When AI Makes Invention Cheap, Who Collects the Toll?

A factory hires an AI company to improve a production line.

At first, it buys model usage. Words go in, words come out, and the bill follows the token meter. If the answer is wrong, the factory absorbs the loss. If the proposal fails on the shop floor, its engineers start again.

Then the AI company changes what it sells.

Instead of sending the answer back, it runs experiments, builds a prototype, and files patents. Once the method works, the factory receives a license: the machines are yours, the building is yours, but the method is ours. Pay us for every unit you produce.

The same AI was a calculator yesterday.

Today it is a tollbooth.

OpenAI has not announced this plan. It still sells API and enterprise access, including services priced by token usage. This essay is about the next move: if a model company can connect intelligence to laboratories, proprietary data, and intellectual-property law, why remain satisfied with selling computation once?

The most valuable thing may not be the answer.

It may be whether anyone can avoid paying for it.

A Token Sells Use. A Patent Sells Dependence.

A token is a unit for metering model use. It is useful in the same way a kilowatt-hour is useful.

But a meter is a weak fortress.

Models multiply. Inference gets cheaper. Customers switch suppliers. What only one company could do yesterday may become ordinary tomorrow. A brilliant answer does not guarantee a returning customer.

A patented process, protected know-how, or a validated technical system behaves differently. If it is difficult to replace, its price does not have to track the cost of discovery. AI may find the method cheaply. The owner can still price the license according to how badly everyone else needs it.

Discovery can become cheaper while permission remains expensive.

That is the load-bearing idea:

Tokens are priced by consumption.

Technology is priced by results.

Exclusivity is priced by dependence.

The first resembles an electricity bill. The last resembles a bridge whose toll begins only after everyone has reorganized life around crossing it.

That is why an AI company may want to become more than a model supplier. Selling tokens gives it a small place inside a customer's research process. Owning or co-owning the outcome gives it a claim on the future value of a medicine, material, or manufacturing process.

But first it must survive the ugly middle.

Between a Beautiful Answer and a Working Invention Lies a Heap of Trouble

An AI can propose a new material in seconds.

Making it may take months. Proving that it is safe, stable, and manufacturable may take much longer.

Between proposal and product sit failed experiments, measurement error, regulation, supply chains, and experienced workers who can look at an elegant design and say, “That will jam the line.” Confidence is not a substitute for validation.

The strongest AI companies of the future may therefore look less like chat apps. They may be part software company, part research institute, part laboratory, and part patent-licensing operation.

Tokens will still be sold at the front desk.

The expensive business will happen in the back: the model proposes, experiments discard the nonsense, engineers make the surviving idea manufacturable, and lawyers build a fence around it.

Once the fence is standing, a few dollars of computation can become years of rent.

A nearby version of this model already exists. In January 2024, Alphabet's Isomorphic Labs announced drug-discovery collaborations with Eli Lilly and Novartis. The disclosed structures included upfront payments, research funding, milestone payments, and potential royalties on future drug sales. Isomorphic said the collaborations had potential value approaching $3 billion, excluding royalties; the upfront payments were $45 million from Lilly and $37.5 million from Novartis.

Those numbers need supervision.

Potential value is not cash received. A research collaboration is not an approved drug. A molecule proposed by a model is not medicine because the press release photographed it nicely.

Still, the invoice has changed.

The customers are not merely paying to use AI. Payment is tied to research progress and the possible value of the outcome.

A Model Does Not Exhale a Patent Certificate

Technology stories enjoy deleting the middle.

An AI proposes something, and the headline calls it an invention. A company files an application, and the next headline says AI has obtained a patent. It sounds as though the machine thought overnight and the patent office arrived at breakfast with a brass band.

That is not how it works.

Patents generally require novelty, an inventive step or non-obviousness, usefulness or industrial applicability, and sufficient disclosure. Filing is not granting. A granted patent is neither permanent nor automatically worldwide.

Inventorship is messy too. Many jurisdictions still require a human inventor. The World Intellectual Property Organization describes an ongoing debate over how much human contribution is necessary in AI-assisted or AI-generated inventions, and who should receive the rights.

Two situations must stay separate.

A customer may use a general model, organize its own research, and validate the result. That does not automatically make the model company owner of the technology. OpenAI's current services agreement says that, between OpenAI and the customer and to the extent allowed by law, the customer owns the output.

Alternatively, an AI company may help define the target, supply specialized models and data, run experiments, and negotiate ownership with a partner. That is no longer the sale of an output. It is a research venture.

Confusing the two would greatly please whoever drafts the contract.

If every useful technology developed with a model belonged to the model provider, customers would not be buying a tool. They would be inviting a silent shareholder into every laboratory. It would skip the night shift and still collect dividends.

Easier Knowledge Can Become Harder to Use

Patents are not cartoon villains. Limited exclusivity can reward risky research and require technical disclosure. Without an expected return, some expensive experiments would never begin.

AI can still enlarge an old problem.

Public education trains researchers. Public money funds science. Society generates data. Engineers, clinicians, and workers accumulate practical knowledge. A model compresses those contributions into a research process, while the companies with the most compute, laboratories, and lawyers are best positioned to claim the result.

Everyone carries bricks to the roadside.

Someone builds a road with them and charges the brick carriers to enter.

This does not make every AI-assisted patent illegitimate. It means that public return, licensing conditions, and access rules should be discussed before exclusive rights are granted, not remembered afterward.

Nor should the phrase “AI invented it” erase the people who made the proposal real.

A model may recommend a new valve. A maintenance worker discovers that it seizes at high temperature. The patent may omit the worker's name. The cost-saving report may omit the worker's wage.

The machine becomes the inventor in the story.

The property, curiously, always belongs to a human corporation.

The Factory Saves Time. Where Does the Time Go?

Suppose an AI-designed process produces the same output with 20 percent fewer labor hours.

The press release says productivity rose 20 percent.

The worker may receive a layoff notice.

Or the gain may become shorter hours, higher wages, safer staffing, and lower prices. Technology creates these possibilities. It does not choose among them.

Institutions and bargaining power do.

If the savings first pay licensing fees, then increase shareholder returns, and only afterward invite workers into the discussion, AI has not improved efficiency for everyone. It has improved leverage for a few.

Whether AI companies continue selling tokens is therefore the small question.

The large question is who may fence off the medicines, materials, and industrial processes that AI helps create. Who may cross? Who names the price? Who contributed knowledge but receives no claim?

Patent offices must test whether the invention is real. Public funders should negotiate public benefit before the research begins. Companies should record the contributions of engineers and frontline workers who validate and improve what the model proposes.

Otherwise invention accelerates while distribution remains stuck in an older machine.

Humanity may be building a system that makes discovery faster.

Using it only to build tollbooths faster would be a bleak waste.

Today we ask what a million tokens cost.

Tomorrow we should ask:

**If the method has already been discovered, why must the people who use it keep paying rent for generations?**