一家新能源大厂，如何撑起全球最大AI算力超级单体？
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR — 中国新能源企业远景科技于 2026 年 8 月 6 日在乌兰察布点亮"星河基地"，宣称建成全球最大的单体 AI 算力楼；其意义在于表明 GW 级 AI 基建的瓶颈已从芯片转向电力、储能与热管理工程。
- 规模与选址：单体建筑超 12 万 m²，园区规划 2GW，目标百万加速卡并行、百万 P 级算力；落地乌兰察布是看中当地风光资源（全域绿电占比 67%），配套专属风电场与直供输电线路接入数据中心。
- 绿电变"算力级"稳定电力：以"预测—调度—执行"闭环实现——天机气象大模型预测风光出力，天枢能源大模型结合发电、负荷、储能与电价做分配寻优，EnOS 物联操作系统下发指令控制风机、储能与数据中心设备并回传遥测；储能负责缓冲绿电供给 GPU。
- 供电与散热架构：采用 SST + BESS 的 800V 直流架构，减少多级 AC/DC 转换损耗、铜用量与占地，以应对单机柜数百 kW 的功率密度；散热则结合风冷、液冷与乌兰察布的自然冷源。
- 效率主张与商业模式：远景称单位面积算力产出约为传统数据中心的 10 倍（企业口径，未经第三方验证）；分工上远景提供绿电、土地、园区及 AI 电力系统，客户自行部署与运营服务器和芯片集群。
- 战略定位：该基地是 2026 年 6 月提出的"戈壁计划"第一步，目标到 2030 年在沙漠戈壁地区建成 5GW 绿色 AI 算力。
注：两份来源并非同一工作——量子位报道的是远景星河基地（即本篇主题），而雷峰网（AI 科技评论）那篇讲的是深圳戴盟机器人（视触觉传感与 Daimon-TWM 世界模型、蚂蚁领投数亿元融资），内容与本标题无关，故未并入正文。
Sources (2)
一家新能源大厂，如何撑起全球最大AI算力超级单体？
TL;DR - Chinese renewable-energy company Envision (远景科技) brought its Ulanqab "Galaxy Base" (星河基地) online on Aug 6, 2026, unveiling what it calls the world's largest single-building AI compute block — a signal that power, storage, and thermal engineering, not just chips, now gate GW-scale AI buildout.
- The facility spans 120,000+ m² inside a campus planned for 2GW capacity, targeting million-accelerator parallelism and million-P-scale compute; site chosen for Ulanqab's wind/solar resources (67% green power region-wide) with dedicated wind farms and direct transmission lines into the DC.
- To convert intermittent renewables into steady "compute-grade" power, Envision runs a predict–dispatch–execute loop: the 天机 weather foundation model forecasts wind/solar output, the 天枢 energy model optimizes allocation against generation, storage, load and price, and the EnOS IoT layer actuates turbines, BESS and DC gear while feeding telemetry back; battery storage buffers green power to GPUs.
- Power delivery uses an SST + BESS 800V DC architecture to cut multi-stage AC/DC conversion losses, copper use and footprint as racks reach hundreds of kW; cooling mixes air, liquid and Ulanqab's natural free cooling. Envision claims ~10x the compute output per unit area versus conventional datacenters (company estimate, not independently verified).
- Business model splits responsibility: Envision delivers green power, land, campus and the AI power system; customers deploy and operate their own servers and chip clusters. It's the first step of the June 2026 "Mission Gobi" plan to build 5GW of green AI compute in desert regions by 2030.
蚂蚁集团领投，老股东超额跟投加注，戴盟两月内连获数亿元融资，以全栈触觉能力破局物理AI
TL;DR - Shenzhen-based Daimon Robotics (spun out of HKUST) raised several hundred million RMB led by Ant Group — its second round in two months — to build a full-stack "physical AI" system anchored on vision-based tactile sensing. It matters because tactile feedback is the missing modality for reliable, production-line dexterous manipulation, and Daimon is vertically integrating sensors, data, and models around it.
- Daimon-TWM, billed as the first tactile-grounded world model, unifies physical reasoning, planning, and instant control; it follows the company's earlier VTLA (vision-tactile-language-action) architecture that added touch to standard VLA models. Claimed cross-embodiment/cross-gripper/cross-sensor portability.
- Demonstrated contact-rich tasks at WAIC — fruit boxing with grip/force adjustment from slip signals, and a pencil-case task (unzip, load, close) with mid-task disturbance replanning at a claimed 95% on-site success rate; pilots cover USB insertion, label application, and fragile-item sorting in 3C and automotive lines.
- Data infrastructure: Daimon-Infinity holds hundreds of thousands of hours of tactile-inclusive multimodal manipulation data (first 10k hours open-sourced, 4.4M+ downloads on ModelScope), plus the RobOmni benchmark; a China Mobile partnership deploys 1,000 collection kits targeting 1M hours/year of real-world data.
- Commercially, Daimon claims to be the first to ship vision-tactile sensors at 10k+ unit scale, serving 200+ customers with recent deliveries to OpenAI, Figure, Physical Intelligence, Skild AI, Meta, BMW, and Google DeepMind. Investor roster (Ant, Inovance, Lenovo Capital, China Mobile/Telecom, China Merchants) spans data supply through manufacturing.