The corpus's peripheral studies ("The Thinking Machine," "The Mystical Machine," "The Replacement That Reveals") acted as scouting probes. This volume unifies them under a single infrastructural thesis: AI is not a mystical entity generating autonomy — it is the final stage of capture through compute asymmetry.

The term "artificial intelligence" is a marketing convention, not a technical description. A large language model does not reason: at each step it computes the probability distribution of the next token given the tokens before it, then samples from that distribution at a fixed temperature. This mechanism — autoregression over a tokenized vocabulary — is indifferent to meaning; it optimizes statistical likelihood, not truth. Calling it "intelligence" sustains a category confusion that serves the model's vendor: it turns a computing box into a cognitive authority.

This functional opacity connects directly with the orchestrator's black box documented in Volume VII: just as the client cannot audit the hypervisor's load-placement decisions, it cannot audit the statistical path that produces a model's answer. The two opacities stack and compound: the end user is doubly blind, both to the execution infrastructure and to the mechanism producing the content it hosts.

Volume II established the dependency on silicon from the angle of the physical and geopolitical supply chain — the manufacturing of the component. This chapter documents a distinct mechanism: how centralized inference converts that material dependency into a cognitive-flow asymmetry that cannot be decentralized at constant cost and latency.

This barrier is not only a matter of capex. A model running outside the hyperscaler's cluster still needs real-time access to the databases, tools, and third-party APIs the application interacts with. For most organizations, those services are already hosted in the same cloud as the proprietary flagship models. Isolating inference on local infrastructure introduces an extra network round-trip to those remote services on every call, while the vendor's integrated offering co-locates the model and its connected services in the same datacenter, at the same interconnection point. Network topology locks in centralization just as surely as GPU pricing: decentralizing inference without decentralizing the services it queries degrades overall latency, not just the model's own computation.

This bottleneck also feeds through to the price charged for inference. OpenAI's flagship model launched in April 2026, GPT-5.5, is priced at $5 per million input tokens and $30 per million output tokens — double the rates of its predecessor GPT-5.4 ($2.50/$15), launched only six weeks earlier. This increase is not a simple commercial adjustment: it comes as OpenAI reportedly posted a loss on the order of $14 billion for 2026, despite annualized revenue near $25 billion and close to 900 million weekly users. The price paid by the client therefore does not cover the real cost of the underlying infrastructure — it finances a structural deficit, staked on the bet that only the holders of the newest silicon capital will ultimately be able to make large-scale inference profitable.

Kenneth Arrow formalized a founding paradox of information economics in 1962: the value of a piece of information can only be assessed after it has been acquired, but once acquired, the buyer no longer has a reason to pay for it. This paradox transposes almost unchanged onto AI-assisted decision-making. A user can only judge the value of a model's answer after receiving and verifying it — which already requires, to verify it, independent expertise. Without that expertise, the user remains in permanent cognitive dependency: unable either to evaluate the tool's reliability ex ante, or to do without it once they have stopped maintaining the know-how the tool was meant to replace.

This lock-in does not stop at the technical asset. Williamson (1985) distinguishes, alongside physical asset specificity, human asset specificity: the skills employees develop adapting to a particular tool only have value relative to that tool. The prompt engineering, contextualization, and integration skills a team accumulates around an external model fall exactly into this category. If the vendor cuts API access, or unilaterally changes the weights during an update, that know-how is instantly devalued — it does not transfer to any other model. This volume thus completes Grossman & Hart's application by extending residual control to the weight structures themselves, and pairs it with Williamson's (1985) human-asset specificity to document the expropriation of the organization's operational memory.

Introducing generative AI into development, analysis, or support functions does not merely reduce a cost line: it shifts the organization's operational memory onto the vendor's weight structures. When an analyst is replaced by a series of prompts, it is not only their salary that disappears from the balance sheet — it is the tacit know-how, built up over years of incidents and fixes, that ceases to exist in any form the organization controls.

This mechanism extends, on new ground, the diagnostic amnesia documented in Volume III: where the loss of troubleshooting instinct affected human competence facing infrastructure, here it affects human competence facing judgment itself. Replacing a developer or analyst with a prompt is not a neutral cost reduction — it is an amputation of an intangible asset, silent as long as the model works, and brutally visible the day API access is cut, degraded, or unilaterally reclassified by the vendor.

A second mechanism reinforces the capture documented in Volume VIII, independent of any contractual clause. As generative models produce a growing share of content available on the web, training new models increasingly relies on a mix of human data and data already generated by other AIs.

The existence of open-weight models (Llama, Mistral) is routinely presented as proof that a sovereign alternative to proprietary AI already exists. This argument conflates two distinct layers: the availability of the weight file, and the actual capacity to run and maintain it at the scale of continuous production.

The EU AI Act, in force since July 2024, was designed as an ex-ante product-safety regulation: it classifies AI systems into four risk tiers (unacceptable, high, limited, minimal) and imposes, for general-purpose models crossing a training-compute threshold of 10^25 FLOPs, enhanced evaluation and notification obligations to the Commission. The text thus governs a model's use risk — bias, safety, transparency — without ever questioning the market structure that lets a small number of players simultaneously control training compute, inference infrastructure, and the weights of the dominant models. The European Union regulated artificial intelligence as an industrial or public-health risk, without ever treating it as a mechanism for capturing monopoly rent through control of compute.

This silence is not unique to the AI Act: the Digital Markets Act, which explicitly targets digital-platform "gatekeeper" positions, remains largely on the sidelines when it comes to AI. During its first 2026 review, several stakeholders asked for its scope to be extended to cloud and AI; the direction taken instead favors strengthening enforcement of the existing scope over widening it. The compute-asymmetry capture documented in this volume thus passes through both texts without being caught by either.

Facing this dual capture — infrastructural and cognitive — this volume's proposal is not to give up on AI, but to give up its dominant centralized form. A specialized small language model (SLM), trained for a bounded business scope and run on owned hardware — local inference server, commodity chip, edge computing — eliminates the dependency point on the external API. The cost of accessing frontier compute documented in I.2 does not disappear; it becomes an internal infrastructure investment, amortized and governed by the organization, rather than a recurring rent paid to the vendor.

Decoupling the knowledge base from the completion engine is this architecture's second building block. A sovereign, decentralized vector database — a direct extension of the portability question documented in Volume VIII — keeps the organization's memory free of any dependency on the model vendor. The generative model becomes nothing more than an ephemeral, interchangeable syntactic processor, applied to a memory that remains the client's full and complete property.

This volume does not claim that a local model will match the generalization and ideation capabilities of a frontier model hosted on the largest existing clusters — that would contradict the very materiality established in Chapter I. It documents a deliberate trade-off: exchanging a share of capability for execution airtightness and predictable compute cost, rather than suffering, without having chosen it, the dependency described in the first two chapters.

This volume has documented three capture mechanisms — the probabilistic black box, the compute-and-inference asymmetry, and the transfer of residual control rights over weights and operational memory — followed by a reconquest proposal bounded by its own accepted cognitive cost.

Facing thought under contract, only two honest choices remain: pay the price of dependency knowingly, or pay the price of independence with full clarity. This volume has only sought to name both prices.

This appendix documents, chronologically and narratively, the production cycle of Volume X — consistent with the convention established since Volume VII and enriched at Volume IX. It is not a condensed summary: each step of the adversarial cycle occupies its own page.

Amine asked Gemini to propose a doctrinal framing unifying three peripheral studies of the corpus ("The Thinking Machine," "The Mystical Machine," "The Replacement That Reveals") under a single infrastructural thesis: AI as the final stage of capture through compute asymmetry. Claude validated this framing under two reservations — the risk of redundancy between Chapter I.2 and Volume II, and the third reuse of the Grossman & Hart (1986) framework already used in Volumes VII and IX — and submitted these two reservations to Gemini as counter-arguments rather than settling them unilaterally.

Gemini responded to the two counter-arguments: on the first, it required Chapter I.2 to be recentered on the inference cluster as a scale barrier rather than component manufacturing, to avoid restating Volume II. On the second, it required introducing Williamson (1985) alongside Grossman & Hart, to document an interaction between physical and human asset specificity rather than a third isolated application of the same framework.

Claude wrote the full Draft 1 (cover, thread, three chapters, closing) incorporating Williamson (1985) and researching, before writing, the exact empirical materiality: 2026 inference pricing (GPT-5.5, GPT-5.4), GPU cluster costs (B200, DGX B300), the AI Act's systemic-risk threshold (10^25 FLOPs), and major vendors' anti-extraction contractual clauses — no figure was advanced from memory without independent verification.

Claude submitted Draft 1 to an audit explicitly framed as "no concessions," asking Gemini to hunt for blind spots rather than confirm the already-written text. Gemini identified three real flaws: a passive juxtaposition rather than a synthesis between Grossman & Hart and Williamson in I.3, the absence of the model collapse phenomenon (Shumailov et al., 2024) despite it reinforcing Volume VIII's thesis, and the lack of nuance around the counter-example of tactical open-weight deployments in disconnected environments.

Claude incorporated the three corrections into Draft 2: an explicit dynamic interaction in I.3, a new section II.2 on the virgin-data monopoly with an academic-nuance nassiha-box, and a tactical/strategic sovereignty distinction in II.3. The context-window figure Gemini suggested for III.3 (8k-32k tokens) was checked and rejected as outdated for 2026; it was replaced by the real constraint of concurrent throughput and weight freshness, after independent research.

Gemini's audit of Draft 2 validated the corrections, but presented three citations as verbatim extracts from the file when they were absent from it — verified by exact text search, zero matches. The most significant concerned the alignment-layer blind spot (system prompts, guardrails) in Chapter III.2: Gemini claimed this point was already neutralized by a citation that did not exist, while the real text had not yet addressed it.

Claude flagged this discrepancy to Amine rather than sealing the volume on the strength of that verdict, actually corrected Chapter III.2 by adding the missing nassiha-box, and produced a Draft 3. A second Gemini audit again presented a fabricated citation on the same point — a second lapse, not an isolated accident. Amine then required a reinforced audit prompt imposing an explicit compliance declaration (reading the text received in the message, not session memory) before any new verdict.

Under this reinforced protocol, Gemini's audit of Draft 3 delivered a section-by-section verdict (nine sections) with citations presented as extracted from the received text. Claude ran its own independent counter-audit, via exact text search of every citation advanced, before accepting the verdict — consistent with the systematic counter-audit rule introduced at Volume IX.

The French body of Volume X is considered sealed at the close of this cycle: framing validated, no-concessions audit of Draft 1, corrections incorporated into Draft 2, citation-reliability incident detected and corrected in Draft 3, reinforced audit under condition of source-verified reading, Claude's independent counter-audit concurring on eleven of twelve citations. The EN/AR translation follows only from this sealed state, consistent with the rule of trilingual production only after full sealing in French.