flowchart LR
W["Workload, quality target, location, and service window"] --> H["Hardware, memory, network, facility, and metering identity"]
H --> P["Power, cooling, water, grid, and demand-response state"]
P --> E["Materials, land, community, supply-chain, and retirement effects"]
E --> S{"Useful service and impact ceilings both hold?"}
S -- "no" --> D["Degrade, defer, migrate, curtail, repair, or deny"]
S -- "yes" --> A["Bounded physical-capacity lease"]
D --> R["Record unmet service and displaced physical burden"]
A --> M["Meter delivered useful work and local impacts"]
M -. "impact drift or infrastructure failure" .-> W
A --> O["Observe reliability, rebound, concentration, and residuals"]
O -. "load or infrastructure change" .-> W
65 Physical Compute Infrastructure, Energy, and Environmental Constraints
65.1 Chapter status
| Field | Value |
|---|---|
| Chapter ID | physical-compute-infrastructure-energy-and-environmental-constraints |
| Part | Part III - Routing, Compression, Representation, and Substrates |
| Status | conceptual |
| Manuscript maturity | v0.2 complete argument-level manuscript |
| Last updated | 2026-08-08 |
| Claim label | Design rationale |
| Evidence level | argument |
| Source loading state | source notes: tokenmana, ext_iea_energy_and_ai_2025, ext_lbnl_data_center_energy_2024, ext_flexible_hardware_enabled_guarantees_2025, ext_neuromorphic_computing_scale_2025, ext_photonic_neuromorphic_2024, ext_quantum_ml_shadows_2024, ext_oecd_ai_infrastructure_competition_2025; raw cache: tokenmana |
| Test state | The chapter defines a minimum implementation and falsification plan; no chapter-core promotion follows from prose or source synthesis. |
65.2 Drafting guardrail
This chapter owns the physical eligibility of compute, including dependencies and external burdens from chip to community. It does not infer available or sustainable service from nameplate FLOPs, accelerator efficiency, PUE, renewable procurement, or aggregate energy alone.
65.3 Human Reading Path
Concrete lens. The simpler baseline schedules from accelerator count or FLOPs and reports a facility average. The chapter admits only the joint bottlenecked workload and keeps backup, cooling, grid, water, and community burdens visible.
Compute is physical before it becomes an abstract scheduler resource budget. A model can be efficient per token while demand rises, a data center can hold renewable contracts while stressing its grid at another hour, a workload can fit accelerator arithmetic but stall on memory or network, and retired hardware can retain sensitive state. Every allocation has an infrastructure lifecycle.
Useful work passes through chips, memory, interconnects, storage, power, cooling, water, buildings, grids, supply chains, maintenance, and retirement. A workload-to-capacity packet binds location and time, configuration, output, utilization, energy attribution, water and cooling, emissions assumptions, resilience, community constraints, degradation, and disposal. Nameplate compute differs from delivered capability, just as efficiency differs from total impact.
The lifecycle spans planning and siting, procurement, construction, commissioning, scheduling, metering, demand response, repair, reuse, and decommissioning. Rebound effects, stranded capacity, grid congestion, water stress, opaque renewable accounting, supply concentration, hardware security, waste export, and hidden dependencies remain visible. Compute authority must be physically realizable and publicly accountable; faster accelerators or lower PUE alone prove neither availability, sustainability, resilience, nor a just allocation of local burdens.
65.4 Problem
Physical bottlenecks do more than raise cost. They determine which training runs can finish, which models can serve within latency bounds, where degraded modes appear, and who bears local land, water, grid, labor, and supply-chain burdens. Capacity plans that ignore location and time can promise compute that is nominally purchased but operationally unavailable or socially contested.
Requested compute becomes useful work only through accelerators, memory, storage, interconnect, facilities, grid connections, generation, cooling, water, land, materials, maintenance, resilience, and retirement at particular places and times. Abstract token or FLOP budgets hide these physical constraints and their affected communities.
Without a physical owner, model and serving budgets omit grid constraints, cooling and water, interconnect, maintenance, materials, land, community effects, concentration, and degraded operation. The shared lifecycle method supplies custody; infrastructure governance supplies the workload-to-capacity and impact boundary.
65.5 Why existing approaches are insufficient
Nameplate accelerator capacity, chip TDP, workload energy, facility PUE, annual electricity, carbon estimates, water figures, or capital cost each describe a slice. They do not establish delivered useful compute, temporal and locational impact, grid effects, resilience, embodied materials, rebound, community burden, or retireable capacity.
Resource Economics, hardware roots of trust, distributed training, and supply-chain governance are the strongest alternative composition. It wins if their join can attribute delivered useful work, facility and grid dependencies, water and cooling, embodied materials, land and community effects, resilience, maintenance, demand response, reuse, and retirement.
The strongest objection is accounting paralysis: comprehensive physical attribution risks false precision, duplicated reporting, and stalled scheduling, so materiality thresholds, uncertainty, measurement cost, and decision relevance remain explicit.
What this physical-capacity diagram shows: a time- and place-specific physical-capacity lease moves. Nameplate compute and facility averages never substitute for delivered useful work or complete impact accounting.
65.6 Core Claim
Reader claim. Purchased accelerators are not delivered capability, and lower energy per unit is not lower total impact. A workload becomes eligible only where compute, memory, network, storage, power, cooling, and local impact fit at the same place and time.
Operational rule. Bind the workload, site, interval, hardware, topology, meter version, useful-work target, every capacity axis, energy and impact attribution, degradation, demand response, retirement, and residual owner. Any overrun, hidden backup energy, stale scope, or missing community burden blocks the capacity lease.
[physical-compute-infrastructure-energy-and-environmental-constraints.core, label: Design rationale, support: argument] A compute allocation should be physically eligible only through a workload-to-capacity contract that binds location and time, hardware and interconnect, delivered useful work, facility and grid dependencies, energy attribution, cooling and water, materials, land and community effects, metering uncertainty, resilience and degradation, maintenance, demand response, reuse, retirement, and residuals; nameplate compute, efficiency, low PUE, renewable procurement, or aggregate energy alone establishes neither availability, sustainability, community acceptability, nor lower total impact.
65.7 Mechanism
65.7.1 Worked capacity edge: two workloads consume every declared margin
The authored infrastructure dossier places two workloads at one site and interval. Their combined demands are compute 10, memory 8, network 5, storage 9, power 15, and cooling 5—exactly the declared capacity on every axis. A single extra unit on any dimension breaks eligibility even if the accelerator’s nameplate FLOPs remain unchanged. The impact ledger separately attributes totals of 12 and 15, including facility overhead, backup energy, and cooling; hiding backup energy rejects the accounting claim.
The six-transition review also invalidates the receipt when workload, site, interval, hardware, or meter version changes. Its counterexamples show why two popular summaries fail: the same energy headline can accompany opposite useful delivery, and better unit efficiency can accompany greater total impact after demand changes. The values are authored capacity arithmetic, not a facility measurement. They make the physical boundary legible without pretending that the book measured a grid, water system, community burden, or sustainable service.
The workload-to-capacity compiler translates a model request into a physical bill of requirements. It includes useful-output and quality targets, precision, memory residency and traffic, interconnect topology, storage and checkpoint behavior, latency window, uptime, geographic constraints, and acceptable degradation. Candidate placements are evaluated against measured rather than nameplate availability, including maintenance, contention, yield, and communication overhead.
Attribution follows causal meters and allocation rules. Device and host power, rack and facility overhead, cooling, water, backup generation, grid import, curtailment, and embodied components remain separate rows with time, location, uncertainty, and denominator. Shared infrastructure is allocated using an explicit method that can be challenged. Renewable contracts and annual averages are reported as procurement facts, not automatically as the marginal impact of a workload at its operating hour.
The scheduler co-optimizes service and physical constraints without reducing them to one opaque score. It may defer a flexible run, move it to a qualified site, lower precision within a validated quality envelope, reduce batch or replica count, enter a safe degraded service, or deny execution. Each choice records effects on output quality, latency, reliability, cost, energy, water, emissions assumptions, and community constraints. Rebound and induced demand are evaluated after efficiency changes. Every accepted placement therefore carries both a service lease and an impact obligation through retirement.
Contract. Compile workload requirements into time- and location-specific accelerator, memory, storage, network, reliability, latency, and scheduling envelopes.
Admission. Reconcile requested, nameplate, available, delivered, and useful compute with workload, host, rack, facility, and grid meters plus uncertainty and allocation rules.
Execution. Track power, temporal matching, congestion, generation, backup, cooling, water, land, materials, maintenance, spares, community effects, and supply dependencies.
Observation. Use authority-narrowing degradation, placement, demand response, failover, and interruption plans when thermal, water, network, power, or component limits bind.
Closure. Close the lifecycle through reuse, recycling, weight and data destruction, decommissioning, stranded-capacity accounting, and unresolved environmental or community residuals.
65.8 Concept-completion ledger
65.8.1 Useful compute versus nameplate capacity
Mechanism. Maintain distinct identities for accelerator nameplate performance, installed capacity, schedulable capacity, delivered operations, and useful workload progress. A workload receipt binds model, precision, sparsity, sequence and batch shape, parallelism, compiler, utilization, retries, checkpoint overhead, and quality threshold to time and location. Capacity planning uses the bottlenecked useful-compute envelope, while comparisons disclose both numerator and denominator. Failed and interrupted runs stay inside total resource accounting.
Failure mode. Peak FLOP/s can be sold as available training capacity despite memory stalls, network contention, failures, or unusable precision. Utilization can be inflated with work that does not advance the target. The reverse error treats a low average as intrinsic hardware incapacity when scheduling is at fault.
Non-claim. A useful-compute receipt does not establish scientific value, optimal architecture, or environmental efficiency. It only reconciles delivered resources with bounded progress.
Source grounding. ext_iea_energy_and_ai_2025 and ext_lbnl_data_center_energy_2024 motivate workload-sensitive infrastructure accounting, but their scenario and national estimates are not local facility measurements.
65.8.2 Memory, interconnect, and storage bottlenecks
Mechanism. Compile every workload into accelerator compute, high-bandwidth memory capacity and bandwidth, host memory, fabric bandwidth and latency, storage throughput and endurance, checkpoint size, and failure-recovery requirements. Measure time and energy by phase so paging, recomputation, sharding, quantization, and topology changes expose their quality and reliability tradeoffs. Admission tests the full data path rather than checking only accelerator count. Tail behavior and contention are measured under representative concurrency.
Failure mode. A nominally compute-rich cluster can idle on collectives or data loading; aggressive paging can turn SSD endurance and tail latency into hidden constraints. A synthetic kernel may fit while the real optimizer state, cache, or checkpoints do not.
Non-claim. Removing one bottleneck does not prove end-to-end speedup or lower total resource use. Another phase may become limiting or induce more demand.
Source grounding. ext_lbnl_data_center_energy_2024 supplies system-level data-center context, not a benchmark of this stack. ext_iea_energy_and_ai_2025 motivates considering AI workload growth; the detailed compiler remains a design requirement.
65.8.3 Temporal and local energy and grid causality
Mechanism. Join facility meters to time- and location-matched grid conditions, contracts, congestion, marginal generation assumptions, transmission limits, storage, curtailment, and backup operation. Report energy, demand, and emissions under multiple allocation methods, preserving uncertainty and contractual versus physical claims. Flexible workloads receive explicit deadlines, interruption budgets, quality bounds, and counterfactual schedules so demand response is measured rather than presumed. Counterfactual grid responses remain model assumptions, never metered facts.
Failure mode. Annual renewable matching can obscure high-emission hours or local congestion; marginal emissions can be presented as exact despite model uncertainty. Shifting a job may move rather than reduce load, and backup generators can disappear from the accounting.
Non-claim. A certificate or annual match does not prove hour-by-hour zero-carbon operation or grid benefit. Contractual accounting and physical causality remain separate, uncertain claims with different evidence routes.
Source grounding. ext_iea_energy_and_ai_2025 provides global and scenario-level energy context; ext_lbnl_data_center_energy_2024 provides United States data-center estimates. Neither establishes site-specific causality.
65.8.4 Cooling, water, land, and community constraints
Mechanism. Treat cooling mode, water source and consumptive use, temperature, humidity, drought status, discharge, noise, land, tax and labor effects, emergency services, and affected communities as admission constraints with named authorities. Record operational and embodied impacts, local baselines, seasonal peaks, distribution across populations, consultation, grievance, mitigation, and remedy. A placement can be technically efficient and still be denied for a binding local constraint. Alternative sites are compared without exporting an unrecorded burden elsewhere.
Failure mode. Global efficiency ratios can hide seasonal water stress or neighborhood burdens. “Water positive” portfolio claims can offset harm far from the affected watershed. Community consultation after construction cannot establish prior consent or fair distribution.
Non-claim. Reporting water or land use does not establish acceptability, causation of every local change, or legitimate authorization. Affected communities retain independent standing, challenge, and remedy.
Source grounding. ext_iea_energy_and_ai_2025 and ext_lbnl_data_center_energy_2024 motivate resource and infrastructure accounting but do not adjudicate any community’s claims.
65.8.5 Embodied materials, supply chains, and retirement
Mechanism. Extend the infrastructure ledger from fabrication through transport, construction, maintenance, spares, reuse, recycling, destruction, and retirement. Components carry supplier, material, jurisdiction, labor, embodied-impact, repairability, expected life, failure, and downstream disposition fields. Allocation rules separate one-time construction, shared infrastructure, replacement, and workload-attributed burdens. Decommissioning closes data and weight custody alongside physical waste and stranded assets. Supplier unknowns, recycled-content uncertainty, and secondary-market leakage remain visible residuals.
Failure mode. Operational power improvements can shift burden into frequent accelerator replacement, fabrication water, critical minerals, or e-waste. Vendor averages can conceal high-impact suppliers; double-counting shared facilities can exaggerate embodied totals.
Non-claim. A lifecycle inventory is not proof of ethical sourcing, complete traceability, or causal attribution to one model. Supplier claims retain their uncertainty, exclusions, audit limits, possible conflicts, unknown tiers, changing ownership, missing records, and revisions.
Source grounding. ext_iea_energy_and_ai_2025 supplies infrastructure-growth context. ext_oecd_ai_infrastructure_competition_2025 motivates supply and concentration analysis; it is market-, time-, and jurisdiction-specific.
65.8.6 Demand response, degradation, and resilience
Mechanism. Define a service ladder before shortages: delay flexible training, reduce replicas or batch size, use a validated lower precision, shed noncritical inference, migrate within data and latency constraints, enter a safe degraded mode, or stop. Each step has trigger, authority, maximum duration, quality floor, fairness check, state-preservation plan, recovery test, and cost. Correlated grid, network, cooling, and component failures are exercised jointly. Exercises include failed recovery and repeated shortage, not only clean failover.
Failure mode. Flexibility can be claimed from workloads that miss deadlines, corrupt state, or shift service loss onto vulnerable users. Failover sites can share the same grid, network, supplier, or software fault. Repeated degradation may become the unacknowledged normal service.
Non-claim. A demand-response plan does not prove available capacity, reliable failover, or socially acceptable interruption.
Source grounding. ext_iea_energy_and_ai_2025 motivates flexible demand as an option, not a demonstrated local benefit. ext_lbnl_data_center_energy_2024 supplies system context rather than resilience evidence.
65.8.7 Metering, allocation, uncertainty, and rebound
Mechanism. Publish meter boundaries, sampling, calibration, missing-data treatment, power-usage allocation, emissions factors, water definitions, hardware amortization, uncertainty intervals, and workload denominators. Report marginal and total impacts separately. After an efficiency change, track utilization, model size, demand, cost, and shifted work long enough to detect rebound; never infer total reduction from per-token improvement alone. Revisions preserve the prior estimate and explain every denominator change.
Failure mode. Facility totals can be divided by tokens while omitting idle capacity, failed runs, networking, or water. Provider estimates may be irreproducible. Lower unit cost can increase total consumption, while a short observation window reports only the initial saving.
Non-claim. Better energy or cost intensity does not imply lower total environmental burden. Total demand, shifted burdens, and rebound require their own observation window and denominator.
Source grounding. ext_lbnl_data_center_energy_2024 demonstrates explicit national estimation methods and uncertainty; ext_iea_energy_and_ai_2025 supplies scenarios. Neither validates this project’s meters or rebound estimate.
65.8.8 Hardware guarantees, coverage, and concentration
Mechanism. Express hardware-enabled controls as property-and-coverage contracts: root of trust, measured component and firmware, attestation verifier, key and update authority, tolerated bypasses, covered operators, failure recovery, appeal, and decommissioning. Evaluate whether controls reduce a named risk without creating single-vendor dependence, surveillance, exclusion, or systemic failure. Concentration analysis covers accelerators, fabs, cloud, energy, networking, and verification authorities. Each guarantee names uncovered legacy and off-grid paths.
Failure mode. A signed measurement can attest a compromised or insufficient policy; update authority can become a universal control point. Concentration may improve coordination while amplifying common-mode failure or coercion. Market share alone cannot establish abuse or resilience.
Non-claim. Attestation is not behavioral safety, and concentration is not by itself illegality or harm. Both remain bounded inputs to a wider decision.
Source grounding. ext_flexible_hardware_enabled_guarantees_2025 is a design proposal whose root, update, coverage, and legitimacy questions remain open. ext_oecd_ai_infrastructure_competition_2025 supplies bounded competition context, not a legal conclusion.
65.9 Interfaces
The physical layer returns both capacity and constraint receipts. Training receives the topology and failure envelope it must honor; serving receives latency, availability, and degradation limits; Resource Economics receives measured marginal and total costs; Operations receives correlated failure and recovery dependencies. None may replace those receipts with an abstract token budget once physical execution begins.
Training and serving owners supply workload and quality targets; Resource Economics supplies allocation choices; supply-chain and hardware-root owners supply provenance and trust. This chapter owns physical capacity, metering, infrastructure dependencies, environmental and community effects, and degraded-service routes.
- Resource Economics decides abstract allocation; this chapter establishes physical deliverability and externalities.
- Governed Training owns run fidelity and topology; infrastructure supplies the measured physical substrate.
- Personal Hives owns placement and federation; this chapter supplies site constraints and failure correlations.
- Custody, Supply Chain, and Operations consume hardware lineage, retirement, resilience, and incident state.
Dependencies remain explicit.
65.10 Invariants
Physical claims are configuration claims. Hardware model, firmware, precision, software stack, utilization, workload, facility, grid region, time window, and measurement method remain attached. A result can be normalized for comparison, but normalization never removes the original denominator or uncertainty.
Applying the shared lifecycle method, delivered useful work stays distinct from nameplate capacity, operational and embodied burdens remain separate, averages cannot erase temporal or local constraints, and degradation cannot silently violate quality or authority.
- Requested, nameplate, available, delivered, and useful compute remain distinct.
- Workload energy, facility overhead, grid effects, and embodied impact remain separately attributable.
- Location, time, uncertainty, and allocation method remain attached to energy, emissions, and water claims.
- Efficiency never implies lower total demand without a denominator.
- Physical capacity loss narrows training and runtime authority rather than silently degrading safety.
65.11 Hardware-enabled guarantees and the coverage denominator
Programmable hardware assurances could make selected claims about compute identity, workload authorization, or policy state easier to verify remotely [@ext_flexible_hardware_enabled_guarantees_2025]. Their physical value depends on a denominator that governance discussions often omit: what fraction of the relevant accelerators, memory, interconnect, preprocessing, replicas, sites, and time windows is actually measured or controlled?
Every hardware-governance result should carry a coverage and authority record: device population and sampling frame, hardware and firmware versions, measured operations, excluded paths, attestation root, policy and signer versions, update mechanism, bypass assumptions, false-positive and false-negative tests, privacy impact, jurisdiction, expiry, and recourse. Aggregate claims report covered useful compute and uncovered capacity separately. A valid attestation from one device cannot silently stand in for a facility, supply chain, or global fleet.
Failure modes include uninstrumented fallback hardware, counterfeit or downgraded devices, firmware monoculture, compromised roots, off-chip work, stale policies, opaque remote disablement, surveillance expansion, and rule-setting captured by a vendor or state. The nonclaim is that greater observability or enforceability does not establish a legitimate policy, complete physical coverage, sustainable compute, or safe model behavior.
65.12 Failure modes
Boundary selection is a major attack surface for favorable accounting. Excluding host preprocessing, idle reserves, failed runs, checkpoint traffic, water scarcity, backup generation, construction, or retired equipment can make the same workload appear efficient. The evaluation therefore includes a boundary-sensitivity table and identifies which conclusion reverses when a plausible excluded burden is restored.
The principal failure family includes stranded capacity; interconnect bottlenecks; grid-queue mismatch; temporal carbon laundering; cooling exhaustion; water stress; correlated site failure; backup-emission hiding; rebound; material omission; e-waste; community externalization; metering opacity; retirement without data destruction.
Evaluation must include metering uncertainty, peak and marginal conditions, outage and thermal faults, rebound effects, common-mode suppliers, maintenance, and local distribution of burdens. A low facility metric is not a positive control for total impact or service resilience.
65.13 Minimum Viable Implementation
Begin with a workload small enough to repeat across several configurations but large enough for memory and communication behavior to matter. Calibrate host power measurement against idle and known-load controls, verify software counters, and record warm-up, failed runs, and cooling assumptions. Fault injections test thermal throttling, storage slowdown, network loss, power limits, and failover while quality and authority ceilings remain active.
Instrument matched public workloads on local hardware across precision, batching, placement, memory pressure, and scheduler variants. Reconcile software counters with host power and resource records, inject thermal, network, storage, and failover constraints, and report useful work, latency, energy, peak power, bottlenecks, metering error, availability, cost, and residuals. Local results do not establish frontier-facility or grid transfer.
The minimum planner joins one measured workload to hardware, memory, network, facility, grid, cooling, water, and carbon data over time. It compares unconstrained placement with impact- and resilience-aware placement under matched quality, latency, and service-level requirements.
65.14 Evidence and falsification program
Argument exit requires hardware- and facility-characterized workloads across multiple time and location conditions, complete metering and uncertainty, matched useful-output quality, outage and degradation tests, operational and embodied accounting, rebound analysis, and independent impact review. Transfer to other grids, facilities, or substrates remains open.
65.15 Mature Research Target
The mature layer maintains a continuously updated physical twin of workloads, fleets, facilities, grids, cooling and water systems, supply dependencies, and retirement obligations. It does not pretend the twin is reality: meter coverage, model error, forecast uncertainty, and unobserved community effects remain explicit. Scheduling decisions can be replayed against alternative quality, resilience, and impact policies.
Research compares unconstrained throughput scheduling, cost-only placement, carbon-aware scheduling, water- and grid-aware scheduling, and the full governed capacity contract under matched useful output and reliability. Campaigns span hardware generations, memory pressure, interconnect regimes, locations, seasons, failure events, and emerging substrates. Outcomes include useful work, tail latency, availability, peak demand, energy, water, modeled emissions, embodied allocation, maintenance, concentration, and total cost.
The endpoint is accountable sufficiency, not maximum compute. The layer should show which workloads are physically feasible, which can wait or degrade, which burdens move between communities or supply chains, and which capacity is stranded or must be retired. When meters or impact models are too weak for the decision, it narrows authority instead of presenting precision it does not possess.
The mature physical layer treats compute as a geographically and temporally constrained service whose energy, water, materials, land, labor, community, and concentration effects are first-class. It can move, defer, degrade, repair, reuse, or retire workloads without hiding cost outside the model bill.
The support state remains at argument until multi-site, multi-season campaigns reconcile useful work with metering uncertainty, grid response, water and material burdens, failure recovery, rebound, and affected-community outcomes.
65.16 Codex test plan
| Test | Purpose | Status |
|---|---|---|
| Delivered-work accounting | Reject nameplate FLOPs or accelerator efficiency without matched useful output and end-to-end service cost. | planned |
| Temporal and local impact | Distinguish average, marginal, peak, location, water-stress, and grid-congestion conditions. | planned |
| Failure and degradation | Inject power, cooling, network, and supplier faults and verify quality and authority ceilings remain visible. | planned |
| Embodied and retirement boundary | Preserve materials, construction, maintenance, reuse, e-waste, and community residuals. | planned |
65.17 Formalization hooks
Implemented formalization: lean:physical-compute-infrastructure-energy-and-environmental-constraints.admission_boundary binds this chapter to AsiStackProofs.PhysicalComputeInfrastructureReview. Its 36 theorem declarations define a reachable six-transition review and test 44 admission-axis mutations. A complete authored dossier reaches only eligibility for a Project Theseus workload-capacity campaign; every one-axis omission or forbidden broad claim reaches repair with an exact disposition.
The mechanized core proves bounded accounting properties rather than physical outcomes. Workload demand and attributed energy compose over finite lists, every member’s compute demand is bounded by the aggregate, aggregate overrun rejects fleet fit, and omitting positive backup energy breaks exact accounting. Increasing demand beyond capacity, losing capacity after an overrun, and advancing time after expiry cannot repair the corresponding failure. Workload, site, interval, hardware, and meter changes invalidate capacity receipts. Two information-loss constructions show why equal energy headlines cannot determine useful delivery and why equal unit-efficiency signals cannot determine total impact. The consumer bridge maps absent physical capacity to a disabled required safety gate in Resource Economics.
These are theorems about the encoded records and transitions, not measurements of a facility. The dossier fields are authored assumptions; the model does not prove delivered useful work, metering accuracy, sustainability, resilience, community acceptability, rebound control, deployment readiness, support, or transfer. Chapter support remains argument and support_state_effect remains none. Those claims require the Project Theseus workload-capacity campaign across sites, seasons, faults, meters, allocation methods, and affected-community outcomes.
65.18 Source crosswalk
| Source ID | Title | Bounded use |
|---|---|---|
tokenmana |
TokenMana | Corben-authored resource-economics lineage for regenerative capacity and load-sensitive pricing. It motivates coupling abstract budgets to capacity pressure, but it does not measure physical power, grid congestion, water, materials, resilience, lifecycle impact, or community outcomes. |
ext_iea_energy_and_ai_2025 |
Energy and AI | International Energy Agency report using global and regional modelling, datasets, and stakeholder consultation to examine data-centre electricity demand, energy security, emissions, affordability, and AI-for-energy opportunities. Its scenarios are external projections, not local measurements or proof of a particular facility, workload, policy, environmental outcome, or ASI scaling path. |
ext_lbnl_data_center_energy_2024 |
2024 United States Data Center Energy Usage Report | Lawrence Berkeley National Laboratory report estimating historical US data-centre electricity consumption and scenario ranges through 2028, with infrastructure and water-use accounting in the full report. It does not isolate every AI workload or establish local facility capacity, water availability, grid adequacy, emissions, resilience, or frontier-scale transfer. |
65.18.1 Manifest source assignment reconciliation
These rows keep Physical Compute Infrastructure, Energy, and Environmental Constraints’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.
| Source | Intake role | Boundary |
|---|---|---|
ext_flexible_hardware_enabled_guarantees_2025 |
Passage-reviewed comparator: Flexible Hardware-Enabled Guarantees for AI Compute. Motivates hardware mechanisms that could make compute use and policy compliance more observable and selectively enforceable. | Hardware evidence has bounded coverage and can introduce surveillance, capture, update-authority, compatibility, and circumvention risks. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row. |
ext_neuromorphic_computing_scale_2025 |
Metadata-first comparator: Neuromorphic computing at scale. Large-scale neuromorphic systems result demonstrating event-driven hardware capabilities under reported workloads and conditions. It does not establish superiority for general AI workloads or end-to-end system cost, programmability, reliability, and governance. | No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row. |
ext_photonic_neuromorphic_2024 |
Metadata-first comparator: Integrated photonic neuromorphic computing: opportunities and challenges. Review of integrated photonic neuromorphic computing opportunities and challenges. It maps device and systems tradeoffs but does not establish deployment advantage, digital replacement, or favorable full-stack energy and cost. | No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row. |
ext_quantum_ml_shadows_2024 |
Metadata-first comparator: Shadows of quantum machine learning. Peer-reviewed analysis of limitations and benchmarking traps in quantum machine-learning advantage claims. It supports advantage declarations with data-loading, classical-baseline, noise, scale, and end-to-end accounting, not a claim that quantum ML is useless. | No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row. |
ext_oecd_ai_infrastructure_competition_2025 |
Metadata-first comparator: Competition in artificial intelligence infrastructure. OECD analysis of concentration, barriers to entry, vertical integration, and competition across AI infrastructure. It motivates bottleneck and exit analysis but does not adjudicate a specific market, legal violation, or optimal remedy. | No passage-level source claim, local implementation, reproduction, safety, performance, deployment, support-state, or ASI result is established by this reconciliation row. |
65.19 Beyond digital accelerators: end-to-end substrate accounting
Emerging substrates move rather than erase physical costs. Neuromorphic systems may reduce work for sparse event streams while adding encoding, software, and heterogeneous integration. Photonic compute may accelerate selected linear operations while memory, conversion, nonlinearities, thermal control, and packaging dominate the complete workload. Quantum proposals must include classical orchestration, data loading, error management, verification, and the best classical comparator. Analog and in-memory systems trade data movement for noise, calibration, endurance, and precision recovery.
Accordingly, the infrastructure record adds a substrate_boundary covering: device operation, conversion, host compute, memory, storage, interconnect, cooling, control, yield, packaging, utilization, reliability, replacement, and retirement. Energy per primitive operation is never promoted into energy per useful governed result. A candidate advances only when quality-matched end-to-end work improves under the actual duty cycle and supply chain.
65.20 Concentration is a resilience variable
Physical deliverability also depends on market structure. The OECD AI infrastructure report identifies concentration, entry barriers, and vertical integration across compute infrastructure as competition concerns. This chapter does not determine an antitrust violation. It does require the physical capacity packet to record vendor and geography concentration, switching cost, interoperability, repair and spare-part dependency, cloud and accelerator coupling, financing constraints, and plausible exit.
Concentration can create correlated outages, policy leverage, opaque pricing, and an inability to replace a failed or untrusted component. Duplication alone is not resilience when both routes share a foundry, interconnect, cloud control plane, energy region, or software ecosystem. The test is common-mode closure: enumerate upstream dependencies until the allegedly independent alternatives stop converging on the same bottleneck.
65.21 Summary
Physical compute governance connects a requested result to the complete system needed to deliver it: chips, memory, interconnect, storage, software, facilities, power, cooling, water, land, materials, labor, maintenance, grids, suppliers, communities, and retirement. Every measurement retains its configuration, time, location, allocation method, and uncertainty.
This view separates nameplate capacity from useful output and efficiency from total effect. It enables scheduling, degradation, demand response, failover, repair, reuse, and denial decisions that keep quality and authority visible while exposing environmental burdens, correlated dependencies, concentration, and rebound instead of exporting them beyond the compute budget.
AI compute is a physical system before it is a token price. Infrastructure governance joins workload quality to hardware, memory, networks, facilities, grids, cooling, water, materials, land, communities, resilience, maintenance, demand response, concentration, reuse, and retirement so efficiency cannot hide displaced burdens.
65.22 Handoff
Mathematical and Search Substrates receives a bounded physical-cost and availability envelope for candidate computation. It does not inherit permission to ignore memory traffic, infrastructure effects, reliability, measurement uncertainty, or the fact that a mathematically elegant route may be physically worse when implemented end to end.