AI Democracy, Infrastructure Oligarchy

The Cheap AI Era Was a Transitional Story

For a brief moment, frontier AI was sold to the public in the language of consumer software.

Twenty dollars a month.

An assistant for everyone.

Intelligence, flattened into a subscription.

That was always a transitional fiction.

The assistant phase made AI legible to mass society. It gave the public a clean interface, a simple price point, and a comforting democratic story: this power is broadly available; anyone can access it; the new intelligence age will arrive through ordinary consumer access. OpenAI still prices ChatGPT Plus at $20 per month, and Anthropic still offers Claude Pro at $20, even as both companies increasingly differentiate heavier usage, priority, and advanced workflows through stricter limits, paid upgrades, or metered structures.[1]

But once models stop being conversational novelties and start behaving like delegated labor, the old democratic wrapper begins to crack.

Agentic AI is not just “better chat.” It consumes long context windows, repeated tool calls, retries, memory, orchestration, coding workflows, and sustained inference over time. That is why the economic form of AI is changing. What looked like a flat software subscription is being pulled toward a different logic: scarce infrastructure must be allocated, prioritized, and governed. Anthropic’s own platform materials explicitly describe spend limits, rate limits, usage tiers, and service tiers, while consumer pricing pages on both sides increasingly distinguish basic access from higher-throughput and priority forms of use.[2]

That shift matters for more than business-model reasons. It matters because pricing is no longer just pricing.

It is becoming a political technology of access.

What the OpenClaw Fight Actually Revealed

The recent OpenClaw controversy made this visible. In early April 2026, reporting described Anthropic as changing how Claude subscriptions applied to third-party agentic tools such as OpenClaw, effectively pushing those workflows out of the normal flat-rate subscription logic and into separate paid usage pathways. The stated rationale was infrastructural: these workflows imposed outsized strain relative to the usage profile the subscription product had been designed to support.[3]

The significance is larger than the policy itself.

The all-you-can-eat phase starts breaking down precisely where AI becomes most economically powerful. A bounded assistant can be sold as software. A high-intensity agent begins to look more like capacity.

This is where the fairness conversation begins, but it is not where it ends.

The obvious complaint is economic: powerful AI is becoming too expensive for ordinary people. That is true, but incomplete. The deeper issue is constitutional. We are watching the emergence of a world in which formal access remains broad while effective access becomes stratified.

Everyone can still get an assistant.

Not everyone gets reliable, high-throughput, high-agency machine labor.

Everyone may touch the interface.

Only some get operational depth.

That is the real divide now opening up in front of us.

The Stack Is Becoming a Class System

Anthropic’s own pricing structure makes the stratification explicit. Claude still has a mass-market Pro tier, but it also now offers Max plans with materially more usage and priority access at high-traffic times. OpenAI’s own plan structure likewise distinguishes across tiers by model access, limits, and premium capabilities. That is not merely upselling. It is the market form of differentiated sovereignty over compute.[4]

And the gating does not stop at price.

The strongest systems are increasingly not just expensive, but selectively distributed. Anthropic’s Project Glasswing is the clearest example. Rather than generally releasing Mythos Preview as an ordinary public consumer product, Anthropic described it as being extended through launch partners and a limited group of additional organizations involved in critical software infrastructure, backed by substantial usage credits.[5]

That selectivity is not arbitrary. Anthropic presented Glasswing in explicitly defensive terms: securing critical software, supporting trusted partners, and channeling high-capability deployment toward vulnerability discovery rather than broad public release. But even when the rationale is security rather than profit, the structural fact remains the same: a smaller admitted class receives closer proximity to frontier capability than the general public.[5]

This is the point many people still do not fully see.

The democratic promise of AI can survive at the interface level while breaking down at the infrastructure level.

You can preserve the rhetoric of universal access while reserving the most consequential forms of machine leverage for a narrower class: large firms, privileged partners, capitalized users, state-adjacent actors, security institutions, and those admitted into high-trust deployment channels. Consumer democracy and infrastructure oligarchy are not opposites. They are perfectly compatible. In fact, they may be the default architecture of the AI age.

The Private Constitutional Layer

This is why I do not think the right frame is simply “the labs are being greedy.”

That is too moralistic, and too small.

The more accurate frame is that frontier intelligence is colliding with scarcity. Compute is finite. Latency is finite. Inference capacity is finite. Operational trust is finite. And whenever a civilization encounters a scarce strategic resource, allocation ceases to be merely commercial. It becomes political, whether anyone admits it or not.

Rate limits are political.

Priority access is political.

Usage tiers are political.

Partner-only deployment is political.

Safety gating is political.

API terms are political.

Not because they are partisan, but because they determine who can turn intelligence into action.

This argument has predecessors worth naming — and worth distinguishing from. Benjamin Bratton’s The Stack argued that planetary-scale computation had become a governing architecture in its own right, with control over computational infrastructure constituting a form of sovereignty not reducible to territorial authority.[6] That framework applies here. But the present AI moment adds something sharper: not just structural distribution of agency, but active, real-time allocation decisions by specific firms deciding who gets closer to frontier capability before any democratic order has decided such authority is theirs to exercise.

But there is now a second question underneath the pricing question: who should be allowed to make those allocation decisions in the first place?

Should access to frontier systems remain primarily in the hands of private labs and cloud platforms, governed through product policy, service tiers, internal safety judgments, and partnership agreements? Or should stronger public institutions play a more direct role in determining how strategically significant AI gets distributed and governed?

This is no longer a theoretical question. RAND’s 2025 report Governance Approaches to Securing Frontier AI explicitly models three distinct governance regimes for frontier AI: government-enforced security standards, government-led developer authorization for federal use, and an industry-led certification body.[7] California’s Transparency in Frontier Artificial Intelligence Act and New York’s RAISE Act, meanwhile, now impose transparency and incident-reporting obligations on frontier developers that treat these systems less like ordinary software and more like public-interest infrastructure.[8]

The argument is not that private firms should simply disappear from the process. The point is that they are already exercising a quasi-constitutional function. They are not merely building tools. They are determining, through pricing, rate limits, partnership channels, and deployment rules, who gets meaningful machine leverage — and they are doing so before a democratic order has decided how such leverage ought to be distributed.

The Return of the Admitted Class

The Glasswing case sharpens this even further.

Project Glasswing did not present Mythos Preview as a general public release. It presented it as selective access routed through trusted institutional channels tied to defensive cybersecurity and critical software infrastructure.[5]

This is exactly what happens when a technology becomes too strategically significant to be treated as an ordinary retail good. An admitted class begins to form.

The admitted class does not have to look like old aristocracy. It can be composed of labs, cloud providers, enterprise customers, security organizations, critical infrastructure actors, and state-adjacent institutions. But the function is similar: some users receive closer proximity to the execution layer than others. Some are treated as participants in deployment. Others are treated primarily as consumers.

And that stratification has a geopolitical dimension the domestic pricing story alone does not capture. The U.S. government has already experimented with tiered access to AI compute at the international level. The AI Diffusion Framework issued in January 2025 divided global access to advanced chips into tiers, with close allies receiving broader access and other countries facing tighter caps or restrictions. That specific rule was later rescinded, but the underlying logic did not disappear. Tiered compute access had already become an instrument of foreign policy.[9]

The point is not which administration’s policy prevailed. The point is that the same structural logic operating within consumer pricing tiers now also appears between nations.

That is why the old rhetoric of mass AI access is no longer enough. Formal inclusion can coexist with practical exclusion.

Why This Does Not Fully Centralize — Yet

None of this means concentration is total or permanent.

Open-weight models can widen access. Competition can reduce prices. Inference can become cheaper. Tooling can improve. Public intervention could also broaden access or impose new transparency requirements. Any argument that assumes a perfectly closed system from here forward is too simple.

But that does not dissolve the underlying issue.

Even if access expands at the margin, the highest-leverage forms of deployment are still likely to remain bottlenecked by compute, trust, integration rights, safety adjudication, and commercial relationships. The important question is not whether AI becomes more visible or more affordable in general. It is whether the strongest and most operationally significant forms of access remain selectively governed.

That is why the current shift should not be dismissed as a temporary business-model hiccup. The pressures producing stratification are not only commercial. They are structural.

Democracy on the Screen, Oligarchy in the Stack

This is the core political formula of the moment.

At the interface, the system remains democratic enough to sustain legitimacy. Large numbers of people can use AI, experiment with it, and feel included in the transition. The public story remains one of broad access and shared technological uplift.[1]

But deeper in the stack, the pattern looks increasingly oligarchic. Access becomes conditioned by budget, rate ceilings, partnership status, enterprise contracts, infrastructure alignment, and institutional trust. The most consequential forms of machine leverage are not denied to everyone else in absolute terms. They are simply made less reliable, less scalable, and less governable for the majority.[2][4]

This is the form power often takes in advanced systems. It does not always announce itself through explicit exclusion. Often it appears as managed inclusion: everyone may enter, but only some may do much once inside.

That is a more modern and more durable hierarchy.

The Fight Is Over Authorization

So yes, this is partly about fairness. But “fairness” is still too weak a word.

What is emerging is a struggle over authorization.

The real issue is not whether people can use AI at all. It is who is permitted to turn frontier intelligence into durable advantage. In practice, that means asking which actors get reliable access to the strongest systems, which actors can deploy them at organizational scale, and which actors remain confined to consuming their outputs through heavily managed consumer interfaces.

The consumer era trained people to think in terms of apps, subscriptions, and chat windows. The next era will be shaped less by visibility than by permission, throughput, integration, and institutional trust.

The democratic promise survives on the screen.

The oligarchic reality forms in the stack.

And the more powerful these systems become, the harder that split will be to hide.

Notes

[1] OpenAI’s pricing materials list ChatGPT Plus at $20/month, while Anthropic’s pricing page lists Claude Pro at $20/month and distinguishes it from higher tiers with greater usage and priority treatment.

[2] Anthropic’s platform and pricing materials describe differentiated usage through spend limits, rate limits, usage tiers, and service tiers, supporting the claim that access is increasingly stratified by throughput and operating mode rather than just raw availability.

[3] Reporting in early April 2026 described Anthropic as changing how ordinary Claude subscriptions applied to third-party agentic tools such as OpenClaw, pushing those workflows toward separate paid usage on the grounds that they imposed outsized infrastructure strain relative to the intended subscription model.

[4] Anthropic’s pricing structure distinguishes Free, Pro, Max, Team, and Enterprise tiers, with Max offering materially more usage and priority access; OpenAI’s pricing likewise distinguishes across plans by model access, limits, and advanced capabilities.

[5] Anthropic’s Project Glasswing materials described Mythos Preview as a selective deployment through launch partners and additional organizations involved in critical software infrastructure, framed explicitly around defensive cybersecurity and backed by substantial usage credits rather than a general public release.

[6] Benjamin Bratton, The Stack: On Software and Sovereignty, MIT Press, 2015. Bratton’s central claim is that planetary-scale computation constitutes a new governing architecture in which control over computational infrastructure produces forms of sovereignty irreducible to legal or territorial authority. The present essay draws on this framework while shifting the focus toward deliberate, real-time allocation decisions by specific frontier firms.

[7] Ian Mitch et al., Governance Approaches to Securing Frontier AI, RAND Corporation, RR-A4159-1, 2025. The report identifies three distinct compliance regimes for frontier AI governance: government-enforced security standards, government-led developer authorization, and industry-led certification.

[8] California’s Transparency in Frontier Artificial Intelligence Act took effect January 1, 2026, requiring frontier developers to publish safety frameworks, report critical safety incidents, and maintain internal governance processes. New York’s Responsible AI Safety and Education Act was signed into law in March 2026 and takes effect January 1, 2027, with similar transparency and incident-reporting requirements.

[9] The U.S. AI Diffusion Framework, issued in January 2025, divided global access to advanced AI chips into tiers. Although later rescinded, it demonstrated that tiered access to AI compute had already become an instrument of foreign policy.