Green AI: Balancing Innovation with Sustainability

绿色AI：平衡创新与可持续发展 预计到2034年，全球AI基础设施将需要7万亿美元的投资，数据中心容量也将从7 GW激增至82 GW，AI的扩张正给区域电网和市政供水带来前所未有的压力。专家小组探讨了这一增长面临的伦理与物理瓶颈，并将数据中心的快速扩建与新建基荷能源长达十年的建设周期进行了对比。尽管DC-Flex负荷转移倡议和神经冷却等AI创新在运营层面提供了一定缓解，但发言者强调，仅凭技术效率提升只能减轻冲击。讨论最终指出，要防止社区抵制和生态承压，需要全国性的政策护栏、合乎伦理的资源管理以及竞争前的跨行业协作。 预计到2034年，全球AI基础设施投资将达7万亿美元，数据中心容量将从7 GW激增至82 GW，对市政电网和供水系统构成前所未有的资源挤压。 (0:04) 数据中心蒸发冷却大量耗水引发了资源属于“市场商品”还是“基本公共品”的伦理争议，若缺乏配套的专有清洁能源与淡化设施，极易招致社区抵制并丧失运营合法性。 (3:46) 为满足用户毫秒级AI推理延迟的需求，电力企业面临在5至10年内重建相当于过去150年电网容量的压力；而天然气（7年）与核电（10年）建设周期漫长，势必加剧供需缺口。 (8:06) AI自身与负载灵活性技术正成为稳定电网的关键手段，通过秒级自主故障响应以及杜克能源与Google合作的“DC-Flex”等时空负载平移试点，有望在不盲目扩建冗余电厂的情况下释放电网潜在容量。 (19:46)

A projected $7 trillion global AI infrastructure expansion is colliding with finite power and municipal water resources, creating severe operational, ethical, and environmental bottlenecks that technical efficiency gains alone cannot resolve, according to industry and academic experts. Rapid Compute Growth Clashes with Utility Timelines (0:04) Mike Obradovitch highlights a Blackstone projection estimating $7 trillion in global AI infrastructure investment through 2034, corresponding to data center power demand surging from 7 gigawatts in 2024 to 82 gigawatts. Rémi Raphael of EPRI reports that power utilities are tasked with replicating roughly 150 years of historical electrical grid capacity within the next 5 to 10 years, with individual utilities like Duke Energy allocating $200 billion to infrastructure in the Carolinas alone to keep pace with facility construction. This rapid expansion conflicts with physical energy development timelines. While consumer intolerance for inferencing latency drives distributed data center buildouts, Raphael notes that constructing a standard natural gas plant requires approximately seven years and a nuclear plant takes a decade. Although an estimated 25 new nuclear plants are needed by 2030 to satisfy demand, zero are currently under construction, threatening to force compute workloads toward regions with preexisting power surpluses. Public Common Goods versus Private Extraction (3:46) Thomas Hofweber frames data center consumption as a direct conflict between viewing resources as market commodities versus inviolable common goods. When facilities consume and evaporate municipal freshwater supplies for cooling or draw heavily on regional electricity, they risk breaching the social compact that treats water and power as shared human lifelines. Hofweber argues that unless technology firms construct dedicated clean generation and desalination facilities alongside facilities, data center operations function as an extractive drain that inflates utility bills and exacerbates local scarcity. Maggie Mosher points out that host communities face severe trade-offs: while rural municipalities often welcome data centers for anticipated tax revenue and jobs, residents face permanent environmental disruptions like continuous acoustic noise and heat dissipation. Hofweber warns that tech companies often leverage economic dominance over vulnerable rural towns, cautioning that if public sentiment turns against localized resource depletion, the industry faces severe regulatory roadblocks and a loss of its social license to operate. AI-Driven Load Shifting and Grid Balancing (19:46) While driving unprecedented power demand, AI is also emerging as a necessary mechanism for grid resilience. Raphael cites an Iberian Peninsula blackout where only 27 seconds elapsed between an initial power disturbance and the total collapse of the grid, demonstrating that human operators cannot manually interpret telemetry and reconfigure grid topologies fast enough to avert catastrophic failures. Autonomous AI response systems are becoming critical to execute sub-minute balance corrections and orchestrate storage assets such as pumped-storage hydroelectricity and utility-scale battery systems. Furthermore, AI facilitates both geographic and temporal load shifting to ease transmission congestion. Rather than continually constructing redundant baseload capacity, operators can migrate computational loads to regions with surplus power or defer non-urgent batch training jobs to off-peak nighttime hours. Raphael notes that real-world pilots, such as the Google and Duke Energy "DC-Flex" initiative in North Carolina, are currently testing real-time workload flexing to synchronize data center draws with available grid headroom. Efficiency Limits and the Policy Imperative (21:49) Panelists caution that component-level improvements—such as Joseph Nordling's computational compression architecture, neural cooling systems, and alternative water sourcing—function merely as band-aids that soften rather than eliminate underlying demand growth. Raphael points out that cooling represents the single largest electricity load category in the United States, meaning operational tweaks cannot overcome baseline physical constraints. In addition, Mosher and Hofweber observe an efficiency paradox: while per-unit computation consumes less energy, the exponential explosion in output volume continues to drive net resource consumption upward. To prevent uneven local negotiations and environmental degradation, the panelists argue for pre-competitive industry collaboration and nationwide regulatory standards. While recent federal initiatives have prioritized 24/7 dispatchable baseload sources like nuclear and gas over intermittent wind and solar, participants emphasize that long-term sustainability requires unified building codes, tax incentives, and cross-sector alignment rather than relying on corporate altruism. The panel emphasizes that engineering and algorithmic optimizations cannot substitute for structured, pre-competitive coordination among technology developers, utilities, and communities. Major unresolved questions remain regarding whether national policy can balance 24/7 baseload reliability with decarbonization targets before regional grids and municipal water supplies face severe localized strain.

研讨会专家指出，全球人工智能基础设施面临资本与算力扩张同物理环境承载力之间的剧烈脱节。Ecolab 与 EPRI 等机构代表强调，若科技行业无法在水利负荷、电网扩容及跨部门协同上建立长效机制，算力扩张将直接威胁公共基础设施并引发深层的社会与监管抵制。 算力激增与电网建设的物理周期错配 (8:06) Mike Obradovitch 引用 Blackstone 的预测指出，到 2034 年全球 AI 基础设施投资将达到约 7 万亿美元，驱动数据中心电力容量从目前的 7 GW 飙升至 82 GW。以局部区域为例，Duke Energy 仅在南北卡罗来纳州就计划投入 2000 亿美元进行电网建设，其主要驱动力几乎完全来自新增数据中心。然而，供需两端在物理扩建节奏上存在严重失衡：现代电网历经约 150 年才逐步建立，而电力公用事业机构目前必须在 5 到 10 年内构建同等体量的容量以满足 AI 算力需求。 Rémi Raphael 进一步指出现实建设周期的硬性约束：常规天然气发电厂的建设平均需要 7 年，而新建核电站的建设周期约为 10 年；行业预估到 2030 年需新建 25 座核电站才能满足电力缺口，但目前处于在建状态的核电站为零。与此同时，终端用户对毫秒级低延迟推断的刚性偏好，导致数据中心必须依赖持续稳定的 24/7 基荷电力，这种消费心理将沉重的环境与基建负担完全转嫁给了上游供电网络。 水利电能的公共品属性与社区负外部性 (3:46) Thomas Hofweber 提出，水资源与电力在本质上属于保障基本生存权与公共秩序的“公共品”，而非单纯由市场动态定价的商品。然而，数据中心在采用蒸发冷却系统时消耗并蒸发大量市政淡水，造成地方水资源的实质性消耗，直接冲击社会契约。此外，数据中心的施工周期远快于配套电厂，新建算力中心往往直接截取区域电网的现有电力，导致局部地区电力稀缺、居民电费激增，并迫使公用事业机构重启高碳排的化石能源进行应急补充。 Maggie Mosher 以堪萨斯州的数据中心项目为例指出，地方社区正承受严重的负外部性困扰：尽管项目带来了税收和就业承诺，但居民不得不面对持续的工业噪音与废热排放，社区宜居性遭到削弱。Hofweber 警告，科技巨头常利用经济优势对财政困难的农村社区强加开采性条款；若缺乏国家层面的统一监管标准，数据中心对地方水利与电网资源的无序抽取将激化公众对抗，危及企业的长期运营合法性。 能效优化的缓解作用及其局限性 (19:46) 面对高能耗瓶颈，AI 正在被反向应用于能效优化与电网负荷调度。Raphael 介绍了 Google 与 Duke Energy 在北卡罗来纳州启动的“DC-Flex”试点，该项目利用时间和地理灵活性，通过将批处理任务推迟至夜间低谷期或调度至未拥塞地区来释放电网冗余容量。学术界也在推进技术革新，例如堪萨斯大学研发的紧凑型计算架构专利与神经冷却网络，旨在从芯片级吞吐与散热环节压降水电消耗。 然而，专家一致认为单纯的技术优化无法逆转生态总负担。Hofweber 与 Mosher 直言，模型压缩、硬件加速以及利用工业废水冷却仅属于“治标创可贴”，无法遏制基线需求的膨胀；单次计算任务的能耗下降更会被暴增的总体输出需求所吞噬，形成典型的效率悖论。此外，由于数据中心需要不间断运行，间歇性风能和太阳能无法单独作为基荷电源，必须依赖抽水蓄能、化学储能以及具备秒级自愈能力的 AI 自动化电网调度系统，以应对类似伊比利亚半岛 27 秒内由微小波动演变为全网崩溃的灾难性断电风险。 前竞争性产业协同与国家监管框架 (31:05) 面对数据中心建设可能导致电网负荷翻倍的严峻预测，Raphael 强调公用事业机构作为区域垄断实体缺乏内部竞争机制，因此依托非营利性机构开展跨行业协同至关重要。EPRI 正在向所有电力机构免费提供数据就绪框架等开源工具，促成全行业在核能及配电网络上的协同研发。Obradovitch 与 Raphael 一致呼吁，科技界与能源界必须将可持续发展确立为“前竞争性领域”，打破企业各自为战的商业壁垒，共同分担电网与冷却标准的基础研发。 在制度设计层面，Mosher 强调单纯依靠企业道德自律无法推动战略转向，必须出台具体的财务补贴、税收调节及准入机制来绑定合规激励。针对政策制定，Raphael 指出当前联邦政策存在过度偏向核电与天然气等可调度基荷、边缘化波动性可再生能源的结构性偏颇。专家呼吁技术人员协助政策制定者打破信息差，制定科学的建筑标准与区域协调政策，避免绿色 AI 沦为口号。 研讨会总结指出，AI 基础设施的可持续性已不再是边缘的公关议题，而是决定整个产业未来能否存续的物理与社会瓶颈。尽管负载柔性化调度和跨学科研发提供了渐进式缓解路径，但在极短的建设窗口期内如何协调超低延迟算力诉求与清洁基荷供应短缺，以及如何建立全国性规制防范对薄弱社区资源的掠夺性开发，仍是尚未解决的核心挑战。

## Background 本内容由 Ai4 于 2025 年 8 月 26 日发布在 YouTube 平台上。视频记录了一场关于“绿色人工智能：在创新与可持续性之间寻求平衡”的专题研讨会，多位来自学术界、电力工业界与产业服务机构的专家共同探讨了在降低数据中心能耗与水耗、优化算力模型以及利用可再生能源的同时，如何确保 AI 创新的持续推进。 ## TL;DR 研讨会围绕 AI 算力爆发引发的电力与水资源危机展开深度剖析。专家们指出，随着数据中心电力需求预计将暴增十倍以上，其对地方电网和水资源的虹吸效应正面临严峻的伦理争议与公共舆论反弹。尽管行业正在尝试通过 DC Flex 等负荷弹性调度、新型散热技术以及模型优化来缓解能耗，但受制于核电等基荷电源漫长的建设周期，单纯的修补难以根治矛盾。与会者一致呼吁跳出零和博弈思维，通过跨学科合作、国家级顶层监管以及将 AI 本身作为电网智能控制工具，共同应对这一系统性挑战。 ## Atmosphere & style 整场研讨会基调理性务实且充满深层忧患意识，既有硬核的工程数据测算，也有深刻的公共伦理拷问。主持人与不同领域嘉宾之间的互动张弛有度，在揭示“基建周期严重落后于算力需求”的严峻现实时直白而犀利，令人警醒；而在探讨产学研跨界协同与技术突破时，又洋溢着建设性的创新热情。整体给人一种在狂热的 AI 浪潮中按下暂停键、冷静审视物理世界底层承载极限的思辨氛围。 ## Key takeaways - [01:40](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=100s) 黑石集团预测到 2034 年全球 AI 基础设施投资将达 7 万亿美元，专用于 AI 数据中心的电力容量将从 7 GW 激增至 82 GW，增幅超 10 倍，地方电网与水资源压力已成为核心瓶颈。 - [03:46](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=226s) 大众对 AI 训练与推理的水电消耗存在严重的认知断层；水资源与电力具有极强的公共品/人权属性，若商业算力过度挤占民生资源，将引发巨大的伦理争议与社会反弹。 - [07:59](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=479s) 电力工业花费 150 年构建的电网体系，因 AI 爆发被迫在未来 5 到 10 年内翻倍扩张；杜克能源仅在两卡罗来纳州就计划投入高达 2,000 亿美元进行基建升级。 - [10:32](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=632s) 公众对 AI 亚秒级极速响应的执念迫使数据中心密集铺设，但能源基建周期漫长（天然气电厂约需 7 年，核电站约需 10 年），导致当前电力供给与算力爆发严重脱节。 - [12:42](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=762s) 数据中心建设速度远快于清洁电厂，直接从现有电网分流电力会推高周边社区电价并可能增加高污染发电应急，亟需探索自给自足的光伏、海水淡化等可持续路径。 - [15:17](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=917s) 产学研正在微计算单元、神经冷却网络等软硬件能效上加速协同，但由于 AI 使得内容生成极度廉价泛滥，其总体能耗仍需警惕反弹效应。 - [18:27](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1107s) 在电网突发故障引发崩溃的数十秒极短窗口内，人类调度员无法应对，必须依靠毫秒级响应的 AI 自动化控制系统来实现电网拓扑自愈与韧性重构。 - [19:46](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1186s) 算力负载具备时间与空间灵活性，像杜克能源与谷歌合作的 DC Flex 项目通过动态调度算力负荷实现削峰填谷，用 AI 解决 AI 自身的部分能耗问题。 - [21:49](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1309s) 单纯优化与负荷转移只是治标的“创可贴”，必须在全美最大的单一用电负荷之一——冷却制冷系统上寻求无水冷却、工业中水利用等根本性突破与跨界协作。 - [24:07](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1447s) 数据中心 24/7 不间断运行的刚需与风光等可再生能源的间歇性存在天然矛盾，实现全天候绿电供应必须依赖抽水蓄能等多能互补体系与秒级智能调度。 - [29:09](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1749s) 数据中心入驻本地社区在创造就业的同时带来噪声与资源压力，推进过程需要居民深度参与决策，并借助政策激励避免科技巨头利用脆弱社区进行“逐底竞争”。 - [37:06](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2226s) 政策与监管不能陷入非此即彼的极端，应统筹核能与燃气轮机等可调度资源与风光等非可调度资源，采取兼收并蓄的全方位能源策略。 - [41:10](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2470s) 算力消耗具有显著的负外部性，各方在商业上可以竞争，但在可持续发展领域绝不能相互竞争，必须将其作为全行业的底层协同基石。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=4s) 研讨会开场与嘉宾背景介绍 主持人 Mike Obradovitch（Ecolab 全球大客户区域副总裁）宣布论坛开幕，本次研讨会的主题为“绿色人工智能：在创新与可持续性之间寻求平衡”（Green AI: Balancing Innovation with Sustainability）。Mike 介绍道，Ecolab 作为全球领先的技术服务商，正在为数据中心领域提供可扩展、高可持续性的冷却与用水解决方案。 随后，三位在各自领域具有深厚背景的嘉宾分别进行了自我介绍： - **Thomas Hofweber**：北卡罗来纳大学教堂山分校（UNC-Chapel Hill）哲学与数据科学教授、AI Project 主任，主要研究 AI 对民主体制的伦理冲击及多项前沿技术议题； - **Maggie Mosher**：堪萨斯大学助理研究教授，长期致力于利用 AI 重塑 K-12 与大学教育生态、AI 辅助评估系统以及用于社交发展的扩展现实（XR）技术； - **Rémi Raphael**：美国电力科学研究院（Electric Power Research Institute, EPRI）首席人工智能官（CAIO）。EPRI 作为北美及全球 45 个国家电力公用事业机构的研发中枢，专注于电力能源系统的技术创新与基础设施升级。 ```mermaid flowchart TD accTitle: 嘉宾专业领域分布图 subgraph Panelists["圆桌研讨嘉宾与专业背景"] M["Mike Obradovitch<br/>(Ecolab / 全球冷却与水资源方案)"] T["Thomas Hofweber<br/>(UNC / 伦理、哲学与数据科学)"] G["Maggie Mosher<br/>(KU / 教育 AI 与 XR 社交发展)"] R["Rémi Raphael<br/>(EPRI / 全球电力电网与 AI 基础设施)"] end M -->|"主持引导"| T M -->|"主持引导"| G M -->|"主持引导"| R ``` --- ### [01:40](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=100s) 算力爆炸背后的资源警报：7万亿美元与数据中心的扩张 Mike Obradovitch 阐述了当前 AI 快速规模化落地背后的核心矛盾。根据黑石集团（Blackstone）近期发布的一项预测，从现在到 2034 年间，全球针对 AI 基础设施的投资总额将高达约 7 万亿美元。这一天量资本的投入意味着，专用于 AI 的数据中心用电容量将从当下的 7 吉瓦（GW）飙升至 2034 年的 82 吉瓦，增幅超过 10 倍。 然而，伴随这些震撼行业的数据而来的，是公众与舆论对于地方电网过载、区域水资源枯竭等生态冲击的强烈担忧。这些环境与基础设施瓶颈直接关系到数据中心是否能够获得当地社区的“运营许可”（License to Operate），并可能根本性制约 AI 的进一步演进。研讨会正式进入第一个核心议题：AI 对电力与水资源的巨大消耗，以及产业与公众对其认知程度的断层。 - **投资规模预测**：2024 至 2034 年间全球 AI 基础设施投入预计达 7 万亿美元； - **电力容量激增**：AI 数据中心电力需求将从 7 GW 攀升至 82 GW； - **社会运营门槛**：地方电网负荷与流域水资源压力正演变为限制算力建设的核心阻力。 --- ### [03:46](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=226s) 水资源是商品还是公共品？伦理冲突与公共舆论风险 针对公众与产业是否真正理解 AI 能耗与水耗的问题，Thomas Hofweber 明确指出，大众目前对 AI 训练与推理所消耗的水电资源规模“完全缺乏认知”。这种认知断层构成了巨大的中期风险——一旦未来由于资源匮乏导致公众舆论对 AI 发生负面反转，可能会直接演变为技术研发与行业发展的巨大阻碍。 Thomas 强调，在诸多能耗维度（如电力、算力、内存、水资源）中，**水资源的伦理属性最为特殊**。关于水资源的定位存在两种根本对立的观念： 1. **纯粹商品论（Commodity View）**：将水视作普通市场商品，需求增加则价格自然上涨，完全由供需关系和价格机制调节； 2. **公共品/人权论（Common Good View）**：认为清洁水源是社区共享的共有财产与基本人权，水资源应首先无条件满足民众饮用、生存和农业灌溉需求，只有超出基本生存红线的盈余部分才可作为商品流通。 Thomas 以北卡罗来纳州的飓风灾害立法为例解释：在自然灾害发生水资源短缺时，法律严禁哄抬物价（Price Gouging），这意味着关键生活必需品绝不能完全交给自由市场调节。当前数据中心采用蒸发冷却技术“将水直接蒸发消耗掉”，本质上是在将全民共享的公共品转化为私营企业的商业利润。若企业企图仅用资本力量垄断或过度开采水资源，将在伦理与合法性层面面临巨大的社会反弹。 ```mermaid flowchart LR accTitle: 水资源属性争议对比 subgraph Commodity["极值观点 A：市场商品论"] A1["由供需关系驱动"] --> A2["价格调节配置"] A2 --> A3["资本优先获取"] end subgraph CommonGood["极值观点 B：公共品/基本人权论"] B1["社区与全人类共有"] --> B2["保障生存与农业为先"] B2 --> B3["严禁危机时哄抬物价"] end ``` --- ### [07:59](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=479s) 历史级电网扩张压力：5-10年重造150年电力工业体系 EPRI 首席人工智能官 Rémi Raphael 接力指出，不仅水是公共资源，**电力同样是现代社会的生命线与公共品**——医院、公共设施以及千家万户的日常生活都高度依赖电网稳定。AI 数据中心对电能的超大规模虹吸，正将整个能源系统推向极限。 为了说明电网面临的空前压力，Rémi 分享了与杜克能源（Duke Energy）首席财务官沟通的一手数据：杜克能源历史上单笔基础设施投资达到 50 亿美元就已是重大事件，然而受数据中心扩张驱动，该公司有史以来首次计划**仅在北卡罗来纳州与南卡罗来纳州（The Carolinas）就投入高达 2,000 亿美元**用于基础设施建设。 Rémi 提出了一个发人深省的时间尺度对比： > 现代电力工业花费了整整 150 年才构建起当今规模的电力网络，但由于数据中心爆发式的电力需求，全行业被迫在未来的 5 到 10 年内，重新建造一个几乎与过去 150 年总和体量相当的电网系统。 这种激进的扩张不仅引发了资金储备与供应链能力的严重质疑，还可能导致电力资源从其他公共部门或传统工业被迫分流。最后，Rémi 引入了用户即时满足感的心理学现象作为对比：现实中人们等电梯最多容忍 20 秒就会感到焦躁，而如今公众在各类生成式 AI（如 Google 搜索、Grok）工具上对亚秒级极速响应习以为常，却极少意识到这背后的电网正在承载多么沉重的物理负荷。 ### [10:32](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=632s) 即时推理的公众诉求与电力基建的残酷现实 谈及人工智能对电网的冲击时，雷米·拉斐尔（Rémi Raphael）向现场观众抛出了一个关于交互耐心的现实问题：当用户在 AI 界面输入提示词后，愿意等待多久才收到回复？现场给出的答案是“即时响应”（instantaneous）。雷米指出，正是公众对毫无延迟的执念，迫使科技公司必须在全国各地铺设能够即时响应的数据中心，因为大众绝不可能接受“请等待 2 分钟或 30 秒再返回推理结果”的妥协。这种刚性需求无法被遏制或削减，给能源行业带来了前所未有的压力。尽管能源领域正在为即将涌入的海量资金而欢呼，但雷米严肃提醒大家：“我们不能仅仅为了运行 AI 就把整个星球推向不堪重负的境地。” 随后主持人追问：能源基础设施的建设速度能否跟上 AI 算力爆发的脚步？如果跟不上，是否会导致算力基础设施外流至其他地区？雷米在短短 45 秒内给出了极其直白的残酷现实：根本来不及。建设一座天然气发电厂平均需要 7 年，而建一座核电站大约需要 10 年。按当前测算，从现在到 2030 年全美需要新增约 25 座核电站，“可大家猜猜看，我们今天动工开建了几座？零座，一座都没有。”与基建停滞并存的，是近期部分地区可再生能源发电量的削减。雷米指出，当前的算力竞争不仅是国内竞争，更是国际竞赛；海外那些拥有充裕剩余电力的国家，将顺理成章地吸引数据中心落地与模型开发。 - **电厂建设周期**：天然气电厂平均耗时约 7 年，核电站平均耗时约 10 年 - **2030 年核电缺口**：预估需求 25 座核电站，当前实际在建数量为 0 ```mermaid flowchart TD accTitle: AI推理需求与能源基础设施建设脱节示意图 subgraph 需求端["用户与算力需求端"] A["公众对 AI 推理即时响应的严苛诉求<br>(无法接受数秒至数分钟延迟)"] --> B["全美范围内密集建设即时响应数据中心"] end subgraph 供给端["能源基建供给现实"] C["天然气电厂建设周期：约 7 年"] D["核电站建设周期：约 10 年<br>(2030 年前需 25 座，当前在建 0 座)"] E["部分可再生能源增量放缓"] end B -.->|"供需严重脱节"| F["能源成为瓶颈 / 算力流向海外富电区域"] C --> F D --> F E --> F ``` --- ### [12:42](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=762s) 算力扩张的伦理困境与社区公共资源挤占 托马斯·霍夫韦伯（Thomas Hofweber）从伦理和资源分配角度深入剖析了算力狂飙背后的社会代价。他指出，建造一座数据中心的速度和难度，远比建造一座专门为其供电的清洁发电厂要快得多、容易得多。其必然结果是：数据中心落成后，只能直接从现有电网中分流电力。这种“插队”行为直接将电力变成更加稀缺的高价商品，大幅推高周边社区的用电成本，甚至可能迫使电网启用更多高污染的化石燃料电厂以应急。 托马斯强调，如果科技巨头在兴建数据中心的同时，配套建设太阳能光伏电站和海水淡化厂，完全使用自给自足的绿色能源与水资源，“那就没有任何人会抱怨，你用可持续的方式增加了清洁产能，问题迎刃而解”。但现实中这几乎从未发生。水资源与电力本质上是属于地方社区的公共资源，但在商业利益驱动下正在被商品化，普通居民往往面临水资源枯竭与电价飙升的双重挤压。主持人随后补充，中东等面临严重水资源危机的地区已经开始探索这一路径，利用充沛的土地与光照资源，将太阳能发电与海水淡化技术直接融入下一代数据中心的设计之中。 --- ### [15:17](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=917s) 绿色 AI 创新：产学研合作与软硬件能耗优化 玛吉·莫舍（Maggie Mosher）将讨论视角转向积极的解决方案。她认为当前最令人振奋的转变，是行业正从过去的“空头口头承诺”走向实质性的“深度跨界合作”。以堪萨斯大学为例，校方正与电力公司 Evergy 以及松下（Panasonic）展开合作，制定了在 20 年内完全依赖风能、不挤占 Evergy 公共电网资源的长期目标。这种与传统能源公司携手探索能效优化的模式，打破了以往单纯买电卖电的对立关系。 在技术研发一线，能效革新正从计算架构与散热两端同时发力。玛吉介绍，她的前中学生弟子约瑟夫·诺德林（Joseph Nordling）近期成功获得了一项专利，能够将微计算单元做得更小、更快，大幅降低单次运算的能耗；堪萨斯大学的苏玛亚·沙马伊博士（Dr. Sumaya Shamayi）则通过神经冷却网络（neural cooling networks）大幅提升散热速度，减少冷却阶段的水电消耗。此外，谷歌、微软等巨头也纷纷将算力部署至苏格兰（利用风能）以及挪威、荷兰（利用水资源与自然冷却）。 在软件与应用层面，AI 本身就是解决能耗调度的重要工具。研究表明，利用 AI 预测和分析电网一天中能效最高的用电时段，其精准度与节能效果远超传统分析方法。针对生成式 AI 的能耗争议，玛吉以图像生成为例进行了辩证分析：单次利用 AI 生成图片的耗能确实远低于传统人工绘制，但由于 AI 使得图像生产变得极度廉价和泛滥，总体能耗是否真正降低仍需警惕反弹效应。主持人随之总结道，AI 训练阶段可以全球追逐低成本可再生能源，但推理阶段必须贴近终端社区，因此端侧与架构侧的革新至关重要。 - **堪萨斯大学绿色承诺**：联合 Evergy 与 Panasonic，立项 20 年内实现 100% 风电供应 - **计算微型化专利**：Joseph Nordling 研发更小、更快的计算单元以压降基础能耗 - **先进散热技术**：Dr. Sumaya Shamayi 主导的神经冷却网络技术加速散热循环 --- ### [18:27](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1107s) 电网毫秒级危机：AI 赋能能源系统的不可或缺性 雷米·拉斐尔针对前述讨论补充强调，自己指出行业痛点并非出于悲观，而是为了直面严峻挑战；人类历史上的每一次危机，最终都是依靠人类创造力得以化解，而 AI 自身恰恰是解决 AI 能耗与电网危机最强有力的工具。 为了证明 AI 实时控制系统的必要性，雷米列举了约两个月前伊比利亚半岛发生的一起严重电网事故。当时电网突发故障，导致 5000 万居民陷入长达约 22 小时的大停电。雷米查阅官方研究报告后指出，从电网功率出现异常波动到整个电网彻底崩溃，全程仅仅耗时“整整 27 秒”。在短短 27 秒的极度时间窗口内，任何人类调度员都根本不可能完成数据读取、因果分析并重新配置电网拓扑来阻止连锁崩溃。唯有具备毫秒级反应能力的 AI 自动化控制工具，才能在电网遭遇突发扰动时实现瞬间重构与自愈，这正是未来电网韧性建设不可或缺的核心力量。 ### [19:46](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1186s) 算力负载的时空灵活性与 DC Flex 试点项目 除了利用人工智能实时维持电网频率、避免负荷削减与级联大停电之外，AI 在能源管理中还有巨大的优化潜力。其中一个关键方向是算力负载的“空间灵活性”（Geographic Flexibility）与“时间灵活性”（Temporal Flexibility）。正如电动汽车车主通常会选择在电价较低、电网压力较小的午夜充电，而不是在晚上 8 点的高峰期充电一样，数据中心的大量批处理任务也可以在时间上推迟至深夜，或在空间上直接转移到电网容量充裕、新能源富余的地区。AI 能够精准预测并动态调度这些跨区域、跨时段的负载，实现最佳时机与批次配置。 这种时空调度并非遥不可及的未来设想，而是正在落地的现实项目。例如即将在北卡罗来纳州杜克能源（Duke Energy）服务区启动的 **DC Flex** 示范项目，谷歌（Google）运营团队深度参与其中，探索如何通过灵活调整数据中心负荷配置与电网取电曲线，从而更高效地运行算力设施。当某一区域电网过载时，系统能够将任务动态调度至其他节点。这种模式恰恰是“用 AI 来解决 AI 自身带来的一部分能耗问题”。 ```mermaid flowchart LR accTitle: Data Center Load Flexibility subgraph S1["调度维度"] A["时间灵活性<br/>（移至夜间低谷期）"] B["空间灵活性<br/>（转移至富余电网节点）"] end subgraph S2["智能调度中枢"] C["AI 负载优化系统<br/>（DC Flex 机制）"] end subgraph S3["电网与算力协同效益"] D["削峰填谷<br/>缓解局部电网过载"] E["释放冗余容量<br/>减少新建算力/电厂需求"] end A --> C B --> C C --> D C --> E ``` - **DC Flex 计划**：即将在北卡罗来纳州杜克能源管辖区域内的数据中心执行，与谷歌团队合作测试数据中心负荷的动态弹性调节。 - **负荷转移效益**：通过在不同数据中心之间平移算力负载，不仅能缓解局部电网瓶颈，还能释放现有电力冗余，从而在一定程度上降低盲目新建更多数据中心或备用发电厂的需求。 --- ### [21:49](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1309s) 根本解决还是“创可贴”：数据中心散热与跨领域协作 面对负荷转移等优化措施，托马斯·霍夫韦伯（Thomas Hofweber）提出了冷静的反思。他指出，无论是推理与训练层面的算法优化、模型架构精简，还是电网负荷转移，本质上都只是“缓解”（soften）了问题，而非从根本上“解决”（solve）问题。面对呈指数级扩张的算力需求总量，哪怕经过层层优化，能源缺口的绝对体量依然巨大而严峻。 雷米·拉斐尔（Rémi Raphael）完全赞同托马斯关于“缓解不等于根治”的看法，并顺着现场偏冷的空调温度开玩笑，抛出了全美耗电量最大的单一负荷来源——**制冷系统（Cooling）**。他指出，在新建更多发电设施之前，必须反思人类是否高效利用了水与碳等自然资源，冷却系统是否同样可以进行弹性调度与重构。 麦琪·莫舍（Maggie Mosher）进一步强调，仅靠单项技术突破（如更快的计算芯片、利用造纸厂工业水源或中水/零水冷却）就像是在贴“创可贴”（band-aid），无法实现全局变革。她呼吁教育界、学术研究机构、工业界企业与非营利组织打破壁垒展开跨界协作，汇聚不同领域的思维模式，共同探索颠覆性的散热与能效方案。 - **全美最大用电负荷之一**：各类建筑与数据中心的制冷系统消耗了巨额电能，具备巨大的弹性化改造与能效优化潜力。 - **替代性冷却设想**：包括探索无水冷却技术、利用造纸厂等工业副产水源或城市中水进行散热，减少对原生优质淡水和高碳电力的直接依赖。 - **多方跨界协作机制**：单一企业或学术界无法独自解决能源水资源危机，必须在跨行业、跨学科的协同中寻找系统性解法。 --- ### [24:07](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1447s) 算力与能源的互利（AI for Power）与 24/7 绿电供应挑战 谈及 AI 如何反哺能源行业（AI for Power）以及加速向可再生能源过渡，曾在新能源企业工作多年的雷米·拉斐尔指出了理想与现实之间的核心鸿沟——**规模与速度**（speed and scale），以及绿电的**间歇性与数据中心 24/7 不间断运行之间的天然矛盾**。单纯依靠太阳能直接驱动算力“在理论上很美好，但在实践中行不通”：当用户在凌晨两点向 AI 提问或给家人发送邮件时，他们需要即时响应，但此时既没有日光，某些地区（如堪萨斯州）也可能恰好无风。 因此，要让数据中心实现全天候平稳运行，必须依赖多元化调峰与储能体系的高效协同。除了化学电池储能外，还需结合抽水蓄能（Pumped Hydro）等物理储能方式——在低谷期将水抽往上游水库，在高峰期放水发电。在这个复杂的混合能源网络中，AI 的核心使命在于以秒级甚至毫秒级的速度进行资源最优调度与需求响应（Demand Response），正如在电网突发故障事件中实现数十秒内的极速恢复。麦克·奥布拉多维奇（Mike Obradovitch）总结道，这绝非可再生能源与传统能源的“非此即彼”（either/or），而是一个必须兼顾基荷稳定与可持续性的“两者兼备”（and）的系统工程。 - **24/7 算力与可再生能源间歇性矛盾**：算力服务要求全天候即时响应，而风光资源受昼夜和天气制约，无法在没有储能和调峰电源的情况下独立支撑连续计算。 - **抽水蓄能机制**：利用上下两座水库的水位差，在用电低谷抽水蓄能、用电高峰放水发电，构成大规模自然储能形式。 - **秒级智能调度**：借助 AI 算力优化多能互补系统，提高极端电网波动下的应急响应能力与清洁能源消纳率。 --- ### [27:34](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1654s) 水能纽带关系与通过全局优化破除资源消耗争议 针对麦克·奥布拉多维奇提出的“水-能纽带关系”（Energy-Water Nexus）以及“数据中心正在榨干当地社区水电资源”的公众负面舆论，麦琪·莫舍从算法创新与宏观资源配置的角度给出了分析。诸如**液态神经网络**（Liquid Neural Networks）等新型模型架构正在探索更高效的计算范式，但更深层的解法在于利用 AI 自身强大的模式识别和全局优化能力，重构资源与算力的空间匹配。 与其在缺水缺电的社区盲目扩张算力引发民意反弹，更合理的路径是借助 AI 的分析预测能力，在宏观层面上摸清全国甚至全球的资源底账——精准识别哪些区域天然具备充沛的风能、光能与水资源，并将算力基础设施科学地布局在自然禀赋契合的地理位置，从根源上化解数据中心对脆弱社区资源的挤占。 - **公众感知痛点**：随着大型数据中心落地，地方社区越来越担忧算力集群对当地淡水供应和电力基础设施造成不可逆的过度消耗。 - **新型算法架构探索**：液态神经网络等自适应、轻量化模型正在被寄予厚望，以降低单位计算任务的能耗开销。 - **资源禀赋与算力布局匹配**：利用 AI 宏观分析统筹水资源与绿电分布，推动数据中心向自然资源具备承载力的地区合理迁移与规划。 ### [29:09](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1749s) 社区影响与激励机制：数据中心落地的地方博弈 Maggie Mosher 探讨了数据中心入驻本地社区所带来的复杂现实与矛盾。她以堪萨斯州多奇市（Dodge City, Kansas）即将迎来的一座数据中心为例：从积极的一面来看，当地社区确实急需创造就业岗位，数据中心的入驻填补了关键的经济与生计需求，让许多居民能够保住住房；但与此同时，居民们也对数据中心带来的巨大噪声、散热以及环境压力忧心忡忡，甚至有人抱怨“未来可能会因为噪声太吵而无法正常享受娱乐生活”。 Mosher 强调，在推进此类项目时，必须让当地居民充分参与到决策讨论中。社区居民最了解本地资源的承载能力，也清楚本地有哪些研究人员正在关注相关议题。此外，她指出推动行业转向可持续发展离不开合理的激励机制（Incentives）——“并非所有人都会单凭道德自觉去做正确的事，很多时候人们做出改变是因为有利益激励推动”。主持人随后顺势将话题延展至全场：除了社区协同与激励外，推动绿色 AI 还亟需哪些跨界协作与最佳实践？ ```mermaid flowchart TD accTitle: Data Center Community Tradeoffs subgraph LocalImpact["数据中心落地本地社区的利弊博弈"] Pros["积极利益：增加就业、稳定住房与税收"] Cons["潜在代价：噪声污染、余热排放、资源消耗"] end subgraph Solutions["推进路径"] Involve["吸纳本地社区深度参与决策"] Incentivize["设计经济与政策激励机制"] end Pros --> Solutions Cons --> Solutions ``` --- ### [31:05](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1865s) EPRI 的跨行业协同模式与电网翻倍挑战 Rémi Raphael 从电力行业的角度分享了电力科学研究院（EPRI）的协作实践。EPRI 扮演着各大公用事业公司（Utilities）研发智囊的角色。由于电力公用事业公司在各自特许经营的地理区域内运营，彼此之间并不存在直接的市场竞争，也不单独建立庞大的研发团队，这使得跨州乃至全球范围内的深度技术合作成为可能。EPRI 正在借鉴伊比利亚半岛等海外电网经验，防止美国在电网承载上重蹈覆辙，其研究范围横跨核能、电网调度优化到配电网络的各个环节。 Raphael 指出，当前全球资本与技术投入正大规模涌入数据中心领域，而数据中心的快速扩张预计将使现有电网的负荷规模直接翻倍。面对如此严峻的挑战，单打独斗无法解决问题。作为非营利组织，EPRI 践行“公共利益（Common Good）”使命，近期与 PwC、埃森哲（Accenture）等机构联合发布了免费的数据就绪度（Data Readiness）评估框架，供所有公用事业机构开放使用，旨在通过共享数据与联合研究为决策者和行业提供扎实的技术建议。 - **电网负荷倍增**：数据中心基础设施的爆发式建设预计将使现有电网容量需求扩大一倍。 - **开放协同成果**：EPRI 与埃森哲、PwC 联合研发数据就绪度框架，无偿向全行业开放。 - **公用事业天然协同**：公用事业机构受区域专营限制，无横向竞争壁垒，具备理想的协同研发基础。 --- ### [33:19](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1999s) 监管与国家级顶层设计的伦理责任 讨论随后转向政策制定与监管者的核心角色。Thomas Hofweber 教授直言，AI 与能源消耗的失衡深刻影响着普通人的生存生活。一旦治理失败，人类面临的将是“更严重的污染、更匮乏的水资源，以及一大堆普通人根本无法从中获益的 AI”，这将带来极其不道德的社会后果。他指出，目前社会对于环境恶化心知肚明，但真正的系统性解决方案仍未出台，仅仅依靠零敲碎打的缓解措施（Mitigations）远远不够，必须通过全国性法规来强制落实根本性解决方案。 Mosher 与主持人补充道，政策制定者本身存在认知盲区，需要行业专家主动向其普及数据中心内部运作与资源消耗现实，引导政府在税收减免、绿色建筑审批等方面设立引导机制，促使原本竞争的企业在基础设施共享上展开合作。Hofweber 进一步犀利地指出，如果缺乏国家层面的统一监管规则，科技巨头在选址时往往会挑选处于经济困境的脆弱社区（Rural communities in economic despair），利用其对就业和税收的迫切渴望来强加有利条件，因此绝不能任由地方博弈演变成“逐底竞争”。 --- ### [37:06](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2226s) 政策现实与可调度电力资源的现实约束 Raphael 针对 Hofweber 提出的国家行动发表了深入见解。他呼吁行业跳出党派政治纷争，坚持以“科学驱动（Science-driven）”的研究来解决 AI 与电力的供需矛盾。他提及近期出台的关于 AI 的行政令行动计划（Action Plan / Executive Order），指出了政策愿景与工程现实之间的巨大差距。 Raphael 剖析道，该行政令细则在规划发电能力时实际上排除了不可控的可再生能源，转而着重强调“可调度资源（Dispatchable Resources）”。他向听众解释了这一核心电力概念：核能与燃气轮机属于可调度资源，能够按需启停并提供 24/7 全天候基荷电力；而风能与太阳能作为非可调度资源（Non-dispatchable resources），其发电输出受自然条件制约，“随风而来、随阴而落”，无法单独支撑数据中心每时每刻平稳运行的高密度负荷需求。 | 资源类别 | 典型能源形态 | 电网特性 | 对数据中心 24/7 负荷的支撑能力 | |:--- |:--- |:--- |:--- | | **可调度资源 (Dispatchable)** | 核能、燃气轮机 (Gas Turbines) | 可人工按需启停、调节功率，提供稳定连续的基荷电力 | **极高**：能够保障全天候不间断供电 | | **非可调度资源 (Non-dispatchable)** | 风能 (Wind)、太阳能 (Solar) | 依赖天气自然条件，发电具有间歇性与波动性 | **受限**：若无大规模储能配套，无法单独满足连续高功耗需求 | ### [38:16](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2296s) 能源视角的统筹：摆脱非此即彼的政策思维 讨论在结束阶段延续了对能源转型路径的反思。发言者指出，很多时候行政命令（Executive Order）或监管政策会简单地倒向某一方，比如强制要求建设可调度资源（dispatchable resources），结果导致策略在两个极端之间摇摆，往往顾此失彼。现实中我们不能只盯着单一方案，而必须采取兼收并蓄的“全方位”（all of the above）能源策略，将不同资源和技术路径综合考量，这本质上是一个看待系统复杂性的视角问题。 随后，主持人顺势切入圆桌的收尾环节。在讨论行将结束之际，他邀请台上每位嘉宾分享各自的最终总结与行动呼吁（call to action）：面对 AI 带来的巨大能耗、数据中心的碳排放以及严峻的可持续性挑战，参会者和行业从业者今天最应该带走的一个核心观点是什么？当前最关键、最应该立即着手推进的应对步骤又是什么？ --- ### [39:05](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2345s) 大胆分享创意与跨界落地：学术界与产业界的必由之路 堪萨斯大学助理研究教授 Maggie Mosher 首先给出了她的建议。她呼吁所有人：“如果你有一些想法，无论它们听起来多么不可能、甚至多么荒谬，都请立刻开始分享出来。”因为纵观历史，往往正是这些起初看似异想天开的想法，最终彻底颠覆并重塑了整个系统的底层运转机制。 紧接着，Maggie 强调光有想法还远远不够，必须把创意送到能够将其落地的实干家手中。她坦率地分享了学术界的局限：在堪萨斯大学，研究团队拥有极其出色的研究成果，工程师也能攻克特定技术环节，但学术界本身并不直接建造真实世界的数据中心，也不可能掌握全球所有自然资源的分布细节。而市场上恰恰有懂得如何统筹全局、具备规模化建设能力的企业与机构。因此，跨学科、跨行业的高效合作与伙伴关系，是实现绿色创新的决定性一步。 ```mermaid flowchart TD accTitle: 从学术前沿创意到产业落地闭环 subgraph 学术端 ["学术端：理论与技术探索"] A["异想天开的突破性创意"] --> B["高校实验室工程研究"] end subgraph 产业端 ["产业端：全球资源与基建"] C["数据中心建设与运营"] --> D["全球自然资源统筹分配"] end 学术端 -->|"跨界伙伴关系与协作"| 产业端 产业端 --> E["重塑绿色 AI 基础设施"] ``` --- ### [39:48](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2388s) 严峻性与社会反弹：被低估的资源瓶颈与公众信任危机 北卡罗来纳大学教堂山分校的 Thomas Hofweber 教授随后提出了更为深层的忧患意识。他指出，目前社会各界对 AI 资源消耗问题的严重程度存在着普遍且巨大的低估。眼下，大型 AI 科技公司手握巨额资金与资源，在整个技术生态中占据绝对主导地位，能够全盘主导行业规则的走向与游戏方式。 然而，Thomas 犀利地警告道，这些巨头自身的未来同样正处于这场危机的威胁之下： - **物理资源极限**：AI 膨胀式的用电与用水需求终将撞上物理基础设施的硬墙，导致电力或冷却资源彻底枯竭。 - **公众舆论反弹**：如果普通民众的基本生活受到能源争夺的影响，公众情绪极有可能迅速恶化，整个社会最终可能会对 AI 这一庞大项目本身产生强烈的抵触与反感。 他直言，无论是公共辩论的广度，还是学术研究与工业界协作的深度，全社会目前对该议题的重视都远远不够，必须以更强的紧迫感、在更宏大的系统尺度上来正面应对这一危机。 --- ### [40:33](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2433s) 算力的负外部性与公共利益：不可忽视的真实能源成本 紧接着发言的嘉宾顺着 Thomas 的逻辑，进一步剖析了数据中心算力背后的“负外部性”（externalities）。他提醒所有人，算力绝不是免费的，使用数据中心实质上是在消耗全社会的“公共品”（common goods）。 为了让这一抽象概念具体化，嘉宾半开玩笑地举了一个虽然略带调侃但发人深省的生动例子：想象一下，当你在凌晨两点毫无顾忌地发送一封电子邮件或调用模型时，在同一电网覆盖下的医院里，可能正有一位病患插着维系生命的医疗电力设备。你随手消耗的电能，与维持生命运转的电力来自同一张公共电网。电力资源并不是无限的私人物品，它关乎公共利益与社会底线，每一个使用算力的人都必须对这种无形的公共消耗保持清醒与敬畏。 --- ### [41:10](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2470s) 主持人总结：可持续发展是一场不应相互竞争的协同之战 在聆听完嘉宾们的发言后，主持人发表了饱含共鸣的闭幕致辞。他提到，自己在与业内各方人士交流时，感受最深的一点是：在商业世界中，大家可以为了争夺同一批客户、同一片土地、同一种商业资源而展开激烈竞争；“但是，可持续发展绝不是我们用来彼此竞争的赛道。” 相反，可持续性是维系整个生态存续的底层支柱，也是所有玩家必须搁置分歧、通力协作才能攻克的难关。面对如此庞大而复杂的系统性挑战，唯有汇聚最顶尖的智慧与头脑，才能催生出真正具有突破性的创新解决方案。 在全场热烈的氛围中，主持人感谢了台上的五位嘉宾带来的深刻洞察。由于圆桌时间已到，现场无法留出提问环节，主持人热情地邀请所有参会人员在散会后移步会场外的走廊，在当天下午乃至次日会议结束前，继续与嘉宾们展开更加深入的线下交流。随后，整场圆桌研讨在掌声中正式落下帷幕。 ## Notable quotes > [07:59](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=479s) 现代电力工业花费了整整 150 年才构建起当今规模的电力网络，但由于数据中心爆发式的电力需求，全行业被迫在未来的 5 到 10 年内，重新建造一个几乎与过去 150 年总和体量相当的电网系统。 — Rémi Raphael > [10:32](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=632s) 我们不能仅仅为了运行 AI 就把整个星球推向不堪重负的境地。 — Rémi Raphael > [10:32](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=632s) 可大家猜猜看，我们今天动工开建了几座？零座，一座都没有。 — Rémi Raphael > [12:42](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=762s) 那就没有任何人会抱怨，你用可持续的方式增加了清洁产能，问题迎刃而解 — Thomas Hofweber > [18:27](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1107s) 整整 27 秒 — Rémi Raphael > [29:09](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1749s) 未来可能会因为噪声太吵而无法正常享受娱乐生活 — Maggie Mosher > [29:09](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=1749s) 并非所有人都会单凭道德自觉去做正确的事，很多时候人们做出改变是因为有利益激励推动 — Maggie Mosher > [33:20](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2000s) 更严重的污染、更匮乏的水资源，以及一大堆普通人根本无法从中获益的 AI — Thomas Hofweber > [37:06](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2226s) 随风而来、随阴而落 — Rémi Raphael > [39:05](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2345s) 如果你有一些想法，无论它们听起来多么不可能、甚至多么荒谬，都请立刻开始分享出来。 — Maggie Mosher > [41:10](https://www.youtube.com/watch?v=9S2j-MT8lLQ&t=2470s) 但是，可持续发展绝不是我们用来彼此竞争的赛道。 — Mike Obradovitch ## Mentioned Maggie Mosher, Manoj Saxena, Rémi Raphael, Thomas Hofweber, Mike Obradovitch, Joseph Nordling, Dr. Sumaya Shamayi, University of Kansas, Responsible AI Institute, Electric Power Research Institute (EPRI), UNC-Chapel Hill, Ecolab, Blackstone, Duke Energy, Evergy, Panasonic, Google, Microsoft, PwC, Accenture, DC Flex, Grok, Liquid Neural Networks ## Video description Maggie Mosher, Assistant Research Professor, University of Kansas Manoj Saxena, Founder and Chairman, Responsible AI Institute Rémi Raphael, Chief AI Officer, Electric Power Research Institute (EPRI) Thomas Hofweber, William R. Kenan Jr. Distinguished Professor of Philosophy and Data Science, Director of the AI Project, UNC-Chapel Hill Mike Obradovitch, AVP for Global High Tech, Ecolab Abstracts As AI continues to drive innovation, it also brings significant energy consumption and environmental challenges. This session will focus on how the AI industry can reduce its environmental impact and create more sustainable, energy-efficient solutions. Panelists will discuss the role of Green AI in optimizing AI models, reducing energy consumption in data centers, and leveraging renewable energy to power AI infrastructure. Learn about emerging technologies and practices that aim to balance the advancement of AI with environmental responsibility. Discover how organizations can integrate sustainability into their AI strategies without compromising on performance or innovation. [Source](https://www.youtube.com/watch?v=9S2j-MT8lLQ)

[7:59] Rémi Raphael: 现代电力工业花费了整整 150 年才构建起当今规模的电力网络，但由于数据中心爆发式的电力需求，全行业被迫在未来的 5 到 10 年内，重新建造一个几乎与过去 150 年总和体量相当的电网系统。 [10:32] Rémi Raphael: 我们不能仅仅为了运行 AI 就把整个星球推向不堪重负的境地。 [10:32] Rémi Raphael: 可大家猜猜看，我们今天动工开建了几座？零座，一座都没有。 [12:42] Thomas Hofweber: 那就没有任何人会抱怨，你用可持续的方式增加了清洁产能，问题迎刃而解 [18:27] Rémi Raphael: 整整 27 秒 [29:09] Maggie Mosher: 未来可能会因为噪声太吵而无法正常享受娱乐生活 [29:09] Maggie Mosher: 并非所有人都会单凭道德自觉去做正确的事，很多时候人们做出改变是因为有利益激励推动 [33:20] Thomas Hofweber: 更严重的污染、更匮乏的水资源，以及一大堆普通人根本无法从中获益的 AI [37:06] Rémi Raphael: 随风而来、随阴而落 [39:05] Maggie Mosher: 如果你有一些想法，无论它们听起来多么不可能、甚至多么荒谬，都请立刻开始分享出来。 [41:10] Mike Obradovitch: 但是，可持续发展绝不是我们用来彼此竞争的赛道。

[19:46] Rémi Raphael: using AI to solve the problem that AI already created for us [21:49] Rémi Raphael: if we are freezing in this room, you know why [24:07] Mike Obradovitch: not an either/or, it's an and [29:09] Maggie Mosher: No, we're not going to be able to have entertainment because we won't be able to hear anymore. [29:09] Maggie Mosher: I know that that's a word that not a lot of people want to hear, but I do think that not everybody does the right thing because it's the right thing. They do it because there's an incentive for them to do it. [33:19] Maggie Mosher: they also don't know what they don't know. [33:19] Thomas Hofweber: You got to think about like what should be allowed, what should not be allowed instead of simply picking sort of like your weakest community that you can impose the terms on, as it often happens that rural communities that are in economic despair get. [39:05] Maggie Mosher: no matter how impossible or ridiculous they seem, because some of those ideas are the ones that actually end up changing the way the whole structure works. [39:05] Thomas Hofweber: Their own future is also in some way at risk in this problem. And not just that they'd run out of the resources, but that you know, society could turn against the whole project. And so... this problem needs to be addressed much more urgently and at a grander scale than I think it has been, both in the public debate as well as in research collaboration and industry.

[0:04] Good afternoon. Welcome everyone. Thanks [0:05] for joining uh our panel here this [0:08] afternoon on green AI balancing [0:10] innovation and sustainability. I'm I'm [0:12] Michael Bradovich uh area vice president [0:15] for global accounts with with EcoAB. Uh [0:17] for those of you who aren't familiar [0:18] with with EcoAB uh amongst other things [0:21] that we do, we are a global leader in [0:24] providing scalable sustainable cooling [0:26] solutions for the data center market [0:28] segment. So today I'm pleased to be [0:30] joined by my esteemed panelists uh who [0:33] are going to engage in a very healthy [0:37] discussion around uh sustainability of [0:40] AI. And so I will allow them to each [0:42] introduce themselves starting maybe with [0:44] Thomas. [0:45] >> Uh my name is Thomas Hoffer. I'm a [0:47] professor of philosophy and data science [0:49] at the University of North Carolina [0:50] Chapel Hill. Um I work on various AI [0:53] related issues including ethical impact [0:57] of AI on democracy and um various other [1:00] more technical issues. [1:03] >> I'm Maggie Moer. I'm an assistant [1:04] research professor at the University of [1:06] Kansas and I also work on many AI [1:08] projects, some being assessments, [1:10] reimagining education, K through2 and [1:12] college education with AI and I do [1:15] extended reality for social development. [1:19] >> Good afternoon everyone. Uh my name is [1:21] Remy Rafael. I am the chief AI officer [1:24] of APRI. APRI is the electric power [1:27] research institute. We are the R&D I [1:29] would say arm for all utilities that [1:31] work in power in North America and [1:33] internationally as well in 45 countries. [1:37] >> Awesome. Thank you uh all of you. Um so [1:40] again tremendous amount of expertise up [1:43] here on on this stage and and discussing [1:46] this really complex issue that that we [1:48] um have in front of us specifically that [1:51] as we continue to see AI deployments at [1:54] scale increase. So do we also see [1:57] concerns arising with their [1:59] environmental footprints. um to this [2:02] point I think I saw a Blackstone article [2:05] uh that is forecasting about 7 trillion [2:09] in AI infrastructure investment between [2:12] now and 2034. That equates to a data [2:17] center or AI data center capacity [2:19] increase of of 7 gawatt this year to 82 [2:23] gigawatts uh in 2034. And while we I [2:26] think all see articles with oppressive [2:29] headlines similar to those, we're also [2:32] starting to see an increase in headlines [2:35] expressing concern about the impacts [2:37] that continued data center expansion [2:41] will have on local power grids, local [2:43] watersheds, and and those concerns and [2:46] challenges really strike to the the core [2:48] of data centers licenses to operate as [2:52] well as further AI expansion. And so uh [2:55] today we're going to tackle the balance [2:58] between AI innovation and [3:00] sustainability. We're going to explore [3:03] some of those challenges, the [3:04] opportunities, and I think more [3:05] importantly the path forward for a truly [3:08] uh responsible AI revolution. And so [3:11] this panel is divided up into a couple [3:13] little subtopics here. And we're going [3:15] to start with uh the one that is I think [3:18] one of the core challenges uh namely [3:20] power and AI's thirst for it. And so I'm [3:24] going to start this first question with [3:26] perhaps Thomas. [3:28] To what extent do you believe the public [3:31] uh and industry truly grasp the energy [3:33] and water demands of current and future [3:35] AI deployments? How does this impact or [3:38] strain current industry's sustainability [3:40] initiatives and what are the immediate [3:42] infrastructure challenges faced? [3:46] >> Okay. Um I mean I feel like that the um [3:50] the amount of um energy and water that's [3:53] being used is not at all recognized in [3:56] the public. And uh I think this is a [3:58] real serious issue both for the industry [4:01] as well as for the public that faces uh [4:04] you know the the lack of that energy on [4:06] the on the market otherwise and also the [4:08] potential um scarcity of water. Um and [4:12] it's also a threat um for the industry [4:14] in the sense that if public opinion [4:15] turns on AI, it could be a major block [4:19] for um growth and development in this [4:21] area. So um it's probably not in the [4:25] near term, but in the midterm, I think [4:27] there's a real potential that this is [4:29] one of the obstacles for progress. So I [4:31] think it's really important to [4:32] understand where this is coming from and [4:34] um and what some of the uh solutions [4:37] are. And so I think one of the key [4:38] things and maybe I'll I'll I'll bring [4:40] this up now is that the it's really [4:44] important to understand that this is all [4:45] about um water and energy. Um and water [4:49] I think is maybe the more plausible [4:51] candidate to make the case that there is [4:52] a the real issue here. So everybody [4:55] wants to make the systems more efficient [4:57] and less energy consuming and less you [5:00] know you want to use less memory and you [5:01] want to use less water and and and all [5:03] these dimensions but for some of them um [5:06] water in particular as opposed to memory [5:08] there's a real issue about um uh the [5:11] status of water. Um so one way to think [5:14] about it on the one extreme end is this [5:16] is just a commodity. You know, water is [5:18] something that, you know, you can buy [5:20] and if more water is demanded, then the [5:23] price goes up. Um, but that's just a [5:25] market uh, you know, reacting u to the [5:27] demand for water. There's nothing [5:29] special about it, right? Versus on the [5:30] other end of the spectrum, you think [5:32] that water is really a common good. It's [5:34] not at all a commodity. It's something [5:36] that is tied to uh something that's [5:38] owned by the community, that's tied to a [5:41] human right for clean water. um that [5:44] water primarily belongs to everybody and [5:47] only potentially what's the surplus like [5:50] what's goes beyond what's needed by by [5:53] people for farming and and you know and [5:56] drinking water and so on and so forth [5:57] only that can be seen as a commodity. So [6:00] um right now you know there is a a good [6:03] amount um towards this the second one I [6:07] would say you know so for example um [6:09] it's a it's very common to think that [6:11] there are some limits to the market for [6:13] certain things that are essential goods [6:15] like you know when a hurricane comes [6:16] through not here but in North Carolina [6:18] where I'm from you know we have [6:20] hurricanes if the hurricane comes [6:21] through you know price gouging is [6:23] illegal so if water becomes scarce you [6:25] cannot charge more for it so the market [6:27] is uh not supposed to regulate um access [6:30] to water. So this is the potential uh [6:33] that you know could happen in the in the [6:36] u demand for water for AI data centers [6:38] for cooling that essentially you know [6:40] boil it away. Um you can you can talk [6:44] about the latest innovations maybe [6:46] later. Yeah. Um but um so that you [6:48] really what the AI data centers are [6:50] using they're using a common good for [6:51] their own profit. They're trying to have [6:53] market forces determining it and that's [6:56] illegitimate. So this is this is the the [6:58] kind of view the kind of position that [7:00] we got to think about. Um like what is [7:02] really the status of water? What are the [7:05] restrictions for using it for private [7:07] businesses? Um and then similar concerns [7:11] considerations also apply to energy and [7:13] the potential you know impact on the [7:14] environment global warming and all this [7:16] but but water I think is a very [7:18] plausible case for that it's really a [7:20] common good that uh should be uh not [7:23] completely controlled by market forces. [7:26] Great great some great points there and [7:28] I think I'll maybe come back to you in a [7:29] few here to to maybe expand on some of [7:31] the ethical components involved there as [7:33] well. But I think you hit on a key point [7:35] which is the fact that there is a lot of [7:37] narratives currently getting managed on [7:40] behalf of the industry and not all of [7:43] those are favorable for the industry. So [7:45] it's really really important that we're [7:46] getting out ahead of uh those uh those [7:50] narratives that are being managed around [7:53] the natural resource consumptions of the [7:55] data center market segment of AI [7:57] infrastructure. And so I guess Remy to [7:59] to come over to you maybe your your view [8:02] and your insights from a maybe an energy [8:04] sector challenge or perspective. [8:06] >> I'm not a water expert. I'm glad that [8:08] you you covered the water part of it. [8:10] But uh speaking of the common good and [8:12] you know you got to think as well about [8:13] power being a common good. So yes water [8:15] exists and you know we have rights and [8:17] and and so on but also you know we have [8:21] infrastructure that rely on power. You [8:23] know hospitals you know um churches and [8:26] and so on you know our livelihood is [8:28] also dependent on on power and arguably [8:30] I would say that's also a common good [8:32] and it's being strained I would say with [8:35] the use of uh of power for data centers. [8:38] So I don't think people really grasp the [8:40] magnitude of the challenges ahead of us. [8:42] And I believe when you said it's$7 [8:44] trillion dollars, that's not a that's [8:46] not a little bucket of money. That's a [8:48] lot. That's like major economies [8:50] combined together the investment in in [8:53] in energy. I was speaking two days ago [8:56] with uh Duke Energy, the the CFO of Duke [8:58] Energy. For the first time ever, Duke [9:00] Energy is going to invest $200 billion [9:03] in infrastructure in the Carolinas. In [9:06] the Carolinas only. $200 billion. I [9:08] remember that in the past when they [9:10] would invest $5 billion, we have like a [9:12] big deal. Look, five billion $200 [9:14] billion. What is the source? Data [9:17] centers. They're building data centers [9:19] again and again. And the same story that [9:21] I hear from across in the industry um [9:24] just look at it from a perspective. It [9:26] took us about 150 years to build the [9:30] power the power grid that we have today. [9:32] We need in the next 5 to 10 years to [9:36] build almost the same size of a grid [9:38] that we built in 150 plus years. So I [9:43] don't know if we are ready for that. I [9:44] don't know we have enough money to to do [9:46] that. And the challenge is major in [9:48] front of us and it's going to take power [9:51] or the the use of power from other [9:53] sources. you know you have to uh diverge [9:56] that to data centers to ensure that you [9:59] know they are operating and also how the [10:02] the public perceive this is an important [10:04] perspective. I was talking to somebody [10:06] over over the lunch today and I said how [10:08] long it took you actually to wait for an [10:10] elevator to come. What what is your you [10:13] know patience with that and they [10:15] responded to me about 15 seconds. I said [10:18] the the statistic says 20 seconds. [10:20] people are willing to wait 20 seconds [10:22] before that the elevator arrives and say [10:24] okay I'm not frustrated how long you are [10:27] waiting for the your Google search or [10:30] when you are going on grock or others [10:32] before the answer will come back what is [10:34] your your your patience with it I [10:36] >> think it's instantaneous [10:37] >> instantaneous great instantaneous which [10:40] means that we have to build data centers [10:42] operating instantaneously [10:44] all around the country because we are [10:46] not us as a public we're not ready to [10:49] say Okay, slow it down, you know, do [10:52] inferencing and do actually uh uh [10:54] results to me in in about 2 minutes from [10:57] now or 30 second now or so go on. So [11:00] that demand is not going to be we're not [11:02] going to be able to shut that demand, [11:04] reduce the demand or change it. It's a [11:07] major challenge again from the energy [11:09] industry. Everybody's celebrating tons [11:11] of money is going to come to energy. But [11:13] just think about it too. We don't need [11:15] to build everything to burden the planet [11:17] just to run AI. I'm just gonna leave [11:19] with that. I think it's really big deal [11:21] for us to think about. [11:22] >> Some good points. I guess one real quick [11:24] followup for you. Do you think we're [11:26] going to be able to build fast enough to [11:27] keep up with demand and what are the [11:30] ramifications if we don't? Does that [11:31] mean we're going to start seeing data [11:34] center and AI infrastructure relocating [11:36] to the other regions and what are some [11:37] potential implications there? [11:39] >> I don't know how much of the audience [11:40] have talked about that, but the simple [11:42] answer [11:43] >> in 45 seconds at the most. [11:45] >> 45. No, we don't have we don't have [11:47] actually the time. It takes on average [11:49] seven years to build a uh natural gas [11:52] fired plant. Okay. It takes about a [11:55] decade to build a nuclear plant. We need [11:58] about 25 nuclear plant from here to [12:01] 2030. How many we have start building [12:03] guys today? Zero. None. We have none. [12:07] We're not ready. So what's going to [12:09] happen is we going to and recently we [12:12] are also reducing uh generation from [12:14] renewable energy as well. So the [12:17] challenge is real and you know whoever [12:19] win that competition to some extent [12:21] right now we are competing for AI [12:23] nationally and internationally other [12:25] countries they have excess energy and [12:27] whoever have excess energy is going to [12:28] have the data centers they're going to [12:30] have the model they're going to build [12:31] it. So I'm not pessimistic but I don't [12:34] think that we have the time or um [12:36] >> absolutely a whole separate section on [12:38] that topic. Right. So no really great [12:40] insights and I guess Thomas coming back [12:42] to you because I wanted to maybe peel [12:43] back a little bit further on kind of AI [12:47] and and resource demand because it can [12:49] be a little bit of a prickly and thorny [12:50] issue specifically as we continue to [12:53] push for greater compute to support AI [12:56] continued innovation. Um that that's one [12:59] thing, but when we when that push for [13:01] greater compute potentially results in [13:04] or results in strains on natural [13:06] resources and communities, [13:09] are there ethical considerations that [13:11] should guide how uh resource allocation [13:14] is made for AI? [13:14] >> Right. Abs. Absolutely. Right. Because [13:16] um like what what Remy was just saying, [13:18] so it's uh maybe surprising but but true [13:20] that it's easier to build the data [13:22] center than the energy that supplies the [13:24] data center and um and faster. So but so [13:27] the data centers will come. So that [13:29] means they will draw the energy from [13:31] what already exists and that means they [13:33] will take it away from the from the [13:35] other one you know it will the energy [13:37] will be a commodity the prices will go [13:39] up um it will be more scarce um and the [13:42] fastest ones to build are likely going [13:44] to be like gas powered plants and things [13:46] like that that um are more polluting [13:48] right so so there would be very little [13:50] problem with energy if the AI companies [13:53] would build you know solar panel farms [13:55] along with the data centers at the same [13:57] time and just use that energy, nobody [13:59] would complain, right? It would be [14:00] great. You would increase the capacity [14:02] of the energy production in an [14:04] environmentally sustainable way. There [14:06] would be no problem. But that's not [14:08] going to happen. That's that and that's [14:09] why we have a problem, right? Because [14:11] there will be increased pollution. There [14:12] will be increased uh demand on energy. [14:15] Um and water will be the same thing. I [14:17] mean water, you can't just build solar [14:19] panels for water, I guess. But uh but [14:21] let's say deselination plants, right? if [14:23] they they built solar panels, [14:25] desalination plants, and the data center [14:27] all at the same time, problem solved. [14:30] Um, but because that's not going to [14:31] happen, there is going to be a problem. [14:33] And the problem is tied to that, you [14:35] know, that there will be a draw on [14:38] something that arguably is um a [14:40] community res resource that um will [14:43] become a commodity more and more and [14:45] will be taken away from those who taken [14:48] away in a sense that will be that will [14:49] priced out, the water will be scarce and [14:51] so on and so forth. Um, so there are [14:53] definitely serious problems here. [14:54] >> Some interesting points you raise and I [14:56] think that we we're starting to see some [14:58] exploration about that in in outside of [15:00] the US in some regions like the Middle [15:02] East where where water scarcity is real [15:03] thing where they do have the resources [15:05] and the land and available for renewable [15:07] powers where del and solar um power is [15:12] what they're leveraging for the data [15:14] centers of the future in terms of their [15:15] designs. Now, now Maggie, coming to you [15:17] because I I don't want it to all be doom [15:19] and gloom because I think there is a a [15:21] circularity flip to this this discussion [15:25] where AI in itself can also be part of [15:27] the the solution to address resource [15:29] management and and optimization and and [15:32] so I wanted to get your feedback as to [15:33] what are some of the most promising [15:34] areas you're seeing with respect to AI's [15:37] ability to reduce water, energy, carbon [15:40] footprints. Yeah, I would say that the [15:42] biggest promising area is we've we've [15:43] had a lot of companies making promises [15:45] but not a lot of partnerships. And [15:47] lately we're getting a lot better at [15:48] partnering. So the University of Kansas [15:50] is partnering with EverG and we're [15:52] partnering with Panasonic and we've made [15:54] a commitment to being completely wind [15:57] dependent in 20 years um and to not use [16:00] EverG's resources. But we partnered with [16:02] an energy company and you would think [16:04] that that would sound kind of [16:05] ridiculous. They would want us to be [16:06] using their money and their energy. but [16:09] getting those partnerships of how can I [16:11] do this in a more efficient way. Um, [16:14] we've also been doing some computational [16:16] pieces. I have a student who I had in [16:18] middle school who just got a patent, [16:19] Joseph Nordling, um, just got a patent [16:21] on making every single computational [16:24] piece much smaller and faster so that it [16:26] doesn't take up the energy that it [16:28] should take up. We have Dr. Sumaya [16:30] Shamayi at the University of Kansas and [16:33] she has figured out ways to get [16:35] everything to go so fast that the [16:37] cooling that we're using with neural [16:39] cooling networks and all of the [16:40] different pieces that we're putting [16:42] together, the cooling can happen much [16:43] faster. So, we're looking at water and [16:45] we're looking at energy. Google, [16:48] Microsoft, they've gone to places like [16:49] Scotland so we can use their [16:51] >> their wind. They've gone to places like [16:53] the Netherlands or Norway so we can use [16:55] water. So finding locations where we can [16:59] find those renewable sources is [17:01] important, but also finding the minds [17:03] that can come up with other ideas [17:06] because there are some things that [17:07] actually happen more efficiently using [17:09] AI than traditional methods. Like when [17:11] we're talking about how do we decide [17:13] what times of day are the most energy [17:15] efficient times to use energy AI is [17:17] actually more accurate research is [17:19] showing then not just more accurate more [17:22] energy efficient in figuring out when [17:24] those best times were than actually [17:26] using the resources we used before. So [17:29] we're actually saving energy in that [17:30] aspect. Certain things like image [17:32] creation that happens very quickly with [17:36] AI. It uses less energy than it used to [17:38] take for people to create those images. [17:40] But now we're creating more and more and [17:42] more images. So is that actually going [17:44] to end up being saving energy? Well, [17:47] research shows it does for that one [17:48] time, but if I do it constantly, will [17:50] it? And so it's trying to get those [17:52] minds together to come up with different [17:54] solutions that I think is really [17:55] important. [17:56] >> I think you hit on some some great [17:57] pieces. I think for the the initial [17:59] stages of AI where we're doing a lot of [18:00] the training, we can go chase after the [18:02] renewable power. We can go chase after [18:04] locations where it's available, where [18:06] water sources are available to use for [18:08] cooling infrastructure. But when we [18:10] start transitioning to that inferencing [18:11] stage and having to come back to the [18:13] communities, that's when we're going to [18:15] need to see the development of some of [18:16] the approaches of just doing things [18:18] differently. And so I guess this this is [18:20] I think probably a important question. [18:22] So Remy or Thomas, I don't know if you [18:23] have anything that you wanted to add [18:25] otherwise we'll we'll move on. [18:27] >> I want to go back to some extent to to a [18:29] point that I'm not that pessimistic [18:32] even maybe I came across this way but [18:34] I'm trying to be realistic with with the [18:36] challenge ahead of us. Not not to say [18:38] that there's no solution. There are [18:40] solutions and you you just mentioned [18:41] actually human creativity has has shown [18:44] us you know throughout history that it's [18:47] always the solution for problems that we [18:48] have in front of us and you we have we [18:50] have some solution from from from today [18:52] that we can show I just want to give an [18:55] example that just shows you that what we [18:57] can do with AI that can help us as well [18:59] solve the problem that AI already [19:01] created for us to some extent. Um if [19:04] some of you recall about two months ago [19:07] the the Iberian Peninsula had actually [19:09] blackout and 50 million people went in [19:13] the black for I don't remember 22 hours [19:15] or something like that it took 27 second [19:19] from the fluctuation actually of power [19:22] until that the blackout 27 second [19:24] exactly I looked at the study so [19:27] basically there is no human could have [19:29] actually changed any configuration of [19:32] the grid in order to prevent the the [19:35] blackout. So when we use AI and we have [19:38] AI in our hands and the tools like this [19:40] can react to some disruption to the to [19:43] the grid instantaneously. That's what we [19:46] want that to to happen and help us [19:48] actually maintain the grid and not shed [19:51] resources and have the cascading effect [19:52] that happened in the in the blackout. So [19:54] that's a good use uh for for AI as well. [19:57] I mean many other good uses of of AI [19:59] too. Another example is and you talked [20:02] about that to some extent the the ge [20:04] geographic flexibility. So how you can [20:07] move load from one place to another [20:08] load. You have also the temporal [20:10] flexibility. You can move it to midnight [20:13] to uh you know later at night. We all do [20:15] it if you have electric vehicle. I [20:17] charge it at midnight not at uh 8:00 [20:19] p.m. or so. So you can have that [20:21] temporal flexibility that we can use and [20:23] AI can help us to optimize those [20:25] temporal slash uh uh geographic [20:28] flexibility to do it in the right time [20:30] in the right moment the right batch load [20:32] and some of those examples are real so [20:35] we're not talking about hey 50 years [20:36] from now or 10 years from now that would [20:38] be too late too but so some examples of [20:41] uh uh a program it's called DC flax is [20:44] going to be executed in about two months [20:46] from now or so on a data center in North [20:49] Carolina and and the Duke Energy [20:51] actually uh in Duke Energy service [20:53] territory I would um and um and uh the [20:56] Google team that that operate the data [20:58] center they are very interested to see [21:00] how they can flex their data center so [21:03] change the load change the the [21:05] configuration of of the load that they [21:07] need to take from the from the grid and [21:09] uh and and operate the data center in in [21:11] a better way. You can also move the load [21:14] from one location to another location. [21:16] So when you have overloaded grid or uh [21:19] you know uh in one location you can use [21:21] it somewhere else. So there's so many [21:22] ideas and that's also using AI to solve [21:25] the problem that AI already created for [21:27] us to some extent. [21:28] >> I think those that [21:29] >> there's some hope in the future. Please [21:30] don't get too pessimistic from me. [21:33] >> I I think those are some good examples. [21:34] I I I think we I've started to see some [21:36] of those studies as well wherein just by [21:38] load shifting at some of the data [21:40] centers there there may even is the [21:42] potential to be able to need to build [21:44] less data centers because some power is [21:46] is become more available Tom. [21:49] >> Yeah. So I mean just so it doesn't [21:52] really solve the problem. It it softens [21:54] it a little bit, right? So yeah. So but [21:56] then there's also the question about how [21:57] do we actually solve the problem like [21:58] fundamentally and that's still open as [22:02] as far as I can tell. So all these other [22:03] mitigation efforts you know efficiency [22:06] in inference time in training in uh [22:09] architectures and so on and so forth [22:11] they they soften it a little bit but the [22:13] the problem is is right in front of us [22:16] and it's facing us even with all the [22:18] softening 100% I agree with you soften [22:20] the problem doesn't solve it because the [22:21] magnitude of the problem is much much [22:23] larger but I want to give another [22:24] example if you don't mind is one of the [22:27] largest actually power consumption in [22:29] the US anybody want to guess what that [22:31] is cooling if we are freezing in this [22:33] room. You know why cooling is actually [22:36] the largest load in the US. Can we flex [22:41] cooling? Can we also use that? I mean [22:43] before again my perspective is we are [22:46] using the generation using resources [22:49] that they're natural like you know water [22:51] you know carbon and all of that but we [22:53] are creating power that are we consuming [22:56] that efficiently and effectively and we [22:58] need to think about all the above and [23:00] building new generation as well [23:02] efficient generation and so on sorry [23:04] Maggie if you want [23:04] >> no you're good I was going to say the [23:05] same thing is and can we come up with [23:07] creative ways for cooling other than [23:08] using energy like how are we looking at [23:11] the problems in front of us I think [23:13] that's where that collaboration comes [23:14] in. I don't think we're going to get [23:16] there, like you said, if we're not [23:17] collaborating. I think that there need [23:19] to be people in the education field who [23:20] are working with people in industry and [23:22] in the different companies and in the [23:23] nonprofit world to solve these problems [23:26] because it's going to take all of our [23:28] different minds working together because [23:30] I think very differently than each of [23:32] you do, which I appreciate the way you [23:33] think because when we were preparing for [23:35] the panel, it made me think differently. [23:37] It made me think, okay, I I understand [23:39] some of the concerns and I also [23:41] understand that some of the fixes like [23:43] the computation making it faster, [23:45] stronger, quicker solves a piece but not [23:48] the whole thing or using paper mills for [23:50] the water sources solves a section of [23:53] it. Yeah, we are using zero water or [23:55] waste water to cool. Those things are [23:58] possibilities, but do they really solve [24:01] the problem? Like you were saying, [24:02] Thomas, it it's like a band-aid. So, how [24:05] do we not just do a band-aid? [24:07] >> Absolutely. And we'll have you maybe [24:09] touch on that topic in a little bit more [24:10] depth, but I guess Remy, as we've kind [24:14] of transitioned into this optimization, [24:15] innovation discussion, and you touched [24:17] upon some examples of that that [24:19] circularity piece that we we discussed, [24:22] you know, for, you know, power for AI, [24:24] AI for power if if you will. Um, you [24:27] know, and some specific examples around [24:30] load shifting. [24:31] Are there any other examples that exist [24:34] where AI can be leveraged to improve the [24:37] energy sector's efficiency and [24:38] accelerate a transition or maybe more [24:41] renewables particularly in response to [24:43] some of this data center growth? [24:45] >> Yeah, I mean many examples exist today. [24:48] However, I don't think we have them up [24:50] to scale of and the magnitude and so [24:52] it's really uh speed and scale that you [24:54] need to think about here. I don't think [24:56] they are at the scale that we we need to [24:58] deploy them in the future. Um I spent [25:00] about five or six years in renewable [25:02] energy companies. So I'm very passionate [25:04] about renewable energy as well. Uh I [25:07] really believe that we need to uh uh to [25:09] build the tools and the capabilities [25:11] that can help us. You know the challenge [25:12] with renewable again I I I love the idea [25:14] Thomas said hey let's let's put solar [25:16] and let's operate data centers. It's [25:18] good in theory. It doesn't work in [25:20] practice and I'll tell you why. The [25:21] reason is these are intermittent you [25:23] know resources and data centers runs [25:25] 24/7. So you need to run the data center [25:28] most likely 247 because you're going to [25:30] be at 2 o'clock in the morning sending [25:32] an email to your grandma or whoever else [25:34] and you want that instantaneous you know [25:36] response at 2:00 in the morning there is [25:39] no sun and probably let's say you're [25:40] sitting in Kansas and that day there is [25:42] no wind so what we are relying on it's [25:45] going to be other sources of generation [25:47] that for that data specific data center [25:49] to operate so there's like ranges of [25:52] things that we need to put together in [25:54] order to to supplement you know the the [25:56] needs for for those data center to [25:58] operate 247 and when we're thinking [26:00] about how AI can help with that too is [26:03] you need to have efficient dispatching [26:06] of uh diverse resources you know battery [26:10] storage pump hydro and so go on if you [26:13] don't know what pump hydro is basically [26:14] you have two lakes you pump the the [26:16] water to the upper lake when you have [26:18] low low demand and when you have high [26:20] demand you you let the water cool from [26:22] the upper lake to the lower lake so you [26:24] generate through that so that's also [26:26] another form of storage is natural form [26:28] of storage. So it's not only u you know [26:31] uh we're talking about chemical storage [26:32] and and and things like that. So there [26:34] are so many ideas and we need to use the [26:37] the AI in many of these ideas to see how [26:39] we can optimize do demand response load [26:42] shifting and the example that I [26:44] mentioned about the 27 second you know [26:46] the IBM peninsula's you know blackout is [26:49] just an example of how AI can response [26:52] respond very quickly to to events like [26:54] that and that's what we need tools to [26:56] help us with this and we need to scale [26:58] them. Piggybacking on what Remy just [26:59] said, I think he also said in what he [27:01] said, but without the words, is [27:02] sustainability. Like we have to be able [27:04] to sustain all of those resources that [27:07] he just mentioned like, okay, we've got [27:08] the wind, we've got the water, we've got [27:09] these uh abilities, but how long will [27:12] each of those last? So, [27:14] >> not a not a either or, it's an and, you [27:16] know, and. So, that that consistent base [27:18] load is going to be important. And I [27:19] think prior to your your your comment [27:22] earlier, I think I misspoke. wasn't that [27:24] through load shifting they would be able [27:26] to reduce the amount of data centers is [27:27] is maybe reduce the need for additional [27:30] power generation and consumption by [27:31] freeing up that that additional load. Um [27:34] and so Maggie I guess coming coming to [27:37] you um talking about innovation [27:40] optimization you know beyond model [27:42] optimization what what innovative [27:44] approaches are you seeing within data [27:46] centers uh and cooling systems to reduce [27:48] water and energy how are companies [27:50] addressing the energy water nexus are [27:53] they addressing the energy and water [27:54] nexus how should they be addressing the [27:56] the energy uh water nexus um and more [28:00] specifically the corresponding [28:02] perceptions that are being created right [28:04] now about data centers being resource [28:06] drains in local communities. [28:08] >> Yeah, I think that all of that is a [28:10] really big question allin one. Um, but [28:13] things like [28:13] >> give you a little time. I know things [28:15] like liquid neural networks we've talked [28:17] about are are a good idea and they have [28:20] a place but I think we also need to [28:23] really consider things that um [28:27] optimization essentially we really need [28:29] to optimize what we are doing so that we [28:31] can be more efficient when we're using [28:34] any source of energy wherever it's [28:36] coming from and that still requires a [28:38] lot of collaborative thinking it [28:40] requires a lot of minds to come up with [28:42] new ways with AI because essentially [28:45] it's better at figuring out certain [28:46] things than we are faster than we can. [28:48] Um, so how can we optimize not just [28:52] optimize our um our collaboration with [28:56] each other to figure these things out to [28:58] figure out not just how can we [28:59] computationally go faster but to figure [29:01] out where are the resources where are [29:04] the places that have wind already and [29:06] that have water already and what do they [29:09] need in terms of their communities [29:11] because we're also putting we have one [29:12] going into Dodto Kansas um soon. I don't [29:15] know if that's part of yours or not, but [29:17] we have one coming in. And essentially, [29:19] that community does need jobs. That [29:22] community does have needs that this [29:24] center is filling. But then there's the [29:26] worry about the noise, the um the heat, [29:29] the And so there's both ends of people [29:32] really trying to figure out, is this [29:34] going to be helpful for the community? [29:36] Is it going to hurt the community? And [29:37] there's both sides saying, yes, we're [29:39] finally going to be able to keep all [29:40] these houses. We'll have jobs. No, we're [29:43] not going to be able to have [29:45] entertainment because we won't be able [29:46] to hear anymore, you know. So, there's a [29:48] lot back and forth, but we need to [29:50] involve people like that in these [29:52] discussions and um because they know [29:55] their community better than we do and [29:56] they can tell us specifically this is [29:59] how our community has these resources [30:01] available. These are the people in our [30:02] community doing research on this very [30:05] subject. Maybe you should consider [30:07] those. And then I do think we are going [30:08] to have to incentivize. I know that [30:10] that's a word that not a lot of people [30:12] want to hear, but I do think that not [30:14] everybody does the right thing because [30:16] it's the right thing. They do it because [30:17] there's an incentive for them to do it. [30:19] >> Saying there's alternative motives out [30:20] there sometimes. [30:22] >> So, I I really think it's all part part [30:24] of the plan. [30:25] >> No, I think you're starting to hit some [30:26] on some good pieces here with respect to [30:29] the collaboration. I want to open this [30:30] up to to everybody too because, you [30:33] know, I think we identified one gap. I [30:34] think that there's some synergies that [30:36] can be done with closer collaboration to [30:38] local with local communities where where [30:40] these deployments are are taking place. [30:43] I guess for for all of you are there any [30:46] other best practices additional [30:49] collaboration that is needed to be able [30:51] to advance sustainable AI further faster [30:55] and and are there or things [30:57] organizations can do to embed [30:58] sustainability into their development [31:01] life cycle? [31:04] I go [31:05] >> I I can I can tell you my perspective. [31:07] So we we work for APR. APRI is the [31:09] electric research uh uh power institute [31:12] and we are to some extent the research [31:14] arm for every utility company. So [31:16] utilities they have geography that they [31:19] cut as you know their own monopolies. So [31:22] nobody comes and just have another wire [31:24] line and you know in in your city and so [31:27] go on. But they don't do research, they [31:29] don't do R&D and they don't compete [31:32] against each other. But at least we [31:34] collaborate with each other across as I [31:36] said across the the states all in the uh [31:39] in the country and around the world as [31:40] well. The the study that I mentioned [31:42] about you know the Iberian peninsula is [31:44] just to learn from it. So how we cannot [31:46] have the same problem in the US and and [31:48] so go on. So we do a lot of [31:50] collaboration across the industry for [31:52] the research for power and for grid [31:54] optimization, grid management all the [31:57] way from nuclear to uh uh to [32:00] distribution to to some extent and [32:02] definitely we have a lot of work right [32:03] now in data centers because all the [32:06] money right now is going to data centers [32:07] and the build of data centers going to [32:09] double the size of the grid that we [32:10] have. So I agree with you collaboration [32:12] is going to be essential is actually in [32:14] our DNA in APR's DNA because we [32:17] collaborate when utility will bring [32:19] their own data and they give us their [32:20] data we work with them we do the [32:22] research and we come up with solutions [32:24] that we put out there just I want to [32:26] give a little bit uh an example if you [32:28] go to a website we just publish actually [32:30] data readiness that was uh a work [32:33] between uh PWC Accenture and you know [32:36] other uh and it's free you go and you [32:38] know actually data readiness for uh [32:41] that's available for other utilities [32:42] that they they contributed to to the [32:44] research and it's available for those [32:46] that did not contribute to research as [32:47] well. We believe there is a uh common [32:49] good and that's the mission of every [32:51] Abian nonfor-profit organization to help [32:54] of how we can solve these challenges and [32:56] make proposals and help as well some of [32:59] the I would say policym I'm not going to [33:01] get to the policy I don't think that's [33:02] something that we can take [33:03] >> that's coming next that's coming next [33:04] >> we could talk hours about that [33:05] >> we have one question on that [33:07] >> yeah but we will make recommendations [33:09] for adoption of these ideas and and we [33:12] need we need all the the collaboration [33:14] that we can get to to succeed with with [33:16] this challenge Perfect. And so I guess [33:19] perfect segue and switching switching [33:21] gears a little bit from from going [33:22] through some of and discussing some of [33:24] the collaboration, the innovation, the [33:25] like to discussing a different type of [33:28] collaboration, policy collaboration, [33:30] collaboration with regulators um and how [33:33] that comes into play. And so this will [33:35] be the last question before I ask some [33:36] of uh your wrap-up comments. But I guess [33:39] starting maybe with with Thomas uh what [33:42] role do you believe regulators and [33:43] policy makers must play in advancing uh [33:47] in green AI? And maybe to incorporate a [33:50] little bit of a a twist from from your [33:52] perspective, is there are there ethical [33:54] responsibilities? Are there certain [33:55] other responsibilities that lie with [33:57] organizations and regulators on this [33:59] front? [34:00] >> Oh, absolutely. Yeah, of course. because [34:02] this is an issue that affects people's [34:04] lives profoundly. So, um, so if this [34:06] goes badly, you know, then we'll have a [34:09] lot more pollution, a lot less water, [34:10] and a lot of AI that maybe not very many [34:12] people really benefit from. I mean, it's [34:15] to be seen, right? So, if it goes badly, [34:16] that would be a very undesirable and [34:18] unethical outcome. Um, so regulators [34:21] must play a crucial role there. I mean, [34:23] ideally, somebody should solve this [34:25] problem and then regulate society such [34:28] that that solution will be implemented, [34:30] right? So um the problem is that um [34:34] there isn't a lot of movement in that [34:35] direction right so that sort of for for [34:37] like a real solution there's mitigations [34:39] and stuff like that but a real solution [34:41] um is not on the table as far as I can [34:43] tell it's just um you know we're facing [34:46] this problem everybody knows it's coming [34:48] it's going to get worse um but there is [34:51] no real proper solution to it but so [34:53] that's what we need to find we need to [34:54] find a solution and then we need to [34:56] implement that solution through [34:58] regulation um or some other some other [35:01] form of you know some other close to [35:04] regular policy or some other you know [35:06] society level um solution implementation [35:09] makes sense. Uh Maggie any any comments [35:11] there? [35:12] >> Yeah I would also say it's up to all the [35:13] people in this room all the people in [35:15] this building all of the people who came [35:16] to this conference to help our policy [35:19] makers because they also don't know what [35:21] they don't know. So um they know that we [35:24] have these resources. They know the [35:25] resources are going away, but they don't [35:27] know how they could get certain [35:30] industries to collaborate with each [35:31] other, give incentives for taxes, give [35:33] incentives for buildings, structures. Um [35:35] because essentially building all of [35:37] these places to house the data, that's a [35:39] huge carbon issue. Um so there's a lot [35:42] of pieces to this that if we just [35:44] collaborated and some of the industries [35:46] came together and weren't all competing [35:47] against each other, if there was some [35:49] policy around that, it would save us [35:51] energy. [35:52] I think that that education is a is a [35:54] big piece. And what we've started to see [35:56] is we've we've started to bring together [35:58] data center operators and local [36:00] governmental officials to better [36:02] understand and and make their way into a [36:04] data center and understand what happens [36:06] in that uh not many windowed very large [36:10] uh and and sometimes unfortunately not [36:12] very beautiful uh large building. And so [36:14] I think that's really one of those [36:16] initial first steps to to really help [36:18] starting facilitate more meaningful [36:20] conversations of what solutions look [36:22] like moving forward. [36:23] >> I mean right so I mean I feel like that [36:25] um the co collaboration with local [36:28] communities is very important but often [36:30] of course um the data center providers [36:33] select the communities so they are [36:35] selecting ones that are you know will [36:37] give the most favorable terms. I mean [36:39] they selected not only because they have [36:40] excess energy and water to give but also [36:43] because they need the money and they [36:44] need the jobs and and so on and so [36:45] forth. So to really have a proper [36:47] solution I think it should be at a at a [36:49] national level. You got to think about [36:50] like what should be allowed, what should [36:52] not be allowed instead of simply um you [36:55] know picking sort of like your weakest [36:58] um community that you can impose the [37:00] terms on as as it often happens that you [37:02] know like rural communities that are in [37:04] economic despair get [37:06] >> Thomas I cannot agree more than what a [37:08] said [37:09] >> I think it's if anybody have read the [37:11] executive order that came about 10 days [37:13] ago [37:14] >> u I think it's called winning AI or [37:17] something like that I can't remember the [37:18] exact actually the executive [37:20] around AI and specifically it's you know [37:23] we are nonpartisan organization and you [37:25] know we should get out of actually [37:27] solving AI and power problem of being [37:30] what politics and what side of you are [37:32] at we should have just science-driven [37:35] you know research work focus on these [37:37] things because we need to solve problem [37:38] we don't need to fight each other about [37:40] this and you know people change their [37:41] mind and so go on but it's uh it's a [37:44] it's a action plan around AI that's the [37:46] name of the executive order and it's [37:48] it's good but when you look into the [37:50] details actually it excluded renewables [37:52] from what we should do in order to power [37:54] plants and it called for dispatchable [37:57] resources so I don't want to geek on you [37:59] guys dispatchable that mean that's [38:01] something that you can turn it on like [38:02] nuclear you can u uh you know that it's [38:06] going to be there 24/7 I can turn a a [38:08] gas turbine whenever I want I cannot [38:10] dispatch wind or solar they are you know [38:13] non-dispatchable resource they come when [38:15] they come and they go when they go so [38:16] the executive order said no we should [38:18] build it is dispatchable resources. [38:20] Again, it's it's driving from from one [38:22] source to one side to another and we're [38:25] not looking at all the above even that [38:26] we need all the above. So, again, it's [38:29] it's a perspective. [38:31] >> Very very interesting comments here. So, [38:33] in the the closing minutes here, uh you [38:35] know, as we we wrap up towards the end [38:37] of our discussion, I guess I'd like to [38:39] go through and and have each of you [38:42] provide your final thoughts. Um what is [38:44] one key takeaway that or call to action [38:47] that those here with us uh that should [38:50] be left with today? Um what is the most [38:53] important step or or rec recommendation [38:56] now that we can take to start addressing [38:58] overall carbon footprints sustainability [39:01] of of AI associated data centers and the [39:03] like. So Maggie, [39:05] >> yeah, I would say that the one step I [39:07] would recommend is if you have some [39:09] ideas, start sharing those ideas, no [39:11] matter how impossible or ridiculous they [39:14] seem because some of those ideas are the [39:15] ones that actually end up changing the [39:17] way the whole structure works. Um, and I [39:20] also think bringing those ideas to the [39:22] people who can make it happen because we [39:24] have lots of research at the University [39:26] of Kansas. We have engineers who can do [39:27] a certain piece, but we don't have [39:29] people who are building these centers. [39:32] We don't have people who understand all [39:33] of those pieces, all of the natural [39:36] resources we have in the entire world [39:38] where they're available, but there are [39:40] people who can do that. And so again, [39:42] that partnership collaboration is going [39:44] to be so important. [39:45] >> Absolutely, Thomas. [39:48] >> Um yeah, I mean my sense is that this [39:51] problem is kind of vastly under [39:54] underestimated for its severity um from [39:57] from all different sides. So, so right [39:58] now, uh the AI companies have all the [40:01] power because they have all the money [40:02] and stuff like that, but um so they [40:04] impose a lot of the terms how things go, [40:06] but um their own future is also in some [40:09] way at risk in this problem. And not [40:11] just that they'd run out of the [40:12] resources, but that you know, society [40:14] could turn against the whole project. [40:16] And so um so I would really think that [40:19] you know this problem needs to be [40:20] addressed much more urgently and at a at [40:24] a grander scale than I think it has been [40:26] both in the public debate as well as in [40:28] you know research collaboration and and [40:30] industry [40:32] and [40:33] >> I think something similar to what what [40:34] Thomas is mentioning there I think we [40:36] need to think about externalities [40:38] basically. So uh it's uh it's not free. [40:42] There's some some common goods that's [40:44] being consumed with with us using you [40:46] know the you know data centers. And [40:49] think about it you know when you are uh [40:51] sending that email at the 2:00 in the [40:53] morning. I'm sorry if I'm making that [40:54] that that example or that joke that [40:56] there's somebody sitting in a hospital [40:58] that they also plug to a uh a power [41:02] related device and that that power is [41:05] the same power that you use and they use [41:06] as well. So it's a common good. Just [41:08] think about that as well. [41:10] >> Yeah, some some really great comments. I [41:12] I know that when I've spoken to a lot of [41:14] folks that the thing that is really [41:16] resonating with me is that in this [41:18] space, everyone in this room, there's [41:20] there's there are a number of things [41:21] that everyone can compete on. They can [41:23] compete on for the same customers, same [41:25] land, same resources, whatever it may [41:27] be. Sustainability is not one of those [41:29] things that we should compete on. th [41:31] those are that is one key core tenant is [41:33] something that we need to come together [41:35] to help address and help solve together [41:38] because it is very very complex and it's [41:41] going to require a lot of smart minds [41:43] coming together to develop some really [41:44] really novel uh solutions to address. [41:46] So, I'd like to thank our panelists here [41:49] for their their valuable insights here [41:51] today. And and while we're kind of up [41:53] against time and don't have time for [41:54] questions, if you want to find uh the [41:57] team here, these panelists, I'm sure [41:59] they'll be happy to have some [42:00] conversations with you out here in the [42:01] hallway later on this uh this afternoon [42:04] or or even maybe tomorrow as as this [42:06] conference concludes. So, so thank you [42:08] for attending. Appreciate it. [42:13] Guys,