Where Value Leaks, It Pours

Ask a room of executives where their growth engine loses the most value, and they point downstream.

The product that missed.

The scale-up that burned through its round.

The patent suit that bled a quarter.

These losses are real, and they are also the wrong place to look. The largest losses in any value engine happen upstream, in selection: which questions get asked, which founders get backed, which careers get built. And they are nearly invisible, because they are counterfactuals. The study never reproduced. The outlier never funded. The talent never deployed. Nothing that failed to happen reaches a balance sheet, so nobody is paid to find it.

Start with what we can see, because that is the part the system handles well. A failed launch gets killed. A weak hire gets managed out. A lost case gets booked and absorbed. Visible failure triggers the feedback that corrects it; the very fact that you can see the loss is what keeps it contained. Markets and managers are competent pruners of the errors they can observe.

The upstream losses get no such correction, because they leave no trace. And they are not small. The clearest measure we have comes from biomedical research, the most heavily instrumented innovation system on earth. Two Oxford researchers, writing in The Lancet, estimated that around 85% of research investment is avoidably wasted: lost to the wrong question, the poorly designed study, the result that goes unpublished or unusable. Set against a global biomedical research spend approaching a quarter of a trillion dollars a year, that is on the order of $200 billion evaporating annually, before the knowledge is ever fit to use. Most of that loss is upstream, in choices made before a single experiment runs.

Why would a sophisticated, well-funded system waste most of its fuel? Look at what it rewards. A novel, positive result advances a career. A careful replication of someone else's work does not. So roughly half of preclinical findings cannot be reproduced by anyone else, at an estimated cost of $28 billion a year in the United States alone. The system is not malfunctioning. It is performing exactly to its incentives: it optimises for the thing it counts, publications, over the thing it actually needs, which is truth.

That pattern is the whole story, and it repeats at every stage of the value engine.

In capital, around half of venture investments lose money, and a tiny fraction produce nearly all the returns; in one large dataset, 6% of deals generated 60% of the gains. The high loss rate is not the leak. It is by design, the price of a power law. The leak is in the filtering. A selection function tuned to recognise the patterns of past winners will systematically misprice the outliers it exists to find, because a true outlier looks, almost by definition, like a mistake.

And the input to that function is narrow before it even begins. Companies founded entirely by women received about 2% of US venture funding in 2023, and closer to 1% in 2024. Whatever you believe about the cause, the strategic consequence is not in dispute: under a power law, where a single extreme winner pays for everything, drawing your candidates from a thin slice of the distribution is a reliable way to lower your odds of finding the one outcome that matters. You cannot back the tail if you never let it into the room.

At every gate, a legible proxy beats the illegible real thing. The proxy can be counted, funded and promoted. The real thing is none of those, until much later, if ever.

Step back, and the mechanism is singular. At every gate in the system, the measurable stand-in beats the thing it stands for.

The same trade, five times over

• Publish beats reproduce

• Pattern-match beats outlier

• Scale beats search

• Patent beats practise

• Credential beats build

None of these is stupidity. Each is a rational response to what is countable, and therefore rewardable. The proxy can be measured, defended, funded and put in a deck. The real thing cannot, not yet. So the proxy wins, every time, at every stage. And the friction that was quietly carrying the actual value, the slow replication, the unfashionable founder, the unglamorous search for fit before the scale-up, gets stripped out as inefficiency.

This is the same dynamic, scaled to an entire economy, that produces frictionless, confident, sourceless machine-written content. Smoothness there is not a marker of quality; it is a marker of the friction, and the information, that has been removed. The principle holds at every level: friction taken out of a system is usually value taken out with it. The smooth proxy wins, and the thing it was standing in for slips away unnoticed.

Which is why the leak never closes. A loss you can see generates pressure to fix it. A loss that exists only as a counterfactual, the breakthrough a better-selected portfolio would have produced, the company that was never funded, generates no pressure at all, because there is no line item for it and no one whose job depends on it. The system is stable in precisely the wrong place. It will go on leaking, efficiently and indefinitely, until someone decides to measure the negative space.

The selection audit

01 What does our selection function actually reward?

Take one real decision: a hire, a funded project, a greenlit product. Trace it back to the thing that justified it. Was it the credential, the familiar pattern, the confident projection? Or the underlying substance? If you cannot tell the difference, your filter cannot either.

02 Where are we counting the proxy instead of the thing?

Every metric stands in for something you actually want. Find the places where the two have quietly diverged: utilisation standing in for value created, pipeline for revenue, papers for knowledge, activity for progress. Each gap is a place value leaks out unrecorded.

03 What would we need to measure to see the negative space?

You will never get a clean figure on the counterfactual. But you can instrument the inputs: the breadth of your deal flow, the share of your bets that are genuine outliers rather than pattern-matches, the ratio of search to scale in early work. You manage what you measure, which is the whole problem. So measure the thing you are currently letting leak.

The visible failures will keep getting fixed. They always do; that is what visibility is for. The real question is whether you are willing to spend attention on the losses that never send you an invoice, because those are the large ones. The engine is not slow and expensive because the good ideas ran out. It is slow and expensive because, at every gate, it prefers what it can see.

Until next time, swim upstream, there's gold up there!

Editor,

Ideas for a Better World