{
  "schemaVersion": "1.0",
  "kind": "nilu.article",
  "id": "datacenters-sistema-fisico-demanda",
  "slug": "data-centers-how-they-work-and-why-they-are-expanding",
  "locale": "en",
  "sourceLocale": "pt-br",
  "localizationStatus": "translated",
  "title": "Data centers — how they work and why they are expanding",
  "summary": "A guide to servers, racks, workloads, power, cooling, and the physical supply chain that constrains data-center expansion.",
  "author": "NiLu",
  "publishedAt": "2026-07-29T00:00:00.000Z",
  "updatedAt": "2026-07-29T00:00:00.000Z",
  "canonicalUrl": "https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding/",
  "representations": {
    "html": "https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding/",
    "markdown": "https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding.md",
    "json": "https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding.json"
  },
  "bodyMarkdown": "# Data centers — how they work and why they are expanding\n\nA guide to servers, racks, workloads, power, cooling, and the physical supply chain that constrains data-center expansion.\n\nAuthor: NiLu\nLocale: en\nPublished: 2026-07-29\nUpdated: 2026-07-29\nCanonical: https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding/\nMarkdown: https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding.md\nJSON: https://blog.nilu.pro/en/blog/data-centers-how-they-work-and-why-they-are-expanding.json\n\n## In brief\n\n- A data center combines IT equipment with power, cooling, connectivity, security, and operations.\n- Component, server, rack, and facility power are different measurements.\n- AI racks can be very dense without representing the typical facility.\n- Capacity depends on several industrial layers; the bottleneck can shift between them.\n\n## Contents\n\n- [The system](#o-sistema)\n- [Types and metrics](#tipos-e-metricas)\n- [AI architectures](#arquiteturas-de-ia)\n- [Benchmarks and efficiency](#benchmarks-e-eficiencia)\n- [Dependency chain](#cadeia-de-dependencias)\n- [What is still not measured](#o-que-ainda-nao-e-mensurado)\n\n<a id=\"o-sistema\"></a>\n\n## The system\n\nA data center concentrates information-technology equipment and the infrastructure required to keep it powered, connected, cooled, and available. The IT layer includes servers, accelerators, memory, storage, and networking. The facility adds power conversion, UPS, distribution, redundancy, cooling, physical security, and the building. [[1]](#ref-1) [[2]](#ref-2)\n\nCloud is not an application equivalent to AI, video, or databases. It is a delivery and operating model in which different workloads share infrastructure. [[3]](#ref-3)\n\n<a id=\"tipos-e-metricas\"></a>\n\n## Types and metrics\n\nFacilities may be enterprise, colocation, hyperscale, edge, or specialised. These categories depend on the question being analysed and do not necessarily form mutually exclusive groups. [[2]](#ref-2)\n\nIt is also necessary to distinguish:\n\n- component power;\n- server power;\n- rack power;\n- total facility power;\n- consumption over time;\n- computational work performed.\n\nTDP does not measure server or site consumption. PUE describes the ratio between total facility energy and IT energy within a defined boundary; it does not measure software efficiency, workload value, or full environmental impact. [[1]](#ref-1) [[4]](#ref-4)\n\n<a id=\"arquiteturas-de-ia\"></a>\n\n## AI architectures\n\nSystems with a high density of accelerators change the rack. The NVL72 configuration integrates compute trays, NVLink switches, power, and liquid cooling, with an approximate power of 120 kW in NVIDIA documentation. It is a specific configuration. [[5]](#ref-5)\n\nIn the 2025 Uptime survey, 82% of 709 operators said that the highest density deployed in their facilities remained below 30 kW. The survey measures the highest density per responding facility, not the distribution of all racks. High-density products and the typical rack therefore belong to different populations. [[6]](#ref-6)\n\n<a id=\"benchmarks-e-eficiencia\"></a>\n\n## Benchmarks and efficiency\n\nMLPerf Inference measures performance on defined models and scenarios. When a power submission is included, the system's AC power is measured during execution. Comparisons require compatible models, precision, scenarios, software, and configurations. [[7]](#ref-7) [[8]](#ref-8)\n\nA benchmark result should not be converted directly into annual consumption, average utilisation, or the energy use of the entire data center. [[7]](#ref-7) [[8]](#ref-8)\n\n<a id=\"cadeia-de-dependencias\"></a>\n\n## Dependency chain\n\nCapacity depends on a sequence that may include foundries, advanced packaging, HBM, networking chips, optics, servers, racks, cooling, electrical equipment, construction, and commissioning. Corporate reports demonstrate dependencies and limited supply at some stages, but do not provide a comparable global average for lead times. [[9]](#ref-9) [[10]](#ref-10) [[11]](#ref-11)\n\nThe bottleneck can shift as one layer expands faster than another.\n\n<a id=\"o-que-ainda-nao-e-mensurado\"></a>\n\n## What is still not measured\n\nThere is no robust public source for globally dividing capacity among AI, video, databases, web services, analytics, HPC, storage, and other workloads. The canonical research treats this absence as Uncertainty, not as permission to approximate shares through proxies.\n\nThere is also no comparable global measurement of:\n\n- server and accelerator utilisation;\n- idle power;\n- rack-density distribution;\n- lead times by layer;\n- installed versus usable capacity.\n\nAdoption proxies will not be presented as physical or energy shares.\n\n## References\n\n1. <a id=\"ref-1\"></a>[U.S. EPA ENERGY STAR — Data center rating FAQ and PUE definition](https://nepis.epa.gov/Exe/ZyPURL.cgi?Dockey=P100TSZE.TXT). U.S. EPA. Accessed July 24, 2026.\n2. <a id=\"ref-2\"></a>[LBNL — 2024 United States Data Center Energy Usage Report](https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report). Lawrence Berkeley National Laboratory. Accessed July 24, 2026.\n3. <a id=\"ref-3\"></a>[NIST — The Definition of Cloud Computing (SP 800-145)](https://csrc.nist.gov/pubs/sp/800/145/final). NIST. Accessed July 24, 2026.\n4. <a id=\"ref-4\"></a>[NVIDIA — H100 Tensor Core GPU specifications](https://www.nvidia.com/en-us/data-center/h100/). NVIDIA. Accessed July 22, 2026.\n5. <a id=\"ref-5\"></a>[NVIDIA — DGX GB Rack Scale Systems Hardware Guide](https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html). NVIDIA. Accessed July 24, 2026.\n6. <a id=\"ref-6\"></a>[Uptime Institute — 2025 Global Data Center Survey](https://uptimeinstitute.com/about-ui/press-releases/uptimes-15th-annual-global-data-center-survey-results-shows-both-commitment-and-hesitancy). Uptime Institute. Accessed July 24, 2026.\n7. <a id=\"ref-7\"></a>[MLCommons — MLPerf Inference benchmark documentation](https://docs.mlcommons.org/inference/). MLCommons. Accessed July 24, 2026.\n8. <a id=\"ref-8\"></a>[MLCommons — MLPerf Inference power measurement methodology](https://docs.mlcommons.org/inference/power/). MLCommons. Accessed July 24, 2026.\n9. <a id=\"ref-9\"></a>[NVIDIA — FY2026 Form 10-K](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm). NVIDIA / SEC. Accessed July 24, 2026.\n10. <a id=\"ref-10\"></a>[AMD — FY2025 Form 10-K](https://www.sec.gov/Archives/edgar/data/2488/000000248826000018/amd-20251227.htm). AMD / SEC. Accessed July 24, 2026.\n11. <a id=\"ref-11\"></a>[Arista Networks — FY2025 Form 10-K](https://www.sec.gov/Archives/edgar/data/1596532/000159653226000013/anet-20251231.htm). Arista / SEC. Accessed July 24, 2026.\n"
}