The Thesis

AI’s biggest scientific opportunity may not be inventing everything from scratch. It may be making humanity’s abandoned research searchable again.

Agentic systems can increasingly scan old papers, failed experiments, shelved drug programs, patents and datasets at a scale humans cannot. That could turn forgotten R&D into a portfolio of scientific options.

The deeper shift is even more important: if AI makes hypotheses abundant, the scarce resource moves downstream. Labs, experiments, clinical access, proprietary data, intellectual-property rights and scientific provenance become more valuable.

In other words, AI may make thinking cheap long before it makes proving cheap.

Artificial intelligence is usually framed as a machine for creating something new.

New code.

New images.

New molecules.

New business models.

New discoveries.

But one of the most interesting economic uses of AI may be almost the opposite.

It may become extraordinarily good at finding value in things we already created and then forgot.

Human civilization has accumulated an enormous scientific archive. Pharmaceutical companies have abandoned drug programs. Universities have published millions of papers. Researchers have generated negative results that rarely receive attention. Startups have collapsed while leaving useful technology behind. Patents have expired. Databases have been retired. Entire research programs have been shelved because funding, management or commercial priorities changed.

Some of those projects deserved to die.

Others may simply have been born too early.

AI agents could make it possible to tell the difference.

Science Has a Memory Problem

Scientists do not suffer from a shortage of information.

They suffer from too much of it.

No researcher can read every paper in a field. No pharmaceutical executive can personally revisit every compound their company abandoned over the previous 30 years. No engineer can compare every failed prototype against every technological advance that occurred after it was cancelled.

This creates what I think of as a coordination tax.

Before anyone can build on old research, somebody has to discover that it exists.

Then they have to understand it.

Find the underlying data.

Recreate the methodology.

Locate the code.

Work out why the original project stopped.

Check whether the patents are still relevant.

Determine whether the science has changed since then.

That process can be expensive enough that nobody bothers.

The research is not necessarily worthless.

It is economically invisible.

AI agents change the cost structure of that search.

When a Scientific Paper Becomes an Agent

One of the most important developments in agentic science is the idea that research papers do not have to remain static documents.

Researchers at Stanford have demonstrated systems that can turn papers into interactive agents capable of explaining methods, accessing underlying resources and performing parts of the analysis described in the research.

More interestingly, agents built from different papers can interact.

That creates a new form of scientific infrastructure.

A paper is no longer only something another scientist reads.

It can become something another machine can query.

Eventually, large research archives may operate less like digital libraries and more like networks of specialized scientific workers.

Imagine thousands of research agents, each representing a paper, dataset, methodology or experimental result.

One understands a particular protein.

Another understands a genomic dataset.

Another knows a clinical-trial result.

Another has access to a failed compound from 2011.

Another understands a manufacturing technique that did not exist when that compound was originally investigated.

The interesting discoveries may come from connecting those pieces.

That is where things get economically significant.

The Scientific Graveyard Is Huge

A 2026 analysis of pharmaceutical development identified more than 5,500 drug programs that had apparently been deprioritized after reaching human clinical testing.

That number should not be misunderstood.

It does not mean there are 5,500 hidden cures sitting in filing cabinets.

Drug development fails for good reasons.

Some compounds are unsafe.

Some simply do not work.

AI cannot negotiate with biology.

But commercial drug development is not a perfect sorting machine either.

Programs also disappear because companies merge, funding runs out, management changes strategy or another drug suddenly becomes the corporate priority.

Sometimes a trial is poorly designed.

Sometimes patient populations are too broad.

Sometimes a useful biomarker has not yet been discovered.

Sometimes the technology required to deliver the drug efficiently does not yet exist.

A therapy rejected in 2010 is being evaluated against the scientific world of 2010.

The world of 2026 may contain better sequencing, better diagnostics, better protein modelling, better delivery mechanisms and vastly more biological data.

The molecule might be unchanged.

The surrounding possibility space is not.

AI Could Turn Dead Projects Into Options

This suggests a different way to value abandoned research.

Instead of classifying projects as either alive or dead, they may be better thought of as options.

Most expire worthless.

A small percentage become interesting again when conditions change.

AI agents can lower the cost of constantly re-evaluating those options.

Imagine a pharmaceutical company with 10,000 archived research programs.

Historically, reviewing all 10,000 might make no economic sense.

But if AI can cheaply perform a first-pass review, reconstruct the scientific history, compare each asset against recent research and eliminate 99% of them, the economics change dramatically.

You do not need AI to discover 10,000 viable drugs.

You need AI to find the 20 programs that deserve another human conversation.

Then perhaps five deserve a new experiment.

Maybe one survives.

That is still potentially enormous value recovered from an asset that previously looked dead.

This is the Dormant IP Harvest.

But Then Something Strange Happens

If AI agents become very good at generating and rediscovering scientific hypotheses, the economics of science begin to invert.

Ideas stop being the scarce resource.

Testing becomes scarce.

Today, researchers spend enormous time generating good hypotheses.

In a mature agentic system, thousands or millions of plausible ideas could be produced continuously.

But the physical world does not run at inference speed.

Cells still need time to grow.

Chemicals still need to react.

Animals and human participants cannot be simulated away in every field.

Microscopes, sequencing machines, clean rooms and laboratory robotics still have finite capacity.

Clinical trials still take real time.

That produces what I call the Scientific Scarcity Inversion.

AI makes thinking cheaper faster than it makes reality cheaper.

Welcome to the Proof Economy

The most valuable scientific companies of the next decade may therefore not necessarily be the companies generating the most ideas.

They may be the companies capable of proving or killing those ideas fastest.

That means laboratory automation becomes more important.

Self-driving laboratories become more important.

High-quality proprietary datasets become more important.

Clinical-trial networks become more important.

Biological samples become more important.

Scientific provenance becomes more important.

And strangely, failed experiments become more valuable.

Why?

Because if an AI system is exploring millions of possibilities, knowing which routes already failed can dramatically reduce the search space.

Humans tend to celebrate successful experiments.

Machines may place enormous value on well-documented failure.

A negative result says:

Do not spend compute here again.

Do not spend laboratory time here again.

Do not repeat this mistake.

That information has economic value.

Why Old Corporate Data May Become an AI Moat

There is another implication that I think markets are underestimating.

Two pharmaceutical companies might each possess 30 years of research history.

Company A has structured data, clean metadata, digitized laboratory records, clear patent ownership and well-documented reasons why projects succeeded or failed.

Company B has thousands of PDFs, spreadsheets on obsolete systems, handwritten laboratory notebooks and incomplete records from employees who retired 15 years ago.

The underlying scientific history might be comparable.

The AI value is not.

Company A owns machine-readable institutional memory.

Company B owns an archive restoration project.

This could create a new valuation gap between companies that merely possess intellectual property and companies whose intellectual property can actually be consumed by machines.

The model itself may eventually become replaceable.

The memory layer may not.

Crypto Has a Role, But Not the Simplistic One

There is also an interesting crypto angle here, although it is more subtle than putting research papers on-chain.

Agentic science creates a provenance problem.

Suppose 5,000 agents collaborate on a scientific question.

Agent 812 reads one paper.

Agent 2,414 combines that idea with a dataset.

Another modifies the hypothesis.

Another runs code.

Another generates a candidate explanation.

A final system consolidates everything.

Who contributed what?

Which dataset was licensed?

Which researchers deserve attribution?

Who owns the resulting intellectual property?

Can somebody prove that an agent was authorized to access particular data?

Can another researcher reproduce the chain of reasoning?

These are not abstract questions.

As scientific work becomes more machine-mediated, proving origin and authorization becomes increasingly important.

Cryptographic provenance, verifiable credentials, machine-readable licences and tamper-evident research histories could become useful infrastructure here.

The value is not “blockchain solves science.”

It does not.

The value is that machines increasingly need verifiable records when millions of automated actions begin contributing to economically important outcomes.

The Next AI Bottleneck Will Keep Moving

This follows a pattern we are seeing across the AI economy.

At first, the bottleneck was models.

Then chips.

Then high-bandwidth memory.

Then data centers.

Then electricity and grid access.

Scientific AI will probably behave similarly.

First, the excitement is about intelligence.

Then everyone gets access to enormous quantities of machine intelligence.

The bottleneck moves downstream.

Who has the proprietary data?

Who controls the laboratory?

Who can run the experiment?

Who has regulatory access?

Who owns the IP?

Who can verify the result?

Who has the provenance?

Who can commercialize it?

That is where value starts concentrating.

The Archive May Be More Valuable Than We Think

The biggest mistake would be assuming that agentic science means every abandoned idea becomes valuable.

It does not.

Most scientific dead ends will remain dead ends.

The opportunity comes from search economics.

AI reduces the cost of asking one enormous question:

What have we already tried that might deserve another look?

That question can be applied across pharmaceuticals, materials science, batteries, energy, mathematics, chemistry, engineering and software.

Humanity has accumulated an extraordinary amount of unfinished work.

Until now, much of it has been effectively inaccessible because human attention was too scarce to keep revisiting everything.

Agents change that.

The next major scientific breakthrough may come from a model generating something nobody has ever seen before.

But it might also come from something less dramatic.

A machine could discover that somebody was almost right 17 years ago.

The technology was wrong.

The timing was wrong.

The market was wrong.

The experiment was incomplete.

Then the world changed.

And suddenly the forgotten idea is worth opening again.

The next AI gold rush may not only be about creating the future.

It may be about mining the past.

Originally published on Decentralised News