Appendix C
Empirical Dependencies
Aa
The framework rests on empirical claims about physics, infrastructure, markets, and institutions. If any proves materially wrong, its predictions require revision. This appendix enumerates the claims, their sources, their magnitudes, and the conditions under which they would be falsified.
Each claim below is stated so that contrary evidence can force revision. Without such conditions, the framework would be insulated from failure. The claims are ordered by their proximity to the argument's core.
C.1 — Thermodynamic Foundations
These claims establish the physical basis for treating computation as energy conversion.
Claim 1.1: Landauer's limit sets a floor on irreversible computation.
Statement: Resetting one initially unbiased bit without retaining the discarded information, under the model’s thermal conditions at temperature , requires dissipating at least joules of energy, where is Boltzmann's constant. At room temperature (300K), this equals approximately joules per bit.
Source: Landauer, R. (1961). "Irreversibility and Heat Generation in the Computing Process." IBM Journal of Research and Development, 5(3), 183–191.
Status: This is a theoretical result under a specified physical model of erasure. Applying it requires the operation, information being erased, and thermodynamic conditions to satisfy the model’s assumptions. It is not a fixed energy charge for every computational operation.
Relevance to framework: Establishes that irreversible computation has an irreducible thermodynamic cost. It does not establish a market price for a workload, an energy-to-Bitcoin conversion rate, or a universal commercial hurdle. Those depend on hardware, site, operator, capital, and market conditions.
Claim 1.2: Practical computation remains far above the Landauer floor, but no single gap applies to every workload.
Statement: Practical processors expend far more energy than per operation. A logic transition, memory access, floating-point operation, token, and model call are not the same physical event, so none can be assigned one canonical ratio to a bit-erasure bound.
Source: Koomey, J. et al. (2011). "Implications of Historical Trends in the Electrical Efficiency of Computing." IEEE Annals of the History of Computing, 33(3), 46–54, together with workload-specific semiconductor and system measurements. Koomey's historical series does not turn unlike operations into one contemporary distance-to-Landauer statistic.
Magnitude that matters: Workload-level energy efficiency and system overhead determine useful engineering headroom. An ultimate bound can show that physics permits improvement without predicting which improvements are feasible, economical, or near.
Update frequency: Annual. Track IRDS publications, MLPerf benchmarks, and academic efficiency studies.
Falsifier: If workload-level measurements show that further efficiency is constrained chiefly by unavoidable physical dissipation rather than architecture, memory, interconnect, fabrication, reliability, or economics, the "room to run" claim requires revision.
Claim 1.3: The historical efficiency trend has been rapid; its continuation is workload-dependent.
Statement: Koomey and colleagues found that computations per kilowatt-hour roughly doubled every 1.57 years across the historical systems they studied. Later estimates indicate slower improvement in some periods and workloads. The series is descriptive, not a physical law or a forecast for model inference.
Source: Koomey, J. (2011), updated in subsequent publications through 2020.
Magnitude that matters: Continued efficiency gains can lower the cost of a fixed workload while rising demand increases total energy use. Either effect must be measured at a declared workload and quality level.
Falsifier: Persistent stagnation in task-level efficiency across major architectures would weaken the claim that engineering progress keeps expanding the feasible set.
C.2 — Infrastructure Constraints
These claims test where and when physical constraints bind deployment. They do not establish a universal ordering among software capability, demand, finance, equipment, interconnection, and institutional permission.
Claim 2.1: Interconnection queues are deep and project timelines have lengthened.
Statement: Lawrence Berkeley National Laboratory reported roughly 2,600 GW of generation and storage capacity seeking U.S. grid interconnection at the end of 2023, with typical project timelines having lengthened materially.1 Queue capacity is proposed capacity, not an inventory of projects likely to be built.
Source: Rand, Joseph, et al. "Queued Up: 2024 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2023." Lawrence Berkeley National Laboratory, April 2024. ERCOT and PJM public queue data.
Magnitude that matters: Gross queue depth and lengthening timelines are evidence of congestion and intense demand for interconnection. They do not show that every queued project is viable or that interconnection caused each delay. The stronger causal claim requires project-level evidence on withdrawals, duplication, financing, equipment, permits, and the interval from request to energization.
Update frequency: Quarterly. Queue data is publicly available from ISOs and RTOs.
Falsifier: This evidence would cease to support a persistent interconnection constraint if project-level records showed materially shorter request-to-energization intervals, rising completion rates, and delays attributable chiefly to withdrawal, duplication, finance, equipment, or permitting rather than interconnection review and network work. No queue-to-installed-capacity ratio alone can decide the question.
Claim 2.2: Transformer lead times have extended to 2-3 years for large power transformers.
Statement: The U.S. Department of Energy's 2024 report documented multi-year lead times for some large power transformers and described manufacturing capacity, materials, and specialized labor among the constraints.2 This is a dated equipment-market observation, not a permanent law of data-center construction.
Source: U.S. Department of Energy. "Large Power Transformer Resilience: Report to Congress." July 2024. Utility integrated resource plan filings from PJM, MISO, and CAISO (2024-2025).
Magnitude that matters: Transformers are long-pole items in grid expansion. If lead times returned to 12 months, datacenter capacity additions could accelerate significantly.
Falsifier: If new manufacturing capacity brings lead times below 18 months for standard orders, this specific bottleneck eases.
Claim 2.3: Announced data-center capacity is an uncertain indicator of delivered capacity.
Statement: Company announcements and utility filings indicate a large pipeline of proposed data-center load. Announced capacity, contracted load, interconnected capacity, and operating demand are different objects and must not be summed as though each were built.
Source: Company announcements and utility integrated-resource and interconnection filings. A release claim requires a dated, deduplicated project ledger; no single aggregate is treated here as canonical.
Magnitude that matters: The share of proposed load that becomes firm and energized affects whether power and interconnection remain scarce complements.
Update frequency: Semi-annual. Track utility filings and company earnings disclosures.
Falsifier: If proposed projects are cancelled or fail to reach firm interconnection and construction milestones at high rates, announcement-based demand projections require revision. The relevant rate must be estimated from an explicit cohort rather than assumed in advance.
C.3 — Bitcoin and Mining Economics
These claims test whether Bitcoin mining supplies a relevant outside option for particular sites and operators, and whether Bitcoin supplies useful settlement properties. They do not establish a universal compute floor.
Claim 3.1: Bitcoin mining can supply a site- and operator-specific electricity outside option.
Classification: Conditional derivation plus empirical conjecture.
Statement: An operator with suitable mining hardware, interconnection, operating capability, and market access can estimate the net return from directing a marginal unit of power to Bitcoin mining. For that operator at that site and time, the estimate may be an opportunity-cost comparator for another workload. It is not a universal compute price and does not apply to power or capital that cannot be redirected on relevant terms.
Derivation: Appendix A.1 separates mining revenue, mining operating cost, return on mining capital, BTC acquisition return, secured funding cost, and collateral-lock opportunity cost. The relevant comparator is the feasible net return on the operator's next-best deployment, not gross mining revenue.
Magnitude that matters: The comparator matters only when the mining alternative is technically available, capacity-constrained resources are genuinely substitutable, switching costs are included, and the expected net return is high enough to alter allocation.
Falsifier: If operators with a credible mining option do not reallocate power or capital when the measured return gap persists after switching costs, the proposed allocation mechanism is weak. The claim also does not apply where mining is unavailable or the resources are not substitutable.
Claim 3.2: The China mining ban resulted in hash rate migration, not elimination.
Statement: Cambridge’s May 2022 account reported a sharp fall in network hash rate following China’s June 2021 prohibition and an almost complete recovery to pre-ban levels by December 2021. Its pool-based location data showed growth abroad and a resurgence of reported activity within China. The authors cautioned that proxy use could obscure miners’ locations; the map did not track each machine’s journey.
Source: Cambridge Centre for Alternative Finance, “Bitcoin mining – an (un)surprising resurgence?”, 17 May 2022, sections on hash-rate recovery and methodological trade-offs. The account uses network estimates and aggregated mining-pool location data.
Magnitude that matters: This episode shows that one major jurisdiction's prohibition coincided with relocation and network recovery. It does not establish that a coordinated multi-jurisdictional prohibition would have the same result.
Update frequency: This is a historical claim. Ongoing relevance depends on whether subsequent prohibitions follow the same pattern.
Contrary evidence: Sustained suppression where miners lack accessible substitutes would weaken a general resilience hypothesis. That outcome would not undo the recovery observed in this episode; nor does this episode establish immunity to a coordinated prohibition.
Claim 3.3: Mining has migrated across jurisdictions, but geographic estimates remain incomplete.
Classification: Time-sensitive empirical estimate.
Statement: The Cambridge Bitcoin Electricity Consumption Index and mining-pool disclosures reported geographically distributed activity after the 2021 China prohibition. The public location estimates depend on pool-reported or inferred data, exclude some activity, and should not be read as an exact census. This statement is limited to the sources available through 2024.
Source: Cambridge Centre for Alternative Finance, Cambridge Bitcoin Electricity Consumption Index Mining Map, data available through 2024.3 Pool disclosures are secondary corroboration, not a complete denominator.
Magnitude that matters: Geographic and ownership concentration affect exposure to coordinated shutdown, censorship, and supply-chain action. Jurisdiction share alone does not measure pool control, firmware dependencies, hardware concentration, or state influence.
Falsifier: Reliable evidence of sustained operational concentration under one effective control domain would weaken the geographic-resilience argument. No universal percentage threshold is asserted without a complete and consistently measured denominator.
Claim 3.4: Consensus-attack estimates measure a narrow protocol threat.
Classification: Model-based estimate with a narrow scope.
Statement: Published cost models estimate the resources required to obtain enough hash power for a reorganization or double-spend attack under stated hardware, power, duration, and market assumptions. Estimates are scenario-dependent and change with hash rate, equipment availability, electricity prices, attacker strategy, and defender response.
Source: Nuzzi, Waters, and Andrade, “Breaking BFT: Quantifying the Cost to Attack Bitcoin and Ethereum,” Coin Metrics / SSRN 4727999, February 2024.4 The estimate is dated to that model and its input assumptions.
Scope: The estimate concerns Bitcoin consensus manipulation. It does not measure key theft, custodian insolvency, coercion, legal seizure, exchange exclusion, oracle failure, collateral double pledge, or collectibility. Those risks can dominate an assurance arrangement even when a consensus attack is uneconomic.
Magnitude that matters: The model is relevant when a promised consequence depends on finality against chain reorganization. It is not a general security floor for agent-held assets.
Falsifier: If the assumptions used by a deployment materially understate rentable hash power, hardware access, attack duration, or defender response, that deployment's finality analysis fails. Even a robust consensus estimate cannot validate the other threat classes listed above.
C.4 — AI Capability and Cost Trajectories
These claims establish the pace and direction of AI development relevant to economic substitution.
Claim 4.1: Reconstructed training compute rose steeply across important model generations.
Statement: Published reconstructions place GPT-3 training at approximately 3.14 × 10²³ FLOP and document a steep historical rise in the compute used for notable machine-learning systems.5 Estimates for undisclosed systems remain reconstructions rather than measurements, and a historical trend does not fix the scale of the next training run.
Source: Patterson et al. (2021) for GPT-3 energy estimates and Sevilla et al. for historical training-compute reconstructions.6 Company disclosures remain incomplete, which limits comparisons among particular frontier systems.
Verification note: Epoch AI has revised their estimates over time as methodology improves and new information surfaces. Any specific FLOP figure cited here should be verified against the current published dataset at time of reading. The relevant claim is the order-of-magnitude increase across generations, not any single model's precise training compute.
Magnitude that matters: The historical record supports a race between efficiency and scale. It does not establish that every increase in capability requires more absolute energy, or that scale will continue to absorb every efficiency gain.
Verification standard: This claim would be considered verified if two or more independent sources (company disclosure, Epoch AI estimate, academic reconstruction) agree on order of magnitude for a given model generation.
Update frequency: Per model generation (roughly annual for frontier labs). Epoch AI updates their dataset continuously.
Falsifier: Repeated capability gains at materially lower quality-adjusted training compute would defeat the stronger claim that frontier development necessarily absorbs efficiency through scale.
Claim 4.2: Quality-adjusted inference prices have fallen rapidly in some model classes.
Statement: Posted API prices have fallen rapidly for several classes of inference while capability and context length have improved. Simple token-price comparisons remain imperfect because model quality, batching, latency, caching, context, and output mix change at the same time.
Source: Dated provider pricing pages and independent benchmark archives. Any numerical comparison must preserve the archived date, model tier, token direction, context, and service conditions.
Magnitude that matters: Persistent quality-adjusted price decline would strengthen the claim that some cognitive services commoditize. Provider list prices alone cannot establish economy-wide commoditization or the margins available after deployment costs.
Cross-reference: The gap documented in Claim 1.2 shows that current devices are far from this particular lower bound. It does not establish the realizable engineering headroom, its rate of improvement, or future inference prices, which may be constrained by memory movement, communication, fabrication, reliability, and demand long before Landauer binds.
Update frequency: Quarterly. Pricing changes are publicly announced.
Falsifier: If inference pricing stabilizes or increases for frontier capability (not merely for legacy models), commoditization may be slower than assumed.
Claim 4.3: Open-weight systems can compress some closed-model capability leads.
Statement: Open-weight releases have approached or surpassed earlier closed systems on selected benchmarks. There is no single defensible lag across domains, deployment constraints, safety properties, modalities, or model generations.
Source: Benchmark results from HELM, MMLU, HumanEval; academic and industry publications.
Magnitude that matters: Where reproducible capability gaps narrow and switching is practical, model-level rents face pressure. Persistent differences in reliability, inference cost, tooling, data, distribution, or institutional approval may preserve rents even when benchmark scores converge.
Update frequency: Per major open-weight release (roughly quarterly).
Falsifier: If open systems fail to close economically relevant capability gaps across successive releases, or cannot be deployed under comparable conditions, the proposed pressure is weaker than the benchmark record suggests.
C.5 — Labor Market and Automation Sequencing
These claims state the historical evidence and empirical tests needed to evaluate automation sequencing. They do not establish V/C as an independent causal law.
Claim 5.1: New tasks offset part of displacement in historical U.S. data.
Statement: Acemoglu and Restrepo find that new task content contributed materially to U.S. employment growth over the periods they study, offsetting part of the displacement associated with automation. Their result does not establish that reinstatement dominated every transition, that service-sector absorption had one cause, or that the same balance will persist.7
Source: Acemoglu and Restrepo (2019).8
Magnitude that matters: The estimate establishes that new-task creation has been economically important, while leaving its future magnitude open. The current argument turns on whether comparable task creation remains durable when some new cognitive tasks can also be automated.
Relevance to framework: Chapter 20 presents the adjustment case. This claim bears on it.
Claim 5.2: Baumol sectors (healthcare, education, personal services) absorb an increasing share of employment.
Statement: Health, education, and other labor-intensive services have absorbed a growing share of employment during important periods. Category definitions and dates materially affect the measured share, while automation and institutional reform can alter their future capacity to absorb labor.
Source: Bureau of Labor Statistics employment statistics; Baumol and Bowen (1966)9.
Magnitude that matters: The size of the Baumol reservoir determines how much labor displacement these sectors can absorb. If the sectors are saturated (growth slowing or reversing), absorption capacity diminishes.
Update frequency: Annual. BLS data is publicly available.
Falsifier: If Baumol sector employment growth slows to below overall employment growth for a sustained period, the reservoir mechanism weakens.
Claim 5.3: Verification cost helps explain deployment sequence after other constraints are included.
Statement: Across tasks with comparable capability, integration, expected loss, regulation, demand, and authority, lower credible verification cost should be associated with earlier durable delegation. V/C is a proposed explanatory variable, not an ordering law and not a substitute for the other terms.
Observable indicators: Compare time from credible capability demonstration to routine production use across matched task classes. Measure what counts as verification in each domain, who bears error, whether correction arrives before harm, and whether the deployer possesses authority to act.
Source: Chapter 16 develops the V/C heuristic and its limitations. Deployment histories and regulatory records supply the evidence needed to test it; examples in the body are illustrations rather than present confirmation.
Magnitude that matters: V/C earns independent explanatory status only if verification cost accounts for meaningful variation after competing transaction-cost and institutional variables are modeled. If those variables explain the same ordering without a distinct verification term, V/C remains useful vocabulary rather than a separate empirical contribution.
Update frequency: Continuous. Track comparable production deployments, not announcements or pilots alone.
Falsifier: Across a sufficiently varied matched sample, verification cost has no stable association with durable delegation after capability, integration, expected loss, remedy, regulation, demand, and authority are included, or the sign of the association is systematically reversed.
Claim 5.4: New tasks may themselves encounter automation pressure quickly.
Statement: Some newly created tasks may face automation pressure before they mature into durable occupational categories. Robotic-process configuration and prompt-centered work are suggestive cases, but neither yet supplies a clean cohort from which a general reinstatement lag can be estimated.
Source: Employer records, occupational classifications, product histories, and longitudinal task data would be needed to establish the claim. Current examples remain illustrative.
Magnitude that matters: If new task families are repeatedly automated before workers and institutions can reorganize around them, reinstatement will absorb less displacement than its historical record suggests. One short-lived occupational label would not establish the pattern.
Falsifier: Longitudinal evidence showing stable or lengthening intervals between task emergence and material automation pressure would defeat the compression hypothesis.
Claim 5.5: Large deployment can make inference the dominant lifetime energy demand.
Statement: Training is concentrated, whereas inference recurs with use. A heavily used model may therefore consume more energy in inference over its deployed life than in training. The result depends on utilization, model turnover, serving efficiency, hardware, and the accounting boundary.
Source: Patterson et al. (2021) document the importance of inference in a particular production setting.10 Broader estimates require workload and fleet data that providers seldom disclose.
Magnitude that matters: The lifetime energy account helps identify where efficiency improvements could save consumption. It does not identify the binding deployment constraint: a particular site can lack adequate power or equipment even when serving accounts for less total energy than training.
Falsifier: If credible measurement shows training energy exceeds inference energy for a given model generation even after 24 months of deployment, the inference-dominance claim fails for that generation.
C.6 — Institutional and Market Behavior
These claims concern how economic actors respond to the transition.
Claim 6.1: The portability of an API does not settle enterprise switching cost.
Classification: A distinction among costs, with an empirical hypothesis about their magnitude.
Statement: A compatible interface can spare an enterprise from rewriting a connection while leaving work to establish that another service performs its task adequately. Evaluation suites, prompts, retrieval systems, safety controls, data arrangements, latency requirements and organizational approval can make that work consequential. The extent depends on what must move and what can be reused.
Source and scope: Chapters 10 and 22 distinguish obtaining an artifact or service from making productive use of it; Chapter 23 distinguishes an available alternative from the rights, costs and commitments involved in taking it. The comparison with established enterprise software would require a defined sample of migrations. No such cross-enterprise cost estimate is established here.
Hypothesis to test: For enterprises with acceptable alternatives, reusable interfaces and supporting materials reduce the cost and time of changing providers. Compare completed migrations for sufficiently similar purposes and required quality, recording changed evaluation, integration, operating and contractual requirements. A redesigned workflow may be the means of switching, rather than a reason to exclude the observation.
Magnitude that matters: Lower switching costs can give customers a more credible alternative and put pressure on a provider’s terms. The effect on its margin also depends on competing provision, differentiation, costs and contracts. A difficult migration can sustain dependence without making the incumbent a profitable investment.
Contrary evidence: If comparable migrations obtain little saving from reusable interfaces and materials because other requirements dominate, portability explains little of the measured switching cost. Persistent renewal alone cannot decide this: customers may stay because the existing service remains preferable. Conversely, inexpensive completed migrations weaken an account that treats the surrounding requirements as an enduring barrier in that setting.
Measurement: Record the enterprise, task, alternatives, completion standard, dates, costs and time. Keep API compatibility distinct from successful migration and from the provider’s subsequent return.
Claim 6.2: Stablecoin issuers maintain blacklists and have frozen addresses.
Statement: Major stablecoin issuers (Circle, Tether) maintain address blacklists and have frozen funds at the request of law enforcement or under their own compliance policies.
Evidence: Blacklist and freeze functions, issuer announcements, legal records, and on-chain transactions document the capability and its exercise. Counts and balances change with chain, contract version, date, and measurement method, so the proposition does not depend on one unarchived total.
Source: Issuer contract documentation and transparency reports, legal filings, and reproducible dated on-chain queries.
Magnitude that matters: Issuer-controlled freezing creates a counterparty and governance dependency for any holder, including a computational actor. Its practical severity depends on custody, legal recourse, diversification, and the role of the stablecoin in the arrangement.
Status: This is an established fact. The question is whether freezing extends to agent-held collateral at scale, a contingency that depends on how issuers classify and respond to non-human-controlled addresses.
Claim 6.3: Collateral requirements depend on loss, detection, collection, and enforceability.
Classification: Conjecture.
Statement: Collateral can deter defection or fund restitution only to the extent that violations are detected, an authorized process can impose the consequence, the pledged asset is collectible, and the collateral remains available. Persistent identity and future rents may reduce required capital in some arrangements, but no universal 150% or 100% ratio follows.
Source: Appendix A.3 states the incentive and claimant-restitution conditions separately. DeFi lending is not treated as a direct analogue: liquidation collateral secures market-price exposure under a specific protocol, not arbitrary off-chain performance.
Magnitude that matters: Required capital rises with covered loss, defection gain, tail exposure, low detection or collection probability, illiquidity, correlation, and weak legal or technical control. It may fall with independently verified performance, credible future-rent loss, or stronger collection mechanisms.
Falsifier: If disclosed detection, adjudication, collection, exposure, and correlation variables do not explain observed capital requirements better than a fixed collateral-ratio rule, this model adds little.
Claim 6.4: A complete bilateral quote graph is quadratic; actual coordination need not be.
Classification: Conditional combinatorial result.
Statement: A complete undirected graph among participants contains bilateral relationships. Quoting each participant against one shared reference uses participant-reference quotes. Real markets can instead use sparse graphs, intermediaries, auctions, clearinghouses, or hierarchical aggregation, so a common benchmark is neither mathematically necessary nor sufficient for scalable coordination.
Source: Appendix A.2 provides the conditional graph comparison.
Falsifier: If a shared benchmark does not reduce search, comparison, or hedging cost in measured deployments, its proposed coordination benefit is absent. Large-scale coordination without one would falsify any stronger necessity claim.
C.7 — Contingent Claims Requiring Monitoring
These claims are forward-looking and will be resolved by events.
Claim 7.1: A Bitcoin-native secured capital curve may emerge.
Classification: Forward-looking conjecture.
Statement: Market participants may publish multiple tenor-specific curves for Bitcoin-denominated or Bitcoin-secured funding. Any such curve would embed counterparty, custody, liquidity, collateral, venue, and legal risks; it would not be a universal risk-free rate.
Observable indicators:
- Appearance of standardized instruments (options, futures, duration-hedged products) with liquid markets across multiple tenors
- Publication of reference rates by credible benchmark administrators
- Adoption of term structure rates in agent service pricing
Falsifier: If no sufficiently liquid, reproducible, and governance-disclosed curves emerge, the proposed Bitcoin-native pricing profile fails. Agent coordination could still scale through other rails, references, auctions, or bilateral arrangements.
Claim 7.2: Agent-mediated commerce may become a material transaction category.
Classification: Forward-looking conjecture.
Observable indicators:
- Transaction volume attributable to autonomous agents (measured by on-chain activity, API logs, or market disclosures)
- Emergence of agent-native services with no human counterparty requirement
- Collateral locked in agent-controlled smart contracts
Threshold rationale: No universal percentage separates novelty from materiality. A release benchmark must name the market, denominator, observation method, and institutional consequence before assigning a threshold.
Falsifier: If systems capable of delegated economic action remain confined to pilots or trivial transaction shares across well-defined markets, the claim fails at the scale proposed. The clock begins only when the denominator and measurement protocol are registered.
C.7.X — Geopolitical Boundary Conditions
The framework assumes market allocation of compute resources. These claims establish the boundary conditions under which that assumption fails.
Claim 7.3: Leading-edge semiconductor fabrication is geographically concentrated.
Statement: Production at the most advanced commercial logic nodes is concentrated in very few foundries and heavily exposed to Taiwan. The precise share changes with the node, product class, date, capacity measure, and definition of "leading edge," so no timeless percentage describes the chokepoint.11
Source: TSMC investor presentations and earnings calls; Semiconductor Industry Association data; IC Insights and TrendForce market share estimates.
Magnitude that matters: A disruption affecting capacity that cannot be replaced within the relevant planning horizon would constrain frontier systems even if designs and demand remained intact. Geographic diversity matters only when substitute facilities can produce the required chips at usable yield and scale.
Update frequency: Annual. Track quarterly earnings disclosures from TSMC, Samsung, and Intel Foundry Services.
Falsifier: Commercially substitutable capacity distributed across independent geopolitical regions, with enough yield and volume to absorb a major disruption, would weaken the chokepoint claim.
Claim 7.4: U.S. export controls treat advanced compute as a strategic asset.
Statement: Beginning in October 2022, U.S. export-control rules restricted specified advanced-computing items and semiconductor-manufacturing equipment for China, with later rules revising coverage and compliance. The legal categories are technical and changing; the defensible inference is that advanced compute and its production stack have become objects of strategic control.
Source: Federal Register entries for BIS rules; Commerce Department public statements; Congressional testimony.
Status: This is an established fact. The question is whether controls remain, tighten, or relax over time.
Relevance to framework: Export controls demonstrate that states treat compute as strategic. If controls expand to include inference hardware, training clusters, or model weights, the market allocation assumption may fail for frontier capability.
Claim 7.5: State capture of apex compute is observable through specific signals.
Statement: If states move to capture apex compute infrastructure, the movement will manifest through observable signals before full capture occurs.
Observable signals:
- Licensing requirements for training runs above specified compute thresholds (some jurisdictions have proposed this)
- Mandatory government access to frontier model weights before deployment
- Nationalization of fabrication or datacenter infrastructure
- Compute allocation quotas or priority systems subordinating commercial use to state-designated objectives
- Extension of export controls to inference hardware and model deployment
Status: Some jurisdictions have adopted reporting, registration, procurement, or export-control measures. Those interventions do not by themselves establish comprehensive state allocation of compute.
Magnitude that matters: States are already principals in some procurement and allocation decisions, not merely friction. The scope boundary is narrower: where administrative allocation displaces price-mediated choice for a material class of compute, the market sequence proposed in the body does not govern that class.
Threshold rationale: No universal market-share threshold distinguishes regulation from administrative allocation. An evaluation must specify the controlled resource, jurisdiction, substitutability, duration, and ability of affected users to route around the intervention.
Falsifier: Sustained public allocation of a material compute class would falsify any account of its deployment based chiefly on decentralized price signals. Conversely, continued private allocation under contestable general rules would weaken the stronger capture scenario.
Claim 7.6: The framework's predictions are jurisdiction-contingent under bifurcation.
Statement: If compute allocation bifurcates between market-governed and state-governed jurisdictions, the framework's market predictions apply to the former but not the latter. V/C ordering and site- or operator-specific opportunity comparisons depend on price signals governing deployment decisions. Where administrative allocation replaces price signals, different mechanisms operate.
Relevance to framework: This is not a falsification condition but a scope limitation. Chapter 26 develops this contingency in detail.
Update frequency: Monitor for state-action signals annually. Flag jurisdictions that transition from market governance to administrative allocation.
C.8 — Update Protocol
The claims in this appendix should be reviewed annually. For each claim:
- Verify the source remains current and the magnitude remains accurate.
- Check whether falsification conditions have been triggered.
- Update magnitudes, sources, and dates as new data becomes available.
- Add new empirical dependencies as the framework develops.
- Flag claims referencing prior-year data for priority refresh (claims citing "2024" data in a 2026 review require verification or update).
Claims whose falsification conditions trigger do not automatically invalidate the framework. They require analysis of whether the failure is local (affecting one prediction) or structural (requiring revision of core mechanisms).
Versioning: Each annual review should increment a version number and append a changelog noting which claims were updated, added, or flagged for attention. The revision history enables tracking of how the empirical foundation evolves alongside the framework.
This appendix was reviewed on September 4, 2026 (Version 1.4). Updates should preserve the claim history while allowing the structure itself to change when a claim no longer belongs.
Version 1.4 changelog: Removed unsupported universal thresholds and unarchived point estimates; separated historical trends from forecasts; reclassified open-model lag, reinstatement lag, switching cost, inference energy, agent commerce, and state allocation as conditional claims; and replaced a fixed foundry-share claim with a node- and capacity-specific concentration test.
Version 1.3 changelog: Replaced the universal mining-floor claim with a site- and operator-specific opportunity-set claim. Narrowed Claim 3.4 to consensus reorganization and double-spend threats. Separated collateral deterrence from restitution, made benchmark scaling conditional, and reframed Bitcoin-native capital curves as risky, optional profiles.
Version 1.2 changelog (superseded by Version 1.3): Revised Claim 3.4 methodology using a $10–20 billion scenario estimate and described it as a general security floor. Version 1.3 removes that interpretation because consensus-attack cost does not measure key, custody, legal-seizure, or collectibility risk.
Version 1.1 changelog (historical framing, superseded): Added Claim 3.4 as a “51% attack cost as security floor,” plus Claims 7.3–7.6 on geopolitical boundary conditions. Version 1.3 retains the geopolitical claims and rejects the security-floor generalization.
Future reviews should preserve failed predictions as evidence scars rather than quietly moving their dates or denominators. A surviving claim earns continued attention, not immunity from a harder test.