AI in Power Systems: From Insight to Deployment With Electric Power Research Institute

Ai4

AI-assisted summary of AI in Power Systems: From Insight to Deployment With Electric Power Research Institute — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background Published on YouTube by Ai4 on August 14, 2026, this presentation features Rémi Raphael, Chief AI Officer at the Electric Power Research Institute (EPRI). According to the description, the session explores the dual transformation shaping the energy sector: the surging power demands required to expand artificial intelligence and the opportunities to deploy AI to modernize, optimize, and secure power infrastructure.

## TL;DR The presentation frames the fundamental interdependence between artificial intelligence and electrical infrastructure, warning that exponential data center power demand is on a collision course with multi-decade grid construction timelines. To harness AI safely without risking catastrophic grid failure, EPRI launched the Open Power AI Consortium and the Safer AI initiative to develop domain-specific models, robust governance frameworks, and mandatory operational kill switches for control-room automation. The talk emphasizes ratepayer neutrality to ensure hyperscalers directly fund multitrillion-dollar grid upgrades rather than shifting costs onto household consumers. Finally, addressing physical supply chain realities—such as seven-year lead times for natural gas turbines—the session highlights the need for pragmatic self-regulation, cross-sector alignment, and closed-loop cooling architectures.

## Atmosphere & style The session has a pragmatic, grounded, and cautionary atmosphere that balances technological enthusiasm with the unforgiving physical physics of bulk power grids. Speaking directly in an intimate conference room without a roaming microphone, Raphael delivers a reality check on AI scaling, using vivid warnings of 27-second regional blackouts and the absolute necessity of engineering hardware kill switches. The ensuing Q&A creates an engaging, candid exchange among industry practitioners grappling with trillion-dollar investments and multi-year supply chain gridlocks.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) The fundamental reality of modern computing is the interdependence between power and AI: there is no AI without power just as there is no AI without GPUs. - [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) EPRI operates as an independent, non-profit R&D organization funded by private industry with an annual research budget of roughly $500 million across ~45 countries. - [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) Unlike China, which maintains excess power generation capacity, the United States operates with virtually zero spare reserve margins, creating strict trade-offs for new data center loads. - [04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) A severe timeline mismatch exists between AI growth and grid infrastructure: data center power consumption (currently 2-4% of US electricity) is doubling or tripling in 4-5 years, while new power plants take 10 to 20 years to build. - [04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) Data center architectures require operational flexibility across three tiers: Macro (~1 GW hyperscale campuses), Micro (5-10 MW regional edge sites), and Nano (sub-kilowatt embedded edge devices). - [06:22](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=382s) The Open Power AI Consortium unites ~300 member organizations, including NVIDIA, Microsoft, and Amazon, to share data, validate energy use cases, and build domain-specific power engineering foundation models. - [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) EPRI established the S.A.F.E. framework (Safety, Accountability, Fairness, Explainability/Reliability) to prevent catastrophic hazards when deploying autonomous intelligence across zero-margin power systems. - [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) Electrical grids operate with extreme physical fragility on millisecond timescales, as demonstrated by an Iberian Peninsula transformer failure that blacked out 50 million people in just 27 seconds. - [11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) While utility executives initially resist autonomous grid management, AI deployment across control rooms is inevitable within 5 to 7 years, necessitating strict tiered guardrails and non-negotiable physical kill switches. - [13:56](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=836s) Significant semantic and cultural divides persist between tech and power industries: terms like 'AI alignment' baffle utility executives, and 'AI safety' is often misinterpreted as physical occupational safety. - [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) Modernizing the power grid to accommodate AI demands an estimated $2 trillion to $3 trillion in capital, which must be funded directly by tech hyperscalers under ratepayer-neutral structures rather than retail consumers. - [18:46](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1126s) Physical supply chain bottlenecks, such as a 7-year backlog for natural gas turbines, severely constrain grid expansion regardless of available capital.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) EPRI and the Strategic Interdependence of Power and AI

The fundamental reality underlying modern artificial intelligence is simple: "There's no AI without power, right? As much as there is no AI without GPUs." The Electric Power Research Institute (EPRI) operates as an independent, non-profit R&D organization dedicated to bridging the gap between theoretical research and real-world deployment across the global energy sector. With an annual research budget of approximately half a billion dollars ($500 million) funded entirely by private industry rather than government grants, EPRI serves utility members across roughly 45 countries—from China and South Africa to Canada and South America. Because EPRI does not compete in the commercial energy market, it serves as an industry-wide collaborative engine sharing research and best practices across the entire value chain, from nuclear safety and generation down to the distribution lines and local pole-top transformers that power everyday household devices.

The conversation around AI and energy centers on what EPRI terms the "AI duality." While society faces a once-in-a-generation opportunity to build the computational infrastructure required for knowledge generation, expanding computing power must be balanced against ecological imperatives. Deploying high-density compute clusters cannot come at the expense of burning the planet, depleting regional water resources, or degrading biodiversity—humanity must avoid walking blindly into the prologue of a science-fiction disaster film. Simultaneously, compute capacity represents a vital geopolitical asset; falling behind in energy infrastructure directly impairs national competitiveness.

```mermaid flowchart TD accTitle: The AI Duality in Power Systems subgraph DUALITY["The AI Duality"] P4AI["Power for AI Meeting Exponential Compute Demand"] AI4P["AI for Power Optimizing Grid Operations & Generation"] end

subgraph CHALLENGES["Infrastructure Bottlenecks"] DC["Data Centers Consume 2-4% Power (Doubling/Tripling in 4-5 Years)"] LEAD["Long Lead Times 10 yrs (Conventional) / 15-20 yrs (Nuclear)"] CAP["Zero US Excess Generation Capacity"] end

subgraph SAFE_DEPLOY["SAFE Deployment Framework"] S["Safety & Accountability"] F["Fairness & Equity"] E["Explainability & Reliability"] end

P4AI --> CHALLENGES CHALLENGES --> AI4P AI4P --> SAFE_DEPLOY ```

- **EPRI Global Footprint:** Non-profit R&D entity spanning ~45 countries, funding over $500M annually across nuclear, generation, transmission, and local distribution. - **Geopolitical Energy Realities:** Unlike China, which maintains significant excess generation capacity, the United States operates with virtually no surplus reserve margins, meaning any unmanaged increase in data center load forces painful trade-offs elsewhere on the grid.

---

[04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) Infrastructure Timelines and Data Center Classification

Meeting the rising power demands of artificial intelligence requires navigating massive disparities in industrial timelines. Data centers currently consume 2% to 4% of total U.S. electricity generation, and this figure is projected to double or triple within the next four to five years alone. In contrast, expanding the physical grid moves at a traditional industrial pace: permitting and constructing a conventional thermal power plant averages 10 years, while a modern nuclear plant takes between 15 and 20 years. Because the U.S. grid operates without spare capacity, new large-scale computing facilities cannot simply plug into existing margins without localized system upgrades or cannibalizing power from other sectors.

To manage this growth, the industry must develop operational flexibility across three distinct tiers of data center architecture:

---

[06:22](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=382s) The Open Power AI Consortium and Real-World Use Cases

To harness AI for power systems, EPRI established the **Open Power AI Consortium**—a research-driven collaborative network of roughly 300 member organizations, including technology giants such as NVIDIA, Microsoft, and Amazon. The consortium deliberately preserves a non-profit, pre-competitive research focus (analogous to the original foundational mandate of OpenAI) and is structured around four primary pillars:

1. Facilitating secure data sharing across traditionally siloed utility operators. 2. Curating and validating hundreds of real-world energy use cases. 3. Building domain-specific foundation models and LLMs tailored for high-voltage power engineering and nuclear operations. 4. Accelerating commercial adoption by matching industry operational problems with vetted startup solutions.

EPRI actively runs rapid-translation programs to commercialize these capabilities. At a recent power industry conference, 150 innovative technology companies delivered focused pitches matching functional AI solutions directly against EPRI's catalog of field use cases, accelerating the deployment of algorithms for asset health, intelligent renewable siting, and real-time load balancing.

---

[08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) AI Safety and Grid Fragility: The S.A.F.E. Framework

Deploying autonomous intelligence to manage high-consequence infrastructure presents severe operational hazards. Placing automated machine learning agents in control of a substation, power plant, or nuclear reactor introduces the catastrophic risk of adversarial manipulation, loss of human override control, or refusal to execute emergency shutdowns. To govern deployment safely, EPRI developed the **SAFE** framework alongside strict mandates for explainability and reliability:

- **Safety (S):** Hard boundaries ensuring autonomous agents cannot be subverted by threat actors, drift outside operating limits, or refuse shutdown commands. - **Accountability (A):** Traceable ownership and deterministic audit trails for all algorithmic decisions across grid dispatch. - **Fairness (F):** Algorithmic safeguards preventing socio-demographic discrimination or inequitable dynamic rate structures based on neighborhood income or location. - **Explainability:** Absolute transparency in operational trade-offs. The model must never make black-box dispatch decisions, such as cutting power to a critical hospital in order to divert electricity to an adjacent high-revenue commercial casino simply because the loss function was configured to maximize gross market revenue. - **Reliability:** Maintaining sub-second frequency equilibrium across interconnected synchronous grids.

Modern electrical grids are far more physically fragile than the public realizes. Grid balance is fundamentally a millisecond-by-millisecond frequency equation; any sudden disruption in generation or transmission can collapse an entire interconnection in seconds. A stark real-world illustration occurred in April on the Iberian Peninsula: a single transformer failure at a solar installation triggered an instantaneous cascading frequency collapse that completely severed power to 50 million people across Spain and Portugal—unfolding from initial failure to total regional blackout in just 27 seconds. Implementing AI across such ultra-low-latency, zero-margin infrastructure requires absolute reliability before autonomous controls can be entrusted with the grid.

[11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) Tiered Safety Frameworks and the Grid Kill Switch

In electrical power systems, generation and consumption must balance instantaneously; if supply and demand fall out of alignment, the entire physical system collapses. This fragile equilibrium becomes increasingly precarious when surplus generation is tight and massive new loads emerge without warning—such as sudden data center surges caused by end users impulsively generating images with tools like ChatGPT. To prevent catastrophic vulnerabilities, EPRI is spearheading the "Safer AI" initiative to establish tiered safety frameworks and guardrails mapped directly to operational risk.

```mermaid flowchart TD accTitle: Tiered AI Safety Framework in Power Utilities accDescription: Flow diagram showing three tiers of AI safety, escalating from customer service to control room deployment.

subgraph Tier1["Low Criticality: Customer Service"] A["Customer Service Chatbots Billing queries, generic triage"] A1["Baseline ethics, tone checks Fairness & accent non-discrimination"] end

subgraph Tier2["Moderate Criticality: Distribution"] B["Distribution-Level AI Local town management (~10,000 residents)"] B1["Substation automated balancing Strict operational guardrails"] end

subgraph Tier3["Critical: Bulk Power Grid"] C["Control Room Autonomous Grid Management Expected within 5–7 years"] C1["Continuous human-in/on-the-loop oversight Mandatory hardware 'Kill Switch'"] end

Tier1 --> Tier2 Tier2 --> Tier3 ```

The stringency of safety guardrails must scale with the physical consequences of an AI failure:

* **Customer Service Level:** Front-office billing assistance does not require complex safety mechanisms, but it must be ethical, robust, and free from algorithmic bias—such as mishandling a caller's inquiry simply because they speak with a foreign accent. * **Distribution System Level:** Managing power flows for a localized community of roughly 10,000 people requires significantly higher operational guardrails to prevent neighborhood-level outages. * **Bulk Transmission & Control Room Level:** While utility executives frequently insist they will never allow AI inside central control rooms, creeping automation makes its eventual arrival inevitable. It starts with minor utility chatbots, followed by incremental operational tasks, until autonomous agents permeate the system. Within the next 5 to 7 years, power grids worldwide will likely be managed directly by AI.

For deep control-room integration, high-level human oversight (human-in-the-loop and human-on-the-loop) is non-negotiable. Crucially, the engineering community must build a physical and architectural fail-safe. If the model misbehaves, operators must possess the unimpeded authority to execute a manual shutdown: "The day that I want to say, 'AI, shut up, I don't want you to be used anymore,' we have built that kill switch, and if we don't build that kill switch, shame on us."

---

[13:56](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=836s) Bridging the Semantic Divide and Industry Self-Regulation

Rather than waiting for federal interventions from Washington to impose rigid mandates, the energy industry is pursuing proactive self-regulation. EPRI is currently collaborating with regional regulatory bodies, state utility commissions, the North American Electric Reliability Corporation (NERC), and the Department of Energy (DOE) to establish native standards from within the sector.

A cultural and linguistic divide complicates these initiatives. When engaging utility CEOs and CIOs, AI practitioners discover that foundational machine learning concepts are frequently misunderstood: * Mentioning **"AI alignment"** prompts utility executives to look baffled, wondering what planet the speaker arrived from. * Uttering **"AI safety"** leads power engineers to assume the discussion concerns physical occupational hazards—such as electrical shocks or field workers sustaining physical injuries on high-voltage equipment—rather than model behavioral drift or unintended algorithmic consequences.

To address these disconnects across planning, operations, and asset management, EPRI publicly launched the **Safer AI** initiative. Backed by EPRI's global network of over 450 member utilities, the core launch coalition features leading US energy companies alongside major technology partners, hyperscalers, and systems integrators—including NVIDIA, Microsoft, SHI, and WWT (World Wide Technology)—while interfacing directly with NERC's regulatory framework.

---

[16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) Q&A: Infrastructure Capital, Ratepayer Neutrality, and Closed-Loop Cooling

Opening the floor for discussion in an intimate room without roaming microphones, the session addressed the practical barriers to scaling AI infrastructure.

```mermaid flowchart LR accTitle: Data Center Grid Upgrade Capital Allocation accDescription: Flowchart illustrating ratepayer neutrality in multi-trillion dollar grid upgrades.

DC["Tech Hyperscalers & Data Centers Massive load additions"] -->|"Directly fund network upgrades"| UP["Grid Capital Investments $2T to $3T required globally"] UP -->|"Protects residential customers"| RP["Protected Retail Ratepayers Seniors & low-income households"] ```

An audience member asked how EPRI reconciles soaring capital expenditure with household affordability when power system expansions require multitrillion-dollar commitments. In response: * **The Multitrillion-Dollar Demand:** Expanding and modernizing the power grid to absorb doubling data center loads will require approximately **$2 trillion to $3 trillion** in capital investments over the coming years. * **Ratepayer Neutrality:** This capital cannot be extracted from ordinary retail customers or vulnerable grandmothers surviving on fixed Social Security checks. Big tech firms and hyperscale data center operators are negotiating directly with electric utilities to absorb network upgrade costs under strict **ratepayer-neutral** tariff structures. * **Water and Thermal Constraints:** When asked about cooling footprints, the speaker clarified that while he is a computer scientist rather than a hydrologist, modern chip designs—such as the NVIDIA Rubin platform and current liquid-cooled architectures—are transitioning toward **closed-loop liquid cooling**, dramatically curbing net water consumption.

---

[18:46](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1126s) Q&A: Supply Chain Bottlenecks vs. the Open Power AI Consortium

A software vendor working with utilities noted that despite ambitious automation goals, utilities cannot secure basic physical hardware, switchgear, or microgrid components. This operational logjam underscores that generation represents only half the challenge; the physical supply chain is strained across every asset class.

* **Severe Turbine Backlogs:** Procuring a standard combined-cycle natural gas generation turbine from heavy industrial manufacturers like GE Vernova or Siemens carries a staggering backlog. Securing an allocation—even with $500 million in available liquid capital—means waiting roughly **seven years** for equipment delivery. As the audience and speaker concurred, the industry faces severe shortages spanning money, materials, lead time, and specialized labor.

The session closed by delineating the relationship between EPRI's broader collaborative ecosystem and the specific Safer AI workstream:

Because EPRI operates as a non-profit, this dedicated workstream relies on targeted capital from participating sponsors, who in turn help sculpt the regulatory guardrails and secure early access to the foundational research before the session adjourned.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) There's no AI without power, right? As much as there is no AI without GPUs. — Rémi Raphael

> [04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) Data centers currently consume 2% to 4% of total U.S. electricity generation, and this figure is projected to double or triple within the next four to five years alone. — Rémi Raphael

> [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) The model must never make black-box dispatch decisions, such as cutting power to a critical hospital in order to divert electricity to an adjacent high-revenue commercial casino simply because the loss function was configured to maximize gross market revenue. — Rémi Raphael

> [11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) Within the next 5 to 7 years, power grids worldwide will likely be managed directly by AI. — Rémi Raphael

> [11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) The day that I want to say, 'AI, shut up, I don't want you to be used anymore,' we have built that kill switch, and if we don't build that kill switch, shame on us. — Rémi Raphael

> [13:56](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=836s) AI practitioners discover that foundational machine learning concepts are frequently misunderstood — Rémi Raphael

> [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) Expanding and modernizing the power grid to absorb doubling data center loads will require approximately $2 trillion to $3 trillion in capital investments over the coming years. — Rémi Raphael

> [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) Big tech firms and hyperscale data center operators are negotiating directly with electric utilities to absorb network upgrade costs under strict ratepayer-neutral tariff structures. — Rémi Raphael

> [18:46](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1126s) Securing an allocation—even with $500 million in available liquid capital—means waiting roughly seven years for equipment delivery. — Rémi Raphael

## Mentioned Amazon, Department of Energy (DOE), Electric Power Research Institute (EPRI), GE Vernova, Microsoft, North American Electric Reliability Corporation (NERC), NVIDIA, NVIDIA Rubin, Open Power AI Consortium, OpenAI, Rémi Raphael, Safer AI, SHI, Siemens, World Wide Technology (WWT)

## Video description Rémi Raphael - Chief AI Officer, Electric Power Research Institute (EPRI)

This presentation explores the dual transformation shaping the future of the energy sector: the growing demand for reliable, scalable power to support the rapid expansion of AI, and the opportunity to apply AI to modernize, optimize, and secure power systems. It highlights how energy infrastructure is becoming a critical enabler of digital innovation, while at the same time AI is unlocking new capabilities across grid operations, asset management, planning, and resilience. The purpose of the presentation is to frame this interdependence as a strategic imperative, encouraging leaders to think holistically about how to invest in infrastructure, data, and talent to both meet emerging energy demands and harness AI as a force multiplier for operational performance and long-term system transformation.

[Source](https://www.youtube.com/watch?v=s7cqDwk9R8o)

中文简报

电力系统中的AI：携手EPRI从洞察迈向部署本次演讲阐明了人工智能与电力基础设施之间根本的相互依存关系，并警告称，数据中心用电需求的指数级增长正与长达数十年周期的电网建设时间线产生冲突。为在避免引发灾难性电网故障风险的同时安全运用AI，EPRI发起了Open Power AI联盟和Safer AI倡议，以开发领域专用模型、稳健的治理框架以及用于控制室自动化的强制性紧急切断机制。演讲强调了保持电费纳税人中立性，以确保超大规模厂商直接出资承担数万亿美元的电网升级改造，而非将成本转嫁给居民消费者。最后，针对天然气轮机长达七年的交付周期等物理供应链现实，会议强调了开展务实行业自律、促进跨领域协同以及采用闭式循环冷却架构的必要性。电力与 AI 呈现出深度的战略相互依存关系，但二者面临严峻的时间线错配：数据中心用电量预计将在 4 到 5 年内翻倍或翻三倍，而传统电厂的建设需要 10 年，核电站更是长达 15 到 20 年。 (0:04) 美国电网几乎在零备用容量的极限状态下运行，而数据中心架构需涵盖约 1 GW 的宏观超级园区、5 至 10 MW 的微观边缘站点以及设备级纳米级边缘三层灵活性。 (4:51) EPRI 联合 NVIDIA、微软和亚马逊等约 300 家机构成立 Open Power AI 联盟，推动跨电力企业的数据共享并构建电力与核能领域的专用基础模型。 (6:22) 由于现代电网以毫秒级频率平衡运行且极度脆弱（如伊比利亚半岛单台变压器故障在 27 秒内导致 5000 万人断电），EPRI 制定了 S.A.F.E. 框架并强调未来 AI 进驻控制室必须配备物理层面的强制断电开关（Kill Switch）。 (8:01) 电网吸纳 AI 负荷将面临数万亿美元的升级投资，各方正推进由科技巨头承担成本的费率中立结构，同时行业还受到燃气轮机等关键硬件长达约 7 年交付周期的严重供应链制约。 (16:17)

Reading summary

According to the Electric Power Research Institute (EPRI), the rapid expansion of artificial intelligence is colliding with long industrial lead times and zero spare capacity on the U.S. electrical grid. EPRI argues that meeting this demand requires multi-trillion-dollar investments insulated from retail ratepayers, coupled with strict safety architectures and operational kill switches before autonomous models can enter utility control rooms. Disparity Between Compute Growth and Grid Timelines (4:51) Data centers presently consume between 2% and 4% of total U.S. electricity generation, a share that EPRI projects will double or triple within four to five years. In contrast, physical utility expansion operates on protracted industrial cycles: permitting and building a conventional thermal power plant averages 10 years, while a commercial nuclear plant takes 15 to 20 years. Because the U.S. power grid operates without surplus reserve margins—unlike China's grid—adding large macro data centers (which can require up to 1 GW, equivalent to a full nuclear plant's output) risks displacing existing loads. Modernizing and expanding the grid to accommodate this compute surge will require an estimated $2 trillion to $3 trillion in global capital investment. To ensure these expenses do not burden residential consumers or low-income households on fixed incomes, tech hyperscalers are negotiating ratepayer-neutral tariff structures to absorb network upgrade costs directly. In parallel, hyperscalers are adopting closed-loop liquid cooling systems to mitigate localized water resource depletion. Pre-Competitive Collaboration and Domain-Specific Foundation Models (6:22) To adapt machine learning for power systems, EPRI established the Open Power AI Consortium, a pre-competitive research initiative with roughly 300 member organizations, including NVIDIA, Microsoft, and Amazon. Operating backed by EPRI's $500 million annual industry-funded budget, the consortium focuses on four structural pillars: establishing secure utility data sharing, curating field use cases, accelerating commercial vendor adoption, and training domain-specific foundation models tailored to power systems and nuclear engineering. To translate models into operational practice, the consortium pairs utility challenges with external tech solutions. At a recent industry conference, 150 technology vendors delivered pitches directly aligned with EPRI's catalog of field use cases, advancing automated solutions for high-voltage asset health, renewable energy siting, and real-time local load balancing. Grid Fragility and the SAFE Governance Mandate (8:01) Bulk electric power grids rely on sub-second frequency equilibrium, making them far more physically vulnerable than typical enterprise IT environments. EPRI points to an incident on the Iberian Peninsula, where a single solar facility transformer failure triggered a cascading regional frequency collapse across Spain and Portugal, leaving 50 million people without power within 27 seconds. Deploying autonomous AI across such low-latency infrastructure introduces severe risks of model drift or operational failure. To manage these risks, EPRI established the S.A.F.E. deployment framework, demanding Safety against adversarial compromise, traceable Accountability for dispatch actions, and Fairness to prevent socio-demographic discrimination in dynamic utility rates. The framework also mandates strict Explainability to prevent black-box algorithms from making hazardous operational trade-offs, such as cutting power to a critical hospital to direct load toward higher-revenue commercial customers. Tiered Operational Risk and the Grid Kill Switch (11:01) EPRI organizes AI deployment into a tiered safety framework calibrated to the physical consequences of failure. Low-criticality customer service chatbots require baseline fairness and accent non-discrimination. Moderate-criticality distribution-level systems managing localized towns of roughly 10,000 residents demand automated balancing guardrails to avert local blackouts. Although utility leadership remains hesitant to permit autonomous algorithms inside high-voltage transmission control rooms, EPRI forecasts that creeping operational automation will make autonomous grid dispatch likely within 5 to 7 years. For critical bulk power systems, EPRI insists that high-level human oversight and architectural hardware kill switches are mandatory, guaranteeing human operators unimpeded authority to shut down misbehaving autonomous models instantly. A persistent semantic divide remains a barrier to implementation: utility executives frequently confuse machine learning concepts like 'AI safety' with physical occupational hazards such as electrical shock. Moreover, software algorithms cannot circumvent physical supply chain constraints, where procuring standard combined-cycle natural gas turbines carries backlogs of roughly seven years regardless of capital availability.

精炼阅读摘要

美国电力科学研究院（EPRI）指出，人工智能的算力爆发正受制于严峻的电力基础设施物理瓶颈；在电网容量裕度极低且扩建周期长达数年的现实下，行业必须在推进跨领域协同的同时，通过分级治理与强制物理停机机制确保高危电网的安全运行。算力激增与电网建设周期的严重错配 (4:51) 现代人工智能发展构成了EPRI所称的“AI双重性”：高密度计算集群既是维系国家竞争力的核心资产，其扩张又必须受到生态承载力与水资源保护的刚性约束。目前美国数据中心消耗全美2%至4%的电力，且预计将在四至五年内翻倍甚至增至三倍。然而，与保有显著过剩发电容量的中国不同，美国电网几乎没有闲置备用裕度，新增算力负荷无法凭空接入。与此同时，电网的物理扩建周期极为漫长：新建传统火电厂审批与建设平均耗时10年，商用核电站需15至20年，而通用电气（GE Vernova）或西门子等重工业厂商的天然气轮机订单积压已长达约7年，单纯投入资本亦无法绕过供应链瓶颈。为缓解电网接入压力，行业正尝试将数据中心划分为三个层级推进灵活部署：一是规模约1吉瓦（GW）的宏观数据中心，其负荷相当于一整座商用核电站，必须配套专属发电设施与高压输电互联；二是5至10兆瓦（MW）级的微型数据中心，能耗相当于5,000至10,000户家庭，用于分布式任务处理和局部电网平衡；三是设备级或智能电表级的纳观数据中心，通过电表端GPU实现家庭边缘侧就地推理。电网瞬态脆弱性与SAFE安全治理框架 (8:01) 现代电网属于高物理脆弱性系统，供需平衡维系于毫秒级的频率稳定方程，任何突发扰动均可在数秒内造成互联电网的连锁崩溃。例如，此前伊比利亚半岛一处光伏电站的单个变压器发生故障，在短短27秒内便引发全域频率级联瓦解，导致西班牙与葡萄牙境内5,000万人遭遇全面大停电。随着ChatGPT等终端生成式AI带来难以预测的瞬时负荷激增，盲目引入自主算法控制变电站或核反应堆将带来遭受恶意篡改、模型漂移以及拒不执行停机指令等灾难性风险。针对关键基础设施的高危属性，EPRI提出了SAFE安全框架并辅以严格的可解释性与可靠性指标：在安全性（Safety）上设立硬性边界，杜绝自主代理脱离运行限制；在可问责性（Accountability）上建立确定性的调度决策审计追踪；在公平性（Fairness）上防止算法因社区收入水平形成差别定价或歧视；在可解释性（Explainability）上严禁将电网调度交由黑盒模型处理，防范算法因追求市场收益最大化而做出切断关键医院供电以优先输电给高收益商业机构的恶劣决策；在可靠性（Reliability）上确保亚秒级的同步频率稳定。分级风险防护与硬件物理停机开关 (11:01) EPRI牵头的“Safer AI”倡议主张安全防护规范必须与算法失误所导致的物理后果严格对应。在前端客户服务层，聊天机器人主要聚焦伦理合规与口音平等待遇，防范服务偏见；在配电系统层，涉及约10,000名居民的局部电网自动化调度需要设立严格的操作防线，防范社区级停电；而在主干输电网络与中央调度室层面，尽管电力公用事业管理层普遍对AI进驻持谨慎态度，但伴随自动化逐步渗透，EPRI预估5至7年内全球电网将广泛由AI直接参与运行。对于深入调度中枢的自主模型，系统必须保留高级别的人类监督机制（人机在环与人机监控）。工程界必须构建完全独立于软件层的物理硬件断电机制（Kill Switch）；一旦算法表现异常，调度人员必须具备毫无阻碍的绝对控制权手动终止AI运行，这是关键基础设施交由智能代理管理不可逾越的底线。行业自律协同与万亿扩建的费率中立 (16:17) 由于电力企业高管常将“AI安全性”误解为防触电等物理工伤，且对“AI对齐”等术语缺乏认知，EPRI正联合北美电力可靠性公司（NERC）、美国能源部（DOE）及英伟达、微软等科技巨头推进自律性标准。EPRI设立了由约300家机构构成的“开放电力AI联盟”（Open Power AI Consortium），聚焦数据共享、实际案例库验证与高压工程领域模型预研发；同时设立专门出资的“Safer AI”工作流，专项研制电网安全防护标准与治理框架。在成本与资源层面，为承载翻倍的AI电力负荷，全球电网现代化扩建预计需耗资2万亿至3万亿美元。为避免这笔巨额支出波及依赖固定社保养老金的普通家庭，超大规模数据中心与科技巨头正与电力公司协商，依据“费率中立”（Ratepayer-Neutral）原则由科技企业直接承担电网升级费用。针对高散热负荷引发的水资源忧虑，英伟达Rubin平台等新一代芯片架构正加速转向闭环液冷设计，以从工程源头遏制对区域水资源的消耗。 EPRI阐明了算力扩张与电力工业之间难以割裂的共生纽带：AI的规模化部署不仅受限于能源供给，还受制于重型发电设备长达数年的交付周期与电网毫秒级频率平衡的物理极限。尽管科技资本出资的“费率中立”模式与闭环液冷技术有助于缓解财务与生态阻力，但重工业供应链的天然滞后以及关键控制室中对硬件级物理停机开关的刚性依赖，决定了AI重塑电力系统的进程必须在严苛的工程自律与安全红线内审慎展开。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台上。内容为美国电力科学研究院（EPRI）首席 AI 官 Rémi Raphael 的主题演讲，主要探讨了能源领域正在发生的双重转型，即支撑 AI 算力爆发所需的电力基础设施扩建，以及利用 AI 现代化改造并保障电网系统的战略路径。 ## TL;DR 演讲深刻剖析了“AI 与电力的双重性（AI Duality）”：算力激增面临着全网缺乏冗余发电容量、常规电厂与关键发电设备动辄 7 至 20 年漫长交付周期的残酷现实，并提出通过“费率中立”由科技巨头承担升级成本。与此同时，利用 AI 反哺电网调度必须敬畏电力系统毫秒级平衡的极端脆弱性，绝不能让算法脱离工程安全底线。为此，EPRI 联合科技与电力产业发布了 Safer AI 倡议，力推在未来自动化调度中建立严格的“紧急断路开关（Kill Switch）”与行业自律治理标准。 ## Atmosphere & style 演讲风格务实沉稳且极富前瞻性，充盈着重工业工程师面对数字狂热时的理性质朴与警醒感。演讲者以风趣而尖锐的行话打破行业壁垒，从“不能为了 AI 烧毁地球”到西班牙 27 秒大停电的震撼警示，构建起一条从宏大愿景跌落至冰冷物理规律的情感弧线。聆听过程中既能感受到千载难逢的技术变革浪潮，又时刻伴随着设备排期长达数年、毫秒失衡即全网崩溃的现实紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) AI 与电力系统存在深度的双向共生关系（AI Duality）：AI 算力扩张极度依赖电力供应（Power for AI），而电力系统也急需 AI 技术来优化调度与提升效率（AI for Power）。 - [02:08](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=128s) 数据中心能耗目前已占全美用电的 2% 至 4%，预计未来 4 到 5 年内将翻 2 至 3 倍，而美国全网当前缺乏富余冗余电能，新增负荷面临挤占存量用电的风险。 - [02:08](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=128s) 电力基础设施建设周期极其漫长，常规发电厂建设需约 10 年，核电站建设需 15 至 20 年，难以迅速匹配 AI 算力的爆发式增长。 - [04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) 算力基础设施可划分为宏观（~1 GW）、微观（5–10 MW）和纳米（家庭边缘端）三级架构，电网需针对不同层级建立灵活的协同调节机制。 - [06:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=377s) EPRI 发起成立包含 NVIDIA、Microsoft 等 300 家成员单位的“开放电力 AI 联盟”，聚焦数据跨界共享、场景库建设及电力行业垂直大模型的开发落地。 - [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) EPRI 提出了针对电力工业的 SAFE(R) 安全评估框架，涵盖安全（Safety）、问责（Accountability）、公平（Fairness）、可解释（Explainability）与可靠性（Reliability）。 - [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) 电网系统具有严苛的毫秒级供需动态平衡特性，局部故障扰动在极端情况下可在数十秒内引发跨区域大面积电网瘫痪。 - [11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) AI 安全护栏机制需因地制宜分级建设，从客服端的防歧视与合规，到配电网的高容错，再到核心调度控制室的极高安全标准。 - [13:00](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=780s) 预计未来 5 到 7 年内 AI 将深度渗透全球电网核心调度室，底层系统必须保留人在回路（Human-in-the-loop）监督与物理/逻辑紧急断路开关（Kill Switch）。 - [14:21](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=861s) 电力管理层与前沿 AI 领域之间存在严重认知鸿沟，EPRI 正式联合科技巨头与监管机构推出 Safer AI 倡议以确立统一的工业安全规范。 - [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) 为满足 AI 需求未来电网需投入 2 至 3 万亿美元，行业正践行“费率中立”原则，确保升级改造成本由科技大厂兜底而非转嫁给普通居民。 - [18:02](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1082s) 电网大扩张面临全方位供应链危机，如订购联合循环天然气发电轮机的排队等待期已长达 7 年，同时面临设备物资与专业人才短缺。 - [20:14](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1214s) Open Power AI 属于涵盖广泛场景的大型产业联盟平台，而 Safer AI 则是其专注于电网 AI 安全标准、对齐规范与切断开关等命题的专项攻坚子倡议。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) EPRI 的使命与 AI 双重性法则演讲者以一句引人深思的行业洞察开场：“你我之间有个公开的秘密：没有电力就没有 AI，就像没有 GPU 就没有 AI 一样。”为了给后续的讨论提供背景，他首先介绍了美国电力科学研究院（Electric Power Research Institute，简称 EPRI）。EPRI 是一家非营利性的科研与产业转化机构，其核心使命绝非仅仅停留在发表理论论文，而是致力于将先进研究真正部署到现实世界的电力工业中。EPRI 的网络遍及全球约 45 个国家，从中国、南非、加拿大到南美洲均有业务布局；其研究纵深涵盖了从核电生成与核能安全，一直延伸到城市街道电线、变压器以及家庭充电插座后方整套复杂的配电网络。 EPRI 每年投入约 5 亿美元的研发经费。与许多依赖政府补贴的机构不同，这笔巨额预算完全由私营电力行业全额资助。由于 EPRI 在电力市场上保持中立、不与任何电力企业竞争，因而能够作为整个行业的协作中枢，将行业资金汇聚起来攻坚关键技术，并将研究成果开放共享给所有会员单位。然而，今天讨论的核心并非 EPRI 自身，而是“AI 与电力的双重性（AI Duality）”——一方面，AI 的爆发极度依赖庞大的电力供给（Power for AI）；另一方面，EPRI 的使命则是利用 AI 反哺电网，生产并调度更多、更高效的电力（AI for Power）。 - **全球覆盖范围**：业务拓展至全球约 45 个国家和地区。 - **年度研发预算**：每年约 5 亿美元（$500M），资金全部来自私营电力工业资助。 - **业务研究跨度**：从核电站安全运行覆盖至居民末端低压配电系统与变压器。 --- ### [02:08](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=128s) 电力驱动 AI：算力饥渴、地缘竞争与电网建设的现实瓶颈谈及“电力支撑 AI”，数据中心的战略意义不言而喻。AI 绝不仅仅是聊天机器人或个人生产力工具，更是一场划时代的“知识生成革命（Knowledge Generation）”。这是属于当代人千载难逢的基础设施建设机遇。但演讲者特别告诫，在加速构建数据中心基础设施的同时必须保持清醒与克制，绝不能为了发展 AI 而“把地球烧毁”，沦为科幻灾难片的前传；必须将水资源消耗、生物多样性保护以及生态环境放在核心位置。与此同时，AI 已经成为极具分量的地缘政治战略资产，甚至在某种程度上具有武器级别的竞争力，各主要大国都在争相抢占制高点。为了在 AI 竞赛中保持优势，美国必须快速扩充发电容量以承载激增的数据中心负荷。根据 EPRI 的测算，目前数据中心已消耗全美 2% 至 4% 的电力供应，而在接下来的 4 到 5 年内（并非十年后），这一能耗比例就将翻倍甚至翻三倍。然而，电力基础设施的建设周期极其漫长：在北美，新建一座普通发电厂平均需要耗时 10 年，而新建一座核电站则需要长达 15 到 20 年。更为严峻且鲜为人知的事实是，美国目前并不存在富余的冗余发电能力（Excess Energy）。这意味着，如果不尽快新建发电厂，向数据中心输送的每一度电都必须从其他现有产业或民生领域硬生生剥离出来；而与此同时，竞争对手（如中国）却拥有大量的富余能源储备，这构成了巨大的地缘基建挑战。 - **短期用电激增**：数据中心能耗目前占全美 2%–4%，预计 4–5 年内将翻 2 至 3 倍。 - **电站建设周期**：常规电厂建设需约 10 年，核电站建设长达 15–20 年。 - **能源供需现实**：美国全网缺乏富余发电冗余容量，算力增量面临直接挤占存量负荷的风险。 --- ### [04:51](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=291s) 数据中心三级架构与 AI 赋能电力系统针对支撑 AI 发展的算力基础设施，演讲者将数据中心划分为三个不同层级，并强调电网必须针对不同规模建立灵活的协同调节机制： ```mermaid flowchart TD accTitle: 数据中心三级分类架构与电网互动 subgraph DataCenterTiers["数据中心三级分类架构"] Macro["宏观数据中心 (Macro DC) 容量: ~1 GW (相当于1座核电站)"] Micro["微观数据中心 (Micro DC) 容量: 5–10 MW (相当于5,000–10,000户家庭)"] Nano["纳米/终端数据中心 (Nano DC) 部署于家庭智能电表旁 (边缘GPU本地推理)"] end subgraph PowerAILoop["AI 与电力双向赋能闭环"] PowerGen["电力系统供给 (Power for AI)"] --> DataCenterTiers DataCenterTiers --> AICompute["AI 算力与算法模型"] AICompute --> GridOptimization["AI 反哺电网 (AI for Power) 新能源选址 / 并网消纳 / 毫秒级动态平衡"] GridOptimization --> PowerGen end ``` 在理清算力分层的同时，议题迅速切换到双重性的另一侧——“AI 赋能电力（AI for Power）”。AI 技术能够为能源行业带来前所未有的效率提升，包括更科学的新能源场站选址（Siting）、可再生能源大规模并网消纳分析、全网实时动态平衡以及资产健康管理等诸多关键场景。 --- ### [06:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=377s) 开放电力 AI 联盟与行业模型落地为了加速 AI 在电力行业的落地，EPRI 发起创立了“开放电力 AI 联盟”（Open Power AI Consortium）。演讲者风趣地提醒观众：“千万别把它跟 OpenAI 搞混了，我们做的是纯粹的科研探索——就像当年的 OpenAI 一样，并且在可预见的未来绝不会转为营利机构。” 目前该联盟已汇聚了约 300 家成员单位，涵盖了包括 NVIDIA、Microsoft、Amazon 等在内的全球科技巨头与能源领军企业，聚焦于四大核心支柱： 1. **数据跨界共享**：打破电力行业内部以及电力与科技巨头之间的数据壁垒。 2. **应用场景库沉淀**：从产业实际痛点中提炼数百个经过验证的高价值落地用例。 3. **电力垂直大模型构建**：研发针对电力行业特定物理机理与工程标准的领域专用大语言模型（Domain-Specific LLMs）。 4. **技术转化与生态落地**：搭建供需对接平台。例如在大会前夕的电梯演讲专场中，有 150 家创新科技企业针对 EPRI 提出的电力用例展示了定制化解决方案，直接推动电力行业现场评估与采纳。 --- ### [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) 电力 AI 的 SAFE 安全框架与毫秒级电网脆弱性将 AI 引入关键基础设施必须直面极端安全挑战。演讲者严肃指出，绝不能将大型发电厂或核反应堆交由一个可能被第三方恶意操纵、遭遇对抗攻击或拒不执行停机指令的“失控 AI”来接管，否则后果将是灾难性的。为此，EPRI 构建了针对电力工业的 **SAFE（R）** 安全可信评估框架： | 维度 | 英文定义 | 电力系统核心要求与底线红线 | |:--- |:--- |:--- | | **S** | **Safety** (安全性) | 确保系统不被外部恶意劫持、不发生控制权丧失，具备硬性故障安全（Fail-Safe）停机能力。 | | **A** | **Accountability** (问责性) | 算法决策链路透明可追溯，明确自动化操作中的权责归属。 | | **F** | **Fairness** (公平性) | 杜绝算法歧视；严禁基于用户地域、社会背景进行差别化电费定价或歧视性供电配给。 | | **E** | **Explainability** (可解释性) | 决策逻辑必须可解释；绝不允许算法为了“追求收益最大化”做出切断隔壁医院供电、转而向高利润赌场输送电力的荒谬决策。 | | **R** | **Reliability** (可靠性) | 保障电网在毫秒级频率扰动下的系统级稳定性与抗干扰能力。 | 最后，演讲者揭示了电网运行鲜为人知的极端脆弱性：整个电力系统的稳定性全靠全网频率在毫秒级维度上严丝合缝地维持供需平衡，一旦频率发生剧烈震荡，整张电网可能瞬间崩溃。他引用了一个真实的警示案例：去年 4 月西班牙某光伏电站发生故障，起初仅仅是一台变压器出现异常扰动，随后连锁反应迅速席卷整个伊比利亚半岛，导致西班牙和葡萄牙全境电网解列瘫痪，近 5,000 万人瞬间陷入停电。从最初的一点故障扰动扩散到 5,000 万人失去电力，整个过程仅仅耗时 **27 秒**。这种瞬时蔓延的严苛物理特性，正是电力系统对 AI 安全与可靠性要求容不得半点差错的根本原因。 ### [11:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=661s) 电力供需平衡挑战与 Safer AI 的分级安全框架电力系统运行的核心在于发电与负荷之间微秒级的严格动态平衡，一旦供需失衡脱节，整个电网系统就会面临崩溃的风险。随着新型数据中心的爆发式增长，电力系统的备用容量被严重压缩，任何人在终端使用 ChatGPT 生成图像等突发推理请求，都会传导为电网负荷的瞬时波动。为了应对这种脆弱性，EPRI 启动了 Safer AI 倡议，旨在针对电力系统的差异化应用场景，确立不同等级的 AI 安全规范与护栏机制。安全护栏的建设需要因地制宜：对于公用事业公司的客户服务系统（如用户打电话查询昨日账单明细），并不需要部署最为严苛复杂的工业级安全系统，核心诉求在于符合商业伦理、防止算法偏见与非歧视性对待——例如不能因来电者带有特定口音就区别对待；然而，当 AI 模型被用于管理小镇（如 10,000 人口规模）的配电网侧时，系统就必须建立起更高强度的安全护栏与边界约束，确保局部电网不会因调度算法波动而发生供电故障。 ```mermaid flowchart TD accTitle: AI 安全防护分级体系 subgraph S1["低风险层：用户与服务端"] A["客户服务 / 账单查询"] -->|核心要求| B["伦理合规与防歧视 / 基础防护"] end subgraph S2["中风险层：边缘与配电网"] C["城镇配电网管理（如万人小镇）"] -->|核心要求| D["高等级安全护栏 / 局部容错机制"] end subgraph S3["极高风险层：电网调度核心"] E["中控室系统调度（Control Room）"] -->|核心要求| F["人在回路监督 + 物理/逻辑紧急断路开关"] end ``` ### [13:00](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=780s) 电网核心控制室的 AI 渗透与“紧急断路开关”（Kill Switch）在电网的核心调度控制室（Control Rooms），许多公用事业公司的高管最初都持排斥态度，明确表示绝不允许将 AI 引入关键调度环节。然而技术渗透往往遵循渐进式的演进路径——从辅助性的小型聊天机器人开始，逐步拓展到边缘业务，再延伸到核心流程。演讲者预测，在未来 5 到 7 年内，不仅是美国，全球范围内的大多数电网都将高度依赖 AI 系统进行自动化管理。伴随这一趋势，必须在系统底层建立起极高标准的人类介入与监督机制（Human-in-the-loop / Human-on-the-loop），并且必须为 AI 设立绝对可靠的物理或逻辑“紧急断路开关”（Kill Switch）。演讲者强调，行业必须具备在突发异常时一键将 AI 从电网中剥离并让其“闭嘴”（"shut up"）的硬性能力，“如果我们没有设计出这个切断开关，那就是我们的失职”。同时，EPRI 正与北美电力可靠性委员会（NERC）及美国能源部（DOE）展开对话，力求在华盛顿出台强制性外部指令之前推动电力行业的自我监管与标准先行。 - **核心观点**：电网调度引入 AI 是不可逆的趋势，预计 5-7 年内将普及全球。 - **治理原则**：坚持行业主导的自律监管体系，避免脱离行业实际的官僚化行政监管。 - **底线机制**：必须构建能够瞬间剥离 AI 调度权限的 Kill Switch，确保最终控制权在人手中。 ### [14:21](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=861s) 行业认知鸿沟与 Safer AI 倡议的正式发布在与电力行业的众多 CEO 和 CIO 沟通时，演讲者发现电力管理层与前沿 AI 领域之间存在巨大的认知代沟。当向电力高管提及“AI 对齐”（AI Alignment）时，对方往往一脸茫然、如同“看着外星人一般”；而提及“AI 安全”（AI Safety）时，电力系统管理人员甚至会误以为是在讨论现场工人的工伤安全（Safety）事故，而非算法模型的安全性与价值对齐。为了弥合这一认知鸿沟并建立统一的行业准则，EPRI 在演讲当天上午正式发布了 Safer AI 倡议的新闻公告。依托 EPRI 全球 450 多家成员单位的网络，该项目汇集了包括英伟达（NVIDIA）、微软（Microsoft）等科技巨头与超大规模云厂商（Hyperscalers），以及系统集成商 SHI、WWT 和监管主体 NERC。该联合体旨在共同定义并测试适用于电网全流程的 AI 治理与安全框架。 ```mermaid flowchart LR accTitle: Safer AI 协同生态系统 subgraph EPRI["EPRI 核心生态"] M["450+ 电力成员单位"] end subgraph Tech["科技与基础设施伙伴"] N["NVIDIA / Microsoft / 超大规模云厂商"] S["SHI / WWT 系统集成商"] end subgraph Reg["监管与行业自律"] R["NERC / DOE 政策对话"] end EPRI --> SaferAI["Safer AI 工业安全标准与框架"] Tech --> SaferAI Reg --> SaferAI ``` ### [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) 现场问答：2-3 万亿美元电网投资与“费率中立”保障现场观众首先提问：面对 AI 带来的巨大用电需求与资本支出缺口，从研究角度如何评估并保障各方利益相关者的用电“可负担性”（Affordability）？演讲者补充了一项关键数据：为了满足数据中心翻倍的用电需求，未来数年内电力基础设施需要进行高达 2 到 3 万亿美元的巨额资本投资。关于“这笔巨资由谁来买单”的问题，目前正在推进的原则是“费率承担者中立”（Rate-payer neutral）。数据中心运营商及科技巨头正与电力公司达成协议，由科技企业全额承担电网扩容与升级费用，绝不能让依赖有限社保金和微薄薪水生活的普通大众（"grandmas living on social security"）来承担因 AI 扩张造成的电网改造成本。 - **投资规模**：电网基础设施升级需投入约 20,000 亿至 30,000 亿美元。 - **成本分摊原则**：全面贯彻“费率中立”，电网升级改造成本由科技大厂兜底。 ### [18:02](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1082s) 现场问答：液冷闭环技术与长达七年的设备供应链危机随后有观众提问关于数据中心的水资源消耗问题。演讲者澄清自己是计算机科学家而非水资源专家，但指出当前硬件正朝着闭环液冷技术发展，例如英伟达后续架构（如 Rubin 等芯片）所采用的闭环液冷系统（Closed-loop liquid cooling），致力于在运行中实现循环利用、不额外消耗水资源。另一位来自软件公司的观众提出了严峻的供应链与物资短缺问题：许多公用事业公司连建设配电管理系统和微电网的基础物资都无法采购齐全，面对这种现实困境 AI 部署如何落地？演讲者完全赞同这一痛点，并举例指出：如果当前计划投资 5 亿美元建设一座联合循环天然气电厂（Combined cycle natural gas plant），前往通用电气（GE）或西门子（Siemens）订购发电燃气轮机，目前的排队等待周期长达整整 7 年。当前的电力大扩张面临的是资金、设备物资、交付周期和专业人才的全方位供应链瓶颈。 | 瓶颈维度 | 现状与核心制约 | |:--- |:--- | | **重型发电设备** | 联合循环天然气发电轮机（GE / Siemens）订货等待期长达 **7 年** | | **电网配电设备** | 配电系统物资与关键硬件严重短缺，影响微电网与基础设施铺设 | | **资源与冷却** | 传统风冷水耗高，依赖新一代芯片向闭环无耗水液冷技术演进 | | **要素短缺** | 资金、时间、工程技术人才与实物物料全维度紧缺 | ### [20:14](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1214s) 现场问答：Open Power AI 与 Safer AI 的架构关系及展望现场最后一个问题聚焦于 Open Power AI 联盟与 Safer AI 倡议之间的定位区别与任务分工。演讲者解释称，Open Power AI 是一个拥有 300 多家成员单位的大型产业联盟平台，涵盖了电力系统中各种各样的 AI 应用场景与泛化课题；而 Safer AI 则是在该大联盟框架下孵化出的一个专属攻坚子倡议。它由一部分愿意提供专项资金支持、深度参与治理架构设计的先锋成员组成，专门集中精力攻克电网 AI 安全标准、对齐规范与切断开关等核心命题。作为非营利性科研机构，EPRI 正通过这一模式筹集研发资金，率先出资并深度参与的合作伙伴将优先从科研成果中受益并共同定义未来的行业规范。最后，演讲者在致谢中结束了本次演讲。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=4s) 你我之间有个公开的秘密：没有电力就没有 AI，就像没有 GPU 就没有 AI 一样。 — Rémi Raphael > [02:08](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=128s) 绝不能为了发展 AI 而“把地球烧毁”，沦为科幻灾难片的前传；必须将水资源消耗、生物多样性保护以及生态环境放在核心位置。 — Rémi Raphael > [02:08](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=128s) 在北美，新建一座普通发电厂平均需要耗时 10 年，而新建一座核电站则需要长达 15 到 20 年。 — Rémi Raphael > [06:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=377s) 千万别把它跟 OpenAI 搞混了，我们做的是纯粹的科研探索——就像当年的 OpenAI 一样，并且在可预见的未来绝不会转为营利机构。 — Rémi Raphael > [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) 绝不能将大型发电厂或核反应堆交由一个可能被第三方恶意操纵、遭遇对抗攻击或拒不执行停机指令的“失控 AI”来接管，否则后果将是灾难性的。 — Rémi Raphael > [08:01](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=481s) 从最初的一点故障扰动扩散到 5,000 万人失去电力，整个过程仅仅耗时 27 秒。 — Rémi Raphael > [13:00](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=780s) 在未来 5 到 7 年内，不仅是美国，全球范围内的大多数电网都将高度依赖 AI 系统进行自动化管理。 — Rémi Raphael > [13:00](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=780s) 如果我们没有设计出这个切断开关，那就是我们的失职。 — Rémi Raphael > [16:17](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=977s) 绝不能让依赖有限社保金和微薄薪水生活的普通大众（grandmas living on social security）来承担因 AI 扩张造成的电网改造成本。 — Rémi Raphael > [18:02](https://www.youtube.com/watch?v=s7cqDwk9R8o&t=1082s) 如果当前计划投资 5 亿美元建设一座联合循环天然气电厂，前往通用电气或西门子订购发电燃气轮机，目前的排队等待周期长达整整 7 年。 — Rémi Raphael ## Mentioned Rémi Raphael, Electric Power Research Institute (EPRI), Open Power AI Consortium, Safer AI, NVIDIA, Microsoft, Amazon, SHI, WWT, NERC (北美电力可靠性委员会), DOE (美国能源部), OpenAI, ChatGPT, GE (通用电气), Siemens (西门子), Rubin ## Video description Rémi Raphael - Chief AI Officer, Electric Power Research Institute (EPRI) This presentation explores the dual transformation shaping the future of the energy sector: the growing demand for reliable, scalable power to support the rapid expansion of AI, and the opportunity to apply AI to modernize, optimize, and secure power systems. It highlights how energy infrastructure is becoming a critical enabler of digital innovation, while at the same time AI is unlocking new capabilities across grid operations, asset management, planning, and resilience. The purpose of the presentation is to frame this interdependence as a strategic imperative, encouraging leaders to think holistically about how to invest in infrastructure, data, and talent to both meet emerging energy demands and harness AI as a force multiplier for operational performance and long-term system transformation. [Source](https://www.youtube.com/watch?v=s7cqDwk9R8o)

Source quotes

[0:04] Rémi Raphael: 你我之间有个公开的秘密：没有电力就没有 AI，就像没有 GPU 就没有 AI 一样。 [2:08] Rémi Raphael: 绝不能为了发展 AI 而“把地球烧毁”，沦为科幻灾难片的前传；必须将水资源消耗、生物多样性保护以及生态环境放在核心位置。 [2:08] Rémi Raphael: 在北美，新建一座普通发电厂平均需要耗时 10 年，而新建一座核电站则需要长达 15 到 20 年。 [6:17] Rémi Raphael: 千万别把它跟 OpenAI 搞混了，我们做的是纯粹的科研探索——就像当年的 OpenAI 一样，并且在可预见的未来绝不会转为营利机构。 [8:01] Rémi Raphael: 绝不能将大型发电厂或核反应堆交由一个可能被第三方恶意操纵、遭遇对抗攻击或拒不执行停机指令的“失控 AI”来接管，否则后果将是灾难性的。 [8:01] Rémi Raphael: 从最初的一点故障扰动扩散到 5,000 万人失去电力，整个过程仅仅耗时 27 秒。 [13:00] Rémi Raphael: 在未来 5 到 7 年内，不仅是美国，全球范围内的大多数电网都将高度依赖 AI 系统进行自动化管理。 [13:00] Rémi Raphael: 如果我们没有设计出这个切断开关，那就是我们的失职。 [16:17] Rémi Raphael: 绝不能让依赖有限社保金和微薄薪水生活的普通大众（grandmas living on social security）来承担因 AI 扩张造成的电网改造成本。 [18:02] Rémi Raphael: 如果当前计划投资 5 亿美元建设一座联合循环天然气电厂，前往通用电气或西门子订购发电燃气轮机，目前的排队等待周期长达整整 7 年。

Source quotes

[0:04] Rémi Raphael: There's no AI without power, right? As much as there is no AI without GPUs. [4:51] Rémi Raphael: Data centers currently consume 2% to 4% of total U.S. electricity generation, and this figure is projected to double or triple within the next four to five years alone. [8:01] Rémi Raphael: The model must never make black-box dispatch decisions, such as cutting power to a critical hospital in order to divert electricity to an adjacent high-revenue commercial casino simply because the loss function was configured to maximize gross market revenue. [11:01] Rémi Raphael: Within the next 5 to 7 years, power grids worldwide will likely be managed directly by AI. [11:01] Rémi Raphael: The day that I want to say, 'AI, shut up, I don't want you to be used anymore,' we have built that kill switch, and if we don't build that kill switch, shame on us. [13:56] Rémi Raphael: AI practitioners discover that foundational machine learning concepts are frequently misunderstood [16:17] Rémi Raphael: Expanding and modernizing the power grid to absorb doubling data center loads will require approximately $2 trillion to $3 trillion in capital investments over the coming years. [16:17] Rémi Raphael: Big tech firms and hyperscale data center operators are negotiating directly with electric utilities to absorb network upgrade costs under strict ratepayer-neutral tariff structures. [18:46] Rémi Raphael: Securing an allocation—even with $500 million in available liquid capital—means waiting roughly seven years for equipment delivery.

Original transcript

[0:04] just kind of a secret between you and I. There's no AI without power, right? [0:08] As much as there is no AI without GPUs. So, let's let's share with you some of [0:15] things that I came here today to talk about. But let me talk about APRI. And if you have [0:20] not captured the title, APRI is the electric power research institute. [0:24] We are nonfor-profit organization that focus on uh research and deployment of research applications [0:31] in the real world. So we don't just write paper and um and have theoretical research but our [0:36] interest is to see how we can take this and put it in application with with power industry all [0:41] across. We have members around uh 45 countries. We are present all the way from China to South Africa [0:48] to uh Canada to our neighbors in South America and so on. Uh our research spans from nuclear [0:54] generation, nuclear safety all the way to uh power system distribution. And if you don't know what [1:00] that means, that means the wires that runs in your street. That means that little transformer [1:05] all the way that comes to your home when you plug actually uh uh your phone to charge. You don't [1:11] know what that powers comes from, but there all of the system behind it. We do the research around [1:15] all of that. That's what it's exactly about. Uh we spend about half half a billion dollar a year [1:21] $500 million in research and these are fund funds not from the government not they are all privately [1:27] funded through all the industry. So what we do we are the the entity or the institute that support [1:33] the industry together and the industry because we are not uh competing in the power industry. [1:39] So uh they fund our research and we we share that knowledge all across uh our members. [1:46] So uh I did not come here to talk about uh about appri that will be another story but uh I want to [1:52] talk about AI and AI for power. So I like to put this representation here about the AI duality and [1:57] as I said there is no AI without power and my job at the institute is to ensure that we can use AI [2:04] to produce and generate more power. So I'm going to start with talking about power for [2:09] AI and why it is important. So u obviously data centers are essential but data centers [2:16] and you know AI are essential not just for us to uh to chat and have chat bots and have you [2:22] know productivity tools are so important as well but this isn't important for knowledge generation [2:29] and we have once in generation I would say opportunity to uh to powering knowledge and [2:35] to power knowledge to create that infrastructure that's needed for us to uh to build data centers [2:40] and make them actually capable of u you know realizing the benefit that we have from AI [2:46] we should do it but we should be very cautious of how we should implement that so we should not [2:51] burn the planet just to generate AI you know I I don't want to be in prologue of one of those [2:56] science fiction movies and and all of us doing you know uh blindly following you know the the things [3:02] that happening and and destroying the planet uh at the end of the day so we need to do it with with [3:07] think about water thinking about biodiversity thinking our plan and things like that that [3:12] they are very essential. But at the same time, we need to ensure that we are building enough [3:16] power capacity in the US to compete in AI and right now as you know that uh AI is geopolitical, [3:24] I would say assets and and and AI on its own is to some extent could become a uh a uh a a weapon [3:32] on its own and you know countries are competing against that. So at the US we need to compete and [3:37] if we need to compete we need to build more data centers and power we need to bring more power to [3:43] AI and this is some of the estimates that you can see actually some of the research app we produce [3:48] to show how many terowatt hours that we need to generate. So today actually the the power sector [3:55] we use about 2 to 4% of our power generation for data centers and that's going to double or triple [4:02] double or triple in the next four to 5 years not in the next decade. We're talking in the [4:08] next couple of years. Do you know how long it takes to build a power plant on average? [4:14] 10 years. How long did it take to build a nuclear plant? 15 to 20 years. [4:19] It takes forever to do that. So we need to do it now. We need to go fast otherwise we're not we're [4:24] not we're not going to be able to compete in this space and I would say there something that not a [4:29] lot of people knows that but we do not have excess generation in the US. We don't have excess energy [4:36] and when we put more data centers in uh you know to to to absorb more energy that we have that [4:41] means we have to take energy from somewhere else if we not going to produce more energy. On the [4:45] opposite side China has a lot of excess energy. just keep this in the back of your mind. So if we [4:51] think about competition there uh data centers so I'm I'm talking about power and AI there's [4:57] three type of data centers I'm not going to go in details of these and there's a lot of flexibility [5:02] we can create around data centers so we have the macro data centers the the micro data centers [5:08] that they are more smaller micro data center five to 10 gawatt I'm sorry 5 to 10 megawatt a macro [5:14] data center is 1 gawatt generation one one gawatt is a nuclear plant uh size 5 to 10 is a 5 to 10 [5:22] actually uh thousand people's home consumption. That's kind of the mic uh the macros micro side [5:28] and the nano actually data centers these are could be in your home this next to your meter. [5:34] The thinking right now is people putting even uh GPUs on your meters and to help you actually [5:39] consume or or run actually some inferencing on your side and we need to create flexibility. So [5:44] how we should run these data centers to to operate uh uh not only in a good way but also efficiently. [5:51] I'm going to go very quickly on this. So we talked about power for AI and I can spend [5:55] two hours talking only about how we can power the AI revolution. But what I would like to talk to [6:00] you about is also how we can use AI to generate more power. So the opportunity that we can use AI [6:06] to find efficiency opportunity of where we should uh have better sighting you know renewable energy [6:12] integration how we can balance the system and many many application talking about application [6:17] I have a slide to to to to share about that but before getting there uh appri has created this [6:23] consortium it's called it's called open power AI consortium don't confuse that with uh open AI okay [6:30] it's open power AI consortium we are research and we focus on research. This is what open AI [6:35] used to be in the past as many of you probably know. So we're still doing research. We're not [6:40] we're not going for profit anytime soon or any in in the foreseeable f future. We're [6:45] about 300 members. We bring actually technology companies you know um uh the Nvidia of the world, [6:52] the Microsoft, Amazon and and and others working together to do four things. uh try to share data [6:59] across the power industry develop actually u some use cases and I will talk about use case [7:04] in a minute we do um uh as well we are building domain specific model uh using LLM and so go on [7:11] and please use the the QR code if you are interested to uh to learn more about open power AI [7:17] trying to move this so I talk about use cases we have collected hundreds of use cases in the [7:23] power industry and I know that's not easy to read it just for people can zoom on But you can see we [7:28] have collected these. This is one example of those use cases that we are implementing right now. Um I [7:34] I just actually uh drove here uh last night and we had actually an elevator pitch where we have [7:39] 150 companies that they showed uh application of some of these use cases and and found solutions [7:46] to show to the power industry. So we had a large conference for the power industry where they were [7:50] talking about this. So we brought those use cases to innovative companies and they're coming with [7:55] solutions and we brought those solutions to uh to the industry to start to start adopting them. [8:01] uh the one thing that I would like to spend in the next 10 minutes before I open for any question if [8:06] you guys have is related to AI safety and u as I said there is no AI without power and if we [8:14] are doing more using more AI to generate power that's a good thing but how can you ensure that [8:19] the AI they're using is safe the last thing I want to do is put an AI system running a power [8:25] plant or or running a nuclear plant and then that system that you are running is [8:31] it's it's capable to be manipulated by a third party or like an adversary [8:36] or it could actually do whatever it wants to do. I mean the last thing you want is a system to uh to [8:41] have loss of control or the system actually refused to shut down or something like this [8:47] that would be catastrophic in the in the power industry. So for adoption of AI [8:51] in the power industry, we trying to build this framework and our framework is focused on [8:56] we took the acronym of safe and we tried to broke it down. Uh S for safety, A for accountability, [9:03] F for fairness. I cannot have an AI system that actually treat people differently. Nobody want to [9:07] be treated differently. Say based on you know your background, where you live or neighborhood, I'm [9:13] going to charge you more money than somebody else because you can pay more of something like this. [9:18] That cannot we cannot build an AI system that does that in the power industry. explanability. [9:22] We cannot have a system that make a decision actually to shut down a hospital because you told [9:27] it actually that you know maximize for revenue and decided it's better to run a casino the next door. [9:32] Let's shut down the hospital. Let's divert the power to the casino. It makes more me more money, [9:36] right? We cannot have an AI system that does something like that. So that's also something [9:41] important for us in the power industry to implement the minute that we bring AI in uh [9:45] in our industry. And the last thing is related to reliability. So uh so essentially you have an [9:50] AI that you want to ensure that it's keeping the the grid reliable and um we had we had a [9:56] second off conversation we don't know of actually how the power industry or the power system is not [10:03] as I would say stable as everybody wants to believe and I'm telling you this as openly here [10:10] we are that this short at any moment from having a fluctuation of power I mean it's [10:14] old it's a frequency if you have any disruption of frequency the whole system could collapse [10:19] and we seen this last year. I want to give you just very quick anecdote if uh if if [10:24] people follows world news. So um last summer in uh last April actually in Spain there was a [10:31] a a a power outage. It started actually from one transformer in in one uh uh solar plant [10:37] and it propagated actually shut down power for 50 million people. the whole Iberian peninsula, [10:42] Spain and Portugal shut down and it took from the first moment that disruption happened and to the [10:48] 40 to 50 million people without power 27 seconds. That's all the time that it took. [10:55] So that's why power system are not perfect. It's instantaneously at the millisecond matching demand [11:01] and supply and if one of these goes out of whack of the other, the whole system will collapse. [11:07] So it's very essential that we build an AI system that ensure the reliability of the system [11:12] especially when you don't have extra generation to to support the system when you have new data [11:17] centers and somebody going and uh deciding I want to generate this image with uh chat [11:22] GPT or any other tools that you have there. So we cannot be that uh vulnerable to uh to these uh uh [11:28] fluctuations. So what we are trying to do with safer AI actually is build those uh capabilities [11:34] and and identify the requirements uh you know we need to have different level of uh [11:40] uh requirement to understand how AI system should be safe. So uh for some tools that [11:45] I would say they are only customer service when you call a utility say hey uh yesterday I [11:50] want to know about my bill what happened to my bill. I don't need to have the most sophisticated [11:55] AI system there. I don't need to have the most sophisticated actually AI safety associated with [12:00] that. I just want it to be ethical. I want it to be non-discriminatory with with people [12:05] calling in. I I have an accent so it's like oh this guy coming from somewhere else I can just [12:09] tell them whatever it is or uh right we want to be sure that uh the system is is safe from those [12:15] characteristic. But if you take a different system and you put it to manage just say a distribution [12:21] side a a small town you you want to have that in in in a higher safety and higher I would say uh [12:30] guardrail that should be created around it not necessarily everything but I mean this is a town [12:34] of 10,000 people what type of guardrail that you need to have you need to build this type of guard [12:39] rail especially on the distribute side but when you go all the way I'm going and put a system [12:44] that just say God forbid even that many executive at the utility sites telling me we don't want to [12:50] put AI in our control control rooms that's where you control actually the grid but trust me they [12:56] will come they will come sooner than later they will start with little chat bots here and there [13:00] then they're going to move it actually oh we do this here nicely what about we do this too and [13:04] then let's what about we do that it's like really kind of uh this slippery slope and I believe that [13:10] within the five to probably five to seven years you're going to see the power grid not only in us [13:16] mostly around the world will be man managed by AI. So how you want to ensure that the power grid is [13:22] behaving the right way and is doing the the right thing for you. So at that level you need high [13:26] level of actually human control human in the loop on the loop and you want to have I love to to call [13:33] this we need to design the kill switch of taking AI off the grid. Not a lot of people want to talk [13:41] about that but we have to have that ability. So the day that I want to say AI shut up I don't want [13:46] you to to to to to be used anymore. We have built that kill switch and if we don't build that kill [13:51] switch shame on us. That's what's really important for us. So we want to do it in our industry and we [13:56] don't want the regulators to come yet to build this for us. We're working actually with North [14:01] America regulatory commissions. We started as well talking with DOE about this this initiative. [14:07] So we want to ensure that we build it. We want to self-regulate as an industry compared to [14:11] uh somebody you know in Washington and others in DC decide actually that uh you need to do this [14:17] this and that. So we want to self-regulate and that's what's really essential for us. [14:21] So here I'm going to walk through this very quickly. We have talked with CIOS and CEOs of uh [14:27] uh most uh power industry leaders and most of them are trying to do something but they don't have the [14:34] full picture. They don't know what exactly. I mean when I mention the word AI alignment you [14:39] can see these guys looking at me like but which planet this guy come from? I mean the terminology [14:43] that we use in in the power industry. I say AI safety they think that I'm talking about somebody [14:47] being injured. That's like I'm not talking about somebody being injured. I'm talking about an AI [14:51] safety and and and alignment. So this is kind of another slide just to show the different aspects [14:59] of the power grid that needs to be uh u managed and regulated by this this type of framework. [15:05] And I will spend the last minute on uh on on our announcement and then and I open it for question [15:10] if anybody wants to ask question about that. So as of this morning we have a press release that [15:15] we announced that uh our safer AI initiative is actually uh is public and we have this [15:22] partnership today this part these partnership are growing as I mentioned EPRI had is has 450 plus [15:28] members around the world. These are only US companies represented and and uh the conversation [15:34] between only US company versus European company and Middle Eastern company and other being part of [15:39] it that's a different one. I will keep this to the side but we have also technology company Nvidia uh [15:44] Microsoft uh and and and other hyperscalers as well and and providers that they are engaged [15:50] with us uh SHI WT and NERK is uh is uh North America Electric Regulatory Commission. So this [15:59] is actually uh it's regulated actually by Congress but uh they have a responsibility to regulate the [16:05] power industry in in North America and we are working with them as well to be part of uh of [16:10] this initiative. So with that said I'm going to open for a couple of minutes for questions. [16:17] Absolutely. Let me see if I figure out how to do this one. Yeah. [16:21] I don't know. Do we have a microphone for people or you can shout. I mean it's a small [16:27] room and you know we we're glad that it's a small conversation. Yes, please go ahead. [16:30] Talk about safety reliability in the in the power uh sector. Obviously you sit at the [16:35] intersection of mult different stakeholders. How are you all thinking from an affordability lens [16:42] like it seems like hey there's a lot of gap people are going to invest capital and demand is growing. [16:49] How do you assess from a research standpoint, hey, what the affordability scene is going to look like [16:54] for different stakeholders? I I love what you're saying. Let's talk about affordability. And I did [16:58] not mention a key statistic that I should have mentioned when I showed actually the data center. [17:02] The investment in power industry or in the power system in the next couple of years. I [17:07] said we need to uh to to double actually the the demand from the data centers. [17:12] The number is about two to3 trillion dollars in the power infrastructure that need to be invested [17:19] and for this two to three trillion dollars who's going to pay for it is it you and I that we just [17:24] you know uh or is it the data centers or is it the technology company so some of the work that [17:30] is ongoing right now empowering AI is u uh most of I would say data centers and data center builder [17:37] working with the power companies to say it should be it should be uh rate payer neutral. So they [17:43] are going to pay for the network upgrade. Those you know two to3 trillion dollars is not going [17:48] to come from you know uh you know the grandmas that's living on actually uh social security [17:54] and you know limited paycheck it's going to come from these big companies that they are going to [17:58] to make those investment in order to upgrade the grid. But yeah very very good point. Thank you. [18:02] Oh I see a lot I'm sorry we start right here please. As loud as you can because we don't have [18:06] the microphone. So you have numbers on water. You mentioned water and land. water. Do you have any [18:13] I don't have any slide about water and I'm not a water expert. I'm a computer scientist [18:17] but definitely there is a lot of development around around water too and um I would say close [18:22] cycle data centers. I think um uh right now Nvidia microchip, Reuben and others are liquid cooled [18:29] but they are also Nvidia promising actually it's liquid cooled it could be run in in a closed loop [18:35] so basically no additional water reuse uh will be needed for that but I'm not a water expert [18:41] in no means please I'm sorry the person holding microphone then we'll come to you quickly yes [18:46] hey uh so my experience I deal with a lot of utilities uh we're a software company um [18:51] they are finding it already almost impossible to get the materials for the distribution management [18:58] when you talk about AI management and and and uh you know in the op centers and uh so [19:03] they're they're stuck they they want it they see the need especially with micro grids etc. um [19:10] what's the realistic like the the realistic situation? We they they can't get the materials [19:16] to do the work, right? And I and I listen to groups like WSP partnering with Maximo [19:21] and they're looking at all these different these things that they're working on. Um I [19:25] don't I don't I guess my my concern is it's not just generation, it's actually the grid and the [19:30] complexity and the material availability. The whole grid and the supply chain is a challenge [19:35] as well. And I totally agree with you. If you if you ask I want to build actually a combined cycle [19:40] you know natural gas plant in the next five years the wait list right now it's about seven. [19:46] So if you say I want to build one and I have a $500 million I want to spend in that you go to the [19:51] big company GE Seammens and others they're going to sell you in seven years I will give you that [19:55] generation turbine what you going to do from from now to then we need to do something else. [20:01] I wish I have more money so I can invest in that thing but I don't know. So go with your question [20:05] please. It's not just materials. It's it's money and time and people too, right? We just like all [20:09] the things. All the above. Absolutely. Will you contrast uh safer AI and OPAI? What what are [20:15] the com what are the two different missions or how do they compare to each other? So open open power [20:19] AI is a consortium where we bring everybody under the consortium and we have many many [20:24] initiative under it. So like the use cases and so go on. The safer AI is one of the initiative that [20:29] we found a smaller group from the big consortium the 300 plus members. We we found a smaller group [20:35] that's decided we want to be dedicated to focus on this and to work on safer AI. I mean we need [20:41] we are we are raising money to to to do that effort. We are non forprofit so but that doesn't [20:45] mean that we will work work for free right so we need to raise money in order to work on this [20:50] initiative and that's what we're doing actually with safer AI. So um and the people that they're [20:54] going to first spend their money and benefit from the result not only benefit from result they're [20:59] going to help us actually uh structure that will be the people that they are supporting that this [21:04] research at this point in time and I believe I have no more time here but really appreciate [21:09] you coming here this morning and I hope you enjoy the rest of your conference. Thank you everyone.