Inference Is Adderall for Computers

Training builds the mind. Inference is the dose. And the state will eventually regulate the dose.

Adderall does not install a second brain. It does something more subtle and more consequential: it pushes existing hardware toward higher sustained output. The user experiences that as sharper focus, faster retrieval, more throughput.

AI inference plays a similar role for silicon. Training builds the model. Inference is what turns a static weight file into live cognition. A frontier model sitting inert in a datacenter is latent capability. Inference is the dose that administers it.

That is why people keep misreading the politics of AI. They argue about models as if the central question were whether they are open or closed, biased or unbiased, helpful or harmful. Those fights matter. But they sit one layer above the deeper reality: inference is becoming a mass-distributed cognitive enhancer. And governments do not leave mass-distributed cognitive enhancers entirely ungoverned.

The Supply Chain Already Looks Pharmaceutical

We already know what regulation looks like when a product enhances cognition at scale. The state does not merely police the final retail experience. It governs the stack.

At the precursor layer sit the fabs, chip designers, and hyperscalers. This is the chemical base of the system: the semiconductors, the memory, the interconnect, the datacenter footprint. Advanced-chip export controls already treat leading-edge AI compute less like an ordinary commodity and more like a strategic precursor, while the CHIPS and Science Act is an explicit attempt to harden domestic semiconductor capacity and supply chains.

At the formulation layer sit the labs. OpenAI, Anthropic, Google, and their peers do not merely "host models." They take raw compute and turn it into products with distinct behavioral profiles. They choose the effective dosage: latency, context, autonomy, tool access, refusals, memory, monitoring, pricing. Model cards, evals, and responsible scaling policies are the rough equivalent of early safety dossiers.

At the distribution layer sits the app ecosystem. The SaaS wrapper, the enterprise integration, the coding copilot, the customer-service stack, the browser assistant -- this is retail. APIs, seat caps, audit logs, age gates, credits, enterprise permissions, rate limits: software is already borrowing the language of controlled distribution, even if it does not yet admit that is what it is doing.

Once you see the full chain, the regulatory endgame stops looking exotic. It starts looking familiar.

Dependency Arrives Before Regulation

Nobody starts taking a cognitive enhancer expecting it to become part of the floor. They take it to get through a crunch, a deadline, an exam, a season of overload. The dependency curve shows up later, after a new performance baseline has already colonized everyday life.

That is exactly how inference spreads.

First it is optional: code review, drafting, research, triage, summarization, support. Then it becomes embedded. Teams redesign workflows around instant cognition. Managers start budgeting around AI-assisted output. SLAs, staffing models, and margins quietly assume the model is always there.

At that point, inference is no longer a feature. It is part of the operating environment.

A human can stop taking Adderall and eventually restabilize. A company that has rebuilt itself around 100,000 API calls a day cannot simply "take a break." Its withdrawal symptom is not discomfort. It is operational failure.

That is the political hinge. Regulation gets harder, not easier, once the enhancer becomes infrastructure.

The Real Risk Is Capability Without Friction

The pharma analogy gets serious at the frontier.

We do not tightly control stimulants because focus itself is evil. We control them because powerful cognitive enhancement, distributed without maturity, accountability, or friction, destabilizes institutions faster than institutions can adapt.

A college kid with Adderall can fake competence for an exam. Annoying, but bounded.

A teenager with effectively unrestricted frontier inference can run industrial-scale persuasion, synthesize operational knowledge across domains, automate phishing and social engineering, produce misinformation with the texture of expertise, and scale forms of exploitation they do not fully understand. The problem is not merely intelligence. It is intelligence multiplied by speed, replication, anonymity, and reach.

That is why point regulation will fail. Regulate chips but not APIs, and the risk migrates to distribution. Regulate apps but not model behavior, and unsafe capacity leaks through another channel. Regulate outputs but not access economics, and you subsidize low-value misuse at scale.

The whole stack matters because the whole stack is the delivery mechanism.

Open Weights and the Home-Lab Problem

Then there is open weights, which is where the analogy becomes genuinely uncomfortable.

The case for sovereign compute is real. No serious society should want all of its cognitive infrastructure routed through a handful of API gateways. Local inference means resilience. It means privacy. It means autonomy.

But autonomy has a shadow. The argument "I should be allowed to run my own model" is structurally similar to the argument "I should be allowed to make my own compound." It is not absurd. In many domains it is correct. But it also collapses the distance between legitimate self-provision and uncontrolled capability diffusion.

The reason this is not yet a full-blown panic is physics. Frontier capability is still bottlenecked by capital, power, cooling, and chips. You can run strong models on consumer hardware. You cannot casually reproduce a frontier hyperscale cluster in a garage.

So the likely equilibrium is not a blanket ban. It is tiering. Tolerance at the edge. Control at the frontier. Home labs survive. The hard stuff stays behind licensing, monitoring, and controlled access.

The Misdiagnosis Has Already Started

You can already watch politics reaching for the issue from the wrong end.

The backlash to AI data centers has become visibly bipartisan, driven by worries over electricity costs, water use, land, secrecy, and local powerlessness. In February 2026, Bernie Sanders publicly backed a national moratorium on data-center construction. That same month, Denver leaders moved to seek a one-year pause on new data centers. In December 2025, Ron DeSantis proposed empowering local governments to reject hyperscale facilities.

That matters. It shows the state can already feel the pressure.

But the policy instinct is still mostly infrastructural. Moratoriums. Zoning fights. Utility arguments. Water disputes. Construction pauses.

Those are real tools for a real problem. They are also incomplete.

They treat the datacenter as the object of regulation because the building is visible and the token is not. But that is like responding to a drug crisis by focusing only on warehouse permits. The warehouse matters. The substance moving through it matters more.

Inference is not literally a Schedule II drug under the Controlled Substances Act. But the controlled-substance analogy fits its future governance better than the utility analogy does. The state will not stop at pipes and buildings. It will end up caring about who can produce frontier inference, who can distribute it, what safeguards attach to it, what audit trail it leaves, and which uses require tighter access control.

Zoning is the opening skirmish. It is not the final settlement.

Token Monetary Policy

This is where the argument stops being metaphorical.

The Fed exists because cheap money distorts allocation. When credit is underpriced, capital floods into low-discipline uses and the real economy pays later.

Cheap inference threatens a similar distortion. Not because tokens are money, but because tokens are metered claims on scarce physical systems: electricity, transformers, cooling, water, interconnection, land, and time on advanced chips. That is not a poetic statement. It is an industrial one.

You can already see the infrastructure pushback forming. Utilities are signing demand-response deals to curb data-center consumption during peak stress. Grid operators are warning that AI demand is outpacing new supply in key regions. Microsoft has publicly promised measures intended to keep data-center power and water costs from spilling onto surrounding communities.

When the price of tokens falls faster than the price of the physical bottlenecks beneath them, the economy gets a false signal. It starts treating digital cognition as if it were nearly free even when the substrate is not. The result is predictable: more synthetic output, more automation theater, more marginal use cases that make sense privately but not socially, and more pressure on the commons that makes all of it possible.

Call the policy response a Federal Token Rate if you want the memorable phrase. Underneath, it would really be compute scarcity pricing: a dynamic cost on large-scale inference that rises when grids, water systems, or chip supply are tight and falls when spare capacity exists.

That is not some alien policy category. It is the same logic behind peak electricity pricing, congestion pricing, and licensed access to scarce bandwidth. Once a resource becomes economically central and physically constrained, laissez-faire stops being a principle and starts being a subsidy.

The Endgame

This is the part much of the AI world still refuses to see.

Inference is not just software. It is industrial throughput disguised as text.

That is why the eventual regulatory destination will look less like ordinary SaaS oversight and more like controlled access to a powerful enhancer. The trigger will not be philosophical discomfort with machine intelligence. It will be something more mundane and more powerful: price spikes, water fights, grid stress, regional shortages, and the dawning realization that society allowed a new layer of cognition to scale before deciding how much of it it actually wanted, where it should live, and who should get first claim on it.

Every meaningful resource follows the same arc: discovery, overuse, conflict, governance.

Inference will follow it too.

Only faster.

Because that is what stimulants do.