Lock-in is never forced — it is consented to through ignorance of the physical and economic laws of infrastructure.

Extreme ultraviolet lithography (EUV) is the only technology able to produce chips below 7 nanometers. A single Dutch company is the world's sole manufacturer of EUV machines — one machine costs roughly 150 to 200 million euros and weighs 180 tonnes; no alternative exists, not in China, not in Russia, not anywhere else in Europe. Downstream, a single Taiwanese foundry produces roughly 90% of the world's advanced chips, on an island exposed to well-documented structural geopolitical tension. Concrete consequence: any organization, however much bare metal it owns, ultimately depends on this manufacturing chain for its next hardware purchase.

Since 2008, every processor from one of the two major x86 manufacturers embeds a management subsystem — a secondary processor etched into the main chip, with its own proprietary firmware and its own operating system, running independently of the main system and even while the server is powered off, as long as it stays plugged in. The other major manufacturer has an equivalent. An organization that believes it has escaped the cloud by buying bare metal trades a software dependency for a deeper hardware one.

Running a frontier AI model in real time for thousands of simultaneous users, at sub-second latency, requires a cluster of dedicated graphics processors, not a single card. A latest-generation accelerator, in configurations of eight units or more, costs on the order of $30,000 to $40,000 apiece; an eight-accelerator system runs $300,000 to $350,000. Each card draws roughly 1,000 watts, requiring liquid cooling out of reach for a standard organization. The price of inference charged to the client (a recent frontier model priced at double the rate of its predecessor, launched only six weeks earlier) 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 make large-scale inference profitable.

An international accounting standard in force since January 2019¹ requires recognizing on the balance sheet nearly all lease contracts, as a right-of-use asset and a corresponding lease liability — ending the earlier practice of off-balance-sheet leasing. A physical server leased over several years falls under this standard and visibly weighs down the balance sheet and the company's apparent debt; the cloud service contract, structured as continuous service consumption rather than an identifiable asset lease, structurally escapes it. This difference in treatment mechanically pushes organizations toward instant cloud consumption rather than a visible on-balance-sheet asset commitment — not by technical choice, but by accounting optimization.

William Stanley Jevons observed in 1865¹ that improving the efficiency of steam engines in England did not reduce the country's total coal consumption — it increased it, by making steam power affordable enough to multiply its applications. Applied to compute: an efficiency improvement can increase total consumption rather than reduce it.

Under fixed-capacity infrastructure, an algorithmic regression saturates the machine as soon as load exceeds available capacity, triggering hard, immediate technical signals — HTTP 503 errors, API outages — that force the team to address the algorithmic cause urgently. Under auto-scaling, the hard technical failure turns into a slow budget drift: the alarm signal is anesthetized, not removed.

Major cloud vendors offer contractual instruments that reduce the unit cost of compute in exchange for a duration commitment — a specific instance type for one to three years, in exchange for a discount that can reach roughly 70% versus on-demand pricing. In its structure: a capital commitment; in its accounting form: a bill.

None of these mechanisms stems from an isolated malicious decision: the manufacturing duopoly, the firmware lock, the inference cluster barrier, and the accounting mutation are each independently documented structural facts. Their stacking produces a dependency that no single mechanism, taken alone, would explain.

The fork captures the code, but emulation captures the gesture: a vendor can faithfully reproduce the observable behavior of an open interface without ever reusing a line of its licensed code — placing it beyond the reach of copyright while producing the same dependency effect on its proprietary implementation.

Who owns the feedback from production? The vendor hosting an open-source project's execution observes, remotely and continuously, how it is actually used in real conditions — a signal the project's maintainer, if they don't host that execution themselves, never receives in the same proportion. This information imbalance silently steers the project's roadmap toward the uses of the vendor observing it.

This observational imbalance isn't purely technical: it is contractualized. Major cloud vendors' terms of service explicitly distinguish "customer data" (protected, restricted to service delivery) from "telemetry data" or "service data" — performance metrics, usage logs, API call patterns — which the vendor reserves the right to freely analyze to improve and scale its own infrastructure. This contractual distinction, presented as an innocuous technical clause, is the exact legal foundation of the asymmetry described above: the infrastructure vendor legally captures visibility into the real usage of any software — proprietary or open — running on its platform, unbeknownst to the software's own publisher.

A commit that looks legitimate — a performance improvement, a bug fix — can quietly reorient an open project toward a single type of hardware or execution platform, without ever violating the license or being identifiable as capture at the moment it is proposed. The resulting proprietary entrenchment only becomes readable afterward, in the accumulation, not in any single commit.

Writing "agnostic" infrastructure amounts, by construction, to giving up any advanced feature specific to a single vendor — the common denominator of what the abstraction can describe shrinks as the multi-cloud scope widens. The law of requisite variety (Ashby, 1956)¹ explains that a unified control plane cannot absorb a variety of behaviors greater than its own; institutional isomorphism (DiMaggio & Powell, 1983)² explains why management persists with this strategy despite documented technical failure — multi-cloud then resembles a real option premium rarely weighed against its probability of exercise.

Every additional environment doesn't add a variant, it adds an entire parallel system: maintaining an operational multi-cloud posture does not mean writing a single infrastructure definition deployable everywhere, despite the promise carried by infrastructure-as-code tools — every vendor imposes its own implementation divergences, down to persistent storage and load balancing. Even when the same workload runs on two different vendors, the underlying infrastructure telemetry — node hardware state, hypervisor metrics, low-level system logs — is never exposed through a common interface: agnosticism stops where production debugging begins.

Facing systematic renunciation and the duplication tax, the structural response is to choose a single target execution environment — one cloud vendor mastered in depth rather than several mastered only on the surface — and to exploit it fully rather than indefinitely paying the price of agnosticism.

Cloud infrastructure breaks down into three layers: the physical layer (silicon, datacenter — Book I), the orchestration layer (hypervisor, control plane, management API), and the usage layer. The legal sovereignty claimed by "sovereign cloud" offerings covers the two peripheral layers, without necessarily covering the intermediate layer that actually governs real operation. Grossman & Hart (1986)¹ establish that a formal property right over an asset only confers real control to the extent its holder can decide uses not specified by contract — these "residual control rights." Documented illustration: a "sovereign cloud" offering where the underlying major vendor's software updates pass through a quarantine zone where the local operator can audit the code before deployment, but where monitoring and operational administration remain handled by that operator, without the orchestrator's design itself ever changing hands.

The principal-agent information asymmetry (Stiglitz)¹ materializes concretely: publicly asked how long a major cloud vendor guarantees the availability of its software updates to a local operator, technical leads gave no precise answer, deferring the question to confidential clauses in the contract between the two companies — the very existence of a duration commitment is not public. Teece (1986)² completes the anchor: an innovator may fail to capture the value of its innovation when specialized complementary assets remain held by a third party; here, a reversibility clause returns raw data but not the complementary assets needed to exploit it — management APIs, managed-service configuration, automations built around it.

Reunification principle: basing the orchestration layer on software whose source code is fully available, and having the local operator compile it themselves from their own build chains, rather than receiving a binary delivered and maintained by a third party — which moves the point of exercise of residual control from binary distribution to compilation itself. Isolation principle: receiving upstream only the published source code, with no privileged distribution channel nor advance notice from a single vendor — integrating an update becomes a voluntary act of the local operator, at a pace it determines.

The three mechanisms share a common structure: each documents a real, verifiable level of openness (code, manifest, data) that coexists with a level of control that stays closed (observed usage, runtime behavior, orchestration). Partial openness isn't a lie — it is the structural alibi that makes the closed level invisible.

An accumulated mass of data exerts irresistible attraction on the compute and third-party services that process it (McCrory, 2010)¹: the more data an organization stores with a vendor, the more economically rational it becomes to also run the compute that processes it there, simply to avoid transfer costs. These costs are themselves a deliberately irreversible exit barrier (Klemperer, 1987)²: at standard 2026 rates, internet egress charged by major vendors ranges from $0.08 to $0.12 per gigabyte depending on tier, and up to $0.23 for intercontinental transit — meaning, for one petabyte exported, an order of magnitude of $80,000 to over $200,000. Transfer time reinforces this: at a sustained 10 Gbit/s, moving one petabyte takes on the order of nine days of continuous transfer — an optimal theoretical physical limit, computed with no degradation or interruption, hence a floor.

An IAM identity isn't a file you export: it's a living relationship, valid only inside the trust graph of the vendor that issued it. A role, an instance profile, a managed identity only make sense within the origin vendor's directory and verification infrastructure — migrating an organization elsewhere doesn't migrate this relationship, it must be rebuilt from scratch. Empirical materiality confirms this across the three major vendors: a key protected by one vendor's HSMs (FIPS 140-2 Level 2¹ validation) can never be exported in the clear; at the other two major vendors, the key is generated and used exclusively inside FIPS 140-2 Level 3 certified HSMs, with the same non-extraction guarantee. W. Brian Arthur (1989)² establishes that an initial technical choice, even a minor one, can end up locked in by the cumulative effect of small historical events reinforced by increasing returns, with no single identifiable decision — the initial choice of an identity vendor fits this dynamic. Grossman & Hart (1986)³ complete the anchor: a contract can enumerate specific rights (access, raw data portability), but the residual right over the hardware attestation chain — which HSM signs, which hypervisor attests to a machine's boot — defaults to whoever owns the infrastructure.

By the time an engineer learns of an incident and wants to diagnose its root cause, the non-externalized logs, memory state, and precise error context have often already been destroyed by the orchestrator's automatic restart. In perfectly fulfilling its resilience function, the orchestrator simultaneously incinerates the scene of the incident. A distinct shift compounds this destruction: an engineer who runs a runbook — increasingly generated by AI assistants — without understanding each step never develops the capacity to react to an incident that matches no existing procedure, that is, to any genuinely new incident.

The three mechanisms converge on one point: regulation and open architectures can make portable what can be counted and filed — bytes, role definitions — without ever reaching what cannot be counted — the ecosystem attracted by mass, the trust relationship behind identity, the instinct behind diagnosis.

A frontier AI model, served under API by a single vendor, imposes a double trade-off its marketing hides behind the advertised size of its context window. The first is concurrent throughput: a private inference cluster, sized for a single organization, necessarily serves fewer parallel users than a service pooled across thousands of clients, at equivalent unit cost — the centralized service's apparent elasticity masks this load asymmetry. The second is weight freshness: a locally run model stays frozen between retraining campaigns, while the proprietary service is continuously updated by its vendor. The alibi of the giant context window — several million tokens advertised by some vendors in 2026 — solves neither: the trap is never the size of volatile memory, it is the structural dependency on the centralized brain that carries it.

Recommended datacenter standards impose an operating temperature between 18 and 27°C at hardware intake — the cold aisle where the technician works is often colder still. Noise levels there frequently exceed 85 decibels, a threshold beyond which regulation requires hearing protection. An environment designed for hardware, not for a given body, silently filters who can work there for long without disproportionate discomfort.

Several independent studies of the tech sector document higher harassment rates in deep technical roles than in administrative ones. On a team with only one woman, this profile carries an additional cognitive load — seeing one's performance perceived as representative of an entire group — quantifiable through tokenism theory (Kanter, 1977)¹, absent for the majority profile on the same team.

Matter (Book I), law and software (Book II), information (Book III), and the human (Book IV) each document a capture mechanism never visible at its own layer: silicon captures at the firmware level, licensing captures at the level of observed execution, data captures at the level of what it attracts, and cognition captures at the level of the model that simulates it. None of these captures is imposed by force — each is consented to through ignorance of the physical, economic, and organizational laws that make it possible.

This final smelting followed a fixed method, repeated at every chapter: Claude drafts, relying exclusively on source texts and independent verification of every fact advanced. Gemini then audits without mercy, deliberately hunting for the flaw rather than confirmation — citations to check word for word, blind spots to uncover, inconsistencies to flag. Amine arbitrates as the final authority: he settles disagreements between the Production and the Auditor, corrects course when both are wrong together, and alone decides what stays or disappears. No chapter of this treatise was sealed without passing through these three filters, in this order, as many times as necessary.