Leidos Deploys AI to Speed US Grid Expansion as Data Centers Surge

Will AI Destroy or Save the Power Grid? The Race to Design the Grid of the Future With Leidos — Ai4

Ai4

Specialized AI engineering tools must automate transmission and substation design to overcome crippling labor deficits as surging data center demand threatens a power grid backing 85% of US GDP.

AI-assisted summary of Will AI Destroy or Save the Power Grid? The Race to Design the Grid of the Future With Leidos — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by Ai4 on August 18, 2026, this presentation features Josh Wepman, VP and CTO of Energy and Infrastructure at Leidos. The accompanying description highlights how the US power grid faces unprecedented strain from AI data centers, renewables, and aging infrastructure, framing artificial intelligence as both a key driver of the grid's crisis and the essential tool required to overcome severe workforce and resource bottlenecks.

## TL;DR Josh Wepman examines the paradox of artificial intelligence as both the primary driver of an impending power grid crisis—with data center power demand projected to jump to 12%–17% of total national electricity by 2030—and the vital technology required to solve it. Facing severe shortages in workforce, capital, and time to double national transmission capacity and fortify aging infrastructure that underpins over 85% of US GDP, traditional manual engineering workflows are no longer sustainable. While generic text-based language models cannot parse complex electrical schematics, specialized physics-informed grid foundation models and agentic workflows are unlocking massive efficiency gains, such as accelerating transmission design from 150 hours per mile to 150 miles per hour and slashing engineering labor by 30% to 50% across generation, substation, and distribution programs.

## Atmosphere & style The presentation carries a pragmatic, urgent, and technically authoritative tone that balances sobering warnings of an impending energy bottleneck with tangible engineering optimism. Wepman commands the stage with conversational warmth and dry humor, moving seamlessly from relatable personal anecdotes—such as cutting the house power to observe the frantic chaos of his nieces and nephews—to rigorous industrial metrics. The talk delivers a crisp emotional arc, transforming what initially seems like an inevitable grid crisis into a clear, energizing call to action for human engineers augmented by specialized AI systems.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=tHeO16N0AHk&t=4s) Artificial intelligence creates a central paradox for energy infrastructure: it is both the primary demand driver pushing the electric grid toward crisis and the essential engineering tool required to solve it. - [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) Modern daily life and economic activity exhibit near-total real-time dependency on electric power, with an estimated zero percent of an average person's day untouched by continuous grid delivery. - [04:13](https://www.youtube.com/watch?v=tHeO16N0AHk&t=253s) The US power grid operates as the largest interconnected machine ever built, coordinating ~7.5 trillion kWh of generation annually across 500,000 miles of transmission lines and 55,000 substations. - [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) Data center power consumption is projected to surge from ~5% of national generation today to 12%–17% by 2030, equivalent to drawing half the power consumed by the entire US industrial base. - [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) Meeting future demand requires an estimated 70% increase in total power generation and an 'all-of-the-above' mix of natural gas, nuclear, and renewables to manage two-sided grid volatility. - [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) Accommodating new generation and load profiles requires doubling the nation's 500,000 miles of high-voltage transmission lines by 2030, and potentially tripling them by 2040. - [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) Grid expansion is severely bottlenecked by societal siting resistance (the 'BANANA' mentality: 'Build absolutely nothing anywhere near anyone') and a persistent labor shortage in specialized substation engineering. - [12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) Modernizing the grid is constrained by a five-fold deficit in workforce, time, capital (with costs reaching $3 trillion), supply chains, and raw materials, putting over 85% of US GDP at risk. - [13:44](https://www.youtube.com/watch?v=tHeO16N0AHk&t=824s) Standard text-based frontier language models cannot solve grid engineering challenges because they lack the topological reasoning, spatial awareness, and physical world models encoded in electrical schematics. - [13:44](https://www.youtube.com/watch?v=tHeO16N0AHk&t=824s) The long-term technological path relies on physics-informed grid foundation models capable of direct latent-state inference, replacing brute-force computation of hundreds of millions of linear equations. - [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) Agentic engineering systems have compressed transmission line design velocity from 150 hours per mile down to 150 miles per hour, condensing multi-year projects into an afternoon. - [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) Practical AI deployment is already cutting engineering labor by 50% in front-end power generation studies and 30% across substation layout and distribution grid hardening programs.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=tHeO16N0AHk&t=4s) Leidos and the AI-Energy Challenge

Josh Wepman introduces himself as the Chief Technology Officer for the energy business at Leidos, an $18 billion technology integration firm headquartered in the Washington, D.C. area. Explaining the origin of the company's unusual name, Wepman notes it was clipped directly from the center of the word "kaleidoscope" to reflect the company's philosophy of viewing complex problems from multiple angles. Within Leidos operates a $1 billion power engineering practice staffed by roughly 6,000 power systems engineers who are tasked with designing the physical backbone of the energy ecosystem—ranging from generation, substations, and transmission to local distribution, gas infrastructure, and distributed power systems dedicated to data centers.

Reflecting on his five years leading innovation and technology across this engineering workforce, Wepman underscores that the current artificial intelligence boom is placing unprecedented strain on the electrical system. The massive energy demands of the AI era are introducing severe operational friction and threatening to destabilize the existing grid. Setting the stage for the discussion, Wepman posits a central paradox: artificial intelligence represents both the primary accelerant of the grid's impending energy crisis and the very tool engineers must rely on to solve it.

- **Leidos Corporate Scale**: $18 billion technology integrator based in the D.C. area; name derived from the core of "kaleidoscope". - **Power Engineering Division**: A $1 billion business unit employing 6,000 power systems engineers handling generation, high-voltage transmission, distribution, and specialized data center power integration.

---

[01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) The Invisible Dependency: A Thought Experiment

To illustrate how profoundly modern society relies on real-time power, Wepman invites the audience to walk through a simple thought experiment tracing the first 90 minutes of their morning routine. From the moment an alarm rings on a smartphone in the dark, to switching on overhead lights, taking a hot shower, riding an elevator down to a hotel lobby, and brewing coffee, virtually every basic action directly or indirectly depends on a functional electric grid. Industry studies confirm that approximately 0% of an average person's daily life remains untouched by real-time power delivery.

```mermaid flowchart LR accTitle: Morning Electrical Dependency Chain subgraph Routine["Morning Routine"] A["Phone Alarm"] --> B["Lighting & Shower"] B --> C["Elevator & Food Prep"] end subgraph GridLayer["Grid Real-Time Power"] D["100% Real-Time Power Delivery"] end Routine -.-> GridLayer ```

Wepman reinforces this dependency with a personal story about his 24 nieces and nephews. Playing the role of the "fun uncle," he recounts occasionally sneaking over to the breaker panel just before the children arrive home from school and pulling the main switch. The result is instant chaos: with Wi-Fi, computers, televisions, and basement lights cut off—and the inability to stare blankly into an illuminated refrigerator—the children come "absolutely unglued." Within ten minutes, they are slumped on the stairs in despair, professing utter boredom until the breaker is flipped back on with apologies. This total reliance demonstrates that electric power is no longer just a utility, but an absolute baseline requirement for modern existence.

---

[04:13](https://www.youtube.com/watch?v=tHeO16N0AHk&t=253s) Anatomy of the Grid: The Largest Machine Ever Built

Demystifying the infrastructure required to keep modern life running, Wepman breaks down the four core structural components of the North American power grid, alongside the invisible digital nervous system that coordinates them:

```mermaid flowchart TD accTitle: Anatomy of the North American Power Grid subgraph Generation["1. Power Generation (2025: ~7.5 Trillion kWh)"] G1["Natural Gas (< 50%)"] G2["Renewables (< 25%)"] G3["Nuclear (< 20%)"] G4["Coal (~15%)"] end

subgraph BulkTransport["2. Transmission Grid"] T["500,000 Miles of High-Voltage Lines & Towers"] end

subgraph Transformation["3. Substations"] S["55,000 Substations (Transformers, Breakers, Metrology)"] end

subgraph LocalDelivery["4. Distribution Infrastructure"] D["Neighborhood Lines (Residential, Commercial, Industrial, Data Centers)"] end

subgraph ControlSystem["Control Plane (Digital Nervous System)"] C["SCADA / ADMS / DERMS Monitoring & Balancing"] end

Generation --> BulkTransport BulkTransport --> Transformation Transformation --> LocalDelivery ControlSystem -.-> Generation ControlSystem -.-> BulkTransport ControlSystem -.-> Transformation ControlSystem -.-> LocalDelivery ```

In 2025, the power generation fleet generated approximately 7.5 trillion kilowatt-hours (kWh) of electricity—where one single kilowatt-hour equals roughly the energy needed to run a standard dishwasher cycle. Natural gas accounted for just under half of that production, renewables made up slightly less than a quarter, nuclear supplied just under 20%, and coal contributed around 15%. This immense volume of power is routed across 500,000 miles of high-voltage transmission lines carried by giant steel towers and pylons, stepping through roughly 55,000 substations containing switchgear, transformers, and metrology before filtering into local distribution networks that serve homes, factories, government facilities, and distribution-connected data centers.

Wepman reflects on the sheer improbability of the entire system functioning seamlessly. Coordinated behind the scenes by real-time SCADA, Advanced Distribution Management Systems (ADMS), and Distributed Energy Resource Management Systems (DERMS), the grid represents a massive, interconnected machine maintained around the clock by a dedicated workforce of line workers and engineers. To an observer from the 1300s, an ordinary person controlling unseen physical forces simply by flicking a switch would look like a sorcerer. The miracle is not that power outages happen during extreme storms, but that such an overwhelmingly complex machine works almost 100% of the time.

---

[07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) The AI Explosion and the Race to Scale Transmission

The central crisis facing the power industry is not merely that the grid is large, but that it must now expand faster across all segments simultaneously than ever before in history. Today, data centers account for roughly 5% of the 7.5 trillion kWh consumed annually in the United States. By 2030, data center power consumption is projected to surge to between 12% and 17% (roughly 15%), while long-term forecasts out to 2050 remain entirely unpredictable. Reaching 15% of total national electric consumption means data centers alone will draw energy equivalent to half of the entire American industrial base—encompassing every chemical plant, manufacturing facility, and desalination plant that took a century to construct—built out in barely a decade.

Meeting this demand will require an estimated 70% increase in total power generation, demanding an aggressive "everything we can get our hands on" strategy encompassing nuclear, renewables, natural gas, and potentially legacy coal. While clean renewables drive energy costs down and approach zero carbon emissions over their lifecycle, their intermittent output—producing too much power when conditions are optimal and far too little when they are not—upends the traditional operating model where steady generation cleanly followed varying consumer loads. To manage this two-sided volatility, the nation will have to double its 500,000 miles of transmission lines by 2030, and potentially triple them by 2040, an unprecedented civil and electrical engineering undertaking.

- **Data Center Demand Surge**: Growing from ~5% of national generation today to 12%–17% by 2030, equaling half of the entire U.S. industrial footprint. - **Generation Expansion**: Forecasting an approximate 70% rise in electricity generation requirements to supply the load. - **Generation Mix Imperative**: All energy sources must be deployed simultaneously, including natural gas, nuclear, and renewables. - **Grid Volatility**: Unpredictable renewable generation and fluctuating modern loads create two chaotic, independent variables across the grid. - **Transmission Expansion Targets**: 500,000 existing transmission miles must double by 2030 and potentially triple by 2040.

[10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) Siting Bottlenecks, the BANANA Phenomenon, and the Substation Labor Shortage

Building out transmission is not simply an electrical engineering challenge involving corona effects and physical hardware; it is fundamentally a techno-economic and societal dilemma. Siting new high-voltage lines runs into severe public resistance. Beyond standard NIMBYism ("Not In My Backyard"), the industry frequently collides with the "BANANA" mentality—"Build absolutely nothing anywhere near anyone." Consequently, the industry constantly gets in its own way at the exact moment it must expand physical capacity faster and at a larger scale than ever before in history.

```mermaid flowchart TD accTitle: Grid Expansion Challenges Across Sectors subgraph Siting["Siting & Transmission"] A["Transmission Lines"] --> B["NIMBY & BANANA Resistance"] end subgraph Substations["Substation Bottlenecks"] C["Substation Terminations"] --> D["Capacity Must 2x-3x"] D --> E["Grotesquely Labor Intensive"] E --> F["Chronic Engineer Shortage"] end subgraph Distribution["Distribution Grid"] G["5.5M Miles of Plant"] --> H["Brittle Infrastructure"] H --> I["Fortification & Undergrounding"] end ```

Every new transmission line must terminate somewhere, requiring the nation's substation plant to double or triple in capacity. This constitutes one of the steepest hurdles in the energy sector because substation engineering is "grotesquely labor intensive." While labor-heavy engineering was historically a profitable services model, the system breaks down when the workforce runs dry. Engineering job boards across the industry show openings for 15 substation engineers at virtually every firm—vacancies that have remained unfilled for a year and a half because the talent pipeline simply does not exist. Further downstream, the 5.5 million miles of existing distribution lines are increasingly brittle, demanding massive labor and capital expenditures for hardening, pole replacements, and undergrounding.

- **BANANA Acronym**: "Build absolutely nothing anywhere near anyone." - **Substation Expansion Target**: Requires a 2x to 3x increase in overall substation capacity. - **Distribution Plant**: Spans roughly 5.5 million miles requiring fortification and undergrounding. - **Talent Drought**: Substation engineering vacancies routinely sit open for 18 months due to a lack of trained specialists.

---

[12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) The Five-Fold Deficit and the Stakes of Grid Failure

The fundamental threat facing the power sector is not growth itself, but the unprecedented *pace* of growth. Engineering organizations are attempting to solve this historic expansion while lacking almost every critical input: workforce, time, capital, supply chains, and raw materials. No academic pipeline exists to rapidly supply electrical grid engineers, and existing engineering workflows rely on long lead times that the modern demand curve cannot accommodate. Financing this overhaul is equally constrained; the overall modernization bill could reach $3 trillion, an amount that cannot simply be absorbed by ratepayer utility bills when consumers are already feeling the pressure of rising costs.

```mermaid flowchart LR accTitle: Five Core Deficits Threatening Grid Expansion subgraph Deficits["Core Deficits"] direction TB D1["Labor Deficit"] D2["Time Constraints"] D3["Capital Limits ($3T Cost)"] D4["Supply Chain Delays"] D5["Raw Material Shortages"] end Deficits --> Risk["Grid Instability / Failure"] Risk --> GDP["Threatens >85% of Real-Time US GDP"] ```

Failure to overcome these deficits carries catastrophic national consequences. Real-time electric power underpins the entire modern economy, with an estimated 85% or more of United States GDP fundamentally dependent on grid stability. Because more than 85% of economic activity ceases if the power system collapses, failure is simply not an option, forcing engineering firms to rethink how physical engineering work is executed.

- **Estimated Modernization Cost**: Up to $3 trillion across generation, transmission, substations, and distribution. - **Economic Vulnerability**: An estimated 85%+ of US GDP is directly reliant on real-time electric power delivery. - **Core Resource Bottlenecks**: Shortages across specialized personnel, schedule runways, rate-base capital, supply chain bandwidth, and raw materials.

---

[13:44](https://www.youtube.com/watch?v=tHeO16N0AHk&t=824s) The Limits of Generic AI and the Path to Grid Foundation Models

For the past century, grid engineering has relied on experienced senior engineers huddling over complex single-line and three-line schematics, applying decades of unwritten intuition, logic, and physical reasoning stored entirely in their heads. Generic frontier language models are incapable of replicating this work out of the box because they were trained on tokenized text rather than electrical schematics and topological world states. Standard AI tools can recite the theoretical physics of electric power, but like a physician who knows medical textbooks without examining the patient, theoretical knowledge is useless without deep context on discrete physical assets.

```mermaid flowchart TD accTitle: Evolution of Engineering AI Systems subgraph Current["Current State: Tactical Task Compression"] A["Bloated Workflow Segments"] --> B["Task-Specific Agentic Scripts"] B --> C["Linear Equation Solvers"] end subgraph Target["Target State: Physics-Informed Foundation Models"] D["Multi-Modal Schematics & World Models"] --> E["Cooperative Agentic AI"] E --> F["Grid Foundation Models"] F --> G["Direct Latent State Inference"] end ```

To close the gap, engineering teams are transitioning from basic process-compression scripts to cooperative agentic systems and physics-informed grid foundation models. Today's state-of-the-art tools compress bloated process steps using foundation models and specialized agents. However, full scalability requires systems capable of reading non-text schematics, learning directly from human engineering actions, and understanding physical world models. Instead of running brute-force computational simulations that solve 100 million linear algebra equations sequentially, future grid foundation models will directly infer physical system behavior from learned latent states.

- **Schematic Complexity**: Single-line and three-line diagrams contain domain-specific topological logic not represented in standard web-scraped token datasets. - **Limitations of Enterprise AI**: Text-focused language models cannot navigate spatial, electrical, and regulatory constraints without specialized world models. - **Computational Paradigm Shift**: Moving from solving 100M+ sequential linear algebra equations to direct latent-state physical inference.

---

[16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) Real-World AI Breakthroughs Across Generation, Transmission, and Distribution

While the sheer scale of the energy transition can feel like an "enormous perfectly predictable, inevitable calamity," practical deployment of agentic engineering systems is already producing dramatic productivity gains. In power generation, agentic workflows applied to Front-End Engineering Design (FEED) studies have eliminated 50% of the required engineering labor, enabling teams to design twice as much gas and generation infrastructure for data center clients with half the workforce.

```mermaid flowchart LR accTitle: Productivity Gains from AI-Assisted Grid Engineering subgraph Generation["Generation & Data Centers"] G1["FEED Studies"] --> G2["50% Labor Reduction / 2x Throughput"] end subgraph Transmission["Transmission Siting & Design"] T1["Historical: 150 Hours / Mile"] --> T2["AI Agentic: 150 Miles / Hour"] end subgraph Substations["Substation Engineering"] S1["Manual Schematic Layouts"] --> S2["30% Labor Extracted via Collaboration"] end subgraph Distribution["Distribution Grid Hardening"] D1["Pole Hardening / Replacements"] --> D2["30% Labor Saved (Path to 60-90%)"] end ```

The most dramatic breakthrough has emerged in transmission line engineering. Projects that once required up to 150 hours of tedious manual drafting and route calculation per single mile can now be designed at a rate of 150 miles per hour using optimization algorithms and agentic reasoning systems—condensing what was once a four-year engineering project into a single afternoon. Similar collaborative agentic tools are removing 30% of human labor from substation engineering, and delivering immediate 30% labor savings in distribution pole replacement and storm-hardening programs, with clear developmental roadmaps toward 60% and 90% reductions. AI will not rescue the power grid autonomously, but human engineers augmented by specialized agentic systems will provide the capacity needed to secure the energy transition.

- **Transmission Breakthrough**: Shifted design velocity from 150 hours per mile to 150 miles per hour. - **Generation & FEED Velocity**: 50% labor extraction allowing double the data center power design capacity. - **Distribution Roadmap**: Immediate 30% labor savings on high-volume pole hardening, scaling toward 60% and 90%. - **Closing Perspective**: "I don't think AI will save the power grid alone, but I think together we will."

## Notable quotes > Electric power is the basis of the American economy and our way of life. — Josh Wepman

> [04:13](https://www.youtube.com/watch?v=tHeO16N0AHk&t=253s) The US power grid is the largest machine ever built and delivers these capabilities real-time. — Josh Wepman

> [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) absolutely unglued — Josh Wepman

> [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) everything we can get our hands on — Josh Wepman

> [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) Build absolutely nothing anywhere near anyone. — Josh Wepman

> [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) grotesquely labor intensive — Josh Wepman

> [12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) We lack the workforce, capital, cost controls, materials, and time to get it all done. — Josh Wepman

> [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) enormous perfectly predictable, inevitable calamity — Josh Wepman

> The bad news is, AI factories are accelerating these challenges. The good news is, AI is actually helping us solve them too. — Josh Wepman

> [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) I don't think AI will save the power grid alone, but I think together we will. — Josh Wepman

## Mentioned Josh Wepman, Leidos, Ai4, SCADA, Advanced Distribution Management Systems (ADMS), Distributed Energy Resource Management Systems (DERMS), Front-End Engineering Design (FEED)

## Video description Josh Wepman - VP & CTO Energy and Infrastructure, Leidos

Electric power is the basis of the American economy and our way of life. The US power grid is the largest machine ever built and delivers these capabilities real-time. But the power grid is under stress from a myriad of challenges including: AI driven datacenters, intermittent renewables, storage, transportation electrification, ageing infrastructure, severe weather and cyber threats. The grid of the future will need to grow and evolve to accommodate these needs, and as an industry we lack most of the things we need to keep pace. We lack the workforce, capital, cost controls, materials, and time to get it all done. The challenge of building a clean, flexible, reliable, and cost-efficient grid is mammoth. So somehow AI is going to have to scale the impact of our accessible workforce, accelerate optimal design and construction, and reimagine operating and defending the grid. The bad news is, AI factories are accelerating these challenges. The good news is, AI is actually helping us solve them too.

[Source](https://www.youtube.com/watch?v=tHeO16N0AHk)

中文简报

数据中心需求激增，Leidos部署AI加速美国电网扩建激增的数据中心需求对支撑美国85% GDP的电网构成威胁，专用AI工程工具必须实现输电与变电站设计自动化，以克服严重的劳动力短缺。 Josh Wepman探讨了人工智能的一大核心悖论：它既是导致电网危机迫在眉睫的主要推手（预计到2030年数据中心用电需求将激增至全美总用电量的12%至17%），也是化解该危机不可或缺的关键技术。在支撑美国85%以上GDP的老旧基础设施改造以及全美输电容量翻倍的目标面前，人力、资金和时间的严重匮乏使传统的人工工程设计流程难以为继。尽管通用的文本语言模型无法解析复杂的电气原理图，但融入物理知识的专用电网基础模型与智能体工作流正释放出巨大的效率提升潜力，例如将输电设计速度从“每英里耗时150小时”大幅提升至“每小时处理150英里”，并在发电、变电和配电项目中削减30%至50%的工程人力需求。人工智能正给能源基础设施带来核心悖论：它既是将电网推向危机边缘的首要负荷驱动力，也是解决该危机不可或缺的核心工程工具。 (0:04) 预计到 2030 年，数据中心用电量占美国总发电量的比重将从约 5% 激增至 12%–17%，相当于全美工业基础用电量的一半，需要总发电量增加约 70%，并将现存 50 万英里高压输电线路翻倍。 (7:24) 电网现代化受制于人力、时间、高达 3 万亿美元的改造成本、供应链与原材料这五重短缺，加上选址阻力与变电站专业工程人才荒，正威胁着全美超过 85% 的 GDP 安全。 (12:06) 通过引入专用 Agentic 智能体工作流，输电线路设计效率已从以往每英里耗时 150 小时跃升至每小时可设计 150 英里，同时在发电与变电配电工程中实现了 30% 至 50% 的人力精简。 (16:31)

Reading summary

According to Leidos energy CTO Josh Wepman, the rapid expansion of artificial intelligence is placing unprecedented demand on the electric grid, creating an energy crisis that can only be resolved by applying specialized, physics-informed AI tools to accelerate power engineering. The Anatomy and Societal Reliance on the Grid (4:13) Modern economic activity depends almost entirely on real-time power delivery, with Wepman estimating that more than 85% of United States GDP would halt if the power system collapsed. In 2025, the power generation fleet produced roughly 7.5 trillion kilowatt-hours, sourced primarily from natural gas (under 50%), renewables (under 25%), nuclear (under 20%), and coal (around 15%). This energy moves across 500,000 miles of high-voltage transmission lines, through roughly 55,000 substations, and across 5.5 million miles of local distribution lines. The system operates as a single interconnected machine regulated continuously by SCADA, Advanced Distribution Management Systems (ADMS), and Distributed Energy Resource Management Systems (DERMS) to balance supply and fluctuating consumer loads. Data Center Loads and Multi-Deficit Bottlenecks (7:24) Data centers currently consume about 5% of U.S. electricity generation, but Wepman projects this share will reach 12% to 17% (roughly 15%) by 2030—an amount equal to half the entire existing American industrial base built in roughly a single decade. Meeting this surge requires an estimated 70% increase in power generation, alongside doubling the existing 500,000 miles of transmission lines by 2030 and potentially tripling them by 2040. Achieving this expansion is constrained by a five-fold deficit in workforce, schedule runways, capital, supply chains, and raw materials, with the full modernization bill estimated at up to $3 trillion. Siting new high-voltage lines faces extreme public resistance described as the "BANANA" mindset ("Build absolutely nothing anywhere near anyone"). Simultaneously, the required doubling or tripling of substation capacity is hindered by severe labor shortages, with specialized substation engineering roles routinely remaining vacant for 18 months. The Shortcomings of Generic Language Models (13:44) Standard commercial AI models cannot independently solve power engineering challenges because they are trained on web text rather than electrical schematics, spatial topological rules, and physical world constraints. Wepman notes that while generic models can recite theoretical physics, grid engineering historically relies on human intuition applied to single-line and three-line diagrams. Bridging this gap requires developing specialized grid foundation models capable of learning from human design workflows and reading non-text diagrams. The long-term architectural goal is shifting away from solving 100 million sequential linear algebra equations in brute-force simulations toward directly inferring physical grid behavior from learned latent states. Early Productivity Gains Across Power Sectors (16:31) Engineering teams are already deploying specialized agentic tools to compress labor-intensive design stages. In transmission engineering, optimization algorithms and agentic systems have shifted design velocity from up to 150 hours per mile to 150 miles per hour, condensing multi-year design timelines into an afternoon. Efficiency gains extend across the rest of the power sector. Agentic Front-End Engineering Design (FEED) workflows have cut engineering labor by 50% for data center generation projects, effectively doubling throughput. In substation layout engineering and distribution pole hardening, collaborative tools have eliminated 30% of manual labor, with development targets aiming for 60% to 90% reductions in distribution tasks. Wepman emphasizes that AI will not rescue the grid autonomously. Instead, the viability of the ongoing energy transition depends on human engineers leveraging domain-specific agentic tools to overcome severe physical, regulatory, and labor constraints.

精炼阅读摘要

Leidos能源业务首席技术官Josh Wepman指出，人工智能的爆发式增长正在对北美电网施加前所未有的负荷压力并带来失稳风险，但基于物理规律的专业智能体工程AI系统也是克服多重资源赤字、加速电网扩容重建的核心工具。数据中心负荷剧增倒逼电网翻倍扩容 (7:24) 根据Wepman的分析，数据中心用电量目前约占全美年发电总量（2025年约为7.5万亿千瓦时）的5%，但预计到2030年将攀升至12%至17%（约15%）。这一增量相当于美国在一个世纪中建立起来的全部工业基础用电规模的一半，却需要在短短十年内完成布局。为支撑这一庞大需求，全美发电量预计需要增加约70%，迫使行业同步推进天然气、核能、可再生能源及传统煤电的综合部署。然而，清洁能源的间歇性特征颠覆了以往发电追随用电负荷的传统运行模式，使电网两侧均出现高度波动性。为了跨区域输送并平衡这类不稳定的电力负荷，全美现有的50万英里高压输电线路必须在2030年前实现翻倍，并在2040年前扩建至目前的三倍，构成前所未有的工程交付压力。五重赤字与现代经济体系的脆弱性 (12:06) 电网扩建的核心阻碍在于前所未有的扩张速度与极度匮乏的现实资源之间的冲突。工程体系面临着专业劳动力、建设周期、投资资本、供应链和原材料等五重核心赤字。据行业估算，电网全面现代化的总资金需求可能高达3万亿美元，在当前用电成本不断攀升的背景下，难以直接由终端用户电费全额承担。物理建设同样遭遇严峻阻力。高压线路选址频繁受到公众的强烈反对，甚至形成“绝不在任何人附近建设任何设施”（BANANA）的社会阻力；同时，变电站工程对人力依赖极高且容量需要扩建2至3倍，但行业内专业变电站工程师极度匮乏，招聘岗位平均空置18个月已成为常态。Wepman强调，美国85%以上的实时GDP直接依赖电网的稳定运行，电力系统的崩溃将导致现代经济全面停摆，因此电网建设必须寻求突破性的工程模式。通用语言模型局限与物理电网模型 (13:44) 通用前沿大语言模型难以直接承担电网系统的专业设计。过去一个世纪中，电网工程高度依赖资深工程师在单线图、三线图等复杂电气图纸中融入的物理直觉与空间推理经验。现有通用模型建立在文本分词训练之上，既无法准确理解非文本的电气拓扑结构，也无法在缺乏空间、电气规程和物理实体约束的情况下进行合理推断。为此，工程行业正在从基础的工作流压缩脚本转向结合物理法则的电网基础模型与协作智能体。新型系统旨在通过学习人类工程操作并建立物理世界模型，直接读取非文本电气图纸。计算范式也正从按顺序求解上亿个线性代数方程的传统数值模拟，转向基于深度物理潜在状态的直接推理，大幅压缩工程推演耗时。智能体工程落地推动设计效率跃升 (16:31) 在实际工程实践中，协作式智能体系统已在发、输、变、配各个业务板块展现出显著的降本增效成果。在发电与数据中心前端工程设计（FEED）阶段，智能体工作流削减了50%的工程人力消耗，使工程团队得以用一半的人员交付双倍的数据中心电力基础设施设计。在输电线路设计领域，传统模式下每单英里线路需耗费高达150小时的手工制图与选线计算，而借助优化算法和智能体推理系统后，设计速度提升至每小时150英里，将长达四年的工程周期压缩至一个下午完成。此外，智能体协作工具在变电站工程中已消减了30%的人工负荷，在550万英里配电网的线杆抗灾加固项目中也即时节省了30%的人力，并规划了未来实现60%至90%人力工时压缩的演化路径。 Wepman总结指出，人工智能无法孤立地挽救电网，但由专业智能体系统武装的工程师群体能够突破传统产能瓶颈，共同推动能源转型的实现。然而，高达3万亿美元的改造成本分担机制、严苛的公共选址阻力以及深层次的物理供应链短缺，依然是该战略落地过程中亟待解决的系统性难题。

Detailed notes — zh

## Background 本内容由频道 Ai4 于 2026 年 8 月 18 日在 YouTube 平台发布。视频由 Leidos 能源与基础设施副总裁兼首席技术官 Josh Wepman 主讲，探讨了在 AI 数据中心暴增、新能源并网、老旧基础设施及劳动力匮乏等多重重压下，北美电网面临的严峻危机，以及如何利用 AI 智能体与物理世界模型重构并拯救未来电网。 ## TL;DR Josh Wepman 剖析了北美电网作为现代文明隐形底座的运作机制，指出数据中心算力激增与间歇性清洁能源并网正对电网造成前所未有的负荷冲击。面对高达 3 万亿美元改造资金缺口、严峻的劳动力断层及极端的社会阻力，传统依靠人工经验与数值方程计算的工程模式已逼近极限。Leidos 正在通过开发协作式智能体与电网基础模型，将隐性工程知识与物理规律转化为推理能力，在发电设计、输电线路规划（提速至每小时 150 英里）及变配电改造中实现了革命性的工时缩减，力求化解 85% 全美 GDP 面临的断电停摆风险。 ## Atmosphere & style 演讲兼具宏大的工业史诗感与风趣切近的叙事魅力，Josh 既能用关掉电闸整蛊侄子侄女的生动轶事将冰冷的基础设施拉回日常生活，又能以冷峻精确的数据直击万亿级电网危机的咽喉。现场情绪从对劳动力与资金“四大瓶颈”的深沉紧迫感，逐步过渡到展示 AI 落地突破时的振奋与自豪，整体呈现出务实、硬核且充满工程乐观主义的行业领袖风范。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=tHeO16N0AHk&t=4s) AI 浪潮带来了空前的用电负荷与电网稳定性挑战，呈现出既是危机导火索又可能是破局关键的双重属性。 - [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) 现代社会对电网存在绝对的刚性依赖，人的一天中脱离电网实时影响的时间占比约为 0%，全美超过 85% 的 GDP 实时依赖稳定供电。 - [04:13](https://www.youtube.com/watch?v=tHeO16N0AHk&t=253s) 北美电网由发电、50万英里高压输电线路、约5.5万座变电站、配电网及实时控制系统构成，2025年发电量达7.5万亿千瓦时。 - [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) 数据中心用电预计到2030年将占全美发电量的12%至17%，相当于全美现有工业用电总负荷的一半，需要未来全美发电规模激增约70%。 - [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) 间歇性新能源的激增导致发电侧与负荷侧双双陷入混沌，要求现有输电线路在极短周期内实现翻倍甚至翻三倍。 - [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) 社会阻力已从“邻避效应”（NIMBY）演变成“BANANA”现象，叠加变电站高度劳动密集与长期用人荒，阻碍电网建设推进。 - [12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) 电网现代化面临人才断层、时间紧迫、3万亿美元资本极限超出费基承载、以及供应链短缺四大系统性瓶颈。 - [13:34](https://www.youtube.com/watch?v=tHeO16N0AHk&t=814s) 传统通用前沿大模型仅掌握电力文字理论，无法理解高度依赖人类隐性知识与非文本符号的复杂电气工程图纸。 - [15:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=936s) 下一代电网 AI 需要协作式智能体（Agentic AI）与世界模型，通过学习物理规律的基础模型取代传统按序解算上亿线性代数方程组的数值范式。 - [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) 实战落地显示，AI 智能体已在发电 FEED 研究中实现人工减半，将输电线路设计效率提速至每小时150英里，并在变电和配电工程中削减30%繁重劳务工时。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=tHeO16N0AHk&t=4s) 幕后造网者：Leidos 与 AI 时代的电力摩擦作为 Leidos 能源业务的副总裁兼首席技术官，Josh Wepman 掌管着这家总部位于华盛顿特区、年营收达 180 亿美元的技术集成巨头旗下规模达 10 亿美元的电力工程业务。很多人对 "Leidos" 这个名字感到陌生，Josh 解释道，这个名字其实截获自万花筒（kaleidoscope）的中段，寓意着他们习惯从多元视角审视并解决复杂问题。该团队拥有多达 6000 名电力系统工程师，专门负责电网发电、输电、变电站和配电网的设计与规划，同时也涵盖了支撑数据中心运转的分布式能源网络及配套的天然气基础设施。他们是解决“建什么、怎么建、如何让庞杂部件协同并接入现有电网”的核心力量。然而，眼下呼啸而过的 AI 浪潮正在给整个系统带来空前剧烈的摩擦与震荡。AI 的巨大算力需求正在动摇现有电网的稳定性，而业界尚未找到轻松化解这一冲击的捷径。既然整个行业都在大谈特谈人工智能，大家就必须直面一个最底层的现实挑战：驱动这些模型的能源究竟来自何方？我们要如何重新构建电网，其间又会遭遇哪些瓶颈？在 Josh 看来，当下的局面极富戏剧性与矛盾感——"AI 既是引发这一系列严峻问题的导火索，但同时也可能是破解该危机的关键钥匙"。 ### [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) 现代生活的隐形支柱与“断电”思想实验为了让听众切身体会电力在日常生活中到底扮演着多么决定性的角色，Josh 抛出了一个直击感官的思想实验：回顾今早睁开眼后的前 90 分钟，你具体做了些什么？在黑暗的清晨，手机闹钟响起，你闭着眼摸索着关掉它，顺便翻看夜间是否有紧急消息；随后起身开灯、洗个热水澡、乘电梯下楼，并买上一杯热咖啡和早餐。无论你是否意识到，在这短短的 90 分钟里，所做的每一件事都直接或间接地依赖于电网的实时稳定输出。大量行业研究表明，在人的一天当中，脱离电网实时影响的时间占比约为 0%——现代人一整天、一整周的活动，无一不建立在持续通电的基础之上，这绝不仅仅关乎雄心勃勃的 AI 愿景，而是现代生存的前提。谈到电力的不可替代性，Josh 还分享了自己作为“爱整活的舅舅/叔叔”与家族中 24 个侄子侄女相处的趣事。他经常在孩子们放学回家的前一刻，悄悄跑到配电箱前一把拉下总闸，然后躲在一旁观察。当孩子们踏进家门，发现生活秩序瞬间彻底崩塌时，场面极具戏剧性：Wi-Fi 断了、电脑黑屏、电视打不开、地下室的灯不亮，甚至连饿都不饿却习惯性地盯着发呆的冰箱都暗淡无光。通常不出十分钟，这帮年轻人就会无精打采地瘫坐在台阶上哀叹：“我好无聊，完全不知道该干嘛了。”直到他赶紧把电闸推回去，连声道歉说“对不起对不起，快回到你们的正常生活中去吧”，混乱才得以平息。正如他所总结的，现代人宛如利用一根手指就能操控大自然伟力的“中世纪巫师”，电力早已不再仅仅是“重要”，它是现代文明得以成立的绝对刚性依赖。 ### [04:13](https://www.youtube.com/watch?v=tHeO16N0AHk&t=253s) 庞大而脆弱的奇迹：北美电网的精密解剖业界各方都在连篇累牍地呼吁“必须改造电网以承载 AI”，但要想改造它，首先得透彻理解这个系统的构成与规模。2025 年，北美的发电系统累计生产了约 7.5 万亿千瓦时（kWh）的电能——如果对这个量级没有直观概念，可以参考运行一次家用洗碗机大约需要消耗 1 千瓦时。从能源结构来看，天然气发电占比略低于一半，可再生能源略低于四分之一，核电占比不到 20%，煤电则占约 15%，这些电力构成了驱动整个经济运转与各类技术大会的物理底座。这套体系依靠四层环环相扣的基础设施将电能安全派发，其工程复杂性令人叹为观止： ```mermaid flowchart TD accTitle: 北美电网架构与电力传输层级 subgraph GEN["能源生成 Fleet"] G1["发电厂: 天然气 / 可再生能源 / 核电 / 煤电"] G2["2025年发电量: 7.5万亿 kWh"] end subgraph TRANS["高压输电与变电"] T1["高压/超高压输电网: 50万英里"] S1["变电站体系: 约5.5万座"] end subgraph DIST["配电网络与末端负载"] D1["配电网: 工商/居民/联邦负荷"] D2["数据中心与工业用电接口"] end subgraph CTRL["实时监测与调控"] C1["SCADA / ADMS / DERMS 实时监控"] end GEN --> TRANS TRANS --> DIST CTRL -.->|平衡调度| TRANS CTRL -.->|双向监控| DIST ``` - **高压输电网络**：拥有超过 50 万英里的高压输电线路，依靠巨大的铁塔与导线，将超高压电能从偏远的发电端运送到负荷中心。 - **变电站基础设施**：全网分布着大约 5.5 万座变电站，集成了断路器、大型变压器与计量监控设备，是整张电网最核心也是最复杂的网络节点。 - **配电基础设施**：负责将变电站的电能降压输送至居民社区、商业楼宇、工业园区与联邦机构，同时也直接接入了一批在配电电压等级下运行的数据中心。 - **实时控制中枢**：隐于幕后的是 SCADA（数据采集与监视控制系统）、ADMS（高级配电管理系统）以及 DERMS（分布式能源管理系统），时刻对瞬息万变的潮流进行测量、平衡与保护。 Josh 坦言，这套系统庞大而复杂，但最不可思议的地方在于它竟然常年稳定工作。按常理而言，如此规模巨大、分布广泛的复杂分布式系统绝不该运作得这般平稳；偶尔停电丝毫不值得大惊小怪，反倒是它“几乎无时无刻不在正常运转”的这一事实，本身就是个奇迹。这完全归功于北美一整套深藏不露的专业工程子体系与文化，每天有成千上万名工程师在幕后默默维护并托举着这张大网。 ### [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) 算力吞噬巨兽：电网规模扩张与间歇性新能源的激突电网所面临的根本危机，并不在于其体量巨大或以往技术演进缓慢，而在于它目前正遭遇前所未有的全方位剧烈扩张。今天，作为承载 AI 算力基石的数据中心，已经消耗了全美 7.5 万亿千瓦时年发电量中的约 5%。展望 2030 年，这一比例预计将飙升至 12% 到 17%（约占 15% 左右）；至于 2050 年的预测，业界给出的所有数字不过是盲人摸象式的猜测。如果数据中心真的吞下全网 15% 的电量，那将相当于全美整整一半的工业用电总负荷——涵盖所有海水淡化厂、大型化工厂、机械制造基地等百年工业资产的耗电总量。美国花费了一百年时间才建立起整套工业用电体系，而同等量级的数据中心建设却要在短短十年内拔地而起。若以此全速推演，未来为了满足这股算力饥渴所必须新增的发电量将达到惊人的 7 万亿千瓦时，相当于全美发电规模直接激增 70%。面对“未来该用什么来发电”的争论，Josh 的回答极为现实：“答案是倾我们所能得到的一切资源（everything we can get our hands on）”。核电必须上，可再生能源必须上，天然气更是不可或缺的主力，甚至连原本计划淘汰的煤电也可能被迫延期退役。可再生能源在此过程中势必占据极大比重，其零碳与长周期降本特性固然对经济大有裨益，但也带来了难以招架的物理挑战： - **发电与负荷的双重混沌**：传统电网中，发电机组根据负荷的波动进行跟随调度，逻辑清晰；但风电、光伏具有“发力时产能过剩，无光无风时产力骤停”的强间歇性，导致负荷侧与发电侧双双陷入相互独立的混沌状态。 - **输电走廊翻倍的硬性要求**：为消纳并跨区域调度这些极不稳定的清洁电能，现有的 50 万英里输电线路到 2030 年必须翻倍，到 2040 年甚至可能需要翻三倍。在极短的年限内实现这一工程规模，在整个人类历史上都从未有过先例，更不用提高压输电线路在技术设计与施工审批上的极高门槛。 ### [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) 社会阻力、变电站劳动力荒与配电网加固在规划输电线路时，除了电晕效应等物理工程难题外，最棘手的阻碍往往来自社会层面——没有人希望高压电线建在自家附近。这不仅涉及众所周知的“邻避效应”（NIMBY，Not In My Backyard），更演变成了“BANANA 现象”（Build Absolutely Nothing Anywhere Near Anyone，即“别在任何人附近的任何地方建任何东西”）。这种社会和经济层面的自我阻碍，使得电网建设不再只是纯粹的工程问题，而演变成了一场必须以前所未有的速度和规模去解决的“技术-经济-社会”综合挑战。所有新建的输电线路最终都必须接入变电站，这意味着全美的变电站处理能力必须翻倍乃至提升至三倍。变电站工程对工程服务公司来说本是一门极具吸引力的大生意，但其最大软肋在于被业内形容为“极度劳动密集”（grotesquely labor intensive）的作业模式。在劳动力充足时这是优势，但在遭遇人才枯竭时便成了行业死穴。如今在各大招聘网站上，类似 Leidos 这样的工程公司普遍常年挂着十几个变电站工程师的空缺，且许多岗位已经挂了一年半仍无人应聘。此外，全美长达 550 万英里的配电网络同样面临极端脆弱性，必须进行大规模地下化改造和结构加固。 - **社会阻力机制**：从 NIMBY（邻避主义）升级为极端的 BANANA 心理（绝对不在任何人附近建造任何设施）。 - **变电站扩容缺口**：变电站总容量需提升 2 至 3 倍，然而工程师职位面临长期断档（单家公司常年空缺 15 人以上且招聘周期超 1.5 年）。 - **配网规模与韧性**：全美拥有 550 万英里配电资产，亟需通过地下化（undergrounding）与结构强化来抵御脆弱性。 --- ### [12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) 增长速度的死局与 85% GDP 的停摆风险摆在电力工程师面前的根本危机并不单纯是电网需要增长，而是“增长的速度”（pace of growth）。为了跟上这一轮由算力中心与电气化驱动的爆炸式需求，工程界面临着“缺人、缺时间、缺资金、缺供应链”的全面困局。在劳动力供给上，高校教育体系并没有对口的人才输送通道；现有的审批与工程流程设计建立在“工期充裕”的传统假设之上，根本无法适应当前节奏；而在资金端，电网现代化总成本预估高达 3 万亿美元，这笔天文数字不可能直接转嫁到普通用户的电费账单（rate base）中；即便具备了人力、时间和资金，全球供应链的原材料与设备短缺同样构成了硬约束。如果这一系列系统性瓶颈无法攻破，电网面临的瘫痪风险将带来灾难性的连锁反应。根据行业评估，全美至少有 85% 以上的 GDP 从根本上实时依赖于稳定供电。正如讲者直言，“如果这里出了差错，美国超过 85% 的经济运转都将陷入停滞，因此失败绝不是一个选项”，必须寻找革命性的工程范式来打破瓶颈。 ```mermaid flowchart TD accTitle: 电网现代化的系统性资源瓶颈与经济影响 subgraph Bottlenecks["电网扩张面临的四大缺失"] A["人才断层 (高校无对口输送/长期空缺)"] B["时间紧迫 (传统流程耗时冗长)"] C["资本极限 (3万亿美元成本超出费基承载)"] D["供应链短缺 (原材料与重型设备紧缺)"] end subgraph Impact["潜在失败后果"] E["电网扩容滞后 / 供电中断"] F["超85%的全美GDP实时依赖电力 经济面临全面停摆风险"] end Bottlenecks --> E E --> F ``` - **全美现代化改造成本**：预计高达 3 万亿美元，已超出常规费率基础（rate base）的承受空间。 - **GDP 依赖度**：全美 85% 以上的 GDP 实时依赖稳定供电。 --- ### [13:34](https://www.youtube.com/watch?v=tHeO16N0AHk&t=814s) 传统工程图纸的困境与前沿大模型的局限过去一百年间，电力工程的核心问题解决模式几乎没有改变：几名经验丰富的资深工程师围坐在一张极其复杂的电气图纸（单线图或三线图）前，调动脑海中积累数十年、却从未写进任何文档里的隐性知识（tacit knowledge）、工程逻辑与推理直觉，共同商讨系统改造方案。然而，这种高度依赖人类隐性知识的模式，正是当今通用前沿大模型（Frontier Models）的盲区所在。前沿基础模型固然将人类互联网上的所有文字进行了词元化（tokenization）和概率建模，但它们学习到的仅仅是“电力的理论知识”，而非解决现实工程问题的实践方法。这就像一个在医学院学完了所有医学教科书的医生，如果对眼前的具体病人一无所知，就无法进行任何实质性的诊断与治疗。当前的通用企业级 AI（Enterprise AI）在面对由非文本符号构成的复杂电气原理图时几乎无能为力。为了打破这种局面，Leidos 开始梳理工程师全流程中最繁琐、耗时最长的环节，针对性地开发专用 AI 智能体工具与优化算法，对传统臃肿的作业链条进行大幅压缩。 --- ### [15:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=936s) 下一代电网 AI：协作式智能体与物理世界模型当前的专用模型优化虽然取得了进展，但仍不足以彻底解决最底层的核心难题。电网工程的未来必须走向更深层的范式转换：一方面需要构建“协作式智能体 AI”（Cooperative Agentic AI），使其能够真正跨越非文本符号构成的复杂工程图纸进行跨模态逻辑推理；另一方面，必须构建能够通过人类专家的经验与操作动作进行学习的“世界模型”（World Models），把工程师无法言传的物理直觉转化为智能推理能力。更关键的变革在于从根本上改变电网计算的数学范式。传统的电网分析需要按顺序求解数以亿计的线性代数方程组来进行系统优化，耗时极长。而未来的“电网基础模型”（Grid Foundation Models）将直接学习整个电力系统的物理规律，通过训练直接捕捉系统的潜在状态（latent state），从而利用模型推理瞬间预测并输出优化结果，实现真正颠覆性的性能跃升。 ```mermaid flowchart LR accTitle: 传统数值计算与电网基础模型的范式对比 subgraph Traditional["传统计算范式"] T1["单线/三线工程图纸"] --> T2["人工经验解析"] T2 --> T3["按序求解上亿个 线性代数方程组"] T3 --> T4["高耗时输出优化结果"] end subgraph Future["电网基础模型范式"] F1["非文字图纸 / 世界模型"] --> F2["协作式智能体AI"] F2 --> F3["电网基础模型 (学习物理规律与隐状态)"] F3 --> F4["毫秒级推理输出最优化方案"] end ``` - **协作式智能体突破**：突破纯文本词元限制，实现对复杂电气图纸与原理图的直接多模态理解。 - **计算机制革新**：以物理知识嵌入的基础模型推理，取代依赖解算 100,000,000 个线性代数方程的传统数值计算。 --- ### [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) 落地战果：从发电、输电到配电的效率跃升面对电网危机，工程界一度感到像是眼睁睁看着一场可预见且无法避开的灾难降临。但通过在真实场景中不断试错，将特定智能体与优化算法投入实战，Leidos 在发电、输电、变电和配电全链路取得了突破性的提效成果：在发电与数据中心配套工程中，团队利用 AI 智能体将前端工程设计（FEED 研究）的人工工作量减少了一半，实现了“用一半的人力完成两倍的项目交付”，极大加速了燃气电站与 AI 数据中心供电基础设施的建设。在输电线路设计领域，传统上由于地形复杂或路径反复修改，设计 1 英里输电线路有时需要耗费 150 个工时；而如今借助专用优化算法与智能体，设计速度竟达到了惊人的“每小时 150 英里”——曾经需要耗时整整 4 年的工程设计项目，现在仅需一个下午即可全部完成。在极其复杂的变电站工程中，通过人类工程师与智能体系统的协作，已经成功消减了 30% 的繁重劳务工时；在以电杆更换和抗灾加固为代表的高频配电工程中，同样已削减了 30% 的人工，并明确规划了未来削减 60% 乃至 90% 人工成本的技术路线图。 | 电网环节 | 传统工程痛点 | AI 介入后的实际成效与演进路径 | |:--- |:--- |:--- | | **发电设施 (Generation)** | FEED（前端工程设计）研究耗时极长，难以匹配数据中心供电扩张 | **人工减少 50%**，相同团队产出翻倍，数据中心与燃气发电交付速度大幅提升 | | **输电线路 (Transmission)** | 复杂地形设计极端繁琐，曾出现 **150 工时/英里** 的低效瓶颈 | 达成 **150 英里/小时** 的设计速度，**4 年工程量在 1 个下午完成** | | **变电站 (Substations)** | 图纸极度复杂，工程高度劳动密集且行业长期面临用人荒 | 采用人机协作智能体，已确立 **减少 30% 繁杂人工** 的落地路径 | | **配电网络 (Distribution)** | 550 万英里资产维护，电杆抗风暴加固及地下化体量庞大 | 目前已 **削减 30% 人工**，未来技术路线直指 **60% 至 90%** 的降本目标 | 讲者在结尾呼吁，希望所有人每天早晨按掉闹钟、打开电灯时，能为电力工程人员保留一份理解与祝愿。AI 固然无法独自拯救电网，但人类工程师与 AI 的深度结合必将守住现代文明的电力基石。 > “I don't think AI will save the power grid alone, but I think together we will.”（我不认为 AI 能独自拯救电网，但我坚信，人机协作的我们将共同拯救它。） ## Notable quotes > [00:04](https://www.youtube.com/watch?v=tHeO16N0AHk&t=4s) AI 既是引发这一系列严峻问题的导火索，但同时也可能是破解该危机的关键钥匙 — Josh Wepman > [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) 在人的一天当中，脱离电网实时影响的时间占比约为 0% — Josh Wepman > [01:36](https://www.youtube.com/watch?v=tHeO16N0AHk&t=96s) 我好无聊，完全不知道该干嘛了。 — Josh Wepman > [07:24](https://www.youtube.com/watch?v=tHeO16N0AHk&t=444s) 答案是倾我们所能得到的一切资源（everything we can get our hands on） — Josh Wepman > [10:32](https://www.youtube.com/watch?v=tHeO16N0AHk&t=632s) BANANA 现象（Build Absolutely Nothing Anywhere Near Anyone，即“别在任何人附近的任何地方建任何东西”） — Josh Wepman > [12:06](https://www.youtube.com/watch?v=tHeO16N0AHk&t=726s) 如果这里出了差错，美国超过 85% 的经济运转都将陷入停滞，因此失败绝不是一个选项 — Josh Wepman > [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) 用一半的人力完成两倍的项目交付 — Josh Wepman > [16:31](https://www.youtube.com/watch?v=tHeO16N0AHk&t=991s) I don't think AI will save the power grid alone, but I think together we will.（我不认为 AI 能独自拯救电网，但我坚信，人机协作的我们将共同拯救它。） — Josh Wepman ## Mentioned Josh Wepman, Leidos, Ai4, SCADA, ADMS, DERMS, NIMBY, BANANA, FEED, Frontier Models, Cooperative Agentic AI, Grid Foundation Models ## Video description Josh Wepman - VP & CTO Energy and Infrastructure, Leidos Electric power is the basis of the American economy and our way of life. The US power grid is the largest machine ever built and delivers these capabilities real-time. But the power grid is under stress from a myriad of challenges including: AI driven datacenters, intermittent renewables, storage, transportation electrification, ageing infrastructure, severe weather and cyber threats. The grid of the future will need to grow and evolve to accommodate these needs, and as an industry we lack most of the things we need to keep pace. We lack the workforce, capital, cost controls, materials, and time to get it all done. The challenge of building a clean, flexible, reliable, and cost-efficient grid is mammoth. So somehow AI is going to have to scale the impact of our accessible workforce, accelerate optimal design and construction, and reimagine operating and defending the grid. The bad news is, AI factories are accelerating these challenges. The good news is, AI is actually helping us solve them too. [Source](https://www.youtube.com/watch?v=tHeO16N0AHk)

Source quotes

[0:04] Josh Wepman: AI 既是引发这一系列严峻问题的导火索，但同时也可能是破解该危机的关键钥匙 [1:36] Josh Wepman: 在人的一天当中，脱离电网实时影响的时间占比约为 0% [1:36] Josh Wepman: 我好无聊，完全不知道该干嘛了。 [7:24] Josh Wepman: 答案是倾我们所能得到的一切资源（everything we can get our hands on） [10:32] Josh Wepman: BANANA 现象（Build Absolutely Nothing Anywhere Near Anyone，即“别在任何人附近的任何地方建任何东西”） [12:06] Josh Wepman: 如果这里出了差错，美国超过 85% 的经济运转都将陷入停滞，因此失败绝不是一个选项 [16:31] Josh Wepman: 用一半的人力完成两倍的项目交付 [16:31] Josh Wepman: I don't think AI will save the power grid alone, but I think together we will.（我不认为 AI 能独自拯救电网，但我坚信，人机协作的我们将共同拯救它。）

Source quotes

Josh Wepman: Electric power is the basis of the American economy and our way of life. [4:13] Josh Wepman: The US power grid is the largest machine ever built and delivers these capabilities real-time. [1:36] Josh Wepman: absolutely unglued [7:24] Josh Wepman: everything we can get our hands on [10:32] Josh Wepman: Build absolutely nothing anywhere near anyone. [10:32] Josh Wepman: grotesquely labor intensive [12:06] Josh Wepman: We lack the workforce, capital, cost controls, materials, and time to get it all done. [16:31] Josh Wepman: enormous perfectly predictable, inevitable calamity Josh Wepman: The bad news is, AI factories are accelerating these challenges. The good news is, AI is actually helping us solve them too. [16:31] Josh Wepman: I don't think AI will save the power grid alone, but I think together we will.

Original transcript

[0:04] All right. As you may have heard, I'm Josh Watman. I'm CTO of the our energy business at Leidos. [0:10] If you've never heard of Leidos, we're a $18 billion company out of DC, technology integration [0:16] company. Inside of that business, we have a billion dollar power engineering organization. [0:22] And that's the that's the organization I leave innovation and technology for. [0:26] If you can't remember the name Leidos, it's clipped from the middle of the word kaleidoscope. [0:31] The idea is we see things from lots of different directions, in case you wondered. [0:35] Our our power engineering business is 6,000 power systems engineers. [0:41] They're the ones that are responsible for designing the generation, transmission, [0:44] and substation and distribution infrastructure. The distributed energy infrastructure that powers [0:49] data centers, all the gas infrastructure that feeds all that. We're the people that have [0:52] to go figure out what to go build, how to go build it, and how to make all that stuff work [0:56] and integrate it into our existing power system. Over the last five years that I've had this job, [1:02] I've got to work with all the different engineers who do that work day in and day out. [1:06] It's amazing what we can get done. But this AI era, this challenge of powering the AI era, [1:12] is really bringing friction to the system. It's destabilizing the power grid that we have today, [1:17] and it's going to take a fair amount of work to figure out how to overcome that. So I thought as [1:21] long as we're talking about AI here, we ought to talk about where does the energy come from? [1:25] How are we going to build the power grid and what frictions are we going to face? What you'll find [1:29] out is I think AI is the cause of and potentially the solution to the problem we're facing. [1:36] So let's let's do a little thought experiment. I find this interesting. How often do you really [1:42] think about the role of electric power in your personal life? And this is something I do with [1:48] people on occasion. Close your eyes or just go into your mind's eye. Think back to when you [1:53] woke up this morning. What are the first handful of things you did? If you're anything like me, [2:00] you know, your alarm went off cuz it was dark out and you you you swung at it to turn it off [2:04] and it turns out mine's on my phone, so I picked up my phone. I looked to see if anything happened [2:08] overnight that I should care about. I got up and I turned on a light and I went and you know, took [2:14] a shower, took an elevator down to the the lobby and went and found some hot coffee and some food. [2:19] In the first 90 minutes of my day and I can guarantee you in the first 90 minutes of your day [2:25] every single thing you did depended on electric power. It depended on that system [2:30] working either directly or indirectly. And if it didn't work, your day would have gone [2:35] started in a radically different way. Um uh Lots of studies have been done. About 0% of your day [2:44] is uninfluenced by real-time electric power in the power grid. Almost nothing you will do all day, [2:51] all week won't depend on electric power. So it's a system that that's pretty important to all the [2:56] things we do, not our AI ambitions. And if you think it's important to you, I'll ask you to spare [3:02] a moment and think about your kids. Um I don't have kids, but I have 24 nieces and nephews. I'm [3:07] a fun uncle. So one of the fun things I get to go do is just before they get home from school, [3:12] I'll go find the the the panel breakers in in my family's house and just throw the main switch. [3:19] And just wait. And watch as the kids come absolutely unglued because nothing in their [3:24] life that they expect is available to them. The Wi-Fi doesn't work, the computer doesn't work, [3:29] the TV doesn't work, the light switch in the basement doesn't work. They can't stare into [3:33] the refrigerator even though they're not hungry. Nothing they do works. [3:38] And if your experience is anything like mine when you do it, I encourage you to do it. [3:43] What you'll find is about 10 minutes in, there are young people sitting on a step going, "I'm bored. [3:50] I don't know what to do." Right? And you'll race and you'll turn the breaker back on and you'll [3:54] say, "I'm sorry. I'm sorry. I'm sorry. Please go back to your regularly scheduled programming." [3:59] Uh so, let's just take a moment and think about just personally, [4:02] whether you like it or not, electric power isn't just important. It's probably a dependency [4:08] for a modern life. If everything went off here, things would change quickly. [4:13] So, if we if we're going to talk about how we're going to modernize the grid, and I [4:17] keep hearing in presentation after presentation we've got to modernize the grid to make AI work, [4:21] that's probably true. It's certainly true. Let's talk about the grid that needs to get modernized. [4:26] Let's talk about the pieces and parts that that comprise or compose the the the the power grid. [4:31] So, first off is power generation. It's all the places we make all the kilowatt hours to power the [4:37] grid and all of the things we do in America and North America. Uh in 2025, the generation fleet [4:44] created about 7 and 1/2 trillion kilowatt hours. If anyone does isn't comfortable with the numbers, [4:51] 1 kilowatt hour is about enough to run your dishwasher for a cycle. 7 and 1/2 trillion [4:58] kilowatt hours. A little less than half of that came from gas. A little less than a quarter of [5:03] that came from renewables. A little less than 20% came from nuclear. Maybe 15% from coal. [5:09] Right? All of those things created the energy that makes all of this stuff happen. If you get out on [5:14] the trade show floor, everything we're talking about, that's where that energy came from. Uh [5:21] uh once we generate all that power, we have a high voltage transmission system that moves that power [5:26] around the nation. You probably don't notice it when you drive down the highway or across country, [5:31] but there's the giant pylons and towers and the huge conductors that move ultra high voltage [5:35] around from the places we want to generate it to the places we want to use it. Uh then we [5:40] have to build substations. Uh there's a there's about 500,000 mi of transmission. there's about [5:47] 55,000 substations. It's the place where all the lines connect, that's where all of our breakers [5:52] and transformers and metrology lives. That's where all the complexity of the system lives. [5:58] And then we have the distribution infrastructure. It's the infrastructure that brings power from [6:01] our substations down your neighborhood. It powers our residential, our commercial, our industrial, [6:06] our federal consumers of power, the people who represent all of the load and including a [6:11] number of data centers connect at the distribution voltages well. And not pictured, but let's spare a [6:17] moment for all of the real-time control systems, the SCADA and the ADMS and the DERM systems have [6:22] to measure and monitor and balance this whole system. But what's remarkable about the system [6:29] is that it works. I don't know that I'm surprised when the power goes out. The more I study it, [6:34] the more I learn, the more I talk to people, I'm kind of amazed it works at all. Huge, complicated, [6:40] distributed systems don't usually work this well. It works almost all the time. The fact [6:44] that it almost always works is unbelievable. The fact that it only breaks occasionally is crazy. [6:52] So so here's the thing to know. This works because there's a whole subsystem and a whole [6:57] subculture in America that does nothing but this all day, every day. You might not see them except [7:02] when a storm rolls through and they come back and replace the fuse on the pole outside your house, [7:08] but there's a whole subsystem of America that does this every single day. If you [7:13] thought back to the 1300s and someone looked at your daily life, you'd appear as a sorcerer [7:18] to them. It'd be bewildering what you do to control a force of nature with your finger. [7:24] Um so the problem isn't that the power grid is big. It's been big for a long [7:30] time. We've dealt with change for a long time. The problem is it's getting bigger, [7:35] faster in all of its parts at once. So let's talk a little bit about the challenge. Data centers, [7:41] the core infrastructure of everything we're talking about over these couple of days, [7:46] uh those data centers represent about 5% of that 7 and 1/2 trillion kilowatt hours of energy [7:52] that we produce every year. That's what they represent today. As we fast forward to 2030, [7:58] they'll represent something around 15%. The forecast is 12 to 17%. No one's really sure. [8:04] If we forecast to 2050, no one has any idea. If someone gives you a number, they don't know. [8:09] They're just They're just guessing. Nobody knows what that's going to look like. Uh but at 15% [8:15] of our power grid uh energy, that is half of the entire industrial load of America. When people say [8:24] data centers are going to consume a lot of energy, half of the entirety of the industrial [8:29] plant in America, desalinization plants, chemical plants, manufacturing plants, [8:35] everything that we do at an industrial scale, they'll represent half of that. Took us 100 years [8:41] to build that industrial plant. We'll build all these data centers in a decade. It's a crazy [8:46] number. If you forecast it all out, the amount of generation we'll have to create to power all this, [8:52] whether it's big centralized plants, distributed power, or anything else, we're going to approach [8:56] 7 trillion kilowatt hours, a 70% rise in the amount of power. So, the next time someone says, [9:02] "What should we use to generate power?" The answer is everything we can get our hands on [9:06] cuz there's no other way to make that number a reality. Will it be nuclear? Yes, it will. Will [9:11] it be renewables? Yes, it will. Will it be coal? Maybe. We'll see. Um I didn't think it'd be gas. [9:17] It turns out it's going to be gas. Um so, uh that number is going to take everything. And something [9:23] to think about is one of the things that's going to comprise a huge portion of that is renewables, [9:29] which is really good because they're clean, their carbon footprint is approaching zero long term, [9:34] uh Uh, and it's actually driving down the cost of energy. When energy gets cheaper, [9:40] we come up with very clever ways of using it to create new value. Um, and everybody knows [9:45] the downside of renewables. When they're producing, they produce a little too much. [9:49] And when they're not producing, they produce a lot too little. That instability to the system is a [9:54] challenge for us. Cuz the old power grid, the the the loads varied, and we have a generation system [10:00] that just followed it. It just worked. Now those two systems are chaotic independent of each other. [10:06] Creates a lot of challenges. There's a lot of ways to solve those challenges, but one of the ways [10:10] we're going to have to solve that challenge is by growing the transmission plant. We have 500,000 mi [10:16] of transmission by 2030, we're going to have to double it. By 2040, we might have to triple it. [10:23] Which is an insane thing to suggest cuz no one's ever done anything like that in that amount of [10:28] time. And transmission lines are kind of hard to design. They're very high voltage. They have [10:32] corona effects. They have all kinds of challenges to them. Um, but one of the big challenges is that [10:36] nobody wants them anywhere near them. Figuring out where on earth to place them turns out to be a [10:42] pretty big part of the problem. Is anyone familiar with the term NIMBY? Not in my backyard. Uh, how [10:47] about bananas? Anybody ever ever heard that one? Build absolutely nothing anywhere near anyone. [10:53] Bananas. It turns out we get in our own way a lot. So it's not just an engineering problem. [10:59] It's sort of a techno-economic societal problem that we have to solve. And all we got to do is [11:03] figure it out at a scale we've never done faster than ever. Uh, all those transmission lines [11:08] have to terminate somewhere. So our substation plant uh uh has to double or triple in capacity. [11:14] This is probably the biggest problem we face cuz uh uh side note, substation engineering, great [11:20] business to be in. Uh, we describe substation engineering as grotesquely labor intensive. Uh, [11:27] which is a great thing if you're in a services business until you run out of people. [11:32] And then nobody knows what to go do. Uh if you look on all the jobs boards, [11:37] every single company like us has an opening for 15 substation engineers. Those job openings have been [11:43] open for a year and a half. There just aren't any. So, it's not like a labor crisis is new to us. [11:50] And then you go to distribution. Uh we already have uh 5 and 1/2 million miles uh [11:55] of distribution plant we got to we got to figure out. Uh we're going to have to fortify that. It's [12:00] too brittle. We'll need to underground lots of it. So, we've got a big challenge in front of us. [12:06] The problem isn't growth, it's the pace of growth. That's what's going to get us. [12:13] And all we lack to solve that problem as a company of engineers is everything we need. [12:19] We don't have enough people. We didn't have enough people before we had to do more than ever. [12:23] Uh we don't have nearly enough people. Nobody studies this in school. There's no pipeline [12:27] of people who are magically going to come solve this for us. But if we had the people, [12:32] we wouldn't have the time. Uh the what the processes we we use to do this assume that [12:36] we have a long time to do it and we don't have a long time to do it. Uh and if we had the time, [12:41] we wouldn't have the money. We've already run out of the rate base. You've People are already [12:45] complaining about the the cost of modernizing the grid. This whole thing might cost $3 trillion. [12:51] That's not going to come out of our out of our bills anytime soon. But if we had the people, [12:56] then we had the time, and we had the money, we don't have the supply chain and they don't have [13:00] the raw materials. All we lack is all the things we need to be successful with this. [13:06] Uh let's circle back to what failure means. And I don't want [13:10] to engage in hyperbole. Nobody wants to be a scaremonger. Uh but anyone have a guess [13:15] at what percent of GDP is fundamentally dependent on real-time electric power? [13:22] I will make you guess. 85% or more. Nobody really knows. It's impossible to calculate, but something [13:27] north of 85% of America stops if we don't get this right. So, we probably need to get it right. [13:33] Failure is probably not an option. So, all right, hopefully I've convinced you that there might [13:38] be a problem here and someone should probably do something. Um so, we're trying to figure that out. [13:44] So, what should be done? Uh let's think about how we do this today. The way we've always solved this [13:50] for 100 years is that a bunch of engineers who are seasoned and know how to do this uh uh huddle [13:55] around a complicated document, right? A single line or three-line diagram and they all look at it [14:01] and try and figure out, what does this thing do? What do we need to do to this system and how do [14:05] we want to change it? And they apply their their wisdom and logic and reasoning that's in their [14:10] heads and not written down anywhere of how to go solve that problem. And then they go figure it [14:15] out. We've been doing this with people for a long time, uh but those schematics we we need to read, [14:22] frontier models don't know anything about them. They weren't trained on those. Uh we put every [14:26] single word written down, we tokenize them and we build systems around them, but all they do is tell [14:32] you the theory of electric power. They don't tell you how to solve a problem. Think of it like a [14:36] Think of it like a a doctor. You can go to medical school and you can learn all about [14:41] medicine, but if you don't know something about your patient, there's not very much you can do [14:45] until you learn something about them. We have to apply it to real discrete problems, not the [14:50] theory of electric power. Um and everything that we've got going on in enterprise AI, [14:55] turns out it's not particularly helpful in this space. So, that doesn't mean we can't [15:00] make progress. We have been making progress. We've been building a system of AI technologies, [15:04] things you would expect, right? You look at the processes our people follow. [15:08] You look at the bloated part, what is the piece of the pie that just takes all the time and labor [15:13] and effort, and you build agents and tools and algorithms and optimizations that solve those [15:18] specific problems that allows us to compress this down to something much smaller that we [15:22] can actually take on. We're making real headway on that. We're applying foundation models. We're [15:27] applying agentic reasoning. We're applying all kinds of technologies that are available to us at [15:32] the moment. The problem is that's not going to get us there. We still have to solve the hardest parts [15:39] of the problem. We're going to have to build an a cooperative agentic AI. We're going to have to [15:45] build tools that can understand terribly complex documents and reason across them, schematics that [15:52] aren't built out of words that the systems we built do today. We're going to have to build world [15:57] models. Our people know how to do this, but there aren't any words to turn into tokens to turn into [16:01] reasoning. We're going to have to learn through the experience and actions of humans. We're still [16:06] not very good at that. And then fundamentally, we're still chasing the way humans do this. [16:11] But that's not how the future's going to be. Grid foundation models will learn the physics of [16:15] systems, so we don't have to optimize a hundred million linear algebra equations in sequence to [16:21] solve a problem. We'll just infer the result because we'll train things correctly and learn [16:26] the latent state. That's where we've got to get to. And we've got a ways to go. [16:31] So, uh if you're anything like me, if you think about all this, you sort of stare it in the face, [16:36] it can be kind of daunting. I'm not sure I always thought we were going to succeed. [16:40] Sometimes it feels like this just an enormous perfectly predictable, inevitable calamity that's [16:46] coming towards you. You just can't get out of the way. And it felt like that for a little while. [16:50] But we've started to make some meaningful progress in solving these challenges. In power generation, [16:57] we've taken half of the labor out of doing feed studies. So, all the data centers that need all [17:02] the gas and and generation infrastructure, we can do twice as much with half the people. [17:07] We found We built agentic systems that allow them to get through the process much faster and just [17:12] design more data centers more quickly and all the other generation fleet we have to go build. [17:17] In transmission, and this one really gets me, uh we used to have [17:23] occasions where it would take us 150 hours to design a single mile of transmission line because [17:29] it was complicated or or frustrating, we couldn't quite get it right. We can now design transmission [17:34] lines through optimization algorithms and agents 150 miles an hour. That's a real number. [17:41] We did a project that once took us 4 years of engineering time in an afternoon. [17:48] Substations, the the the the terribly difficult space of substations, we see a path to get 30% [17:54] of the labor out of that equation by allowing humans to collaborate with agentic systems. [18:01] The agentic systems don't know what to go do, they certainly can't do it autonomously, but they can [18:05] start picking away at the burdensome labor of doing all this work. And in distribution, we're [18:12] already carving 30% of the labor out of doing this terribly high volume work of replacing poles and [18:17] making sure they don't fall over in a storm, we see a path to 60. We may see a path to 90. [18:24] So, uh I think all of this has been possible because we just owned up to the problem. As an [18:30] engineering services business, there was just no chance we were going to succeed without finding [18:35] something radically different to go do. So, we we took on the challenge. We We focused on what [18:41] the challenge is really going to be. We looked for everything in the world that might help, and [18:45] we found ways to try, fail, and find the pieces that would succeed. Lots of things have failed, [18:51] a few of them have worked spectacularly well. So, I'll ask you this, when you wake up tomorrow, [18:58] when you hit your alarm, when you turn on the light, spare a moment for the people who are [19:03] trying to figure this out cuz it's important to everybody. It's important to this industry, [19:08] it's important to the future we want to go uh see together, uh and wish us a little bit of luck cuz [19:13] we'll probably need a little bit of luck to solve this problem. When it comes right down to it, [19:19] I don't think AI will save the power grid alone, but I think together we will. Thank you.