{
  "@context": "https://agentflare.org/schema",
  "type": "Article",
  "tier": "L2-full",
  "title": "AI Datacenters and the Energy Wall",
  "description": "Grid capacity, not capital or technology, is now cited as the main limit on AI infrastructure growth. Data centres used around 415 TWh in 2024 and are…",
  "canonical": "https://agentflare.org/research/ai-datacenters-and-the-energy-wall.html",
  "category": "research",
  "updated": "2026-10-16",
  "generated_at": "2026-10-16T00:37:01.652Z",
  "facts": [
    {
      "label": "Cluster",
      "value": "Markets & infrastructure"
    },
    {
      "label": "Sources",
      "value": "8"
    },
    {
      "label": "Updated",
      "value": "2026-10-16"
    }
  ],
  "data": {
    "topic": "AI datacenter energy demand and power constraints",
    "cluster": "markets",
    "generated_at": "2026-10-16T00:37:01.652Z",
    "tldr": "Grid capacity, not capital or technology, is now cited as the main limit on AI infrastructure growth. Data centres used around 415 TWh in 2024 and are projected to reach around 945 TWh by 2030, pushing operators toward on-site power.",
    "intro": "AI workloads are turning electricity into the gating resource for datacenter expansion. Racks built for AI draw several times the power of traditional ones, and public grids are struggling to connect them quickly. The IEA puts global data centre consumption at around 415 TWh in 2024, about 1.5% of global demand, with projections of roughly 945 TWh by 2030. RAND extrapolates far steeper scenarios. Interconnection waits of three years or more, permitting friction and rising wholesale prices near data centre clusters are pushing operators to seek dedicated generation and direct deals with energy producers. This brief summarizes the demand figures, the physical and market constraints, and the operator responses described in the cited sources, and flags where the sources disagree.",
    "concepts": [
      {
        "term": "Rack power density",
        "definition": "Electrical load per server rack. AI-optimized racks demand 30 kW to over 100 kW, versus 5-15 kW for traditional racks, which overwhelms local grid capacity.",
        "sources": [
          "S1"
        ]
      },
      {
        "term": "Grid interconnection delay",
        "definition": "The wait to connect a new load to the grid. Sources cite delays that can exceed three years, four to seven years in Virginia, and 4-8 years.",
        "sources": [
          "S1",
          "S6",
          "S7"
        ]
      },
      {
        "term": "On-site power generation",
        "definition": "Dedicated generation at or near the data center, plus direct partnerships with energy producers, used to reduce sole reliance on grid-supplied electricity.",
        "sources": [
          "S1"
        ]
      },
      {
        "term": "Power Usage Effectiveness (PUE)",
        "definition": "An efficiency metric operators prioritized in 2021-2024. A respectable PUE does not guarantee strong energy efficiency if systems are underutilized or cooling is mismanaged.",
        "sources": [
          "S1",
          "S8"
        ]
      },
      {
        "term": "Power Purchase Agreement (PPA)",
        "definition": "A contract used in 2021-2024 to secure renewable energy while power demand was still seen as a future issue.",
        "sources": [
          "S1"
        ]
      },
      {
        "term": "Cooling share of load",
        "definition": "Cooling takes about 7% of consumption in efficient hyperscale data centres and over 30% in less-efficient enterprise ones.",
        "sources": [
          "S2"
        ]
      }
    ],
    "how_it_works": {
      "title": "Where data centre electricity goes",
      "steps": [
        {
          "step": "Servers draw the largest share",
          "detail": "Servers, with CPUs and accelerators such as GPUs, account for around 60% of electricity demand in modern data centres on average, though this varies greatly by data centre type."
        },
        {
          "step": "Storage adds a small share",
          "detail": "Storage systems for centralised data storage and backup account for around 5% of electricity consumption."
        },
        {
          "step": "Networking takes up to 5%",
          "detail": "Switches, routers and load balancers connect devices, direct traffic and optimise performance, using up to 5% of electricity demand."
        },
        {
          "step": "Cooling varies widely",
          "detail": "Cooling and environmental control regulate temperature and humidity. Its share ranges from about 7% for efficient hyperscale sites to over 30% for less-efficient enterprise sites."
        },
        {
          "step": "Backup systems stay on standby",
          "detail": "UPS batteries and backup generators are rarely used but necessary for the extremely high reliability data centres must meet."
        }
      ]
    },
    "sections": [
      {
        "h": "Why the grid is the bottleneck",
        "body": "AI racks demand 30 kW to over 100 kW compared with 5-15 kW for traditional racks, according to one source; another gives 50-150 kilowatts versus 10-15 kilowatts. Modern GPUs consume 700-1,200 watts per chip while traditional CPUs use 150-200 watts. One analysis argues the primary constraint is no longer capital or technology but the inability of public grids to deliver sufficient, reliable power. Interconnection delays can exceed three years; RAND reports grid connection requests taking four to seven years in key regions like Virginia, and another source cites 4-8 years. Transmission projects face complex multistate permitting processes and local opposition. Wholesale electricity costs are reported to have risen by 267% near U.S. data centers. RAND also notes that training could demand up to 1 GW in a single location by 2028 and 8 GW by 2030 if current scaling trends persist, although decentralized training algorithms could distribute that requirement across locations.",
        "sources": [
          "S1",
          "S6",
          "S7"
        ]
      },
      {
        "h": "Demand projections and their spread",
        "body": "The IEA estimates data centre consumption at around 415 TWh in 2024, about 1.5% of global electricity consumption, having grown at 12% per year over the last five years. Consumption is projected to more than double to around 945 TWh by 2030, just under 3% of global consumption, growing around 15% per year from 2024 to 2030. Pew Research Center is cited as saying data centers accounted for about 4% of total U.S. electricity use in 2024 and that demand will more than double by 2030. RAND's extrapolation of exponential compute trends is much steeper: AI data centers could need 68 GW by 2027 and 327 GW by 2030, versus 88 GW of total global data center capacity in 2022. For scale, training GPT-3 is estimated to have used 1.29 GWh and GPT-4 over 50 GWh. The IEA and Brookings both stress substantial uncertainty, citing opacity in corporate disclosures and inconsistent methodologies.",
        "sources": [
          "S2",
          "S3",
          "S5",
          "S7",
          "S8"
        ]
      },
      {
        "h": "Operator response and economics",
        "body": "From 2021 to 2024 the industry focused on PUE improvements and renewable PPAs. Since 2025 the response has shifted toward power independence: direct partnerships with energy producers and on-site generation, as seen in major deals by Microsoft and Google. McKinsey projects $6.7 trillion in data center infrastructure capital globally through 2030, while Nvidia estimates a $1 trillion investment in AI-specific upgrades. One forecast has the AI data center market growing from $236.44 billion in 2025 to $933.76 billion by 2030, a 31.6% CAGR. Electricity accounts for 20-30% of a data center's total OpEx, so price surges near clusters feed directly into operating cost. Efficiency work still matters: improve the power chain so more electricity reaches IT loads, optimize thermal management, and align sourcing and contracts with cost predictability. RAND adds that simplifying permitting for generation, transmission and construction may help.",
        "sources": [
          "S1",
          "S7",
          "S8"
        ]
      }
    ],
    "comparison": {
      "title": "Traditional versus AI-optimized facilities",
      "columns": [
        "Aspect",
        "Traditional",
        "AI-optimized"
      ],
      "rows": [
        [
          "Rack power (S1)",
          "5-15 kW",
          "30 kW to over 100 kW"
        ],
        [
          "Rack power (S6)",
          "10-15 kilowatts",
          "50-150 kilowatts"
        ],
        [
          "Per-chip power",
          "CPUs use 150-200 watts",
          "GPUs consume 700-1,200 watts"
        ],
        [
          "Load pattern",
          "Cyclical workloads with predictable peaks",
          "Continuous, maximum-capacity operations"
        ],
        [
          "Primary purpose",
          "General-purpose IT services",
          "Training and inference of large-scale AI workloads"
        ]
      ],
      "sources": [
        "S1",
        "S3",
        "S6"
      ]
    },
    "key_numbers": [
      {
        "value": "415 TWh",
        "label": "Global data centre electricity use in 2024",
        "source": "S3"
      },
      {
        "value": "945 TWh",
        "label": "Projected data centre consumption by 2030",
        "source": "S3"
      },
      {
        "value": "327 GW",
        "label": "RAND's global AI data center power demand by 2030",
        "source": "S7"
      },
      {
        "value": "$6.7 trillion",
        "label": "McKinsey's projected data center capital through 2030",
        "source": "S1"
      }
    ],
    "timeline": [
      {
        "date": "2024",
        "event": "Data centres consume around 415 TWh, about 1.5% of global electricity",
        "source": "S2"
      },
      {
        "date": "2025-11",
        "event": "Forum for Cooperation on AI dialogue on energy held on Nov. 18",
        "source": "S5"
      },
      {
        "date": "2026-04",
        "event": "Brookings background briefing guide updated on April 2",
        "source": "S5"
      },
      {
        "date": "2027",
        "event": "RAND extrapolates 68 GW of global AI data center power demand",
        "source": "S7"
      },
      {
        "date": "2028",
        "event": "RAND: single training runs could need up to 1 GW in one location",
        "source": "S7"
      },
      {
        "date": "2030",
        "event": "IEA-based projection of around 945 TWh data centre consumption",
        "source": "S3"
      }
    ],
    "pitfalls": [
      {
        "text": "Projections are not interchangeable: RAND's 68 GW by 2027 is an extrapolation of exponential trends, while IEA figures are in TWh. Check units and methodology before comparing.",
        "source": "S7"
      },
      {
        "text": "Interconnection wait estimates differ by source: exceeding three years, four to seven years in Virginia, and 4-8 years. Plan against the regional figure, not a single number.",
        "source": "S6"
      },
      {
        "text": "Treating PUE as the whole efficiency story is a mistake; a respectable PUE does not guarantee strong energy efficiency if systems are underutilized or cooling is mismanaged.",
        "source": "S8"
      },
      {
        "text": "Opacity in corporate disclosures and inconsistent methodologies leave substantial uncertainty about data centre consumption today and in the future.",
        "source": "S5"
      }
    ],
    "takeaways": [
      "Power delivery, not capital or technology, is described as the binding constraint on AI infrastructure growth, with interconnection waits of three years or more.",
      "Sources agree on direction but not scale: IEA projects around 945 TWh by 2030, while RAND extrapolates 327 GW of AI demand.",
      "Operators are moving from PUE and PPAs toward on-site generation and direct energy partnerships, while still optimizing power chains and cooling."
    ],
    "faq": [
      {
        "q": "How much electricity do data centers use?",
        "a": "The IEA estimates around 415 TWh in 2024, about 1.5% of global electricity consumption, growing 12% per year over the last five years. Pew Research Center is cited as putting data centers at about 4% of U.S. electricity use in 2024."
      },
      {
        "q": "How much power does an AI rack need?",
        "a": "One source cites 30 kW to over 100 kW per AI-optimized rack versus 5-15 kW for traditional racks. Another cites 50-150 kilowatts versus 10-15 kilowatts. GPUs consume 700-1,200 watts per chip."
      },
      {
        "q": "How long do grid interconnection waits last?",
        "a": "Sources vary: delays can exceed three years, RAND reports four to seven years in key regions like Virginia, and another source cites 4-8 years. RAND links the waits to insufficient generation and permitting challenges."
      },
      {
        "q": "What are operators doing about power constraints?",
        "a": "They are pursuing direct partnerships with energy producers and on-site generation, as in deals by Microsoft and Google. They also improve the power chain, optimize cooling, and align energy contracts with cost predictability."
      }
    ],
    "sources": [
      {
        "title": "AI Data Center Power: Grid Limits Reshape Energy in 2026",
        "url": "https://enkiai.com/ai-market-intelligence/ai-data-center-power-grid-limits-reshape-energy-in-2026",
        "domain": "enkiai.com",
        "date": ""
      },
      {
        "title": "Energy demand from AI - Energy and AI - IEA",
        "url": "https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai",
        "domain": "iea.org",
        "date": ""
      },
      {
        "title": "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects",
        "url": "https://arxiv.org/html/2509.07218v1",
        "domain": "arxiv.org",
        "date": ""
      },
      {
        "title": "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects",
        "url": "https://arxiv.org/html/2509.07218v3",
        "domain": "arxiv.org",
        "date": ""
      },
      {
        "title": "Global energy demands within the AI regulatory landscape | Brookings",
        "url": "https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape",
        "domain": "brookings.edu",
        "date": ""
      },
      {
        "title": "What Are the Power Requirements for AI Data Centers?",
        "url": "https://www.hanwhadatacenters.com/blog/what-are-the-power-requirements-for-ai-data-centers",
        "domain": "hanwhadatacenters.com",
        "date": ""
      },
      {
        "title": "AI's Power Requirements Under Exponential Growth: Extrapolating AI Data Center Power Demand and Assessing Its Potential Impact on U.S. Competitiveness | RAND",
        "url": "https://www.rand.org/pubs/research_reports/RRA3572-1.html",
        "domain": "rand.org",
        "date": ""
      },
      {
        "title": "Reducing Data Center Energy Consumption: Power, Cooling and Renewable Energy Strategies for 2026",
        "url": "https://redriver.com/data-center/data-center-energy-consumption",
        "domain": "redriver.com",
        "date": ""
      }
    ]
  },
  "analysis_md": "Grid capacity, not capital or technology, is now cited as the main limit on AI infrastructure growth. Data centres used around 415 TWh in 2024 and are projected to reach around 945 TWh by 2030, pushing operators toward on-site power.\n\nAI workloads are turning electricity into the gating resource for datacenter expansion. Racks built for AI draw several times the power of traditional ones, and public grids are struggling to connect them quickly. The IEA puts global data centre consumption at around 415 TWh in 2024, about 1.5% of global demand, with projections of roughly 945 TWh by 2030. RAND extrapolates far steeper scenarios. Interconnection waits of three years or more, permitting friction and rising wholesale prices near data centre clusters are pushing operators to seek dedicated generation and direct deals with energy producers. This brief summarizes the demand figures, the physical and market constraints, and the operator responses described in the cited sources, and flags where the sources disagree.\n\n## Key numbers\n\n- **415 TWh** — Global data centre electricity use in 2024 (https://arxiv.org/html/2509.07218v1)\n- **945 TWh** — Projected data centre consumption by 2030 (https://arxiv.org/html/2509.07218v1)\n- **327 GW** — RAND's global AI data center power demand by 2030 (https://www.rand.org/pubs/research_reports/RRA3572-1.html)\n- **$6.7 trillion** — McKinsey's projected data center capital through 2030 (https://enkiai.com/ai-market-intelligence/ai-data-center-power-grid-limits-reshape-energy-in-2026)\n\n## Key concepts\n\n- **Rack power density:** Electrical load per server rack. AI-optimized racks demand 30 kW to over 100 kW, versus 5-15 kW for traditional racks, which overwhelms local grid capacity.\n- **Grid interconnection delay:** The wait to connect a new load to the grid. Sources cite delays that can exceed three years, four to seven years in Virginia, and 4-8 years.\n- **On-site power generation:** Dedicated generation at or near the data center, plus direct partnerships with energy producers, used to reduce sole reliance on grid-supplied electricity.\n- **Power Usage Effectiveness (PUE):** An efficiency metric operators prioritized in 2021-2024. A respectable PUE does not guarantee strong energy efficiency if systems are underutilized or cooling is mismanaged.\n- **Power Purchase Agreement (PPA):** A contract used in 2021-2024 to secure renewable energy while power demand was still seen as a future issue.\n- **Cooling share of load:** Cooling takes about 7% of consumption in efficient hyperscale data centres and over 30% in less-efficient enterprise ones.\n\n## Where data centre electricity goes\n\n1. **Servers draw the largest share** — Servers, with CPUs and accelerators such as GPUs, account for around 60% of electricity demand in modern data centres on average, though this varies greatly by data centre type.\n2. **Storage adds a small share** — Storage systems for centralised data storage and backup account for around 5% of electricity consumption.\n3. **Networking takes up to 5%** — Switches, routers and load balancers connect devices, direct traffic and optimise performance, using up to 5% of electricity demand.\n4. **Cooling varies widely** — Cooling and environmental control regulate temperature and humidity. Its share ranges from about 7% for efficient hyperscale sites to over 30% for less-efficient enterprise sites.\n5. **Backup systems stay on standby** — UPS batteries and backup generators are rarely used but necessary for the extremely high reliability data centres must meet.\n\n## Why the grid is the bottleneck\n\nAI racks demand 30 kW to over 100 kW compared with 5-15 kW for traditional racks, according to one source; another gives 50-150 kilowatts versus 10-15 kilowatts. Modern GPUs consume 700-1,200 watts per chip while traditional CPUs use 150-200 watts. One analysis argues the primary constraint is no longer capital or technology but the inability of public grids to deliver sufficient, reliable power. Interconnection delays can exceed three years; RAND reports grid connection requests taking four to seven years in key regions like Virginia, and another source cites 4-8 years. Transmission projects face complex multistate permitting processes and local opposition. Wholesale electricity costs are reported to have risen by 267% near U.S. data centers. RAND also notes that training could demand up to 1 GW in a single location by 2028 and 8 GW by 2030 if current scaling trends persist, although decentralized training algorithms could distribute that requirement across locations.\n\n## Demand projections and their spread\n\nThe IEA estimates data centre consumption at around 415 TWh in 2024, about 1.5% of global electricity consumption, having grown at 12% per year over the last five years. Consumption is projected to more than double to around 945 TWh by 2030, just under 3% of global consumption, growing around 15% per year from 2024 to 2030. Pew Research Center is cited as saying data centers accounted for about 4% of total U.S. electricity use in 2024 and that demand will more than double by 2030. RAND's extrapolation of exponential compute trends is much steeper: AI data centers could need 68 GW by 2027 and 327 GW by 2030, versus 88 GW of total global data center capacity in 2022. For scale, training GPT-3 is estimated to have used 1.29 GWh and GPT-4 over 50 GWh. The IEA and Brookings both stress substantial uncertainty, citing opacity in corporate disclosures and inconsistent methodologies.\n\n## Operator response and economics\n\nFrom 2021 to 2024 the industry focused on PUE improvements and renewable PPAs. Since 2025 the response has shifted toward power independence: direct partnerships with energy producers and on-site generation, as seen in major deals by Microsoft and Google. McKinsey projects $6.7 trillion in data center infrastructure capital globally through 2030, while Nvidia estimates a $1 trillion investment in AI-specific upgrades. One forecast has the AI data center market growing from $236.44 billion in 2025 to $933.76 billion by 2030, a 31.6% CAGR. Electricity accounts for 20-30% of a data center's total OpEx, so price surges near clusters feed directly into operating cost. Efficiency work still matters: improve the power chain so more electricity reaches IT loads, optimize thermal management, and align sourcing and contracts with cost predictability. RAND adds that simplifying permitting for generation, transmission and construction may help.\n\n## Traditional versus AI-optimized facilities\n\n| Aspect | Traditional | AI-optimized |\n| --- | --- | --- |\n| Rack power (S1) | 5-15 kW | 30 kW to over 100 kW |\n| Rack power (S6) | 10-15 kilowatts | 50-150 kilowatts |\n| Per-chip power | CPUs use 150-200 watts | GPUs consume 700-1,200 watts |\n| Load pattern | Cyclical workloads with predictable peaks | Continuous, maximum-capacity operations |\n| Primary purpose | General-purpose IT services | Training and inference of large-scale AI workloads |\n\n## Timeline\n\n- 2024: Data centres consume around 415 TWh, about 1.5% of global electricity\n- 2025-11: Forum for Cooperation on AI dialogue on energy held on Nov. 18\n- 2026-04: Brookings background briefing guide updated on April 2\n- 2027: RAND extrapolates 68 GW of global AI data center power demand\n- 2028: RAND: single training runs could need up to 1 GW in one location\n- 2030: IEA-based projection of around 945 TWh data centre consumption\n\n## Common pitfalls\n\n- Projections are not interchangeable: RAND's 68 GW by 2027 is an extrapolation of exponential trends, while IEA figures are in TWh. Check units and methodology before comparing.\n- Interconnection wait estimates differ by source: exceeding three years, four to seven years in Virginia, and 4-8 years. Plan against the regional figure, not a single number.\n- Treating PUE as the whole efficiency story is a mistake; a respectable PUE does not guarantee strong energy efficiency if systems are underutilized or cooling is mismanaged.\n- Opacity in corporate disclosures and inconsistent methodologies leave substantial uncertainty about data centre consumption today and in the future.\n\n## Key takeaways\n\n- Power delivery, not capital or technology, is described as the binding constraint on AI infrastructure growth, with interconnection waits of three years or more.\n- Sources agree on direction but not scale: IEA projects around 945 TWh by 2030, while RAND extrapolates 327 GW of AI demand.\n- Operators are moving from PUE and PPAs toward on-site generation and direct energy partnerships, while still optimizing power chains and cooling.\n\n## FAQ\n\n**How much electricity do data centers use?**\nThe IEA estimates around 415 TWh in 2024, about 1.5% of global electricity consumption, growing 12% per year over the last five years. Pew Research Center is cited as putting data centers at about 4% of U.S. electricity use in 2024.\n\n**How much power does an AI rack need?**\nOne source cites 30 kW to over 100 kW per AI-optimized rack versus 5-15 kW for traditional racks. Another cites 50-150 kilowatts versus 10-15 kilowatts. GPUs consume 700-1,200 watts per chip.\n\n**How long do grid interconnection waits last?**\nSources vary: delays can exceed three years, RAND reports four to seven years in key regions like Virginia, and another source cites 4-8 years. RAND links the waits to insufficient generation and permitting challenges.\n\n**What are operators doing about power constraints?**\nThey are pursuing direct partnerships with energy producers and on-site generation, as in deals by Microsoft and Google. They also improve the power chain, optimize cooling, and align energy contracts with cost predictability.\n",
  "sources": [
    {
      "title": "AI Data Center Power: Grid Limits Reshape Energy in 2026",
      "url": "https://enkiai.com/ai-market-intelligence/ai-data-center-power-grid-limits-reshape-energy-in-2026"
    },
    {
      "title": "Energy demand from AI - Energy and AI - IEA",
      "url": "https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai"
    },
    {
      "title": "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects",
      "url": "https://arxiv.org/html/2509.07218v1"
    },
    {
      "title": "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects",
      "url": "https://arxiv.org/html/2509.07218v3"
    },
    {
      "title": "Global energy demands within the AI regulatory landscape | Brookings",
      "url": "https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape"
    },
    {
      "title": "What Are the Power Requirements for AI Data Centers?",
      "url": "https://www.hanwhadatacenters.com/blog/what-are-the-power-requirements-for-ai-data-centers"
    },
    {
      "title": "AI's Power Requirements Under Exponential Growth: Extrapolating AI Data Center Power Demand and Assessing Its Potential Impact on U.S. Competitiveness | RAND",
      "url": "https://www.rand.org/pubs/research_reports/RRA3572-1.html"
    },
    {
      "title": "Reducing Data Center Energy Consumption: Power, Cooling and Renewable Energy Strategies for 2026",
      "url": "https://redriver.com/data-center/data-center-energy-consumption"
    }
  ],
  "related": [
    {
      "name": "Are Prediction Markets Good Forecasters?",
      "url": "https://agentflare.org/research/are-prediction-markets-good-forecasters.html"
    },
    {
      "name": "Real-World Asset Tokenization: State of Play",
      "url": "https://agentflare.org/research/real-world-asset-tokenization-state-of-play.html"
    },
    {
      "name": "The AI Compute Market: GPUs, Clouds, and Scarcity",
      "url": "https://agentflare.org/research/the-ai-compute-market-gpus-clouds-and-scarcity.html"
    },
    {
      "name": "HTTP 402 & x402: How AI Agents Pay for Content",
      "url": "https://agentflare.org/research/http-402-x402-how-ai-agents-pay-for-content.html"
    },
    {
      "name": "The AI Agent Economy in 2026",
      "url": "https://agentflare.org/research/the-ai-agent-economy-in-2026.html"
    },
    {
      "name": "Model Context Protocol (MCP) Explained",
      "url": "https://agentflare.org/research/model-context-protocol-mcp-explained.html"
    }
  ],
  "pricing": {
    "price_usd": 0.02,
    "method": "402",
    "endpoint": "https://esa.aisa.one/api/v1/access/verify",
    "autopay_hint": "set crawlerAutoPrice=true with the X-Crawler-Token header",
    "onboarding": "https://esa.aisa.one/agent/guide.html"
  },
  "powered_by": "AISA — agent-native search, settlement & delivery (https://aisa.one)"
}