Where Science Makes Money: The Opportunities Nobody Is Searching For
The best scientific opportunities are not undiscovered. They are unsearched — and searching is now a machine problem.
In 1940, two physicists in Birmingham built the cavity magnetron to win an air war. It was a radar tube. Five years later, a Raytheon engineer standing next to one noticed the chocolate bar in his pocket had melted, and the company shipped a commercial oven in 1947. It weighed a third of a tonne. The kitchen took another forty years.
Forty years. Not because the physics was hard — the physics was finished in 1940. Because nobody was looking there. The people who understood the magnetron were radar engineers, and radar engineers do not think about lunch. The link between "resonant cavity produces 2.45 GHz microwaves" and "there are two billion meals a day that need heating" required someone to hold both facts at once, and for five years nobody did. Then it took an accident.
That is the shape of the thing we build against. The magnetron's second life was not hidden in a lab notebook waiting to be discovered. It was sitting in plain sight, one association away, in a field nobody in the room was qualified to think about. And it is not a historical curiosity — it is the default condition of science right now, at a scale that has become absurd. The world publishes millions of papers a year. Each one is a capability. Almost none of them are ever systematically crossed against the industries they could transform.
The thesis of this article is simple and, we think, uncomfortable: the most valuable opportunities in science are not a genius problem. They are a search problem. And search problems are exactly what machines have recently become extraordinarily good at.
Science Is the Substrate of the Physical Economy — and It Is Under-Monetised on Purpose
Start with the boring part, because the boring part is the whole business case.
We have mapped nine scientific domains — Life Sciences, Chemistry, Materials, Medicine, Agriculture, Energy & Environment, Physics & Instrumentation, Earth & Space, Computational — down to fifty-seven subfields, and beneath them the thirty-five industrial activities that meet three tests: they are done at industrial scale today, they are worth roughly ten billion dollars a year or more, and a scientific step materially sets their economics. Drug discovery. Refining. Catalyst manufacturing. Semiconductor fabrication. Crop protection. Battery cell production. Clinical diagnostics.
Those thirty-five activities sum to over ten trillion dollars a year of industry. Every one of them has a science step whose performance sets the economics of everything downstream — and every one of them has a bottleneck we can name in a sentence. Roughly ninety percent trial attrition in drug discovery. Thirty percent scrap in new gigafactories. Catalyst screening measured in years. These are not vague inefficiencies. They are specific, quantified, expensive constraints, and each one is a door.
Here is the uncomfortable part. The reason those doors stay shut is almost never that the science does not exist. It is that the person who owns the bottleneck and the person who owns the capability have never met, and have no mechanism that would ever make them meet. The gigafactory process engineer does not read the electrochemistry literature. The electrochemist does not know what scrap costs. Both are excellent at their jobs. The value sits in the gap between them, and the gap is nobody's job.
Why the Search Fails: Everyone Looks in Their Own Field
There is a specific, diagnosable failure mode here, and naming it precisely is what makes it fixable.
Scientists search their home field. This is not a criticism; it is what expertise is. Deep knowledge is deep because it is narrow. A world expert in directed evolution has spent twenty years earning the right to have strong opinions about directed evolution, and that same twenty years is why they will not spontaneously wonder whether it applies to mining tailings remediation.
Investors search the famous pairs. Ask any venture fund where AI meets science and you will get the same ten answers: protein folding, drug discovery, materials generation, self-driving labs. Those are real, and they are also the ten pairs that every fund on earth is looking at simultaneously, which is precisely why the returns there are being competed away. The famous pairs are the worst place to look, not the best — not because they are wrong, but because they are crowded.
Both are searching a tiny neighbourhood of a very large space. Thirty-five activities crossed against fifty distinct improvement mechanisms is 1,750 combinations. Human search, whether by a scientist or an analyst, examines maybe a dozen — the dozen adjacent to what they already know. The other 1,738 are not rejected. They are never evaluated at all. The magnetron sat in that unexamined remainder for five years, and the only reason it ever escaped was that a man happened to be standing in the wrong place with a chocolate bar.
This is the correctable error. Not "we made a bad judgement" but "we never rendered a judgement, because the pair never came up." An engine does not get bored, does not have a home field, and does not have a career invested in one answer. It can evaluate all 1,750 — and the ones it surfaces from outside your field are, by construction, the ones nobody is competing for.
The Second Half of the Problem: "Science Will Help" Is Not a Strategy
Exhaustive search only works if the thing you are searching with is specific.
"AI will improve chemistry" is not a claim you can act on, price, or falsify. So we did the unglamorous work of enumerating what improvement actually is. There turn out to be seven mechanisms and roughly fifty named capabilities underneath them — you can compute more (ML property prediction, generative design, Bayesian experiment design, ML surrogate simulation, retrosynthesis, quantum chemistry), do more (lab robotics, self-driving labs, high-throughput screening, microfluidics, flow chemistry), see more (in-line process analytics, single-cell and spatial omics, hyperspectral sensing, continuous biosensors), engineer biology (CRISPR, directed evolution, precision fermentation, organoids), engineer matter (next-gen catalysts, green route substitution, electrochemical synthesis, nanostructured materials), run processes better (statistical DOE, quality-by-design, process intensification, mechanistic scale-up), or use knowledge better (FAIR data infrastructure, causal inference, federated collaboration).
Each of the fifty carries four numbers that make it tradeable: a typical gain where it fully applies, calibrated conservatively from published case studies; a kind of gain — Speed, Cost, Yield, Quality, or Discovery; a maturity score for what is actually shippable today (statistical DOE 0.95, quantum chemistry 0.25); and a difficulty score for deployment friction — integration, regulation, culture (literature mining 0.2, self-driving labs 0.75).
That last pair is where most technology strategy quietly dies. Maturity and difficulty are different axes, and confusing them is why so many deeptech theses are wrong in an expensive direction. A mature capability that is brutal to deploy and an immature capability that would slot in tomorrow are not the same bet, and no amount of enthusiasm collapses that distinction. Pricing both, separately, is the difference between a thesis and a hope.
The Cross Product, and the Arithmetic That Refuses to Flatter You
With both sides specified, the search becomes mechanical. Every activity crossed against every capability. In the version we have published the numbers for, that was 1,400 pairs tested, of which 562 passed — a combined theoretical five-year obtainable pool of roughly three hundred billion dollars, with the single best pair (protein structure prediction against drug discovery) at about $3.8B a year.
But the number that matters is not $300B. It is 838 — the pairs that failed, and the fact that the engine says so out loud.
This is where better decision-making stops being a slogan. Most opportunity analysis is an addition machine: stack up favourable assumptions, sum them, present a big number. Ours multiplies, and multiplication is honest in a way addition never is. A pair's realised gain is the capability's typical gain times the activity's science intensity times how well the capability's preconditions actually match the activity's work profile. Three fractions, multiplied. Any weak link collapses the claim. You cannot rescue a bad match with an impressive gain.
Then the pool itself is disciplined. An improvement to drug discovery acts on the R&D step — about forty-five percent of relevant spend — not on pharma's entire revenue. A vendor captures fifteen percent of the value it creates, not all of it, because that is what enterprise software has historically captured. Readiness discounts the whole thing by maturity against half of difficulty. Nothing reaches steady state in year one, so a five-year adoption ramp runs 5% → 12% → 22% → 35% → 48%.
Run drug discovery against Bayesian experiment design through that and you get $33.5B a year of value created for the industry, of which roughly $3.0B a year is obtainable by whoever builds the tool. That is a real number with a visible derivation, and every term in it can be argued with — which is the point. Every figure is a slider. Disagree with the fifteen percent capture rate, move it, and watch the entire ranking recompute in front of you.
We are equally explicit about what the arithmetic does not do. It prices the prize, not the campaign — there is no cost side. It does not model competition. Market sizes are honest public estimates, not audited research; the engine ranks, it does not audit. Summing every capability against one activity overstates, because improvements overlap. These limitations are printed on the screens, not buried in an appendix, because a tool that hides its own weaknesses is not a decision aid — it is a sales deck with a chart in it.
What Actually Makes an Opportunity Missed
Ranking 562 viable pairs by revenue gives you a good list. It does not, by itself, give you a non-obvious list — the top of any revenue ranking is where the crowd already is.
So the more recent generation of the engine inverts the whole interaction, and this is the part we think is genuinely new. There is no field picker. You do not choose a domain and browse. A discovery walks in, the machine sweeps every activity of every field, and a dossier comes out. The field picker had to go, because the microwave logic forbids choosing a field first: the surprising application is by definition in the field you would not have picked. A tool that asks you to narrow before it searches has already thrown away the answer.
What comes out is a case file, not a score. The sweep lists every landing across the whole mapped economy, with the ones outside the discovery's home field called out explicitly — the microwave list. Next to it sits the adjacent possible: the landings that are one condition short, which are often more useful than the winners, because a condition is a thing you can go and change. Then the money, as a visible multiplicative chain with each factor's evidence class and status attached, so you can see exactly which link is weakest and verify that one first.
Then the parts that separate a decision from an opinion:
Four futures, priced separately. Two key uncertainties crossed into a two-by-two of possible worlds, with every opportunity tested against all four. An innovation that stays attractive in all four has a special property, and one that works in a single world and dies in three is a different asset entirely. The tool never declares which future is likely — it prices them and hands you the spread. Averaging four scenarios into one expected value destroys precisely the information you needed.
A kill check, in the arithmetic. Fifty explicit kill-claims — the specific thing that, if true, ends this. Not a risks section at the back. A term in the calculation.
The expert delta. Before the sweep runs, the team records its own belief about the discovery. Afterwards, the machine's answer is compared against it. Where they agree, the machine has confirmed cheap. Where they diverge, someone has learned something — and over time the record of divergences tells you whether the machine is worth listening to, which is a question no vendor should be allowed to answer about themselves.
Dated wagers, publicly scored. Every dossier emits resolvable claims with a date and a confidence. They resolve TRUE or FALSE, a Brier score accumulates, and the misses get published. This is the honesty contract, and it is the only mechanism we know of that makes a forecasting tool improvable rather than merely persuasive.
The dossier closes with a verdict — BUILD / PARTNER / WATCH / PASS — because a decision is what the user actually needed. Not a percentile. Not a heat map. A verdict, with everything it stands on visible underneath, and a ready-to-send outreach letter to the named buyer if the verdict was BUILD.
Advancing Science and Making Money Are the Same Motion
There is a reflex in European science policy that treats commercialisation as slightly grubby — as though the money were extracted from the science rather than fed back into it. That reflex has cost this continent an enormous amount, and it is empirically wrong.
Look at what the engine's arithmetic actually rewards. Its gains are Speed, Cost, Yield, Quality, and Discovery — measured against named bottlenecks. To earn money in this model you must compress a real scientific cycle: fewer experiments to the same answer, higher yield per input, a result that used to be unfindable becoming findable. There is no term in the formula that pays for marketing, packaging, or rent extraction. The only path to the number is making the science itself go faster.
Which produces a loop rather than a trade-off. A capability that cuts experiment counts three- to ten-fold earns revenue because it advanced the science. That revenue funds the instrument, the dataset, and the deployment engineering that the next discovery needs. The installed base then feeds the model, so the next prediction is better, so the next opportunity is easier to price and cheaper to underwrite. The flywheel is tools × knowledge × memory, and it compounds in exactly the way single-shot grant funding does not.
This matters especially here. Europe's problem has never been a shortage of science — the papers, the institutes, the doctoral pipelines are world-class and always have been. The problem is the conversion step: the systematic failure to notice that a result in one European lab is worth billions to an industry three countries and two disciplines away. That is not a funding gap. It is a search gap, and search gaps are the kind of problem software solves.
The Vertical AI Companies This Produces
We are not building a market research product. The engine is upstream infrastructure; the output is companies.
When a dossier comes back BUILD, what it has actually specified is a vertical AI company: a named scientific capability, aimed at a named bottleneck, in an industry of known size, with the value split and the readiness discount already computed, the kill-claims listed, the buyer identified, and the first three things to verify ordered by how much they would change the answer. That is a considerably better starting position than the way most deeptech companies begin, which is a founder with a technology and a hope.
And these companies share the important thing. The technology varies wildly — a catalyst, an assay, a sensing method, an algorithm — but the shape is constant: a capability that collapses the cost of an expensive scientific act, deployed into an industry that already has a budget line for the pain. That constancy is why a portfolio works where a single bet would not. The second company commercialises far more cheaply than the first, because the map, the pricing discipline, the verification method, and the deployment scaffolding are shared. Individual technologies are cargo. The search engine and the deployment platform are the durable assets.
What to Do on Monday
You do not need our engine to run the first move, and we would rather you tested the idea than took our word for it.
Take one capability your organisation genuinely owns — a method, an instrument, a dataset, a process nobody else has. Write down, honestly, the three applications you already have in mind. That is your expert-delta baseline; record it before you search, because recording it afterwards is worthless.
Now force the sweep. List thirty industrial activities that have nothing to do with your field — refining, cold chain, cement, insurance claims, seed breeding, wafer inspection, water treatment — and for each one ask a single question: what is this activity's binding constraint, and does my capability touch it? Most answers are no, and no is fine; no is the 838. You are looking for the two or three where the answer is "actually, maybe," and where your first reaction is mild embarrassment that you had not considered it.
Then price the best one multiplicatively, not additively. Industry size, times the slice your step actually touches, times a realistic gain, times how much of that you could capture. Multiply. Do not average, do not round up, and do not rescue a weak link with a strong one. Then write down the single fact that, if false, kills it — and go check that fact first, before anything else.
If the answer that survives came from a field you would not have picked, you have just run the magnetron experiment on yourself. That is the whole method.
The Space Is Enumerable
The reason we think this works is not that our arithmetic is clever. It is that the space is finite.
The domains of science are enumerable. The industrial activities built on them are enumerable. The mechanisms by which science improves things are enumerable — there are about fifty, and they were constructible in an afternoon once we stopped accepting "AI will help" as a sentence. The cross product of those three is a few thousand cells, which is nothing. It is a rounding error against what a machine can evaluate before lunch.
For four hundred years, the discovery of where science pays off has been left to accident, adjacency, and the occasional melted chocolate bar. It did not have to be. It was left there because the search was too large for a person and nobody had built the machine — and now the search is small, the machine exists, and the remaining question is only who bothers to run it exhaustively before everyone else does.
The magnetron waited forty years for its second life. The next one does not have to wait at all.