# American Infrastructure Collection: The Physical Construction of Machine Intelligence ![[resources/american/americanforme-geographic-buildout.jpg]] > [!summary] The shortest account > The **American Infrastructure Collection** is the engineering counterpart to [[collections/American|the American Collection]]. One teaches the character, comportment, and aesthetic of a composed citizen; this one teaches the machine that citizen is now being asked to help build. It follows the stack from the transistor to the town meeting: wafers, lithography, packaging and memory; accelerators, fabrics and racks; cooling, water, power, substations and turbines; concrete, steel, sequencing and the trades; schedulers, serving engines and endpoints; and the civic layer of permits, agreements, apprenticeships and consent. The ambition behind it is plain. Intelligence capacity is survival capacity for a country, and therefore for the family inside it, and this is the infrastructure on which that capacity rests. > [!map] Canonical technical router > [[wiki/AI Infrastructure|AI Infrastructure]] is the vault's canonical infrastructure router and carries the bounded membership cluster. [[wiki/The American Buildout|The American Buildout]] supplies the civic argument; [[wiki/National Compute Capacity|National Compute Capacity]] supplies the strategic measure; [[wiki/What a Data Center Asks of a Town|What a Data Center Asks of a Town]] is the page to hand a neighbor who has just learned that something large is coming. **Enter by layer:** [[wiki/Upstream Industrial Foundation|Silicon]] · [[wiki/Electrical Supply and Reliability|Power]] · [[wiki/Facilities and Thermal Management|Facilities]] · [[wiki/Rack and Cluster Integration|Systems]] · [[wiki/Data Fabric and Storage|Storage]] · [[wiki/Compute and Accelerator Systems|Compute]] · [[wiki/Software and Orchestration|Software]] · [[wiki/Intelligence Delivery|Delivery]] · [[wiki/Cloud Deployment|Deployment]] **Enter by concern:** [[wiki/Data Center Water Use|Water]] · [[wiki/Data Center Noise|Noise]] · [[wiki/Zoning|Zoning]] · [[wiki/Community Benefit Agreement|Agreements]] · [[wiki/Apprenticeship|Apprenticeship]] · [[wiki/Energy Abundance|Energy]] · [[wiki/Building Well|Building Well]] **Adjacent collections:** [[collections/American|American Collection]] · [[collections/Machine Succession|Machine Succession]] · [[collections/The Civilian Intelligence Marketplace|The Civilian Intelligence Marketplace]] · [[collections/Neurotech|Neurotech]] **Source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] · [[articles/America's Neo-Western Renewal from the Acropolis of Athens to Eternity|America's Neo-Western Renewal from the Acropolis of Athens to Eternity]] ## Why this collection exists The bridges of the last American century crossed canyons. The bridges of this one cross something harder to photograph: the distance between a civilization that survives the century and one that does not. That is the register this collection works in, and softening it would be a disservice. The capability that designs a drug designs a pathogen; the reasoning that models an epidemic can engineer one. Work that once needed a national laboratory and a thousand cleared staff is collapsing toward work a determined individual can do alone, which means one person in a room may eventually assemble something worse than the neutron bomb. **There is no version of safety in which the defenders arrive second.** The only answer to a capability that exists is a greater capability built first, by people with reason to protect the world, and kept ahead. Between states the same logic runs at scale. For a few years after 1945 one nation held a weapon no other could answer, and the following half century was written inside that interval. Machine intelligence is broader than a bomb, because it improves everything a country does at once, including its capacity to make the next advance before a rival can; a decisive lead in it is dominion and not market share. A world where that capability is distributed unevenly enough is a world in which one civilization can put the others into a dark age and keep them there, with no appeal, for as long as it chooses. Beneath the human contest sits the older one. Life here has never had a backup. Earth has been struck before and will be struck again, the geological record is a list of extinctions, and civilization currently keeps the entire human pattern, every language and archive and lineage, on one rock in one copy with no means of restoring itself. Machine intelligence is the first instrument plausibly able to change that: to find the object early and move it, to cure what is presently incurable, to carry a planet through a shock, and eventually to place the pattern somewhere else so that one bad morning stops being the end of the story. The sooner that capability exists, the better. Hold those three together and the conclusion is unavoidable. This may be the last invention, since a sufficiently capable intelligence performs the subsequent ones itself; it may also be the last civilization, since the window in which these decisions remain open does not reopen. That is why the unglamorous particulars in this collection, the transformer lead times and the weld cleanliness standards and the apprenticeship intake, carry a weight wildly out of proportion to how they read. And it is why the duty of an American who understands the stakes is to get behind the decision to build and carry a piece of the work: when the call is to build, the answer is to build. Specific terms, high standards, and hard questions belong inside that answer, never in place of it. The second reason is simpler. This stack has become genuinely difficult to understand. A person can know what a GPU is and still have no account of why a rack weighs seven thousand pounds, why the water in the pipes never leaves, why a transformer takes three years, why an accelerator is twelve pieces of silicon, or why the fastest route to more compute runs through a scheduling improvement rather than a purchase order. Everyone involved is learning this, and nobody should have to pretend otherwise. These pages are written so that no component remains a black box after its entry is read: plain language first, every term defined before it is leaned on, mechanism in steps, and the honest evidence posture at the end. ## Three ways to read it **For the neighbor.** Begin with [[wiki/What a Data Center Asks of a Town|What a Data Center Asks of a Town]], then [[wiki/What a Data Center Gives Back|What a Data Center Gives Back]], then the specific concern: [[wiki/Data Center Water Use|water]], [[wiki/Data Center Noise|noise]], [[wiki/Backup Generator|generators]], [[wiki/Electrical Substation|the substation]]. Close with [[wiki/Infrastructure Citizenship|Infrastructure Citizenship]], which is about how to get what you want from the process rather than how to feel about it. **For the learner.** Read the layers in order, bottom to top, following the sequence below. It is designed as a curriculum: a reader who finishes holds a working model of the whole stack and can follow any technical conversation in the industry. **For the builder or official.** Go to the constraints: [[wiki/Supply Chain Chokepoint|Supply Chain Chokepoint]], [[wiki/Interconnection Queue|Interconnection Queue]], [[wiki/Transformer Supply|Transformer Supply]], [[wiki/Electrical Trades|Electrical Trades]], [[wiki/Construction Sequencing|Construction Sequencing]], [[wiki/Capital Intensity|Capital Intensity]], [[wiki/Utilization|Utilization]]. Those seven pages contain most of what decides whether a project lands on schedule. --- ## Layer 1 — The upstream industrial foundation [[wiki/Upstream Industrial Foundation|Upstream Industrial Foundation]] is the layer router. Everything above it waits on a small number of extraordinarily advanced factories, which is why an account of this stack begins in manufacturing rather than in software. | Node | What it explains | | --- | --- | | [[wiki/Silicon Wafer\|Silicon Wafer]] | The 300 mm disc every chip is printed on, and why its diameter stopped growing | | [[wiki/Semiconductor Lithography\|Semiconductor Lithography]] | Printing circuits with light, and the wavelength and aperture limits that govern it | | [[wiki/Extreme Ultraviolet Lithography\|Extreme Ultraviolet Lithography]] | Manufacturing a color of light that does not occur here, in vacuum, from tin plasma | | [[wiki/Chiplet\|Chiplet]] | The yield mathematics that ended monolithic chip design | | [[wiki/Advanced Packaging\|Advanced Packaging]] | Bonding verified dies and memory onto one substrate; the binding constraint of the mid-decade | | [[wiki/High-Bandwidth Memory\|High-Bandwidth Memory]] | Stacked memory beside the logic, and why its generations set accelerator generations | | [[wiki/Memory Wall\|Memory Wall]] | Why moving data, rather than arithmetic, governs the design of everything above | | [[wiki/Foundry\|Foundry]] · [[wiki/TSMC\|TSMC]] · [[wiki/ASML\|ASML]] · [[wiki/Intel\|Intel]] | The firms at the narrow points, and the fabless model that created them | | [[wiki/Supply Chain Chokepoint\|Supply Chain Chokepoint]] | Where the flow narrows, and how the constraint has migrated decade by decade | | [[wiki/Semiconductor Export Controls\|Semiconductor Export Controls]] · [[wiki/Critical Minerals\|Critical Minerals]] | The policy instruments and the materials beneath them | ## Layer 2 — Electrical supply and reliability [[wiki/Electrical Supply and Reliability|Electrical Supply and Reliability]] is the layer router, and electricity is the binding constraint on American compute. Read [[wiki/Energy Abundance|Energy Abundance]] for why demand returned after twenty flat years. | Node | What it explains | | --- | --- | | [[wiki/Utility Interconnection\|Utility Interconnection]] | How a very large load joins the grid, and why energization date governs the project | | [[wiki/Interconnection Queue\|Interconnection Queue]] | The ordered backlog, and the reforms and strategies that shorten the wait | | [[wiki/Electrical Substation\|Electrical Substation]] | The fenced yard a neighbor sees first, and what actually happens in it | | [[wiki/Power Transformer\|Power Transformer]] · [[wiki/Transformer Supply\|Transformer Supply]] | Voltage conversion with no moving parts, and the three-year queue for one | | [[wiki/Switchgear\|Switchgear]] · [[wiki/Medium-Voltage Distribution\|Medium-Voltage Distribution]] | Protection, isolation, and carrying power across a campus | | [[wiki/Uninterruptible Power Supply\|Uninterruptible Power Supply]] · [[wiki/Data Center Batteries\|Data Center Batteries]] | The seconds between a fault and the engines, and the chemistry that buys them | | [[wiki/Backup Generator\|Backup Generator]] · [[wiki/Automatic Transfer Switch\|Automatic Transfer Switch]] | Continuity, testing hours, permitted emissions, and the handover logic | | [[wiki/Redundant Power Feeds\|Redundant Power Feeds]] | The A and B discipline, and the common-mode failures that defeat it | | [[wiki/Busbar Busway\|Busbar Busway]] · [[wiki/Rack Power Distribution Unit\|Rack Power Distribution Unit]] | The last hundred feet, metered at every outlet | | [[wiki/High-Voltage Direct Current Distribution\|High-Voltage Direct Current Distribution]] | Why the densest racks stopped alternating | | [[wiki/Behind-the-Meter Generation\|Behind-the-Meter Generation]] · [[wiki/Turbine Supply\|Turbine Supply]] | Building your own power, and the order book that decides whether you can | | [[wiki/Power Purchase Agreement\|Power Purchase Agreement]] · [[wiki/Demand Response\|Demand Response]] · [[wiki/Grid Inertia\|Grid Inertia]] | Contracting new generation, bending load, and the spinning mass that forgives a shock | ## Layer 3 — Facilities, thermal management and water [[wiki/Facilities and Thermal Management|Facilities and Thermal Management]] is the layer router. [[wiki/Data Center Cooling|Data Center Cooling]] is the single most useful page here: the complete heat path from silicon to sky in five hand-offs. | Node | What it explains | | --- | --- | | [[wiki/Cold Plate\|Cold Plate]] | The metal block on the chip, and the joint a few microns thick that governs performance | | [[wiki/Direct-to-Chip Liquid Cooling\|Direct-to-Chip Liquid Cooling]] | Manifolds, quick disconnects, and how liquid kept its serviceability | | [[wiki/Coolant Distribution Unit\|Coolant Distribution Unit]] | The boundary between the building's water and the servers' water | | [[wiki/Air Cooling\|Air Cooling]] · [[wiki/Rear-Door Heat Exchanger\|Rear-Door Heat Exchanger]] | Forty years of containment discipline, and the ceiling that ended it | | [[wiki/Immersion Cooling\|Immersion Cooling]] | Putting the whole board in the liquid, and the operational trade that follows | | [[wiki/Chilled Water Plant\|Chilled Water Plant]] · [[wiki/Cooling Tower\|Cooling Tower]] | Making cold water, and the water-for-electricity choice at the end of the chain | | [[wiki/Data Center Water Use\|Data Center Water Use]] | Where the water actually goes, and the six numbers a site should publish | | [[wiki/Heat Reuse\|Heat Reuse]] | Selling the heat to a neighbor who wants it | | [[wiki/Power Usage Effectiveness\|Power Usage Effectiveness]] | What the industry's efficiency metric measures, and what it conceals | | [[wiki/Data Center Security\|Data Center Security]] | Perimeter, access, surveillance and procedure | ## Layer 4 — Shell, construction and the trades The building outlives four generations of equipment inside it, which makes these the longest-lived decisions on the site. [[wiki/Building Well|Building Well]] is the standard; [[wiki/Gigawatt Campus|Gigawatt Campus]] is the unit now being built in. | Node | What it explains | | --- | --- | | [[wiki/Data Center Siting\|Data Center Siting]] | Power, fiber, water, geology, labor and consent, and why brownfields won | | [[wiki/Data Center Shell\|Data Center Shell]] | Floor loading, clear height, compartmentation, and designing for the unspecified | | [[wiki/Structural Steel\|Structural Steel]] · [[wiki/Concrete in Data Center Construction\|Concrete in Data Center Construction]] | The frame and the foundation, and the cure time nothing compresses | | [[wiki/Construction Sequencing\|Construction Sequencing]] | Power blocks, long-lead items, and getting the first hall earning early | | [[wiki/Prefabrication and Modular Construction\|Prefabrication and Modular Construction]] | Building the complicated parts indoors, and why that raises quality | | [[wiki/Commissioning\|Commissioning]] | Breaking the building on purpose, under load, before anyone depends on it | | [[wiki/Fiber Route Diversity\|Fiber Route Diversity]] · [[wiki/Structured Cabling\|Structured Cabling]] | Two paths that share no trench, and labelling as a reliability practice | | [[wiki/Electrical Trades\|Electrical Trades]] · [[wiki/Pipefitting and Mechanical Trades\|Pipefitting and Mechanical Trades]] | The licensed craft that installs all of it, and what the work actually is | | [[wiki/Apprenticeship\|Apprenticeship]] · [[wiki/Data Center Workforce\|Data Center Workforce]] | How a journeyman is made, and the two employment numbers stated honestly | ## Layer 5 — Systems, racks and integration [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]] is the layer router. The decisive distinction here is between a server that works alone and a rack that is one machine. | Node | What it explains | | --- | --- | | [[wiki/Compute and Accelerator Systems\|Compute and Accelerator Systems]] | The two integration levels, and why they are different machines | | [[wiki/Rack-Scale Computing\|Rack-Scale Computing]] · [[wiki/Data Center Rack\|Data Center Rack]] | When the rack became the computer, and what the old standard could not carry | | [[wiki/Compute Tray\|Compute Tray]] | The serviceable unit, arriving plumbed and wired and leaving the same way | | [[wiki/Host CPU\|Host CPU]] · [[wiki/PCIe\|PCIe]] | The processor that feeds the accelerators, and the bus that reaches everything else | | [[wiki/Baseboard Management Controller\|Baseboard Management Controller]] | The small computer that works when the big one does not | | [[wiki/OCP Accelerator Module\|OCP Accelerator Module]] | Open hardware standards, and why a published socket is industrial policy | | [[wiki/Server OEM and System Integration\|Server OEM and System Integration]] | Who actually builds the racks, and why the chip vendors bought in | ## Layer 6 — Accelerators The table below is a market survey rather than any operator's portfolio. The accelerator field has four distinct populations, and the entries here sample all of them: merchant graphics processors sold to everyone; custom silicon designed by the largest operators for their own traffic, including a cloud operator's tensor processors, another's training and inference parts, and a third's announced accelerator, physically implemented in several cases by specialist design houses that supply the interconnect and backend expertise; dedicated inference parts from established processor vendors; and independent architectures pursuing wafer-scale construction, deterministic dataflow, very large on-device memory, or meshed cores on open instruction sets. One supplier holds well above ninety percent of data-center share as of October 2026, which is the fact that makes the other three populations strategically interesting. | Node | What it explains | | --- | --- | | [[wiki/Accelerator\|Accelerator]] | What a matrix-arithmetic processor is, and the families competing to supply it | | [[wiki/Inference Accelerator\|Inference Accelerator]] | Why answering requests rewards a different machine than training does | | [[wiki/NVIDIA\|NVIDIA]] · [[wiki/AMD\|AMD]] · [[wiki/Intel\|Intel]] | The merchant suppliers | | [[wiki/NVIDIA HGX\|NVIDIA HGX]] | The eight-accelerator baseboard that became the industry's server building block | | [[wiki/NVIDIA H100\|NVIDIA H100]] · [[wiki/NVIDIA H200\|NVIDIA H200]] | The generation most current models were trained on, and the memory-only refresh that proved where the bottleneck sits | | [[wiki/NVIDIA B200\|NVIDIA B200]] · [[wiki/NVIDIA B300\|NVIDIA B300]] | Two dies acting as one, four-bit arithmetic, and volume measured in millions | | [[wiki/NVIDIA Grace Blackwell\|NVIDIA Grace Blackwell]] · [[wiki/NVIDIA Vera Rubin\|NVIDIA Vera Rubin]] | Coherent host-accelerator coupling, and the announced successor | | [[wiki/NVIDIA RTX PRO 6000 Blackwell\|NVIDIA RTX PRO 6000 Blackwell]] | A card in a slot, for the enormous amount of real work that needs no rack | | [[wiki/AMD Instinct GPUs\|AMD Instinct]] · [[wiki/AMD MI300X\|MI300X]] · [[wiki/AMD MI350X\|MI350X]] · [[wiki/AMD MI450\|MI450]] | The second GPU supplier's line, and chiplet construction at its most legible | | [[wiki/AMD Helios\|AMD Helios]] · [[wiki/AMD EPYC\|AMD EPYC]] | A rack built from open standards, and the server processor whose recovery taught cadence | | [[wiki/Google TPU\|Google TPU]] | The first operator-designed custom accelerator, and when custom silicon pays | | [[wiki/Qualcomm Cloud AI 100 Ultra\|Qualcomm Cloud AI 100 Ultra]] | Energy per answer, from a lineage formed by battery life | | [[wiki/Tenstorrent Galaxy Blackhole\|Tenstorrent Galaxy Blackhole]] | An independent mesh architecture on open instruction sets | | [[wiki/Edge Inference\|Edge Inference]] | Accelerators in phones, cars, cameras and robots, and the physics that requires them | | [[wiki/Hyperscaler\|Operator custom silicon]] · [[wiki/Inference Accelerator\|independent architectures]] | The rest of the field: cloud operators' own training and inference parts, specialist backend design houses, wafer-scale and deterministic-dataflow startups, and reconfigurable logic where volumes are low | ## Layer 7 — Networking and fabrics [[wiki/Scale-Up and Scale-Out|Scale-Up and Scale-Out]] is the page to read first: an AI cluster has two networks, and conflating them is the most common misunderstanding of the architecture. | Node | What it explains | | --- | --- | | [[wiki/NVLink\|NVLink]] · [[wiki/NVSwitch\|NVSwitch]] | The memory fabric that makes many accelerators one, and the crossbar that scaled it | | [[wiki/Infinity Fabric\|Infinity Fabric]] · [[wiki/UALink\|UALink]] | A coherent interconnect reused at three scales, and the open alternative | | [[wiki/InfiniBand\|InfiniBand]] · [[wiki/Ultra Ethernet\|Ultra Ethernet]] | The lossless cluster fabric, and Ethernet's adoption of its principles | | [[wiki/RDMA\|RDMA]] | Writing into another machine's memory without its processor noticing | | [[wiki/Collective Communication\|Collective Communication]] | All-reduce and its relatives, and why one slow device paces a thousand | | [[wiki/Network Fabric Topology\|Network Fabric Topology]] · [[wiki/Congestion Control\|Congestion Control]] | Fat trees, rail-optimized wiring, and what happens when everyone transmits at once | | [[wiki/Data Center Optics\|Data Center Optics]] · [[wiki/DPU and SmartNIC\|DPU and SmartNIC]] | Where copper stops and light begins, and the adapter that runs the plumbing | | [[wiki/NVIDIA Quantum\|NVIDIA Quantum]] · [[wiki/NVIDIA ConnectX-8\|NVIDIA ConnectX-8]] | The switches and the endpoint silicon, with arithmetic performed inside the fabric | | [[wiki/Pensando\|Pensando]] | A programmable data plane, and the migration of cloud control points into silicon | ## Layer 8 — Storage and the data path [[wiki/Data Fabric and Storage|Data Fabric and Storage]] is the layer router. An accelerator waiting on data is the most expensive idle object in the building. | Node | What it explains | | --- | --- | | [[wiki/Storage Tiering\|Storage Tiering]] | The hierarchy from accelerator memory to cold archive, and staging ahead of need | | [[wiki/Parallel File System\|Parallel File System]] | Distributed flash, and why metadata rather than bandwidth is the hard problem | | [[wiki/NVMe SSD\|NVMe SSD]] · [[wiki/NVMe over Fabrics\|NVMe over Fabrics]] | Flash spoken to properly, locally and across the network | | [[wiki/S3-Compatible Object Storage\|S3-Compatible Object Storage]] | Storage as a key-value store, and an interface that became a standard by adoption | | [[wiki/WEKA\|WEKA]] | A distributed flash platform, and the storage field it competes in | | [[wiki/Checkpointing\|Checkpointing]] | How long runs survive failure, and why a checkpoint never restored is a hypothesis | ## Layer 9 — Software and orchestration [[wiki/Software and Orchestration|Software and Orchestration]] is the layer router, and it is where the hardware market is currently being decided. | Node | What it explains | | --- | --- | | [[wiki/CUDA\|CUDA]] · [[wiki/ROCm\|ROCm]] | Eighteen years of accumulated libraries, and the open stack built against them | | [[wiki/PyTorch\|PyTorch]] | Where models are written, and the layer at which hardware becomes substitutable | | [[wiki/OCI Container\|OCI Container]] | Immutable images, real isolation mechanisms, and why accelerator containers are harder | | [[wiki/Kubernetes\|Kubernetes]] · [[wiki/Slurm\|Slurm]] · [[wiki/Job Scheduling\|Job Scheduling]] | Two orchestration traditions, and the scheduler as the highest-leverage software on site | | [[wiki/Observability\|Observability]] · [[wiki/Site Reliability Engineering\|Site Reliability Engineering]] | Error budgets, straggler hunting, and facility telemetry read beside job metrics | | [[wiki/Failure Domain\|Failure Domain]] · [[wiki/Mean Time to Repair\|Mean Time to Repair]] | Blast radius, and why repair speed outranks failure rate at scale | | [[wiki/Supply Chain Security\|Supply Chain Security]] · [[wiki/Confidential Computing\|Confidential Computing]] | Proving hardware and software are what they claim, and computing on data nobody can read | ## Layer 10 — Intelligence delivery [[wiki/Intelligence Delivery|Intelligence Delivery]] is the layer router. This is where most of the world's compute is now spent, and where the economics of the buildout are settled. | Node | What it explains | | --- | --- | | [[wiki/Prefill and Decode\|Prefill and Decode]] | Two physically different phases, one compute-bound and one memory-bound | | [[wiki/KV Cache\|KV Cache]] · [[wiki/Continuous Batching\|Continuous Batching]] | The memory that dominates serving, and the technique that multiplied throughput | | [[wiki/Quantization\|Quantization]] | Fewer bits per number as a bandwidth strategy, and the honest accuracy position | | [[wiki/vLLM\|vLLM]] · [[wiki/SGLang\|SGLang]] | Paged cache management, radix prefix reuse, and why open engines mattered strategically | | [[wiki/Inference Serving\|Inference Serving]] · [[wiki/Inference Request Routing\|Inference Request Routing]] | Turning weights into a service, and making a mixed fleet behave as one | | [[wiki/Inference Endpoint\|Inference Endpoint]] · [[wiki/Retrieval-Augmented Generation\|Retrieval-Augmented Generation]] | The boundary where the stack's complexity is discharged, and grounding answers in a corpus | | [[wiki/Model Router\|Model Router]] · [[wiki/Accelerator Selection\|Accelerator Selection]] | Which model answers, and which silicon runs it | | [[wiki/Managed Inference Service\|Managed Inference Service]] | Buying answers rather than accelerators, and why most institutions start there | | [[wiki/Private Model Hosting\|Private Model Hosting]] · [[wiki/Sovereign AI\|Sovereign AI]] | What a hospital, a bank, or an agency actually requires before it can adopt this | ## Layer 11 — Deployment and market structure [[wiki/Cloud Deployment|Cloud Deployment]] is the layer router. [[wiki/Bare Metal Compute|Bare Metal Compute]] · [[wiki/Neocloud|Neocloud]] · [[wiki/Hyperscaler|Hyperscaler]] · [[wiki/Colocation|Colocation]] · [[wiki/Edge Computing|Edge Computing]] · [[wiki/Google Distributed Cloud|Google Distributed Cloud]] · [[wiki/Server OEM and System Integration|Server OEM and System Integration]] The structure matters nationally. Hyperscale operators have the balance sheets to commit to generation and fabs faster than any public process; specialist operators and colocation distribute that capacity to universities, hospitals, state agencies, manufacturers, and companies of ordinary size. Breadth of access, rather than total installed scale, determines what the country actually gets; see [[wiki/National Compute Capacity|National Compute Capacity]]. ## The economics | Node | What it explains | | --- | --- | | [[wiki/Capital Intensity\|Capital Intensity]] | Where the money goes, and why the building is the cheap part | | [[wiki/Depreciation of Compute Assets\|Depreciation of Compute Assets]] | The short clock on accelerators, and the cascade that makes capability cheap later | | [[wiki/Utilization\|Utilization]] | The four ways to measure it, the seven ways it leaks, and why software beats hardware | ## The civic and cultural layer This is the half the engineering accounts leave out, and it decides whether the buildout is something Americans are proud of or something that happened to them. | Node | What it carries | | --- | --- | | [[wiki/The American Buildout\|The American Buildout]] | What is being built, why it fell to this generation, what choosing well and badly produce | | [[wiki/National Compute Capacity\|National Compute Capacity]] | Intelligence capacity as survival capacity, and the dimensions that actually measure it | | [[wiki/Energy Abundance\|Energy Abundance]] | A century of electricity and prosperity, and the case for abundance over sufficiency | | [[wiki/Industrial Competence\|Industrial Competence]] | Capability held in people, how it thinned, and how it is rebuilt | | [[wiki/The Trades as Civic Inheritance\|The Trades as Civic Inheritance]] | Craft as a civic institution, and what it offers the person who enters it | | [[wiki/Infrastructure Citizenship\|Infrastructure Citizenship]] | Reading the documents, asking the specific question, and taking part in the work | | [[wiki/What a Data Center Asks of a Town\|What a Data Center Asks of a Town]] | An honest inventory, and the ten specific terms worth requiring | | [[wiki/What a Data Center Gives Back\|What a Data Center Gives Back]] | What arrives automatically, what must be asked for, and what it costs to grant | | [[wiki/Building Well\|Building Well]] | The decisions that are nearly free before the concrete and impossible after it | | [[wiki/Gigawatt Campus\|Gigawatt Campus]] | The unit the industry now builds in, and why the gigawatt became the measure | ## Seams into the AI superpower ontology This collection is the component-level floor beneath the vault's existing strategic ontology, and the two are designed to be read together. Each node below carries a reciprocal seam section pointing back into these layers. | Strategic node | Where it becomes physical here | | --- | --- | | [[wiki/AI Stack\|AI Stack]] | The layer routers are this abstraction's concrete form, and [[wiki/Supply Chain Chokepoint\|Supply Chain Chokepoint]] is its inherited-bottleneck mechanism observed migrating decade by decade | | [[wiki/AI Factory\|AI Factory]] | [[wiki/Gigawatt Campus\|Gigawatt Campus]], [[wiki/Rack-Scale Computing\|Rack-Scale Computing]], [[wiki/Data Center Cooling\|Data Center Cooling]], [[wiki/Construction Sequencing\|Construction Sequencing]] | | [[wiki/Energy-Compute Nexus\|Energy-Compute Nexus]] | [[wiki/Interconnection Queue\|Interconnection Queue]], [[wiki/Transformer Supply\|Transformer Supply]], [[wiki/Turbine Supply\|Turbine Supply]], [[wiki/Grid Inertia\|Grid Inertia]], [[wiki/Energy Abundance\|Energy Abundance]] | | [[wiki/AI Economy\|AI Economy]] | [[wiki/Capital Intensity\|Capital Intensity]], [[wiki/Depreciation of Compute Assets\|Depreciation of Compute Assets]], [[wiki/Utilization\|Utilization]] | | [[wiki/Intelligence as a Utility\|Intelligence as a Utility]] | [[wiki/Intelligence Delivery\|Intelligence Delivery]], [[wiki/Managed Inference Service\|Managed Inference Service]], [[wiki/Inference Endpoint\|Inference Endpoint]] and the cost mechanisms beneath them | | [[wiki/Compute Sovereignty\|Compute Sovereignty]] | [[wiki/National Compute Capacity\|National Compute Capacity]], and [[wiki/Accelerator Selection\|Accelerator Selection]] with the open standards that lower [[wiki/Exit Cost\|Exit Cost]] | | [[wiki/Citizen AI Literacy\|Citizen AI Literacy]] | [[wiki/Infrastructure Citizenship\|Infrastructure Citizenship]] and [[wiki/What a Data Center Asks of a Town\|What a Data Center Asks of a Town]] | | [[wiki/Public AI Trust\|Public AI Trust]] | [[wiki/Building Well\|Building Well]], [[wiki/Community Benefit Agreement\|Community Benefit Agreement]], and measurement published rather than promised | | [[wiki/AI Infrastructure Consent\|AI Infrastructure Consent]] | [[wiki/Zoning\|Zoning]], [[wiki/Permitting\|Permitting]], [[wiki/Environmental Review\|Environmental Review]], [[wiki/What a Data Center Gives Back\|What a Data Center Gives Back]] | | [[wiki/NVIDIA Open Agent Safety Platform\|NVIDIA Open Agent Safety Platform]] | [[wiki/Inference Serving\|Inference Serving]] policy enforcement and [[wiki/SGLang\|SGLang]] constrained decoding | | [[wiki/AI Transition Governance\|AI Transition Governance]] | The instruments already exercised weekly: permitting, review, zoning, interconnection, host agreements | | [[wiki/Public Epistemic Capacity\|Public Epistemic Capacity]] | [[wiki/Infrastructure Citizenship\|Infrastructure Citizenship]], where the documents are public before construction and the numbers become enforceable | The connective argument is [[wiki/Compute Sovereignty\|compute sovereignty]] read downward. Sovereignty is the practical ability to substitute rather than the location of a rack, and substitutability is manufactured in specific places: at the socket through [[wiki/OCP Accelerator Module\|open module standards]], at the rack and the fabric through [[wiki/UALink\|consortium interconnect]] and [[wiki/Ultra Ethernet\|standards-aligned Ethernet]], at the software layer through [[wiki/ROCm\|open accelerator stacks]], [[wiki/PyTorch\|framework neutrality]], and [[wiki/vLLM\|open serving engines]], and at the operations layer through [[wiki/Accelerator Selection\|selection]] that makes a mixed fleet usable. Every one of those is an unglamorous engineering decision performing strategic work. ## How this collection relates to the American Collection [[collections/American|The American Collection]] teaches a standard of conduct: [[american/American Ethos - Vision and Doctrine|the doctrine]], [[american/American Character - The American Trust Code|the character]], [[american/American Comportment - The Nine Domains|the comportment]], [[american/American Aesthetic - Design and Material Culture|the aesthetic]]. This collection teaches the works that standard is now being applied to. They are one project read from two directions, and several of the American axioms turn out to be engineering specifications when applied here. [[wiki/Culture as Infrastructure Precondition|Culture as Infrastructure Precondition]] states the relation directly: the cultural work is upstream of the industrial capability, and this buildout is the proper and natural product of that re-centering rather than a separate undertaking that happens to require workers. Large works require a people capable of making them, and the shortages now binding on the American grid are the evidence, since the decades in which practical craft lost standing are the same decades in which apprenticeship intake fell. [[american/American Ethos - Axiom 02 - Competence Is Beautiful|Competence Is Beautiful]] is the argument of [[wiki/Building Well|Building Well]]: the screened yard, the quiet fans, the labelled cable, the published noise measurement. [[american/American Ethos - Axiom 04 - Utility Is Beauty Under Pressure|Utility Is Beauty Under Pressure]] is what [[wiki/Commissioning|commissioning]] tests by failing a building on purpose before anyone depends on it. [[american/American Ethos - Axiom 03 - Trust Is an Aesthetic|Trust Is an Aesthetic]] is why a [[wiki/Community Benefit Agreement|community agreement]] written in numbers with published reporting is worth more than any assurance. [[american/American Ethos - Founding Condition 03 - The Geographic Buildout|The Geographic Buildout]] named this material map before these pages described its components. And [[american/American Ethos - Cultural Thesis 10 - Old Intelligence Reawakened|Old Intelligence Reawakened]] is the trades: practical knowledge that institutional life once transmitted reliably, now needed again in quantity. ## Evidence and interpretation boundary - A vendor's performance or superiority statement is a claim, attributed and graded, until it is independently demonstrated. - An announced system is distinct from a shipping one, and a shipping one from a deployed fleet. - Specifications are cited by generation and date; figures that are not sourced are omitted rather than estimated. - A manufacturer's capacity, an operator's installed base, and a facility's actual consumption are three different quantities. - Analyst projections of spending, market share, and demand are tiered as plausible rather than established. - Published macroeconomic loss estimates for disruption scenarios are models, not measurements. - Where a mechanism is reconstructed generically because the specific implementation is undisclosed, the entry says so. - Time-sensitive 2026 claims are datestamped, because this is a field in which a figure ages within a quarter. The collection describes industry practice rather than any single operator's practice, and the technical substance is presented as the engineering of the field as a whole. ## Router role This is the canonical collection for the physical and institutional construction of machine intelligence in America. The wiki owns the reusable definitions; the linked articles own their arguments; [[collections/American|the American Collection]] owns the standard of conduct; [[collections/Machine Succession|Machine Succession]] owns the question of what inherits civilizational process. This collection owns the machine: what it is made of, how it is built, what it asks of a place, what it gives back, and why a country that intends to survive the century is building it now.