Privileges Already Secured

Crédit Mobilier, open weights, and Locke’s reserved judgment

In July 2026, after Moonshot released Kimi K3 (a Chinese open-weight model that serious technical observers declined to dismiss as a toy), OpenAI’s Head of Strategic Futures, Dean W. Ball, posted a thread that said the quiet part in public.

The model was strong. Open weights were “inherently decelerationist” because they deter further AI capital expenditure. An open-weight-dominant future risked “full AI communism”: intelligence as state-provided digital public infrastructure, which he called a “dystopian hellscape.” The United States need not ban open source. Agencies could issue soft law that creates fear, uncertainty, and doubt (for example, advisory warnings about possible backdoors) so regulated enterprises flee Chinese weights. Such warnings, in the original formulation, “needn’t be that well justified.”

Walk-backs followed. Soft law was framed as prediction rather than advocacy. The posts were said not to be “what OpenAI thinks.” Take the clarifications seriously. Also notice what did not get clarified. The economic object remained capex. The political instrument remained agency-created uncertainty. The dystopia remained the idea that intelligence might be treated as public infrastructure.

That is not mainly a licensing spat. It is a political-protection episode for a closed, capital-intensive buildout. And it is happening in real time.

The historical rhyme is Crédit Mobilier. The philosophical backdrop is John Locke: political power as a trust, and the people’s reserved right to judge when that trust is invaded.

Power on loan

In the Second Treatise, Locke’s legislative power is supreme inside a constituted government, but only as a fiduciary power. It is held for certain ends: the preservation of the community (life, liberty, and property in Locke’s wide sense). When rulers act contrary to that trust, the power is forfeited. It devolves to the people who gave it.

Locke does not license a riot over every bad statute. Where the law still offers a remedy, force is not the answer. Isolated private injuries do not dissolve a government. What moves the argument is a longer pattern: designs against the people’s liberties and estates, force without authority, a “long train of abuses” that makes the direction visible. When there is no earthly umpire left, he says the appeal is to Heaven. In plainer terms, the people reserve an ultimate judgment they never had the power to alienate.

You do not need to want a revolution to need that frame. The peaceful form of Locke’s reserved judgment is scrutiny before losses are socialized. The dangerous form is waiting until the only options left are bailout, amnesia, or rage. The railroad era chose delay. Description arrived after extraction. Rate regulation arrived after overbuilding. The useful question now is whether we refuse that timing while the franchise is still being secured.

The franchise in four beats

The AI infrastructure boom is usefully read as a publicly co-produced franchise. Four features do the work. None of them requires calling frontier models a fraud, or calling anyone Oakes Ames. Mapping a political economy is not an indictment for bribery.

First, the franchise begins with public purpose. The Pacific Railway Act of 1862 manufactured a continental project with bonds and land, justified by mail, military use, and national binding. Today the equivalents are stacked: CHIPS-era industrial policy, export controls, White House theater around Stargate, and dozens of state data-center tax programs that de-risk private construction and select champions. A real public purpose is not a solvent for private extraction. It is often what makes extraction politically cheap, because every critic can be answered with the flag.

Second, the profit center is the build. Crédit Mobilier billed construction far above cost while the operating road struggled. Today chipmakers, clouds, and data-center vehicles book revenue and backlog. Frontier labs often burn cash. Industry capex has, on available syntheses, run far ahead of AI-service revenue. When a senior figure measures open competition by whether it “deters capex,” the metric of success is the continuation of the build.

Third, the same small set of actors sits on both sides of the major contracts. That was Crédit Mobilier’s legal essence: common control of railroad and contractor. It is the financial essence of the present web. Microsoft invests in OpenAI; OpenAI spends heavily on Azure; Microsoft books cloud revenue, takes a revenue share, and holds equity whose value is itself a function of the loop. Nvidia invests in customers who buy its chips. Specialized intermediaries stand between GPUs and labs, financed with GPU-backed debt and take-or-pay capacity. Circularity is not automatically fraud. The civic question is whether circularity is disclosed as circularity, or sold as independent third-party demand for a scarce national asset.

Fourth, when competition threatens rents, the franchise recruits the state. Ames used discounted stock so Congress would “resist any encroachment upon, or interference with, the rights and privileges already secured.” The modern instruments are licensing advocacy, national-security framing, and soft law. George Stigler’s claim that regulation is often acquired by the industry and operated for its benefit, and Bruce Yandle’s “bootleggers and Baptists,” are the theory of the case. Genuine safety fears and genuine China concerns can travel with private rent protection. That coexistence is the point.

Why open weights scare the build

Open-weight models are to a capitalized closed stack what a parallel railroad was to a land-grant franchise: a threat not primarily to the technology, but to the scarcity assumptions that make a subsidized buildout pay.

Put the conflict in one table of facts, not vibes:

If capable models are free or cheap to deploy across competing infrastructure, Microsoft–OpenAI-style loops no longer look like the inevitable route to AI revenue. GPU-backed debt and hyperscale backlog narratives face utilization and pricing pressure. State tax abatements start to look less like investment in “national leadership” and more like support for a few franchisees seeking federal help against alternatives.

Campaigns against open weights (especially Chinese ones) are thus part of the modern political-protection layer. When the product layer commoditizes faster than the buildout can amortize, the coalition must accept thinner rents and possible stranded assets, or recruit the state to recreate scarcity. The Ball episode is that recruitment spoken aloud.

Do not overclaim. A personal thread is not company policy. It is not an Ames-letter equivalent and supplies no public evidence of cash bribes. The asserted capture is structural and discursive: revolving doors, aligned interests, soft law, and a definition of progress tied to frontier-lab capex. That resembles Gilded Age political economy without reproducing the 1872 fact pattern. Treat the initial thread as revealed elite discourse. Include the walk-backs. Keep the triad that survived them: capex as object, soft-law uncertainty as instrument, public-good framing as dystopia.

Public money, private capture

Railroad land grants were subsidy. The modern equivalent is often quieter: a sales-and-use tax exemption, frequently automatic past an investment threshold, often without a sunset, often without a clear public schedule of who received how much.

Independent compilations in 2025 and 2026 counted dedicated data-center incentive programs across most of the country, with some states losing on the order of a billion dollars a year once AI-scale facilities arrived. Permanent employment at finished sites is measured in dozens or low hundreds against billions in capitalized plant. Who captures? The cloud provider, the chip layer through demand, the construction and financing layer, and sometimes the utility rate base. Who pays? The general fund and other ratepayers when electricity is exempted or grid upgrades are socialized.

Locke’s vocabulary helps. Public support creates a trust relation. Munn v. Illinois later said the American version out loud: property devoted to a use that affects the community at large becomes clothed with a public interest. Railroad sponsors took the land grant in the name of public purpose and spent a generation denying the obligation. Watch for the same split now. National-champion rhetoric when the exemption is written. Private-franchise rhetoric (or “AI communism” panic) when obligation or open competition is mentioned.

If “U.S. leadership” means shielding two or three closed labs from open competitors, public-goods language has become private-franchise support: the analogue of grants ostensibly for “the railroad” that disproportionately enriched construction insiders.

What this is not

This is not a claim that useful AI is imaginary. Utilization can be real while payment paths are still partly reflexive. Use does not settle payment path. Independence of demand is an empirical question, not a mood.

This is not a reduction of safety to bad faith. Security risks can be genuine while commercial interest shapes the preferred remedy (opaque soft law instead of transparent standards). Bootleggers and Baptists can ride together.

This is not a prediction of 1873 on a loop. Correct technological foresight is consistent with catastrophic capital allocation. Frontier GPU clusters age in years, not generations. Open weights speed price discovery for intelligence. If capex assumes closed-lab pricing power, successful open models become a demand-side stress test even without an “AI winter.” That reveals fragility in the assumed demand curve. It does not, by itself, prove collapse.

Scrutiny before collapse

Locke’s reserved judgment, translated for a republic that still has courts, commissions, and filings, is a refusal to treat “privileges already secured” as self-authenticating.

Three asks follow.

Make circular demand say its name. Registrants with material AI-infrastructure exposure should disaggregate AI-attributed revenue, backlog, and capex into related-party, highly concentrated, and other. Equity and revenue shares in counterparties that are also customers should sit in the same frame as the demand story. Open competition raises the stakes: it becomes harder to treat related-party cloud spending as pure organic end-user demand.

Stop writing exemptions blind. A sunset, a cap, published beneficial owners at the parent level, real clawbacks, and an end to socializing hyperscale power costs onto residential ratepayers are the minimum conditions of a trust the public can still bargain with.

Force safety and China rules to survive three questions. Is the asserted harm unique to open weights, or a harm of capable models as such? Is the instrument process-regular (transparent evaluations, critical-infrastructure procurement, classified controls aimed at identified military end users), or soft law chosen because it need not be well justified? Who captures the residual rent if the rule works? If the answer to the third is the same small set already living on circular build finance and state tax capacity, the Baptists have a bootlegger problem, and the record should say so out loud.

Large-load tariffs that assign upgrade costs to the causing customer are already the most important AI regulation in the country. They should be backed. A concentrated new use does not get to treat the residual public as its construction company.

Encroachment worth wanting

Oakes Ames did not say he was corrupting the House. He said he was placing stock where it would do the most good, so men of standing would resist interference with privileges already secured. The sentence lasts because it is more precise than “bribery.” It names what a subsidized franchise always needs after the ribbon-cutting: a politics that treats the arrangement as an accomplished fact and the critic as an encroachment.

In 2026 the costume changed. The function did not. Soft-law FUD places career risk on an enterprise general counsel so the market will protect privileges already secured. Capex becomes the definition of progress. Public infrastructure becomes the dystopia, except when the same stack is living on tax capacity and the grid.

Locke’s answer is older and colder. The community retains a supreme power to judge whether the trust is being kept. In a healthy constitutional order, that judgment shows up as disclosure, rate cases, antitrust facts, and the refusal to launder commercial scarcity through national fear. In an unhealthy one, it shows up too late.

The privileges are being secured now. Encroachment, if it is to mean anything honest, had better mean scrutiny.

Adapted from the student Note “The AI Hypocrisy Paradox” and the July 2026 research memo “Protecting the Franchise.” Figures and case developments move quickly; treat reported deal sizes and state fiscal estimates as order-of-magnitude evidence pending primary pin cites. Ball’s thread is cited as revealed discourse, with walk-backs noted.