Locally Owned Reasoning Infrastructure: Economics and Institutional Implications — Research Brief

Full research memo: Locally Owned Reasoning Infrastructure: Economics and Institutions

1. Domain question

What do recent economic, systems, organizational, regulatory, and interoperability findings establish about locally owned reasoning infrastructure, and what would follow if AIOS can combine person-controlled knowledge, increasingly capable local models, selective frontier inference, bounded semantic judgment, deterministic effects, layered memory, verification, and human authority into a lower-cost and more portable system for ordinary personal and organizational reasoning?

2. Executive answer

Current evidence supports locally owned reasoning infrastructure, but not a universal local-only architecture. The defensible unit of analysis is the accepted task outcome rather than the token. Fixed-capability API prices fell extremely quickly through 2024, yet agent designs can multiply calls, energy, and run-to-run cost. Model capability is therefore one input among context, tools, memory, scaffolding, evaluation, integration, and human review.

On-device language-model inference is technically viable for bounded workloads, while cloud services retain strong advantages for bursty demand, frontier capability, long contexts, and shared hardware utilization. The resulting architecture is likely to be hybrid. AIOS sharpens that conclusion by separating custody from execution: canonical files, relationships, provenance, memory, evaluation criteria, and authority may remain locally controlled even when a difficult judgment is routed to a remote frontier model. Dynamic routing by cognitive operation, capability, sensitivity, latency, cost, and assurance need could reduce unnecessary remote inference without assuming that frontier models or centralized training disappear.

The larger implication is conditional but substantial. If purpose framing, context composition, cognitive modes, layered memory, metadata, verification, and integration produce cumulative improvements, local systems may satisfy a large share of ordinary reasoning while lowering retries, rework, and provider dependence. Durable intelligence would move toward people and institutions; remote services would supply selected capability rather than own the complete application relationship. This could support private cognition, portable organizational memory, regulated evidence chains, and peer-distributed domain packages.

Upstream chips, frontier training, cloud capacity, and electricity remain concentrated. Local systems introduce endpoint, maintenance, synchronization, and governance burdens. Technical export does not guarantee semantic portability, and elaborate scaffolding can increase coordination overhead. Open protocols solve interfaces more readily than trust, meaning, authorization, or liability. Complete systems therefore require comparison on accepted-task quality, cost, privacy, energy, switching, assurance, and human authority. That comparison, not model size alone, determines the economic result.

3. Essential findings

Finding 1 — Falling token prices do not establish falling reasoning-task costs

• Finding: Fixed-capability retail inference prices fell rapidly, but the economically relevant measure is cost per accepted and verified task.
• Evidence: Stanford’s 2025 AI Index reports that the cheapest model reaching a 64.8 MMLU threshold fell from about $20 to $0.07 per million tokens between November 2022 and October 2024. Epoch estimates declines from roughly 9-fold to 900-fold per year across selected thresholds. Both use posted prices and benchmarks rather than provider costs or complete workflows.
• Relationship to AIOS: direct mechanism — AIOS places models inside a larger production system, so its economic performance depends on inference, hardware, orchestration, review, rework, compliance, and continuity together.
• Implication: AIOS should be evaluated and budgeted per accepted outcome. A system that uses a smaller model but improves context, verification, and integration may outperform a cheaper call that produces retries or hidden review costs.
• Limits or counterevidence: Benchmark thresholds do not measure professional acceptance, latency, reliability, discounts, caching, or quality-adjusted output. Frontier capability can remain expensive even while fixed capability becomes cheap.
• Exact descent: Full memo §2, “The economic object: ownership, execution, and complements” and §3.1–3.2; Stanford AI Index 2025, Epoch AI price analysis, and Cost-of-Pass.

Finding 2 — Agentic orchestration creates both rebound risk and an efficiency opportunity

• Finding: Architecture can change inference demand by orders of magnitude, making scaffolding a potential source of either savings or rebound.
• Evidence: Kim et al. report tool-using agents making 9.2 times as many calls as a chain-of-thought baseline on average, with selected workflows raising measured GPU energy by more than two orders of magnitude. Bai et al. find roughly thousand-fold token differences between coding-agent trajectories and shorter code tasks, plus up to 30-fold variation across repeated runs. Both are preprints on bounded benchmarks.
• Relationship to AIOS: complicating evidence — AIOS hypothesizes that composed context, memory, bounded actions, verification, and routing reduce retries and hallucination exposure, but the same components can add calls, indexes, review, and background work.
• Implication: The cumulative AIOS bundle could offset agentic rebound if its components remove different failure sources, although the net effect remains unmeasured.
• Limits or counterevidence: Coding and question-answering scaffolds do not represent all work. Caching, batching, review, and model updates change costs. More structure is not automatically better.
• Exact descent: Full memo §3.3–3.4, §12.2, and §12.8; Kim et al., Bai et al., and Kim et al., Nature Machine Intelligence.

Finding 3 — Hybrid local and frontier execution is the economically coherent architecture

• Finding: On-device inference is feasible for bounded tasks, while remote frontier execution remains valuable; custody and execution should therefore be optimized separately.
• Evidence: Xu et al.’s peer-reviewed ASPLOS study tests five 1.8B–7B models on two Qualcomm phones and reports 1.4–32.8 times end-to-end speedups and 1.9–59.5 times lower energy against selected CPU/GPU baselines. MLPerf Client and Mobile now benchmark client language models through 8B parameters.
• Relationship to AIOS: supporting evidence — the results support local execution as a real component of AIOS, while their limits support selective routing rather than local-only doctrine.
• Implication: Repeated, stable, sensitive operations may run locally, with frontier models invoked for difficult judgment, current information, or independent review. Canonical files and accepted ground need not migrate when execution moves.
• Limits or counterevidence: Two phones, small models, selected benchmarks, and a specialized runtime do not establish lifetime cost or broad task coverage. Hardware turnover, thermals, maintenance, and capability gaps can favor remote service.
• Exact descent: Full memo §4 and §12.4; Xu et al., MLPerf Client, and MLPerf Mobile v6.0.

Finding 4 — Local ownership changes application dependence, not upstream concentration

• Finding: Local control of durable knowledge may reduce application-layer dependence without decentralizing frontier training, chips, cloud capacity, or electricity demand.
• Evidence: IEA retrospectively estimates about 485 TWh of global data-center electricity use in 2025 and projects about 950 TWh in 2030; LBNL models US data centers reaching 9.5–15.3% of electricity in 2030. Competition authorities document hyperscaler concentration, technical switching barriers, licensing constraints, and vertical ties.
• Relationship to AIOS: boundary condition — AIOS can relocate the durable intelligence layer while continuing to rely selectively on concentrated upstream infrastructure.
• Implication: The consequential thesis is thinner or more specialized centralized applications, not unnecessary data centers. Portable local state can improve bargaining, continuity, and provider substitution even when frontier compute remains centralized.
• Limits or counterevidence: Energy totals and forecasts are modeled, category boundaries are uncertain, and local workloads can shift rather than eliminate infrastructure. Rebound may increase total compute despite lower remote dependence per task.
• Exact descent: Full memo §5, §11, and §12.9; IEA 2026, LBNL 2026, and UK CMA cloud investigation.

Finding 5 — Semantic portability is more demanding than data export

• Finding: Switching costs include technical movement, model-specific behavioral adaptation, and preservation of institutional meaning and authority.
• Evidence: The CMA and OECD document data movement, proprietary services, retraining, architecture changes, licensing, and egress barriers. Jahani et al.’s preregistered experiments with 3,750 participants and about 37,000 prompts find that user prompt adaptation accounted for about half of the performance improvement on a bounded image-replication task after a model upgrade, but only about 7% on an open creative task.
• Relationship to AIOS: direct mechanism — canonical local files, annotations, provenance, relationships, and evaluations are potential exit assets because they preserve meaning outside one model or provider.
• Implication: AIOS can make model routing and provider switching more credible if it retains behavioral regression tests and semantic stewardship alongside exportable files.
• Limits or counterevidence: Prompt adaptation is not a direct measurement of full semantic switching cost. Local schemas and embeddings can themselves become proprietary or unintelligible, and behavioral migration remains task-dependent.
• Exact descent: Full memo §6; UK CMA, OECD cloud competition report, and Jahani et al..

Finding 6 — Productivity depends on system complements and task boundaries

• Finding: AI can materially improve selected tasks, but gains depend on task fit, evaluation, worker experience, adoption, and organizational redesign.
• Evidence: Cui et al. pool randomized trials across 4,867 developers and estimate 26.08% more completed tasks. Dell’Acqua et al. find faster, higher-quality consultant work inside a capability frontier but a 19-percentage-point correctness decline on an outside-frontier task. Danish administrative evidence finds near-zero earnings and hours effects over the first two years despite adoption.
• Relationship to AIOS: supporting evidence — the studies support treating useful intelligence as a sociotechnical outcome rather than equating model access with organizational productivity.
• Implication: AIOS may accumulate productive complements in files, domain rubrics, evaluation packs, controlled workflows, and human authority. Its value should be measured across complete work and institutional routines, not isolated outputs.
• Limits or counterevidence: Task counts are not full quality-adjusted productivity; studies use particular firms, occupations, and models. Short-run labor outcomes may miss consumer surplus, quality, or delayed reorganization.
• Exact descent: Full memo §7 and §12.6; Cui et al., Dell’Acqua et al., and Humlum and Vestergaard.

Finding 7 — Local custody is an assurance affordance, not assurance itself

• Finding: Regulated auditability requires a reconstructable lifecycle evidence chain, whether execution is local, remote, or hybrid.
• Evidence: The EU AI Act requires risk management, documentation, logging, human oversight, robustness, quality management, and monitoring for covered systems. NIST AI 800-4 synthesizes literature and three practitioner workshops, finding monitoring fragmented across technical, operational, human, security, compliance, and societal dimensions.
• Relationship to AIOS: direct mechanism — exact file operations, source states, model identifiers, permissions, evaluations, diffs, and human approvals can form the evidence chain around bounded judgments.
• Implication: AIOS could support regulated domain systems whose local semantics vary beneath shared evidence and control contracts. This may make workflow-level assurance more portable across model providers.
• Limits or counterevidence: Local endpoints can be insecure or poorly governed. Logs can create privacy and surveillance risks. Neither a transcript, local file history, management certificate, nor model certificate proves substantive fitness for an intended use.
• Exact descent: Full memo §8 and §12.6; EU AI Act, NIST AI 800-4 report, and FDA draft guidance.

Finding 8 — Open protocols enable exchange but do not establish trust or shared meaning

• Finding: Current protocols and provenance standards make attestable domain packages plausible, but certification remains an institutional function.
• Evidence: MCP standardizes access to tools, resources, prompts, and extensions; A2A standardizes capability discovery and task exchange. RO-Crate describes portable files and relationships, while Verifiable Credentials, C2PA, and Sigstore bind claims or provenance to issuers and signatures.
• Relationship to AIOS: supporting evidence — these standards supply components for AIOS domain packages and protocol adapters without dictating the system’s local semantics.
• Implication: Peer communities and institutions could exchange bounded, signed packages while retaining private overlays and choosing local or remote execution. This could support federated expertise and low-connectivity use.
• Limits or counterevidence: Interfaces do not prove semantic agreement, authorization, accuracy, freshness, safety, or liability. Revocation, licensing, governance, and accountable certification remain necessary.
• Exact descent: Full memo §9 and §12.7; MCP specification, A2A specification, RO-Crate 1.3, and W3C Verifiable Credentials 2.0.

Finding 9 — The electricity analogy applies to complements, not to reasoning as a commodity

• Finding: AI resembles electrification in diffusion, complementary investment, and infrastructure effects, but reasoning outputs are not standardized or semantically fungible.
• Evidence: The OECD finds generative AI plausibly has important general-purpose-technology characteristics while emphasizing uncertain diffusion and productivity. Historical electrification research shows that productivity required factory redesign and complementary capital rather than installation alone.
• Relationship to AIOS: useful analogy — AIOS is best interpreted as complementary organizational reconfiguration around model inputs, not as a claim that intelligence becomes a homogeneous utility.
• Implication: If durable context, memory, evidence, and authority become locally portable, model services may diffuse broadly without owning the durable intelligence layer. This is the analogy’s strongest AIOS-relevant implication.
• Limits or counterevidence: Electricity has standardized metering, stable physical behavior, mature safety institutions, and weak semantic switching costs. The analogy cannot justify utility regulation, inevitable productivity, or a grid-versus-device equivalence.
• Exact descent: Full memo §10 and §13, “Foundational lineage”; OECD GPT assessment, David, and Brynjolfsson, Rock, and Syverson.

4. How the evidence refines the AIOS account

1. Accepted-outcome economics: It replaces token price as the main economic metric with a complete cost that includes orchestration, review, failure, compliance, continuity, and exact integration.
2. A three-layer switching model: It distinguishes technical, behavioral, and semantic switching costs, clarifying why canonical files require regression tests and semantic stewardship to become a genuine exit asset.
3. A rebound-versus-complementarity problem: It shows that agentic architectures can multiply demand while the cumulative AIOS system may reduce retries and unnecessary frontier calls; the net effect must be measured.
4. An assurance evidence chain: It specifies the identities, sources, versions, permissions, tool calls, diffs, tests, approvals, and monitoring records needed to turn locality into an assurance affordance.
5. An institutional implication map: It sharpens how local ownership could affect personal cognitive capital, organizational memory, application infrastructure, regulated work, peer knowledge institutions, accessibility, and energy without requiring frontier models or data centers to disappear.

5. Architectural boundaries preserved

1. Person-controlled canonical files and accepted ground: External evidence supplies no basis for moving authoritative files, memory, annotations, or relationships back into provider-owned application state.
2. Human purpose and consequential authority: Productivity findings and standards reinforce accountable human judgment rather than unrestricted autonomy.
3. Multiple self-contained domains: Common evidence controls remain compatible with local semantic variation; distinct domains do not collapse into one universal ontology.
4. The recurring Why–How–What grammar: No reviewed study tests or refutes the Fractal Seed. It remains the internal organizing grammar rather than being redesigned around an external taxonomy.
5. Local-first but model-plural execution: On-device limits do not weaken local ownership, and frontier capability does not imply a local-only rule. Custody remains separate from execution.

6. Where the evidence connects to AIOS

7. Relationships across research programs

1. Benchmark and routing research: Accepted-outcome economics connects routing benchmarks, capability thresholds, privacy classifications, and failure costs to one complete unit of evaluation.
2. Reasoning architecture research: System-level intelligence depends on whether gains from context composition, cognitive modes, planning, verification, and integration combine, overlap, or interfere.
3. Memory and knowledge-evolution research: Multiple-resolution memory, provenance, annotations, and relationship graphs connect to semantic switching through evidence continuity and correction.
4. Human authority and organizational design research: Bounded judgment, permissions, escalation, review burden, and local variation connect economic performance to agency, expertise, responsibility, and institutional control.
5. Interoperability and package research: Protocol semantics, identity, attestations, revocation, and certification governance determine whether peer-distributed domain packages carry trust; technical interoperability alone does not.

8. Priority source set

• March 2025 — reproducible research-organization analysis: Supports the claim that fixed-capability posted inference prices fell by roughly 9-fold to 900-fold annually across selected thresholds, not that full task cost fell at the same rate. Epoch AI
• April 2025 — authoritative synthesis: Supports the observed decline from about $20 to $0.07 per million tokens at a fixed MMLU threshold through October 2024. Stanford AI Index 2025, chapter 1
• June 2025 — preprint: Supports the claim that agent design can multiply calls and measured GPU energy by orders of magnitude on selected benchmarks. Kim et al., The Cost of Dynamic Reasoning
• March 2025 — peer-reviewed systems study: Supports technical feasibility and large speed/energy improvements for selected small models on two smartphone platforms. Xu et al.
• April 2026 — authoritative modeled estimates and scenarios: Supports 2025 global data-center electricity and capital-concentration estimates plus the conditional 2030 trajectory. IEA, Key Questions on Energy and AI
• July 2025 — regulator final report: Supports technical, commercial, licensing, and operational barriers to cloud switching and multicloud use. UK CMA cloud investigation
• April 2026 — peer-reviewed preregistered experiments: Supports model–user behavioral complementarity and task-dependent prompt adaptation across a model upgrade. Jahani et al.
• February 2026 — peer-reviewed randomized field experiments: Supports a pooled 26.08% increase in completed developer tasks across three firms under the studied intervention. Cui et al.
• March 2026 — peer-reviewed preregistered experiment: Supports gains inside a model capability frontier and increased error on one outside-frontier task. Dell’Acqua et al.
• July 2026 — peer-reviewed, vendor-funded controlled benchmark study: Supports the claim that multi-agent coordination benefits depend on task and base-model capability and can impose substantial reasoning-turn overhead. Kim et al.
• March 2026 — authoritative workshop and literature study: Supports the need to connect technical, operational, human, security, compliance, and societal monitoring evidence. NIST AI 800-4 report
• August 2024 — law: Supports lifecycle risk management, documentation, logging, human oversight, robustness, and monitoring duties for covered high-risk systems. EU AI Act
• July 2026 — open protocol specification: Supports standardized model-application access to tools, resources, prompts, and extensions, not semantic or trust equivalence. MCP 2026-07-28
• June 2026 — community recommendation: Supports portable description of files, entities, provenance, actions, and relationships in a knowledge package. RO-Crate 1.3
• June 2025 — authoritative evidence synthesis: Supports treating generative AI as a plausible general-purpose technology while retaining uncertainty about diffusion, productivity, and institutions. OECD

9. Open research and design questions

1. Which accepted-outcome measures best capture the full economic effect of orchestration, review, failure, compliance, continuity, and exact integration?
2. Does “semantic switching cost” name a distinct, measurable barrier beyond technical and behavioral switching costs?
3. Under what conditions does local ownership shift application dependence toward person-owned durable intelligence rather than merely relocating costs?
4. What evidence and governance distinguish an attestable domain package from a certified package under a named assurance regime?
5. Where does the electricity analogy clarify complementarity and institutional reorganization, and where does it fail because reasoning remains task-, context-, and judgment-dependent?