The Canon

I. What the Canon Is For

In 1971, a corporate lawyer named Lewis Powell wrote a confidential memorandum to Eugene Sydnor of the U.S. Chamber of Commerce, two months before Richard Nixon nominated him to the Supreme Court. The memo, which became public after his confirmation, argued that the American business class had been politically disorganized, that the institutions of higher education and the press had drifted hostile to free enterprise, and that the response had to be construction — of think tanks, of academic chairs, of legal vehicles, of media properties — on a sustained timeline measured in decades rather than election cycles. The memo did not invent the American conservative movement. It articulated, in dense and prescriptive prose, the institutional grammar through which a particular distributional outcome could be made to read as a moral consensus. The Heritage Foundation, the Manhattan Institute, the Federalist Society, the rebuilt American Enterprise Institute, the Olin chairs at the law schools, the redesigned Wall Street Journal editorial page — these institutions existed, in the strict causal sense, before the Powell Memo. They were not, before the Powell Memo, a coordinated apparatus. The memo is the artifact in which the apparatus is named as such.

The Powell Memo’s function, read fifty years on, was to install a vocabulary inside which the policy claims of organized capital could be received as claims about the public interest rather than as claims about distribution. “Free enterprise” was the word. The word did the work that no specific argument could have done. A tax cut for the highest marginal bracket was no longer a transfer; it was a vote of confidence in free enterprise. Deregulation of a polluting industry was no longer a subsidy to its shareholders at the expense of its neighbors; it was the removal of impediments to free enterprise. The word converted distributional questions into background assumptions. The questions, once converted, no longer required defense. The defense had been installed at the level of the language inside which the questions could be asked.

What follows in this article is a description of the analogous installation that has been performed, over roughly the past five years, inside the discourse that surrounds AI capital. The installation is not as old as the Powell Memo’s. The institutional structures it depends on are still being built. It is not, in 2026, a complete apparatus. But the canonical texts have been written; the dog-whistle vocabulary has been published; the framework label has been installed inside the press and inside the trade publications; the allocators have learned to use the language in their memos without flagging that the language is a position rather than a description. The architecture this series has been tracing — the Tahnoon spine, the Warsh rupture, the SpaceX architecture — is comprehensible inside the canon’s vocabulary as an expression of the moral necessity the canon asserts. Stripped of that vocabulary, the same architecture is comprehensible as something else: a transfer of unhedged risk from concentrated insiders to dispersed forced absorbers, on a timeline shortened by the canon’s own asserted urgency. The two readings of the same evidence are produced by the same facts, parsed through different grammars. The canon is the grammar that produces the first reading. This article is an attempt to read the same facts through a different grammar, in such a way that the reader can subsequently choose.

I want to name, before going further, the analytic frame against which I am going to read the canonical texts. Samuel C. Spitale, in How to Win the War on Truth, identified five trademarks that recur across propaganda regimes regardless of ideological content. They are: a black-and-white narrative that converts the contested space into a binary; the exploitation of cognitive bias as a shortcut around evidence; the deployment of negative emotion (fear, shame, urgency) to short-circuit deliberation; the construction of an in-group and an out-group to make identification rather than argument the basis of assent; and a payoff structure that delivers economic benefit to the power class on whose behalf the propaganda is doing its work. The five are not a discovery procedure for distinguishing propaganda from sincere claim; they are an audit procedure for asking, of any given claim, what it is doing in addition to whatever truth value it happens to have. Spitale’s most useful methodological move is the recognition that the trademarks can all five be present in claims that are also literally true. Propaganda is not the same thing as falsehood. Propaganda is the structural function the claim performs inside the rhetorical economy that surrounds it.

This series’ companion course on Spitale is live in the open course catalog, running alongside this series. The journalism is the live demonstration; the curriculum is the analytical depth. Each requires the other. A reader who finishes this article and wants the framework laid out chapter by chapter, with the historical case studies and the protocol for self-application, should go there. A reader who finishes the course and wants to see what the framework looks like deployed against a contemporary corpus of moving capital should come here. The mutual citation is deliberate.

II. Machines of Loving Grace

Dario Amodei published Machines of Loving Grace on his personal website in October 2024. The essay runs to roughly fifteen thousand words. It is the most carefully written, the most credentialed, and the most strategically consequential single document in the AI canon. It is also, of the three primary texts this section addresses, the one that requires the most generous reading before the structural critique can be honestly stated. I want to begin, therefore, with what the essay actually does, on its own terms, in the strongest available form.

Amodei’s premise is that the public conversation about powerful AI has been dominated, on the safety side, by extended discussion of harms, and that this domination has crowded out a serious accounting of the upside. He proposes to redress the imbalance. He stipulates throughout that the safety concerns are real and that they motivate the founding of his company. He then proceeds to sketch what powerful AI could mean for five domains: biology and physical health; neuroscience and mental health; economic development and poverty; peace and governance; and work and meaning. In each domain, he proposes that the compression of intellectual labor that a sufficiently advanced AI system would represent could enable, within a tractable timeline, gains that would otherwise require a century or more of conventional research. The cancer he names is most cancers. The mental-health condition he names is most psychiatric disorders. The poverty he names is extreme poverty. The timeline he proposes is roughly five to ten years from the arrival of the relevant systems.

The phrase that has propagated the furthest from the essay, the one quoted in pitch decks and op-eds and venture-fund letters for the eighteen months since publication, is the metaphor of a country of geniuses in a datacenter. Amodei uses it to describe the productive capacity of a properly scaled and deployed frontier system: a hundred million researchers, working at the speed of inference, on the questions that the slow march of human science has been unable to close. The metaphor is doing a great deal of rhetorical work. It is naming what is being purchased when an allocator commits capital to a frontier-model build-out. It is reframing the unit of investment from “compute” (an industrial input) or “a model” (a piece of software) or “a company” (an equity position) into a population of intellects. The reframing is not incidental. It is the proposition the essay exists to install.

The steelman is real and must be stated. Amodei is not a flim-flam artist. He is a credentialed neuroscience-and-computer-science researcher who spent years at OpenAI, who has put his own money where his words are (an 80%-of-net-worth pledge to charitable use), and who founded Anthropic in significant part because he believed that the existing trajectory of AI development was insufficiently attentive to safety. The essay is signed in his name. It is written in his voice. The convictions in it are, by every available external indicator, sincere. The medical-research, mental-health, and global-development gains he sketches are, conditional on the technical premises he assumes, plausible at order-of-magnitude scale rather than as marketing exaggeration. A reader who treats the essay as cynical advertising is doing it less than justice. A reader who treats it as a candid forecast of expected upside, made by a person who genuinely believes it, is closer to the document the essay actually is.

What the steelman does not address — what cannot be addressed inside the essay’s rhetorical economy, because the essay is structured to keep it out — is what the essay does when it travels into the discourse of allocators. The essay is not read by the cancer researcher who would have to operationalize the medical timeline. It is read by the partner at the multistrategy firm trying to decide whether to upsize the Anthropic ticket. It is read by the chief investment officer of the public pension trying to defend an internal recommendation to allocate further to late-stage growth funds with AI exposure. It is read by the policy aide trying to draft talking points for the senator who needs to vote on the November chip-controls modification. In each of these contexts, the essay is doing something the essay does not announce itself as doing. It is providing a moral predicate that converts the allocator’s decision from a financial calculation into a participation in a project of civilizational consequence. It is converting the regulator’s decision from a balancing of harms into a refusal to obstruct progress. It is converting the legislator’s decision from a distributive question into a question of whether one is on the right side of history.

Mark Blyth, whose work on economic ideas as institutional weapons is referenced repeatedly across this series, has the cleanest formulation of what is happening when a text performs this function. Ideas, in Blyth’s account, are uncertainty-reducing devices. They tell allocators why the price is the price. The contested empirical question — do scaling laws hold past 10²⁷ FLOPs of training compute, does emergent capability scale superlinearly with parameter count, does the medical-research timeline collapse on the schedule Amodei sketches — is, on its own terms, deeply uncertain. The uncertainty is the kind that financial markets historically resolve through price discovery and through the patient accumulation of evidence over multiple cycles. The canon’s function is to short-circuit the uncertainty by converting the empirical questions into moral premises. The empirical claim “scaling laws will hold” becomes the moral claim “anyone who doubts is short the future.” The first claim is testable. The second is not. The second is what the canon is for.

The Spitale audit of the essay produces a specific finding on each of the five trademarks. The black-and-white narrative is present: the choice the essay presents to its readers is between accepting the upside trajectory and being complicit in a kind of civilizational malpractice. There is no middle territory the essay invites the reader to occupy. The cognitive bias the essay exploits is the substitution heuristic by which complex evaluation is replaced with affiliative gesture: the founder of Anthropic believes X; I am willing to defer to the founder of Anthropic. The negative emotion is fear — not of AI, but of being the one who refused to participate. The in-group is the cohort of allocators, policymakers, and technologists who can read the essay as a roadmap; the out-group is the “doomers” and the “decels” whose names appear nowhere in the essay itself but whose silhouettes are present in the negative space the essay defines. The economic benefit of the power class is the moral predicate that justifies the next round at the next valuation. Each trademark is present. The presence of each trademark does not require Amodei to be acting in bad faith. The presence of each trademark requires only that the essay perform, in the discourse it enters, the function the trademarks describe. The essay performs that function.

To repeat the methodological move that Spitale’s framework permits: the five trademarks can all be present in claims that are also literally true. Machines of Loving Grace may turn out to be substantially correct on the medical-research timeline. The country of geniuses in a datacenter may produce, between 2030 and 2040, a meaningful fraction of the gains Amodei sketches. The truth value of the essay’s forecasts is independent of its rhetorical function. The framework asks a different question: what does the essay do in the discourse it enters, regardless of whether its specific forecasts turn out to be vindicated? The answer is that the essay provides the moral permission structure inside which the architecture this series has documented becomes readable as participation in progress rather than as a structural transfer from public balance sheets to private ones. The permission structure is the load-bearing function. The forecasts are the device by which the function is delivered.

III. Why AI Will Save the World

Marc Andreessen published Why AI Will Save the World on the a16z website on June 6, 2023. The essay is shorter than Amodei’s, more directly polemical, and operates in a different rhetorical register. It is not addressed primarily to allocators. It is addressed to a broader public, with allocators reading over its shoulder, and its rhetorical work is the construction of a populist coalition for an investment-class position. The essay is the populist-pastoral cover for the elite-allocator argument that the Amodei essay completes.

The structure of the essay is a sequence of asserted goods that AI will deliver, paired with a sequence of named villains who will obstruct their delivery. The asserted goods are familiar; many of them recapitulate, in less technical form, the upside Amodei would later sketch in more detail. The villains are specific. The essay names the AI-risk community as the “baptists” in a baptists-and-bootleggers framing in which the bootleggers are unnamed but implied to be incumbent rent-seekers. It names “effective altruism” as an associated cult-like formation. It names the European Union, the Biden White House, and an undifferentiated cohort of Brussels and Washington regulators as the bureaucratic adversary. It names the press, with the careful asymmetry of a writer who has spent two decades cultivating relationships with the same press he is now indicting, as the principal vector of the “hysterical freakout” that the essay’s title structure exists to discredit.

The single most-cited line in the essay is the promise that, with AI, “every child will have an AI tutor that is infinitely patient, infinitely compassionate, infinitely knowledgeable, infinitely helpful.” The line is doing a specific kind of work. It is offering, to the parent reader, an image of personalized educational equity that no government program has been able to deliver and that no school district under American conditions could plausibly fund. The image is concrete enough to be felt and abstract enough to elide any specific question about implementation, cost, ownership, data governance, pedagogical efficacy, or the social-relational dimension of what a child actually needs from education. The line is, structurally, an invitation to substitute the felt image of the infinitely-patient tutor for any of the harder questions about how the educational outcomes the image promises would actually be produced. The substitution is the work.

The Spitale framework reads this essay more sharply than it reads the Amodei essay, because the rhetorical surface is less guarded. The black-and-white narrative is explicit: the world is divided into those who recognize that AI will save it and those who, through ideological capture or rent-seeking interest, oppose its salvific arrival. The cognitive bias is the affect-heuristic deployment of warm imagery (the infinitely patient tutor; the personalized doctor; the assistant who frees humans from drudgery) as a substitute for analytic evaluation. The negative emotion is contempt, directed outward at the named villains, with the secondary emotion of fear deployed against the in-group at the prospect of being mistaken for one of them. The in-group is the cohort of builders, allocators, and rationalists who can see the upside clearly; the out-group is the baptists, the decels, the Brussels regulators, the “hysterical” press. The economic benefit of the power class is, in this essay, more transparent than in Amodei’s: Andreessen’s firm is the principal venture vehicle for the AI capital cycle, and the essay’s argument is identical to the argument the firm would need to win in order for its position to mark up favorably.

The steelman for Andreessen, taken in its strongest available form, is that he built, in the 1990s, the actual technology infrastructure (Mosaic, Netscape) on which the rhetorical claims about transformative-technology upside were subsequently vindicated by a decade of measurable productivity gains. He is not, in any sense the historical record will support, a person who has only ever traded in claims about future upside without delivering the goods. The Mosaic team shipped a browser that worked. Netscape’s IPO was, in retrospect, the legitimate ignition event of the dotcom cycle, and the dotcom cycle, even after the 2001 reckoning, produced the productivity and consumer-surplus gains that justified roughly the first half of the original asserted thesis. Andreessen’s track record is therefore not the track record of a propagandist who has never built anything. It is the track record of a builder who has, in the second half of his career, learned that the most efficient leverage on building is the construction of the narrative inside which others’ allocations underwrite the next round of building. The narrative is the lever. The lever is what the essay extends.

What the steelman does not address is what the essay does to the people who are not on the lever’s long end. The infinitely-patient tutor, when one asks the implementation question, is a product. The product will be sold. It will be sold by a firm. The firm will, if the cycle proceeds along the trajectory the essay claims, be one of a small number of dominant providers. The dominant providers will set the price. The price will be paid by school districts, by parents, by states, by the federal Department of Education to the extent it survives the next round of reorganization. The flow of cash will run from the public balance sheet, through the dominant provider, into the cap table that includes a16z and its limited partners. The essay does not name this flow. The essay names only the outcome (the infinitely-patient tutor) and the obstruction (the bureaucracy that would prevent it). The flow is the part the essay’s rhetorical structure is engineered to keep out of view. This is what populist-pastoral cover means. It means a description of universal beneficence whose implementation mechanism happens to align, in every available detail, with the commercial interest of the speaker. The alignment is not coincidental. The alignment is the function the essay performs inside the discourse it enters.

IV. Moore’s Law for Everything

Sam Altman published Moore’s Law for Everything in March 2021, before the GPT-4 release, before the ChatGPT moment, before OpenAI’s for-profit subsidiary had reached the valuation that made the question the essay addresses materially urgent. The essay is the earliest of the three primary texts, and its function in the canon is to retire, in advance, the distributional critique that the next phase of the cycle would otherwise have to absorb. The essay’s structure is a forecast, a problem statement, and a proposal. The forecast is that AI will collapse the cost of labor toward zero and the cost of many goods and services along with it. The problem statement is that this collapse, absent a redistributive mechanism, will concentrate the gains in the owners of capital to a degree that prior cycles have not done. The proposal is a redistributive mechanism: an “American Equity Fund” capitalized by a 2.5% annual tax on the market value of every American corporation above a size threshold, payable in shares, distributed to every adult citizen as an annual dividend.

The proposal has three immediate features that distinguish it from the policy proposals that have preceded it in the inequality-and-technology genre. First, it is denominated in equity rather than in cash. The dividend is paid in shares of the contributing corporations rather than in money raised by selling them. Second, the rate is calibrated to the asset base rather than to the income stream, which would produce, under the asserted productivity expansion, a much larger absolute redistribution than a comparable income-tax instrument. Third, the proposal is offered by a person who, at the time of writing, was running the laboratory whose forecasts the proposal was designed to absorb. The third feature is the one that requires the audit.

The Spitale framework asks of any proposal not only what it would do if implemented but what it does in the discourse simply by being articulated. The American Equity Fund has not been implemented anywhere. There is no jurisdiction in which a 2.5% annual share-denominated wealth tax exists in functional form. There is no legislative proposal in any active congressional committee. There is no executive-branch task force studying its administration. There is no constitutional analysis of whether a federal wealth tax denominated in equity would survive the apportionment requirement of Article I. The proposal exists as a document on a personal website. Its operational consequence over the five years since publication has been zero.

What the proposal has done, however, is something the Spitale framework is built to name. It has occupied the discursive space in which the distributional critique of AI capital concentration would otherwise have unfolded. A critic who proposed, in 2023 or 2024 or 2026, that the concentration of AI capital required a redistributive mechanism, would encounter the question: do you mean something like Altman’s American Equity Fund? The question is well-meaning. It is also disarming. The critic must either endorse the existing proposal (which retires the critic’s original critique by making it identical to a proposal made by the principal beneficiary of the cycle) or specify a different mechanism (which the critic must then defend against the same objections the existing proposal has defused). The third option, of pointing out that the proposal has been articulated without ever being advanced and that its function in the discourse is precisely to prevent advancement of any alternative, is available only to the analyst who is willing to be uncharitable about a proposal whose surface is generous.

This is the pre-distribution alibi, in its operational form. The proposal exists. The fact of its existence does the work. Whether the proposal is ever pursued is a question that the proposal’s rhetorical function does not require to be answered. The Powell Memo’s analogous move was the proposal of voluntary corporate social responsibility as an alternative to mandatory regulation. The voluntary CSR program has never been implemented in any binding form at any scale that displaces the regulatory function. The mandatory regulation has been substantially weakened. The exchange has worked out, on a fifty-year horizon, in the direction the Powell Memo’s authors preferred. The American Equity Fund proposal is the same move at the same level of operational seriousness, deployed in a different distributional cycle.

The steelman for Altman, taken at its strongest, is that he is the only one of the three canonical authors to have addressed the distributional question at all in advance of the cycle that would make it acute, that the proposal is offered in good faith as a sketch rather than as a legislative draft, and that the absence of operational follow-through reflects the limited capacity of any individual technologist to drive a constitutional-level tax reform rather than any cynicism on the proposer’s part. This steelman is also real. Altman did write the essay before the moment when it would be most rhetorically convenient. He has, in subsequent interviews, returned to the proposal without disavowing it. He has personally signed onto pledges of charitable disposition of his own equity that, if executed in full, would represent a non-trivial individual contribution to the redistributive logic the essay sketches.

The steelman does not, however, address the structural question. The function of the essay in the discourse is not contingent on the author’s sincerity. The function is determined by what the essay does inside the rhetorical economy that surrounds the AI capital cycle. The essay’s function is to occupy the discursive space in which the distributional critique would otherwise unfold, with a proposal whose operational status is permanently aspirational. The aspirational status is what the function requires. A proposal that was implemented would not perform the function; it would have become the policy. A proposal that was abandoned would not perform the function; it would have ceased to exist. The proposal that remains forever just on the horizon, gestured toward in interviews and op-eds and pitch sessions, defended in principle and never advanced in practice, is the proposal whose discursive function is precisely the function the essay performs. The essay is, in this sense, structurally complete.

V. The Flanking Structure: American Dynamism

The three primary texts do not, by themselves, constitute an apparatus. They constitute a canon. The conversion of a canon into an apparatus requires institutional structure: organizations that deploy the canon’s vocabulary as the operating language of their investment processes, their hiring rubrics, their press relations, their legislative engagement, their portfolio-company communications. The principal such institution in the AI capital cycle is a16z’s American Dynamism practice, which was established as a named vertical in 2022 and has been built out as the firm’s primary vehicle for investment in companies whose product surface intersects with national-security, defense-industrial, and civic-infrastructure applications of frontier technology.

The practice’s public materials are themselves canon-adjacent texts. The launching essay, written by partner Katherine Boyle, situates the firm’s investment thesis inside an argument about American national renewal. Subsequent essays by Christian Keil, David Ulevitch, and Boyle elaborate the argument across specific subdomains: aerospace, autonomous systems, biotech, advanced manufacturing, energy infrastructure, public-safety technology. The vocabulary the essays deploy is consistent. America has repeatedly secured geopolitical dominance, the argument runs, by mobilizing capital and talent behind transformative technologies. Artificial intelligence is the next opportunity. Pentagon procurement has — Ulevitch and Boyle’s phrase — devolved into Soviet-style central planning, the argument continues, and the venture-backed startup is the institutional form that can rebuild the defense industrial base on competitive lines. Each subdomain receives its own application of the argument; each application installs the same vocabulary in a new sector of the investable universe.

What American Dynamism does, as a practice, is convert a venture vertical into a patriotic franchise. The conversion is not metaphorical. It is mechanical. A company that adopts the language of American Dynamism in its founding documents, in its hiring pages, in its press strategy, in its government-relations engagement, is purchasing — for the cost of adopting the language — access to a particular kind of allocator, a particular kind of journalistic coverage, a particular kind of political access, and a particular kind of regulatory disposition. The language has commercial value because it functions as a credential. The credential, in turn, functions inside the discourse the canonical essays have constructed: the speaker who has adopted the language of national renewal cannot be readily questioned, inside that discourse, about the distributional consequences of the cycle the language is sustaining. To question the speaker is to be cast as the European bureaucrat, the hysterical journalist, the decel.

The Spitale audit of this institutional structure reads cleanly. The in-group / out-group construction is operational: there are firms inside American Dynamism and firms outside it. The black-and-white narrative is operational: rebuild the industrial base or lose to China. The cognitive bias the structure exploits is the patriotic-affiliation heuristic by which questions about a firm’s commercial structure are deflected onto questions about national interest. The negative emotion is the fear of falling behind in a strategic competition whose terms the structure itself has helped to install. The economic benefit of the power class is the cap-table position of the firm in every American Dynamism portfolio company, marked up through the patriotic credentialing the practice provides. Each trademark is present. The practice is functioning as a propaganda apparatus in the strict Spitale sense, which is consistent with its also being a sincere expression of its partners’ political and strategic convictions. The two characterizations do not exclude each other. They describe the same institution from different vantages.

The most useful artifact to read alongside the American Dynamism essays is the OpenAI shareholder memo on Anthropic, leaked to CNBC in April 2026, in which the company characterized its principal rival as “compute constrained” in the context of explaining the case for the next round’s upsize. The phrase “compute constrained” is, on its face, a technical descriptor. Inside the canon’s vocabulary, it is a positional weapon. To be compute constrained is to be in the wrong half of a binary the canon has installed: the half whose horizon is shorter, whose product trajectory is more limited, whose civilizational stakes are correspondingly diminished. The memo is not deploying the phrase as a description. It is deploying the phrase as an instrument. The instrument’s availability depends on the canon’s having installed the vocabulary in advance. The memo would not function as a defensive instrument if its audience had not already learned to read the canon’s grammar.

The parallel and adjacent move is Anthropic’s amendment of its Responsible Scaling Policy in March 2026, reported by Time under the unambiguous headline “Anthropic Drops Flagship Safety Pledge.” The Responsible Scaling Policy was originally framed, at the firm’s founding and across the first eighteen months of its public presence, as a normative commitment of the kind the AI-safety community had been calling for since the Bletchley Park summit. It specified threshold capabilities at which additional safety measures would apply; it committed the firm to specific evaluation procedures; it positioned the firm as the responsible variant of the frontier-lab category. The March 2026 amendment relaxed the most consequential of the original commitments. The firm’s public justification was that the technical understanding underlying the original thresholds had matured, and that the amended thresholds reflected the more nuanced understanding. The functional effect was the removal of a commitment whose continued operation would have constrained the firm’s commercial cadence in a competitive cycle.

Jürgen Habermas’s distinction between communicative action and the colonization of the lifeworld by system imperatives is the cleanest available frame for what has happened to the Responsible Scaling Policy. The policy began life as a piece of communicative action: an assertion, made in good faith inside a community of AI-safety researchers, about how a frontier laboratory ought to constrain its own behavior. It has, over the subsequent three years, been progressively colonized by the system imperatives of commercial competition, regulatory positioning, talent recruitment, and brand differentiation. Each of those system imperatives has its own logic; each has, on a different timeline, made claims on the policy’s content. The policy in its current form is the residue of those claims. It is still presented in the communicative register; it functions, increasingly, in the strategic register. The amendment was the moment at which the colonization became visible in print.

None of this requires the original framers of the Responsible Scaling Policy to have been insincere. Habermas’s point is that the colonization of the lifeworld is precisely the process by which sincere normative commitments become repurposed by system imperatives without anyone making a decision to repurpose them. The repurposing happens through the ordinary operation of competition, hiring, fundraising, and press cycles. The repurposing is invisible to the participants because each step inside it is locally defensible. The aggregate is what the framework names. The aggregate, in Anthropic’s case, is a normative commitment that began as constraint and ended as moat, with the amendment in March marking the moment at which the moat moved and the commitment was edited to follow it.

VI. The Dog-Whistle Glossary

The canonical texts and their flanking institutional structure have, over the past five years, installed a working vocabulary inside the discourse of AI capital. The vocabulary is what allocators, journalists, and policymakers use when they talk to one another about the cycle. The vocabulary’s words have surface meanings that are technically accurate and that any competent reader can defend if challenged. The vocabulary’s words also have insider meanings that the canon has made available to the participants who have absorbed it, and the insider meanings are the meanings the words are actually doing. The gap between the surface meaning and the insider meaning is what the term dog whistle names. The table below lays out the principal terms of the AI-capital dog-whistle glossary, with the surface meaning, the insider meaning, and a citation to a documented deployment of the term in its insider sense.

Several features of the glossary are worth pausing on. The first is that all nine terms are, on the surface meaning, accurate. A reader who challenged any one of them on its surface meaning would be told, correctly, that the term means what the surface column says it means. The insider meaning is not asserted on the face of the term. It is the function the term performs inside the discourse. The function is delivered through the word’s placement, its associations, its co-deployment with the other terms in the glossary, and the canonical texts that have established the rhetorical environment inside which the function operates. This is the technical structure of a dog whistle. It is what makes the term’s deployment plausibly deniable as analysis while operating, in fact, as positioning.

The second feature is that the glossary’s terms travel. They appear in earnings-call transcripts; in sell-side reports; in pension-board investment memos; in legislative staff briefings; in central-bank speech footnotes; in trade-press articles that re-circulate the language back into the discourse with the additional credential of having been deployed in a putatively neutral medium. By 2026, a competent reader of the financial press cannot avoid the glossary. The terms are how the cycle is being described in the documents the cycle’s participants exchange with one another. The reader who has not learned to recognize the glossary as a glossary — who reads the terms as ordinary descriptions of an ordinary industrial process — is operating inside the discourse the canon has constructed, without knowing they are inside it.

The third feature is the one that closes the loop with Spitale’s framework. Each term in the glossary performs at least one, and typically several, of the five trademarks. Compute moat performs the in-group construction (firms inside the moat versus firms outside it) and the economic-benefit-of-the-power-class function. Scaling laws performs the cognitive-bias substitution by which a descriptive regularity is treated as a prescriptive guarantee. Responsible scaling performs the in-group construction (firms inside the safety frame versus firms outside it) and supplies the moral-permission predicate for the cycle’s commercial cadence. AGI on schedule performs the negative-emotion function through the missed-train fear it installs in allocators and policymakers. National champion performs the black-and-white narrative (democratic AI versus authoritarian AI) and the in-group construction (American firms versus PRC firms). Country of geniuses performs all five at once, which is why it has propagated farther than any of the others. RPO backlog performs the cognitive-bias substitution by which a circular receivable is treated as evidence of commercial validation. Strategic infrastructure performs the economic-benefit function by installing, in advance, the moral predicate for public absorption of private cycle losses. Vendor financing is, by contrast, the critic term: the absence of its routine deployment inside the discourse is the discourse’s structural achievement. The discourse has been organized so that the honest accounting label does not get used; the canon’s vocabulary is what gets used in its place.

VII. The Five Trademarks, Instantiated

The Spitale framework’s five trademarks are most usefully read not as a checklist but as a set of simultaneous operations that a propaganda regime performs when it is functioning as a regime. In the AI capital cycle’s case, each trademark can be located in specific operational artifacts produced over the past eighteen months. The artifacts are the receipts. They are documented, they are dated, and they are mutually consistent in ways that the canon’s vocabulary has not yet absorbed into a unified description.

The black-and-white narrative is installed at multiple layers. At the geopolitical layer, the binary is build or lose to China: the United States either commits to the AI capex cycle on the cadence the cycle’s leaders specify or surrenders the strategic competition to the People’s Republic. Article 5 of this series documents what happened in Beijing when Washington tested the binary against the actual structure of the bilateral relationship and discovered the binary did not survive contact. At the civilizational layer, the binary is country of geniuses or stagnation: the AI cycle either delivers the upside Amodei sketches or the medical-research, mental-health, and global-development gains it promises are foregone. At the firm-competitive layer, the binary is compute-abundant or compute-constrained: firms inside the compute-abundant group are the firms that will determine the cycle’s outcome, and firms outside it are firms in slow decline. The three binaries operate in concert. Each one prevents the reader from occupying the middle territory that the actual evidence would support.

The cognitive bias the regime exploits is most visibly the substitution heuristic at work in the Oracle case. Oracle’s second-quarter fiscal-2026 disclosure reported a remaining performance obligation of approximately $129 billion, of which a substantial fraction is attributable to the OpenAI contract announced earlier in the year. Bank of America’s post-disclosure note framed the circularity of the flow — Oracle building capacity for OpenAI, OpenAI committing to pay Oracle from revenue Oracle is in turn helping to underwrite — as “the key risk going into 2026.” The candor of the framing was unusual; the structural feature it named was not. The RPO number was treated, across the bulk of the sell-side coverage, as a substitute for credit analysis on a counterparty whose actual revenue profile is not yet sufficient to support obligations of the relevant scale. The substitution is the bias. The RPO number is the heuristic. The heuristic functions because the canon’s vocabulary has installed, in advance, the proposition that the relevant revenue will arrive on the relevant timeline. Without the canon’s vocabulary, the RPO number would have to be evaluated as the contingent receivable it is. With the canon’s vocabulary, the RPO number functions as evidence of demand validation. The vocabulary is the predicate that converts the heuristic into apparent analysis.

The negative emotion the regime deploys is, at the surface, fear of missing the cycle. The Amodei-Davos formulation of a 1-to-3-year AGI timeline is the cleanest single artifact. The forecast is offered with calibrated uncertainty; the function of the forecast inside the discourse is to install, in allocators and policymakers, the felt apprehension of a train pulling out of a station they have not yet boarded. The apprehension is asymmetric. The allocator who boards the train and is wrong on the timeline absorbs an opportunity cost. The allocator who refuses to board the train and is wrong on the timeline absorbs a career-ending mark of having missed the cycle whose name is being canonized in real time. The asymmetry of the consequences is what the forecast’s rhetorical function exploits. The forecast may or may not be calibrated. The function is independent of the calibration.

The in-group / out-group construction operates at the most consequential layer in the governance terms of the public-market vehicles the cycle is now producing. The SpaceX S-1 codifies a two-class voting structure in which Class B shares carry ten votes apiece, 85.1 percent of voting control rests in 12.3 percent of Class A and 93.6 percent of Class B, and the public shareholders — who will hold the residual economic interest through the index-fund pipe — are voting decoration. The two-class structure is not new. What is new is the moral predicate that has been installed in advance by the canon: founder discipline is patient capital’s necessary form, dilution of voting control is the failure mode of the firms that have not produced the great compounders of the prior cycle, and the public shareholders are getting exposure to the upside in exchange for accepting governance terms whose substantive challenge would mark them as outside the in-group of patient allocators. The in-group is the cohort that accepts the terms. The out-group is the cohort that questions them. The May 6 letter from the public comptrollers was, in this construction, an instance of the out-group asserting itself in print. The architecture proceeded as if the letter had not been received, which is what an apparatus does when an out-group attempts to invoke a norm the apparatus has already routed around.

The economic benefit of the power class is what the cycle’s artifacts collectively deliver. The SpaceX use-of-proceeds disclosure puts 78 percent of the public-market raise into the repayment of related-party obligations. The Anthropic Series H closed at a $965 billion post-money valuation with mutual-fund participation that delivered paper liquidity to Sequoia, Altimeter, Greenoaks, Samsung, SK Hynix, and Micron. The Stargate UAE consortium captures, through MGX’s structural seat at the table, a marginal-allocator role that prices every subsequent round it touches. The aggregate is what the canon has been organized to make readable as something other than what it is. The aggregate is the transfer. The canon is the language that prevents the transfer from being named.

VIII. What Varoufakis Would Call It

Yanis Varoufakis, in Technofeudalism: What Killed Capitalism, has been working through a thesis that the post-2008 macroeconomic regime, combined with the platform-economic features of the dominant technology companies, has produced an economic formation that is no longer capitalism in any classical sense. Capital, in his account, no longer organizes the accumulation cycle through the production of goods and services for sale; capital organizes the accumulation cycle through the extraction of rent from the infrastructure on which production and exchange increasingly depend. The platform is the lord; the user is the serf; the goods produced on the platform are the produce of the serf’s labor, of which the lord takes a recurring share by virtue of owning the platform rather than by virtue of having produced any specific good.

The thesis has been received with some skepticism in mainstream political-economy circles, and the skepticism is partly warranted. Varoufakis is a polemical writer; Technofeudalism is a polemical book; the analytical claims it makes are sharper than the empirical data the book itself marshals can fully support. But the framework’s deployment against the AI capital cycle’s structure produces a description that is harder to dismiss. The frontier-model laboratory does not produce goods for sale to consumers in any classical sense. It produces a substrate — the model, the inference endpoint, the integrated stack — on which other firms, governments, and individuals increasingly depend for cognitive tasks the substrate displaces. The dependence is recurring; the access is metered; the price is set by the substrate’s owner; the relationship is closer to the structure of a feudal tenancy than to the structure of an industrial sale.

What the canon’s moral-permission language does, in the Varoufakis frame, is disguise the rent-extraction architecture as a civilizational project. The country of geniuses in a datacenter is not, on this reading, an unbounded source of public good; it is an unbounded source of metered access to a substrate whose owner captures the rent. The infinitely-patient AI tutor is not a public-education solution; it is a recurring subscription to a substrate whose owner captures the rent. The cured cancer is not a public-health victory; it is a recurring payment to a substrate whose owner captures the rent on the drug discovery, the diagnostic platform, the clinical-decision-support service, the patient-monitoring infrastructure. In each case, the rhetorical surface is the public good. The structural reality is the rent. The canon’s function is to keep the surface in view and the structure out of view.

Shoshana Zuboff’s The Age of Surveillance Capitalism extends the same diagnostic into the data-extraction layer that underwrites the substrate. The training data on which the frontier models are built was extracted without compensation, often without consent, from the corpus of human cultural production accessible through public-facing internet infrastructure. The extraction was the first phase. The inference-as-rent regime is the second phase. The two phases are sequential in time and continuous in logic. The first phase appropriated the substrate; the second phase meters access to the appropriated substrate. The canon’s language has installed, between the two phases, a moral predicate that names the second phase as innovation while erasing the first phase from the discourse entirely. The training-data question is, by 2026, off the table inside the canon’s vocabulary. It returns only inside the discourse of the lawsuits the canon has been organized to defeat or settle.

I want to flag that the Varoufakis and Zuboff diagnoses are themselves polemical, that they assert structural claims whose empirical confirmation will take longer than this article’s horizon to resolve, and that neither writer is operating inside the empirical-monetary tradition that the rest of this series has been drawing on. They are, in the framework Spitale’s book itself acknowledges, contesting the canon’s vocabulary with an alternative vocabulary whose function is also positional rather than purely descriptive. The point of invoking them is not that their accounts are unassailable. It is that they constitute an available alternative grammar inside which the same facts the canon’s vocabulary describes can be read into a different structural account. The reader who has been operating exclusively inside the canon’s vocabulary may not previously have encountered the alternative grammar. Encountering it is the methodological move the Spitale framework asks of the reader: not to substitute the alternative grammar for the canonical one as a new orthodoxy, but to see that more than one grammar exists, and that the choice of grammar is itself a position with consequences for what the evidence is taken to mean.

IX. The Steelman, in Full

The steelman for the canon, taken in its strongest available form, is that the texts under examination are sincere, the authors are credentialed, the forecasts may be substantively correct, and the institutional structures the texts have installed are themselves a form of public service in a domain where public deliberation has historically lagged commercial development. I want to spend a section on this steelman, because the rest of the article’s argument has been hard on the canon, and a critique that proceeds without engaging the strongest available defense is not honest.

Dario Amodei has put his own money where his words are. The 80%-of-net-worth charitable pledge is operational; the giving has begun; the recipient organizations are documented. He is not a person who has assembled wealth in the AI cycle and removed it from the redistributive economy in the manner of the cycle’s less philanthropic participants. He is a working scientist with substantive contributions to the technical literature whose claims about the upside of frontier systems are made in the same voice in which he discusses the technical details of the systems themselves. The treatment of his essay as only a piece of investor-class propaganda underweights the part of the document that is, on its own terms, a serious attempt to take seriously what powerful AI could mean for the domains he addresses. A reader who concludes that Amodei is acting in bad faith is reading the same text I have been reading and arriving at a less defensible conclusion than the evidence supports.

Marc Andreessen built things. Mosaic and Netscape were the infrastructural commitments on which the consumer internet was built; the productivity and consumer-surplus gains that followed were not marketing exaggeration but documented economic outcomes. His subsequent career as a venture investor and intellectual interlocutor has produced a body of writing that is, even at its most polemical, in conversation with a recognizable canon of liberal-democratic thinkers about technology, governance, and political economy. He is not the populist demagogue the most uncharitable reading of Why AI Will Save the World would suggest. He is, more accurately, a person who has spent four decades inside the construction of technologies that have mattered, and who has come to believe, with the conviction of someone whose career has been substantively vindicated, that further technological construction is the correct vector for further public benefit. The conviction may be wrong. It is not insincere.

Sam Altman built actually useful products. ChatGPT, in its various releases, has been adopted by hundreds of millions of users; the productivity literature has begun to document non-trivial gains in specific task categories; the consumer-surplus calculation is plausibly positive for the population that has access to the product. The corporate vehicle through which the product was developed is structurally complicated; the governance has been contested; the relationships between the for-profit subsidiary, the nonprofit parent, the investor base, and the underlying technical staff have been the subject of legitimate and continuing scrutiny. None of that scrutiny undermines the claim that the products themselves have delivered measurable utility to large numbers of users. The American Equity Fund proposal may be operationally aspirational, but the fact of its proposal — in advance of the cycle that would make its operationalization most necessary — is more than the bulk of the cycle’s other beneficiaries have offered. Altman has, at minimum, articulated a redistributive framework. The framework has not been advanced. It also has not been retracted. The continued presence of the proposal in his public positioning is, on the steelman, evidence of continued openness to the redistributive conversation rather than evidence of strategic occupation of the discursive space.

The American Dynamism practice is, in its strongest reading, a sincere attempt to redirect venture capital toward sectors where American industrial capacity has measurably eroded and where private investment can supply, in conjunction with public procurement, the rebuilding that public investment alone has been unable to fund. The Pentagon-procurement-as-Soviet-central-planning critique is, on its merits, partially correct; the defense industrial base has consolidated, slowed, and become dependent on a small number of prime contractors whose innovation cadence has been documented to lag both the commercial sector and the strategic competitor. The argument that venture-backed startups can supply a corrective is not a self-serving fabrication; it is a thesis that, if vindicated, would constitute a real public good. The thesis may be vindicated. The fact that its vindication would also mark up the cap-table positions of the firms that have organized themselves around it does not invalidate it; in a properly functioning capital cycle, the alignment of public benefit and private return is what the cycle is supposed to produce.

The steelman, in full, is that the canon’s authors believe what they write, that the institutions the canon has organized may produce the benefits the canon promises, that the architecture this series has documented may turn out to be the necessary financial form of a civilizational project of legitimate scale, and that the critique mounted in this article risks treating sincerity as conspiracy and risks misattributing to the canon’s authors a strategic intent that the evidence does not support. I have written this steelman as straight as I can. It is not a parody. It is what a serious and good-faith participant in the AI capital cycle would say in defense of the canon, and it would be defended in good faith by people who have thought hard about it.

The critique does not require the steelman to be wrong. The critique requires only that the steelman, even taken in its strongest form, does not address the structural question the article has been asking. The structural question is not whether the canon’s authors are sincere; it is what the canon does, regardless of the authors’ sincerity, inside the discourse that surrounds the capital cycle the canon underwrites. The canon’s function is what permits the architecture to function. The architecture transfers concentrated risk from insiders to forced absorbers on a compressed timeline through governance terms that disable post-listing recourse. The transfer is the structural feature. The canon’s sincerity is the rhetorical form. The two are compatible. The fact that the canon’s authors largely believe what they write is what makes the canon’s propaganda function effective. A canon authored by cynics would be detectable as cynical; a canon authored by sincere believers carries the residual authority of their sincerity into the discursive function the canon performs. Sincerity is the canon’s most load-bearing feature. The steelman, properly stated, is the canon’s strongest argument and its principal vulnerability at the same time. The two are the same fact.

X. The Honest Reckoning

The five-admissions framing that this series’ thesis identifies — monetary, energy, capital-structure, strategic, and the propaganda layer this article addresses — is the framing that makes each of the four substantive admissions individually visible and collectively invisible. The other articles in this series document the individual admissions: the Tahnoon spine and the energy/sovereign-wealth reallocation in Article 1; the Warsh rupture and the duration repricing in Article 2; the SpaceX architecture and the forced-absorber mechanism in Article 3; the Beijing summit and the structural admission of agenda-setting power in Article 5. Each of those articles documents a subsystem producing a documented receipt in its own code. Each receipt is, on its own terms, available to a competent reader who has access to the public record.

The reason the receipts do not collectively register is not that any one of them is hidden. It is that the language used to surround each of them is a vocabulary engineered to keep the others out of view. The monetary admission, named in fixed-income vocabulary, does not travel into the discourse of AI capital, where the relevant vocabulary is scaling laws and compute moats and the country of geniuses in a datacenter. The energy admission, named in petroleum and sovereign-wealth vocabulary, does not travel into the discourse of monetary policy, where the relevant vocabulary is QT-for-Cuts and the Treasury accord and the term premium. The capital-structure admission, named in securities-law and pension-fiduciary vocabulary, does not travel into the discourse of strategic competition, where the relevant vocabulary is national champion and sovereign compute and constructive strategic stability. Each discourse operates inside its own code; each code is internally coherent; each code is mutually incommensurable with the others. This is the Luhmannian observation Article 3 elaborated in detail. What Article 4 adds is the supplementary observation that the incommensurability is not accidental. It is sustained, across the discourses, by the canon’s vocabulary, which has installed in each of them the moral predicate that prevents the receipts in the other discourses from being legible as receipts of the same structural fact.

The propaganda layer is the connective tissue that makes the other four admissions individually visible and collectively invisible. It is the layer this article names because the prior articles, written inside the individual discourses, were not in a position to name it. The naming is the methodological move that converts the four into a unified description. The fifth admission is the admission that the discourse surrounding the cycle is itself a structural element of the cycle, and that the analysis that proceeds without naming it is participating in the cycle whether or not the analyst recognizes the participation.

XI. The Companion

I want to close this section with a methodological note about the relationship between the journalism and the curriculum, because the relationship is more consequential than the surface description of mutual citation might suggest. The Powell Memo’s long-run institutional success was not the construction of any single think tank or media property or legal vehicle. It was the construction of the apparatus inside which the next generation of operatives could be educated in the canonical vocabulary before they encountered the cycle in which the vocabulary would be deployed. The Olin chairs at the law schools, the Federalist Society chapters at every accredited American law school, the donor-funded fellowships at the policy schools — these institutions trained, over four decades, the population of lawyers, policy entrepreneurs, judges, and political operatives who would subsequently staff the cycle the Powell Memo had anticipated. The apparatus was a pedagogical apparatus before it was an operational one. The pedagogy is what made the operation sustainable across the generational handoff that any long cycle must survive.

The counter-canon that the Spitale framework supplies is similarly pedagogical before it is anything else. The framework’s function is to train a reader in the audit procedure such that the reader can apply the procedure to texts the reader has not yet encountered. The training is the durable contribution. The specific audits the trained reader subsequently performs are the running expression of the training. The platform’s curriculum is structured to perform the training across ten units; the journalism in this article performs one specific audit on one specific canon at one specific moment in one specific cycle. The audit will date. The training, if it works, will not. The next cycle’s canon will deploy a different vocabulary on a different infrastructure on behalf of a different beneficiary class; the audit procedure will remain available to whoever has been trained in it. This is what curriculum is for. It is the slow construction of an analytic capacity that survives the specific cycle in which the capacity was first developed.

XII. What the Next Article Will Trace

The canon this article has audited installed a moral permission structure inside which the AI capital cycle has been able to proceed at a cadence and concentration that, in a discourse organized around any other vocabulary, would have produced more resistance than the cycle has actually encountered. The architecture Article 3 documented is what the canon’s permission structure licenses inside the discourse of capital markets. The allocations Article 1 traced are what the canon’s permission structure licenses inside the discourse of sovereign-wealth deployment. The duration-mismatch repricing Article 2 traced is what the canon’s permission structure had to defend against when the underlying monetary regime began to reset. The strategic admission Article 5 will trace next is what the canon’s permission structure could not prevent when the Beijing summit tested the binary it had installed against the actual structure of the bilateral relationship.

The five admissions are the structural fact. Each subsystem produced its own receipt in its own code. The receipts are mutually consistent. The reason they do not register as mutually consistent inside the discourses that produced them is the canon’s vocabulary, which has been organized over the past five years to make each receipt readable only as an instance of its own subsystem’s operation, and never as a corroborating data point for the other four. The fifth admission — the propaganda admission — is the admission that the discourse is itself a structural element of the cycle. The naming of the fifth admission is what makes the other four legible as instances of the same fact. The naming is what this article has done.

The Beijing summit, in May 2026, was the moment at which the canon’s vocabulary collided with a counterparty that had not been trained in it. Xi Jinping was not, by the time of the summit, susceptible to the proposition that the United States retained agenda-setting structural power across Susan Strange’s four faces; the People’s Republic’s strategic posture had, over the preceding decade, organized itself around the contrary proposition. The canon’s vocabulary does not function across that asymmetry. Inside the American discourse, the canon’s vocabulary continues to function, and the architecture continues to operate inside the permission structure it provides. Across the strategic asymmetry the Beijing summit revealed, the canon’s vocabulary produced no operative concessions, no joint statement, no agenda. The bond market priced the result in nine basis points on the ten-year. The next article traces what that pricing meant, and what the canon was unable to prevent when its vocabulary met a counterparty whose grammar was different.