scale-Up Options Emerge — SemiEngineering. https://semiengineering.com/multiple-ai-scale-up-options-emerge/
5. Research on Key Technologies and Industry Development Trends of Supernodes (survey of interconnect protocols and domestic protocols) — Chinese Society of Engineering Thermophysics (repost), 2025-2026. https://www.zgcsswdx.cn/info/24724.html
6. Success Story: xAI Colossus Cluster — Supermicro (written by ServeTheHome), 2024. https://www.supermicro.com/CaseStudies/Success_Story_xAI_Colossus_Cluster.pdf
7. Energy and AI — International Energy Agency (IEA), 2025-04. https://www.iea.org/reports/energy-and-ai
8. Implementation Plan for Facilitating the Synergistic Digital-Green Transformation (2026-2030) — Office of the Central Cyberspace Affairs Commission and six other departments, 2026-09 (relayed by Xinhua News Agency). https://www.gov.cn/
9. NVIDIA Spectrum-X / ConnectX-9 related releases (GTC 2026) — NVIDIA. https://www.nvidia.com/en-us/networking/
10. The Llama 3 Herd of Models (background on power fluctuation and training workloads) — Meta, 2024. https://arxiv.org/abs/2407.21783